Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
6 *
7 * Implementation of the Transmission Control Protocol(TCP).
8 *
9 * Authors: Ross Biro
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
20 */
21
22/*
23 * Changes: Pedro Roque : Retransmit queue handled by TCP.
24 * : Fragmentation on mtu decrease
25 * : Segment collapse on retransmit
26 * : AF independence
27 *
28 * Linus Torvalds : send_delayed_ack
29 * David S. Miller : Charge memory using the right skb
30 * during syn/ack processing.
31 * David S. Miller : Output engine completely rewritten.
32 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
33 * Cacophonix Gaul : draft-minshall-nagle-01
34 * J Hadi Salim : ECN support
35 *
36 */
37
38#define pr_fmt(fmt) "TCP: " fmt
39
40#include <net/tcp.h>
41#include <net/mptcp.h>
42
43#include <linux/compiler.h>
44#include <linux/gfp.h>
45#include <linux/module.h>
46#include <linux/static_key.h>
47
48#include <trace/events/tcp.h>
49
50/* Refresh clocks of a TCP socket,
51 * ensuring monotically increasing values.
52 */
53void tcp_mstamp_refresh(struct tcp_sock *tp)
54{
55 u64 val = tcp_clock_ns();
56
57 tp->tcp_clock_cache = val;
58 tp->tcp_mstamp = div_u64(val, NSEC_PER_USEC);
59}
60
61static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
62 int push_one, gfp_t gfp);
63
64/* Account for new data that has been sent to the network. */
65static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
66{
67 struct inet_connection_sock *icsk = inet_csk(sk);
68 struct tcp_sock *tp = tcp_sk(sk);
69 unsigned int prior_packets = tp->packets_out;
70
71 WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(skb)->end_seq);
72
73 __skb_unlink(skb, &sk->sk_write_queue);
74 tcp_rbtree_insert(&sk->tcp_rtx_queue, skb);
75
76 if (tp->highest_sack == NULL)
77 tp->highest_sack = skb;
78
79 tp->packets_out += tcp_skb_pcount(skb);
80 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
81 tcp_rearm_rto(sk);
82
83 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
84 tcp_skb_pcount(skb));
85 tcp_check_space(sk);
86}
87
88/* SND.NXT, if window was not shrunk or the amount of shrunk was less than one
89 * window scaling factor due to loss of precision.
90 * If window has been shrunk, what should we make? It is not clear at all.
91 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
92 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
93 * invalid. OK, let's make this for now:
94 */
95static inline __u32 tcp_acceptable_seq(const struct sock *sk)
96{
97 const struct tcp_sock *tp = tcp_sk(sk);
98
99 if (!before(tcp_wnd_end(tp), tp->snd_nxt) ||
100 (tp->rx_opt.wscale_ok &&
101 ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale))))
102 return tp->snd_nxt;
103 else
104 return tcp_wnd_end(tp);
105}
106
107/* Calculate mss to advertise in SYN segment.
108 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
109 *
110 * 1. It is independent of path mtu.
111 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
112 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
113 * attached devices, because some buggy hosts are confused by
114 * large MSS.
115 * 4. We do not make 3, we advertise MSS, calculated from first
116 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
117 * This may be overridden via information stored in routing table.
118 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
119 * probably even Jumbo".
120 */
121static __u16 tcp_advertise_mss(struct sock *sk)
122{
123 struct tcp_sock *tp = tcp_sk(sk);
124 const struct dst_entry *dst = __sk_dst_get(sk);
125 int mss = tp->advmss;
126
127 if (dst) {
128 unsigned int metric = dst_metric_advmss(dst);
129
130 if (metric < mss) {
131 mss = metric;
132 tp->advmss = mss;
133 }
134 }
135
136 return (__u16)mss;
137}
138
139/* RFC2861. Reset CWND after idle period longer RTO to "restart window".
140 * This is the first part of cwnd validation mechanism.
141 */
142void tcp_cwnd_restart(struct sock *sk, s32 delta)
143{
144 struct tcp_sock *tp = tcp_sk(sk);
145 u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
146 u32 cwnd = tcp_snd_cwnd(tp);
147
148 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
149
150 tp->snd_ssthresh = tcp_current_ssthresh(sk);
151 restart_cwnd = min(restart_cwnd, cwnd);
152
153 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
154 cwnd >>= 1;
155 tcp_snd_cwnd_set(tp, max(cwnd, restart_cwnd));
156 tp->snd_cwnd_stamp = tcp_jiffies32;
157 tp->snd_cwnd_used = 0;
158}
159
160/* Congestion state accounting after a packet has been sent. */
161static void tcp_event_data_sent(struct tcp_sock *tp,
162 struct sock *sk)
163{
164 struct inet_connection_sock *icsk = inet_csk(sk);
165 const u32 now = tcp_jiffies32;
166
167 if (tcp_packets_in_flight(tp) == 0)
168 tcp_ca_event(sk, CA_EVENT_TX_START);
169
170 tp->lsndtime = now;
171
172 /* If it is a reply for ato after last received
173 * packet, enter pingpong mode.
174 */
175 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
176 inet_csk_enter_pingpong_mode(sk);
177}
178
179/* Account for an ACK we sent. */
180static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts,
181 u32 rcv_nxt)
182{
183 struct tcp_sock *tp = tcp_sk(sk);
184
185 if (unlikely(tp->compressed_ack)) {
186 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
187 tp->compressed_ack);
188 tp->compressed_ack = 0;
189 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
190 __sock_put(sk);
191 }
192
193 if (unlikely(rcv_nxt != tp->rcv_nxt))
194 return; /* Special ACK sent by DCTCP to reflect ECN */
195 tcp_dec_quickack_mode(sk, pkts);
196 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
197}
198
199/* Determine a window scaling and initial window to offer.
200 * Based on the assumption that the given amount of space
201 * will be offered. Store the results in the tp structure.
202 * NOTE: for smooth operation initial space offering should
203 * be a multiple of mss if possible. We assume here that mss >= 1.
204 * This MUST be enforced by all callers.
205 */
206void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss,
207 __u32 *rcv_wnd, __u32 *window_clamp,
208 int wscale_ok, __u8 *rcv_wscale,
209 __u32 init_rcv_wnd)
210{
211 unsigned int space = (__space < 0 ? 0 : __space);
212
213 /* If no clamp set the clamp to the max possible scaled window */
214 if (*window_clamp == 0)
215 (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE);
216 space = min(*window_clamp, space);
217
218 /* Quantize space offering to a multiple of mss if possible. */
219 if (space > mss)
220 space = rounddown(space, mss);
221
222 /* NOTE: offering an initial window larger than 32767
223 * will break some buggy TCP stacks. If the admin tells us
224 * it is likely we could be speaking with such a buggy stack
225 * we will truncate our initial window offering to 32K-1
226 * unless the remote has sent us a window scaling option,
227 * which we interpret as a sign the remote TCP is not
228 * misinterpreting the window field as a signed quantity.
229 */
230 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows))
231 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
232 else
233 (*rcv_wnd) = min_t(u32, space, U16_MAX);
234
235 if (init_rcv_wnd)
236 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
237
238 *rcv_wscale = 0;
239 if (wscale_ok) {
240 /* Set window scaling on max possible window */
241 space = max_t(u32, space, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
242 space = max_t(u32, space, READ_ONCE(sysctl_rmem_max));
243 space = min_t(u32, space, *window_clamp);
244 *rcv_wscale = clamp_t(int, ilog2(space) - 15,
245 0, TCP_MAX_WSCALE);
246 }
247 /* Set the clamp no higher than max representable value */
248 (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp);
249}
250EXPORT_SYMBOL(tcp_select_initial_window);
251
252/* Chose a new window to advertise, update state in tcp_sock for the
253 * socket, and return result with RFC1323 scaling applied. The return
254 * value can be stuffed directly into th->window for an outgoing
255 * frame.
256 */
257static u16 tcp_select_window(struct sock *sk)
258{
259 struct tcp_sock *tp = tcp_sk(sk);
260 u32 old_win = tp->rcv_wnd;
261 u32 cur_win = tcp_receive_window(tp);
262 u32 new_win = __tcp_select_window(sk);
263
264 /* Never shrink the offered window */
265 if (new_win < cur_win) {
266 /* Danger Will Robinson!
267 * Don't update rcv_wup/rcv_wnd here or else
268 * we will not be able to advertise a zero
269 * window in time. --DaveM
270 *
271 * Relax Will Robinson.
272 */
273 if (new_win == 0)
274 NET_INC_STATS(sock_net(sk),
275 LINUX_MIB_TCPWANTZEROWINDOWADV);
276 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
277 }
278 tp->rcv_wnd = new_win;
279 tp->rcv_wup = tp->rcv_nxt;
280
281 /* Make sure we do not exceed the maximum possible
282 * scaled window.
283 */
284 if (!tp->rx_opt.rcv_wscale &&
285 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows))
286 new_win = min(new_win, MAX_TCP_WINDOW);
287 else
288 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
289
290 /* RFC1323 scaling applied */
291 new_win >>= tp->rx_opt.rcv_wscale;
292
293 /* If we advertise zero window, disable fast path. */
294 if (new_win == 0) {
295 tp->pred_flags = 0;
296 if (old_win)
297 NET_INC_STATS(sock_net(sk),
298 LINUX_MIB_TCPTOZEROWINDOWADV);
299 } else if (old_win == 0) {
300 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
301 }
302
303 return new_win;
304}
305
306/* Packet ECN state for a SYN-ACK */
307static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
308{
309 const struct tcp_sock *tp = tcp_sk(sk);
310
311 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
312 if (!(tp->ecn_flags & TCP_ECN_OK))
313 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
314 else if (tcp_ca_needs_ecn(sk) ||
315 tcp_bpf_ca_needs_ecn(sk))
316 INET_ECN_xmit(sk);
317}
318
319/* Packet ECN state for a SYN. */
320static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
321{
322 struct tcp_sock *tp = tcp_sk(sk);
323 bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk);
324 bool use_ecn = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_ecn) == 1 ||
325 tcp_ca_needs_ecn(sk) || bpf_needs_ecn;
326
327 if (!use_ecn) {
328 const struct dst_entry *dst = __sk_dst_get(sk);
329
330 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
331 use_ecn = true;
332 }
333
334 tp->ecn_flags = 0;
335
336 if (use_ecn) {
337 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
338 tp->ecn_flags = TCP_ECN_OK;
339 if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn)
340 INET_ECN_xmit(sk);
341 }
342}
343
344static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
345{
346 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback))
347 /* tp->ecn_flags are cleared at a later point in time when
348 * SYN ACK is ultimatively being received.
349 */
350 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
351}
352
353static void
354tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
355{
356 if (inet_rsk(req)->ecn_ok)
357 th->ece = 1;
358}
359
360/* Set up ECN state for a packet on a ESTABLISHED socket that is about to
361 * be sent.
362 */
363static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
364 struct tcphdr *th, int tcp_header_len)
365{
366 struct tcp_sock *tp = tcp_sk(sk);
367
368 if (tp->ecn_flags & TCP_ECN_OK) {
369 /* Not-retransmitted data segment: set ECT and inject CWR. */
370 if (skb->len != tcp_header_len &&
371 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
372 INET_ECN_xmit(sk);
373 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
374 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
375 th->cwr = 1;
376 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
377 }
378 } else if (!tcp_ca_needs_ecn(sk)) {
379 /* ACK or retransmitted segment: clear ECT|CE */
380 INET_ECN_dontxmit(sk);
381 }
382 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
383 th->ece = 1;
384 }
385}
386
387/* Constructs common control bits of non-data skb. If SYN/FIN is present,
388 * auto increment end seqno.
389 */
390static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
391{
392 skb->ip_summed = CHECKSUM_PARTIAL;
393
394 TCP_SKB_CB(skb)->tcp_flags = flags;
395
396 tcp_skb_pcount_set(skb, 1);
397
398 TCP_SKB_CB(skb)->seq = seq;
399 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
400 seq++;
401 TCP_SKB_CB(skb)->end_seq = seq;
402}
403
404static inline bool tcp_urg_mode(const struct tcp_sock *tp)
405{
406 return tp->snd_una != tp->snd_up;
407}
408
409#define OPTION_SACK_ADVERTISE BIT(0)
410#define OPTION_TS BIT(1)
411#define OPTION_MD5 BIT(2)
412#define OPTION_WSCALE BIT(3)
413#define OPTION_FAST_OPEN_COOKIE BIT(8)
414#define OPTION_SMC BIT(9)
415#define OPTION_MPTCP BIT(10)
416
417static void smc_options_write(__be32 *ptr, u16 *options)
418{
419#if IS_ENABLED(CONFIG_SMC)
420 if (static_branch_unlikely(&tcp_have_smc)) {
421 if (unlikely(OPTION_SMC & *options)) {
422 *ptr++ = htonl((TCPOPT_NOP << 24) |
423 (TCPOPT_NOP << 16) |
424 (TCPOPT_EXP << 8) |
425 (TCPOLEN_EXP_SMC_BASE));
426 *ptr++ = htonl(TCPOPT_SMC_MAGIC);
427 }
428 }
429#endif
430}
431
432struct tcp_out_options {
433 u16 options; /* bit field of OPTION_* */
434 u16 mss; /* 0 to disable */
435 u8 ws; /* window scale, 0 to disable */
436 u8 num_sack_blocks; /* number of SACK blocks to include */
437 u8 hash_size; /* bytes in hash_location */
438 u8 bpf_opt_len; /* length of BPF hdr option */
439 __u8 *hash_location; /* temporary pointer, overloaded */
440 __u32 tsval, tsecr; /* need to include OPTION_TS */
441 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
442 struct mptcp_out_options mptcp;
443};
444
445static void mptcp_options_write(struct tcphdr *th, __be32 *ptr,
446 struct tcp_sock *tp,
447 struct tcp_out_options *opts)
448{
449#if IS_ENABLED(CONFIG_MPTCP)
450 if (unlikely(OPTION_MPTCP & opts->options))
451 mptcp_write_options(th, ptr, tp, &opts->mptcp);
452#endif
453}
454
455#ifdef CONFIG_CGROUP_BPF
456static int bpf_skops_write_hdr_opt_arg0(struct sk_buff *skb,
457 enum tcp_synack_type synack_type)
458{
459 if (unlikely(!skb))
460 return BPF_WRITE_HDR_TCP_CURRENT_MSS;
461
462 if (unlikely(synack_type == TCP_SYNACK_COOKIE))
463 return BPF_WRITE_HDR_TCP_SYNACK_COOKIE;
464
465 return 0;
466}
467
468/* req, syn_skb and synack_type are used when writing synack */
469static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb,
470 struct request_sock *req,
471 struct sk_buff *syn_skb,
472 enum tcp_synack_type synack_type,
473 struct tcp_out_options *opts,
474 unsigned int *remaining)
475{
476 struct bpf_sock_ops_kern sock_ops;
477 int err;
478
479 if (likely(!BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk),
480 BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG)) ||
481 !*remaining)
482 return;
483
484 /* *remaining has already been aligned to 4 bytes, so *remaining >= 4 */
485
486 /* init sock_ops */
487 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
488
489 sock_ops.op = BPF_SOCK_OPS_HDR_OPT_LEN_CB;
490
491 if (req) {
492 /* The listen "sk" cannot be passed here because
493 * it is not locked. It would not make too much
494 * sense to do bpf_setsockopt(listen_sk) based
495 * on individual connection request also.
496 *
497 * Thus, "req" is passed here and the cgroup-bpf-progs
498 * of the listen "sk" will be run.
499 *
500 * "req" is also used here for fastopen even the "sk" here is
501 * a fullsock "child" sk. It is to keep the behavior
502 * consistent between fastopen and non-fastopen on
503 * the bpf programming side.
504 */
505 sock_ops.sk = (struct sock *)req;
506 sock_ops.syn_skb = syn_skb;
507 } else {
508 sock_owned_by_me(sk);
509
510 sock_ops.is_fullsock = 1;
511 sock_ops.sk = sk;
512 }
513
514 sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type);
515 sock_ops.remaining_opt_len = *remaining;
516 /* tcp_current_mss() does not pass a skb */
517 if (skb)
518 bpf_skops_init_skb(&sock_ops, skb, 0);
519
520 err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk);
521
522 if (err || sock_ops.remaining_opt_len == *remaining)
523 return;
524
525 opts->bpf_opt_len = *remaining - sock_ops.remaining_opt_len;
526 /* round up to 4 bytes */
527 opts->bpf_opt_len = (opts->bpf_opt_len + 3) & ~3;
528
529 *remaining -= opts->bpf_opt_len;
530}
531
532static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb,
533 struct request_sock *req,
534 struct sk_buff *syn_skb,
535 enum tcp_synack_type synack_type,
536 struct tcp_out_options *opts)
537{
538 u8 first_opt_off, nr_written, max_opt_len = opts->bpf_opt_len;
539 struct bpf_sock_ops_kern sock_ops;
540 int err;
541
542 if (likely(!max_opt_len))
543 return;
544
545 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
546
547 sock_ops.op = BPF_SOCK_OPS_WRITE_HDR_OPT_CB;
548
549 if (req) {
550 sock_ops.sk = (struct sock *)req;
551 sock_ops.syn_skb = syn_skb;
552 } else {
553 sock_owned_by_me(sk);
554
555 sock_ops.is_fullsock = 1;
556 sock_ops.sk = sk;
557 }
558
559 sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type);
560 sock_ops.remaining_opt_len = max_opt_len;
561 first_opt_off = tcp_hdrlen(skb) - max_opt_len;
562 bpf_skops_init_skb(&sock_ops, skb, first_opt_off);
563
564 err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk);
565
566 if (err)
567 nr_written = 0;
568 else
569 nr_written = max_opt_len - sock_ops.remaining_opt_len;
570
571 if (nr_written < max_opt_len)
572 memset(skb->data + first_opt_off + nr_written, TCPOPT_NOP,
573 max_opt_len - nr_written);
574}
575#else
576static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb,
577 struct request_sock *req,
578 struct sk_buff *syn_skb,
579 enum tcp_synack_type synack_type,
580 struct tcp_out_options *opts,
581 unsigned int *remaining)
582{
583}
584
585static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb,
586 struct request_sock *req,
587 struct sk_buff *syn_skb,
588 enum tcp_synack_type synack_type,
589 struct tcp_out_options *opts)
590{
591}
592#endif
593
594/* Write previously computed TCP options to the packet.
595 *
596 * Beware: Something in the Internet is very sensitive to the ordering of
597 * TCP options, we learned this through the hard way, so be careful here.
598 * Luckily we can at least blame others for their non-compliance but from
599 * inter-operability perspective it seems that we're somewhat stuck with
600 * the ordering which we have been using if we want to keep working with
601 * those broken things (not that it currently hurts anybody as there isn't
602 * particular reason why the ordering would need to be changed).
603 *
604 * At least SACK_PERM as the first option is known to lead to a disaster
605 * (but it may well be that other scenarios fail similarly).
606 */
607static void tcp_options_write(struct tcphdr *th, struct tcp_sock *tp,
608 struct tcp_out_options *opts)
609{
610 __be32 *ptr = (__be32 *)(th + 1);
611 u16 options = opts->options; /* mungable copy */
612
613 if (unlikely(OPTION_MD5 & options)) {
614 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
615 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
616 /* overload cookie hash location */
617 opts->hash_location = (__u8 *)ptr;
618 ptr += 4;
619 }
620
621 if (unlikely(opts->mss)) {
622 *ptr++ = htonl((TCPOPT_MSS << 24) |
623 (TCPOLEN_MSS << 16) |
624 opts->mss);
625 }
626
627 if (likely(OPTION_TS & options)) {
628 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
629 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
630 (TCPOLEN_SACK_PERM << 16) |
631 (TCPOPT_TIMESTAMP << 8) |
632 TCPOLEN_TIMESTAMP);
633 options &= ~OPTION_SACK_ADVERTISE;
634 } else {
635 *ptr++ = htonl((TCPOPT_NOP << 24) |
636 (TCPOPT_NOP << 16) |
637 (TCPOPT_TIMESTAMP << 8) |
638 TCPOLEN_TIMESTAMP);
639 }
640 *ptr++ = htonl(opts->tsval);
641 *ptr++ = htonl(opts->tsecr);
642 }
643
644 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
645 *ptr++ = htonl((TCPOPT_NOP << 24) |
646 (TCPOPT_NOP << 16) |
647 (TCPOPT_SACK_PERM << 8) |
648 TCPOLEN_SACK_PERM);
649 }
650
651 if (unlikely(OPTION_WSCALE & options)) {
652 *ptr++ = htonl((TCPOPT_NOP << 24) |
653 (TCPOPT_WINDOW << 16) |
654 (TCPOLEN_WINDOW << 8) |
655 opts->ws);
656 }
657
658 if (unlikely(opts->num_sack_blocks)) {
659 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
660 tp->duplicate_sack : tp->selective_acks;
661 int this_sack;
662
663 *ptr++ = htonl((TCPOPT_NOP << 24) |
664 (TCPOPT_NOP << 16) |
665 (TCPOPT_SACK << 8) |
666 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
667 TCPOLEN_SACK_PERBLOCK)));
668
669 for (this_sack = 0; this_sack < opts->num_sack_blocks;
670 ++this_sack) {
671 *ptr++ = htonl(sp[this_sack].start_seq);
672 *ptr++ = htonl(sp[this_sack].end_seq);
673 }
674
675 tp->rx_opt.dsack = 0;
676 }
677
678 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
679 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
680 u8 *p = (u8 *)ptr;
681 u32 len; /* Fast Open option length */
682
683 if (foc->exp) {
684 len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
685 *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
686 TCPOPT_FASTOPEN_MAGIC);
687 p += TCPOLEN_EXP_FASTOPEN_BASE;
688 } else {
689 len = TCPOLEN_FASTOPEN_BASE + foc->len;
690 *p++ = TCPOPT_FASTOPEN;
691 *p++ = len;
692 }
693
694 memcpy(p, foc->val, foc->len);
695 if ((len & 3) == 2) {
696 p[foc->len] = TCPOPT_NOP;
697 p[foc->len + 1] = TCPOPT_NOP;
698 }
699 ptr += (len + 3) >> 2;
700 }
701
702 smc_options_write(ptr, &options);
703
704 mptcp_options_write(th, ptr, tp, opts);
705}
706
707static void smc_set_option(const struct tcp_sock *tp,
708 struct tcp_out_options *opts,
709 unsigned int *remaining)
710{
711#if IS_ENABLED(CONFIG_SMC)
712 if (static_branch_unlikely(&tcp_have_smc)) {
713 if (tp->syn_smc) {
714 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
715 opts->options |= OPTION_SMC;
716 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
717 }
718 }
719 }
720#endif
721}
722
723static void smc_set_option_cond(const struct tcp_sock *tp,
724 const struct inet_request_sock *ireq,
725 struct tcp_out_options *opts,
726 unsigned int *remaining)
727{
728#if IS_ENABLED(CONFIG_SMC)
729 if (static_branch_unlikely(&tcp_have_smc)) {
730 if (tp->syn_smc && ireq->smc_ok) {
731 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
732 opts->options |= OPTION_SMC;
733 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
734 }
735 }
736 }
737#endif
738}
739
740static void mptcp_set_option_cond(const struct request_sock *req,
741 struct tcp_out_options *opts,
742 unsigned int *remaining)
743{
744 if (rsk_is_mptcp(req)) {
745 unsigned int size;
746
747 if (mptcp_synack_options(req, &size, &opts->mptcp)) {
748 if (*remaining >= size) {
749 opts->options |= OPTION_MPTCP;
750 *remaining -= size;
751 }
752 }
753 }
754}
755
756/* Compute TCP options for SYN packets. This is not the final
757 * network wire format yet.
758 */
759static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
760 struct tcp_out_options *opts,
761 struct tcp_md5sig_key **md5)
762{
763 struct tcp_sock *tp = tcp_sk(sk);
764 unsigned int remaining = MAX_TCP_OPTION_SPACE;
765 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
766
767 *md5 = NULL;
768#ifdef CONFIG_TCP_MD5SIG
769 if (static_branch_unlikely(&tcp_md5_needed.key) &&
770 rcu_access_pointer(tp->md5sig_info)) {
771 *md5 = tp->af_specific->md5_lookup(sk, sk);
772 if (*md5) {
773 opts->options |= OPTION_MD5;
774 remaining -= TCPOLEN_MD5SIG_ALIGNED;
775 }
776 }
777#endif
778
779 /* We always get an MSS option. The option bytes which will be seen in
780 * normal data packets should timestamps be used, must be in the MSS
781 * advertised. But we subtract them from tp->mss_cache so that
782 * calculations in tcp_sendmsg are simpler etc. So account for this
783 * fact here if necessary. If we don't do this correctly, as a
784 * receiver we won't recognize data packets as being full sized when we
785 * should, and thus we won't abide by the delayed ACK rules correctly.
786 * SACKs don't matter, we never delay an ACK when we have any of those
787 * going out. */
788 opts->mss = tcp_advertise_mss(sk);
789 remaining -= TCPOLEN_MSS_ALIGNED;
790
791 if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_timestamps) && !*md5)) {
792 opts->options |= OPTION_TS;
793 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
794 opts->tsecr = tp->rx_opt.ts_recent;
795 remaining -= TCPOLEN_TSTAMP_ALIGNED;
796 }
797 if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_window_scaling))) {
798 opts->ws = tp->rx_opt.rcv_wscale;
799 opts->options |= OPTION_WSCALE;
800 remaining -= TCPOLEN_WSCALE_ALIGNED;
801 }
802 if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_sack))) {
803 opts->options |= OPTION_SACK_ADVERTISE;
804 if (unlikely(!(OPTION_TS & opts->options)))
805 remaining -= TCPOLEN_SACKPERM_ALIGNED;
806 }
807
808 if (fastopen && fastopen->cookie.len >= 0) {
809 u32 need = fastopen->cookie.len;
810
811 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
812 TCPOLEN_FASTOPEN_BASE;
813 need = (need + 3) & ~3U; /* Align to 32 bits */
814 if (remaining >= need) {
815 opts->options |= OPTION_FAST_OPEN_COOKIE;
816 opts->fastopen_cookie = &fastopen->cookie;
817 remaining -= need;
818 tp->syn_fastopen = 1;
819 tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
820 }
821 }
822
823 smc_set_option(tp, opts, &remaining);
824
825 if (sk_is_mptcp(sk)) {
826 unsigned int size;
827
828 if (mptcp_syn_options(sk, skb, &size, &opts->mptcp)) {
829 opts->options |= OPTION_MPTCP;
830 remaining -= size;
831 }
832 }
833
834 bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining);
835
836 return MAX_TCP_OPTION_SPACE - remaining;
837}
838
839/* Set up TCP options for SYN-ACKs. */
840static unsigned int tcp_synack_options(const struct sock *sk,
841 struct request_sock *req,
842 unsigned int mss, struct sk_buff *skb,
843 struct tcp_out_options *opts,
844 const struct tcp_md5sig_key *md5,
845 struct tcp_fastopen_cookie *foc,
846 enum tcp_synack_type synack_type,
847 struct sk_buff *syn_skb)
848{
849 struct inet_request_sock *ireq = inet_rsk(req);
850 unsigned int remaining = MAX_TCP_OPTION_SPACE;
851
852#ifdef CONFIG_TCP_MD5SIG
853 if (md5) {
854 opts->options |= OPTION_MD5;
855 remaining -= TCPOLEN_MD5SIG_ALIGNED;
856
857 /* We can't fit any SACK blocks in a packet with MD5 + TS
858 * options. There was discussion about disabling SACK
859 * rather than TS in order to fit in better with old,
860 * buggy kernels, but that was deemed to be unnecessary.
861 */
862 if (synack_type != TCP_SYNACK_COOKIE)
863 ireq->tstamp_ok &= !ireq->sack_ok;
864 }
865#endif
866
867 /* We always send an MSS option. */
868 opts->mss = mss;
869 remaining -= TCPOLEN_MSS_ALIGNED;
870
871 if (likely(ireq->wscale_ok)) {
872 opts->ws = ireq->rcv_wscale;
873 opts->options |= OPTION_WSCALE;
874 remaining -= TCPOLEN_WSCALE_ALIGNED;
875 }
876 if (likely(ireq->tstamp_ok)) {
877 opts->options |= OPTION_TS;
878 opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off;
879 opts->tsecr = req->ts_recent;
880 remaining -= TCPOLEN_TSTAMP_ALIGNED;
881 }
882 if (likely(ireq->sack_ok)) {
883 opts->options |= OPTION_SACK_ADVERTISE;
884 if (unlikely(!ireq->tstamp_ok))
885 remaining -= TCPOLEN_SACKPERM_ALIGNED;
886 }
887 if (foc != NULL && foc->len >= 0) {
888 u32 need = foc->len;
889
890 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
891 TCPOLEN_FASTOPEN_BASE;
892 need = (need + 3) & ~3U; /* Align to 32 bits */
893 if (remaining >= need) {
894 opts->options |= OPTION_FAST_OPEN_COOKIE;
895 opts->fastopen_cookie = foc;
896 remaining -= need;
897 }
898 }
899
900 mptcp_set_option_cond(req, opts, &remaining);
901
902 smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining);
903
904 bpf_skops_hdr_opt_len((struct sock *)sk, skb, req, syn_skb,
905 synack_type, opts, &remaining);
906
907 return MAX_TCP_OPTION_SPACE - remaining;
908}
909
910/* Compute TCP options for ESTABLISHED sockets. This is not the
911 * final wire format yet.
912 */
913static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
914 struct tcp_out_options *opts,
915 struct tcp_md5sig_key **md5)
916{
917 struct tcp_sock *tp = tcp_sk(sk);
918 unsigned int size = 0;
919 unsigned int eff_sacks;
920
921 opts->options = 0;
922
923 *md5 = NULL;
924#ifdef CONFIG_TCP_MD5SIG
925 if (static_branch_unlikely(&tcp_md5_needed.key) &&
926 rcu_access_pointer(tp->md5sig_info)) {
927 *md5 = tp->af_specific->md5_lookup(sk, sk);
928 if (*md5) {
929 opts->options |= OPTION_MD5;
930 size += TCPOLEN_MD5SIG_ALIGNED;
931 }
932 }
933#endif
934
935 if (likely(tp->rx_opt.tstamp_ok)) {
936 opts->options |= OPTION_TS;
937 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
938 opts->tsecr = tp->rx_opt.ts_recent;
939 size += TCPOLEN_TSTAMP_ALIGNED;
940 }
941
942 /* MPTCP options have precedence over SACK for the limited TCP
943 * option space because a MPTCP connection would be forced to
944 * fall back to regular TCP if a required multipath option is
945 * missing. SACK still gets a chance to use whatever space is
946 * left.
947 */
948 if (sk_is_mptcp(sk)) {
949 unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
950 unsigned int opt_size = 0;
951
952 if (mptcp_established_options(sk, skb, &opt_size, remaining,
953 &opts->mptcp)) {
954 opts->options |= OPTION_MPTCP;
955 size += opt_size;
956 }
957 }
958
959 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
960 if (unlikely(eff_sacks)) {
961 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
962 if (unlikely(remaining < TCPOLEN_SACK_BASE_ALIGNED +
963 TCPOLEN_SACK_PERBLOCK))
964 return size;
965
966 opts->num_sack_blocks =
967 min_t(unsigned int, eff_sacks,
968 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
969 TCPOLEN_SACK_PERBLOCK);
970
971 size += TCPOLEN_SACK_BASE_ALIGNED +
972 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
973 }
974
975 if (unlikely(BPF_SOCK_OPS_TEST_FLAG(tp,
976 BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG))) {
977 unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
978
979 bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining);
980
981 size = MAX_TCP_OPTION_SPACE - remaining;
982 }
983
984 return size;
985}
986
987
988/* TCP SMALL QUEUES (TSQ)
989 *
990 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
991 * to reduce RTT and bufferbloat.
992 * We do this using a special skb destructor (tcp_wfree).
993 *
994 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
995 * needs to be reallocated in a driver.
996 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
997 *
998 * Since transmit from skb destructor is forbidden, we use a tasklet
999 * to process all sockets that eventually need to send more skbs.
1000 * We use one tasklet per cpu, with its own queue of sockets.
1001 */
1002struct tsq_tasklet {
1003 struct tasklet_struct tasklet;
1004 struct list_head head; /* queue of tcp sockets */
1005};
1006static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
1007
1008static void tcp_tsq_write(struct sock *sk)
1009{
1010 if ((1 << sk->sk_state) &
1011 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
1012 TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) {
1013 struct tcp_sock *tp = tcp_sk(sk);
1014
1015 if (tp->lost_out > tp->retrans_out &&
1016 tcp_snd_cwnd(tp) > tcp_packets_in_flight(tp)) {
1017 tcp_mstamp_refresh(tp);
1018 tcp_xmit_retransmit_queue(sk);
1019 }
1020
1021 tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
1022 0, GFP_ATOMIC);
1023 }
1024}
1025
1026static void tcp_tsq_handler(struct sock *sk)
1027{
1028 bh_lock_sock(sk);
1029 if (!sock_owned_by_user(sk))
1030 tcp_tsq_write(sk);
1031 else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
1032 sock_hold(sk);
1033 bh_unlock_sock(sk);
1034}
1035/*
1036 * One tasklet per cpu tries to send more skbs.
1037 * We run in tasklet context but need to disable irqs when
1038 * transferring tsq->head because tcp_wfree() might
1039 * interrupt us (non NAPI drivers)
1040 */
1041static void tcp_tasklet_func(struct tasklet_struct *t)
1042{
1043 struct tsq_tasklet *tsq = from_tasklet(tsq, t, tasklet);
1044 LIST_HEAD(list);
1045 unsigned long flags;
1046 struct list_head *q, *n;
1047 struct tcp_sock *tp;
1048 struct sock *sk;
1049
1050 local_irq_save(flags);
1051 list_splice_init(&tsq->head, &list);
1052 local_irq_restore(flags);
1053
1054 list_for_each_safe(q, n, &list) {
1055 tp = list_entry(q, struct tcp_sock, tsq_node);
1056 list_del(&tp->tsq_node);
1057
1058 sk = (struct sock *)tp;
1059 smp_mb__before_atomic();
1060 clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
1061
1062 tcp_tsq_handler(sk);
1063 sk_free(sk);
1064 }
1065}
1066
1067#define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
1068 TCPF_WRITE_TIMER_DEFERRED | \
1069 TCPF_DELACK_TIMER_DEFERRED | \
1070 TCPF_MTU_REDUCED_DEFERRED)
1071/**
1072 * tcp_release_cb - tcp release_sock() callback
1073 * @sk: socket
1074 *
1075 * called from release_sock() to perform protocol dependent
1076 * actions before socket release.
1077 */
1078void tcp_release_cb(struct sock *sk)
1079{
1080 unsigned long flags = smp_load_acquire(&sk->sk_tsq_flags);
1081 unsigned long nflags;
1082
1083 /* perform an atomic operation only if at least one flag is set */
1084 do {
1085 if (!(flags & TCP_DEFERRED_ALL))
1086 return;
1087 nflags = flags & ~TCP_DEFERRED_ALL;
1088 } while (!try_cmpxchg(&sk->sk_tsq_flags, &flags, nflags));
1089
1090 if (flags & TCPF_TSQ_DEFERRED) {
1091 tcp_tsq_write(sk);
1092 __sock_put(sk);
1093 }
1094 /* Here begins the tricky part :
1095 * We are called from release_sock() with :
1096 * 1) BH disabled
1097 * 2) sk_lock.slock spinlock held
1098 * 3) socket owned by us (sk->sk_lock.owned == 1)
1099 *
1100 * But following code is meant to be called from BH handlers,
1101 * so we should keep BH disabled, but early release socket ownership
1102 */
1103 sock_release_ownership(sk);
1104
1105 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
1106 tcp_write_timer_handler(sk);
1107 __sock_put(sk);
1108 }
1109 if (flags & TCPF_DELACK_TIMER_DEFERRED) {
1110 tcp_delack_timer_handler(sk);
1111 __sock_put(sk);
1112 }
1113 if (flags & TCPF_MTU_REDUCED_DEFERRED) {
1114 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
1115 __sock_put(sk);
1116 }
1117}
1118EXPORT_SYMBOL(tcp_release_cb);
1119
1120void __init tcp_tasklet_init(void)
1121{
1122 int i;
1123
1124 for_each_possible_cpu(i) {
1125 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
1126
1127 INIT_LIST_HEAD(&tsq->head);
1128 tasklet_setup(&tsq->tasklet, tcp_tasklet_func);
1129 }
1130}
1131
1132/*
1133 * Write buffer destructor automatically called from kfree_skb.
1134 * We can't xmit new skbs from this context, as we might already
1135 * hold qdisc lock.
1136 */
1137void tcp_wfree(struct sk_buff *skb)
1138{
1139 struct sock *sk = skb->sk;
1140 struct tcp_sock *tp = tcp_sk(sk);
1141 unsigned long flags, nval, oval;
1142 struct tsq_tasklet *tsq;
1143 bool empty;
1144
1145 /* Keep one reference on sk_wmem_alloc.
1146 * Will be released by sk_free() from here or tcp_tasklet_func()
1147 */
1148 WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc));
1149
1150 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
1151 * Wait until our queues (qdisc + devices) are drained.
1152 * This gives :
1153 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
1154 * - chance for incoming ACK (processed by another cpu maybe)
1155 * to migrate this flow (skb->ooo_okay will be eventually set)
1156 */
1157 if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
1158 goto out;
1159
1160 oval = smp_load_acquire(&sk->sk_tsq_flags);
1161 do {
1162 if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
1163 goto out;
1164
1165 nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED;
1166 } while (!try_cmpxchg(&sk->sk_tsq_flags, &oval, nval));
1167
1168 /* queue this socket to tasklet queue */
1169 local_irq_save(flags);
1170 tsq = this_cpu_ptr(&tsq_tasklet);
1171 empty = list_empty(&tsq->head);
1172 list_add(&tp->tsq_node, &tsq->head);
1173 if (empty)
1174 tasklet_schedule(&tsq->tasklet);
1175 local_irq_restore(flags);
1176 return;
1177out:
1178 sk_free(sk);
1179}
1180
1181/* Note: Called under soft irq.
1182 * We can call TCP stack right away, unless socket is owned by user.
1183 */
1184enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer)
1185{
1186 struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer);
1187 struct sock *sk = (struct sock *)tp;
1188
1189 tcp_tsq_handler(sk);
1190 sock_put(sk);
1191
1192 return HRTIMER_NORESTART;
1193}
1194
1195static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb,
1196 u64 prior_wstamp)
1197{
1198 struct tcp_sock *tp = tcp_sk(sk);
1199
1200 if (sk->sk_pacing_status != SK_PACING_NONE) {
1201 unsigned long rate = sk->sk_pacing_rate;
1202
1203 /* Original sch_fq does not pace first 10 MSS
1204 * Note that tp->data_segs_out overflows after 2^32 packets,
1205 * this is a minor annoyance.
1206 */
1207 if (rate != ~0UL && rate && tp->data_segs_out >= 10) {
1208 u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate);
1209 u64 credit = tp->tcp_wstamp_ns - prior_wstamp;
1210
1211 /* take into account OS jitter */
1212 len_ns -= min_t(u64, len_ns / 2, credit);
1213 tp->tcp_wstamp_ns += len_ns;
1214 }
1215 }
1216 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
1217}
1218
1219INDIRECT_CALLABLE_DECLARE(int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl));
1220INDIRECT_CALLABLE_DECLARE(int inet6_csk_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl));
1221INDIRECT_CALLABLE_DECLARE(void tcp_v4_send_check(struct sock *sk, struct sk_buff *skb));
1222
1223/* This routine actually transmits TCP packets queued in by
1224 * tcp_do_sendmsg(). This is used by both the initial
1225 * transmission and possible later retransmissions.
1226 * All SKB's seen here are completely headerless. It is our
1227 * job to build the TCP header, and pass the packet down to
1228 * IP so it can do the same plus pass the packet off to the
1229 * device.
1230 *
1231 * We are working here with either a clone of the original
1232 * SKB, or a fresh unique copy made by the retransmit engine.
1233 */
1234static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb,
1235 int clone_it, gfp_t gfp_mask, u32 rcv_nxt)
1236{
1237 const struct inet_connection_sock *icsk = inet_csk(sk);
1238 struct inet_sock *inet;
1239 struct tcp_sock *tp;
1240 struct tcp_skb_cb *tcb;
1241 struct tcp_out_options opts;
1242 unsigned int tcp_options_size, tcp_header_size;
1243 struct sk_buff *oskb = NULL;
1244 struct tcp_md5sig_key *md5;
1245 struct tcphdr *th;
1246 u64 prior_wstamp;
1247 int err;
1248
1249 BUG_ON(!skb || !tcp_skb_pcount(skb));
1250 tp = tcp_sk(sk);
1251 prior_wstamp = tp->tcp_wstamp_ns;
1252 tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache);
1253 skb_set_delivery_time(skb, tp->tcp_wstamp_ns, true);
1254 if (clone_it) {
1255 oskb = skb;
1256
1257 tcp_skb_tsorted_save(oskb) {
1258 if (unlikely(skb_cloned(oskb)))
1259 skb = pskb_copy(oskb, gfp_mask);
1260 else
1261 skb = skb_clone(oskb, gfp_mask);
1262 } tcp_skb_tsorted_restore(oskb);
1263
1264 if (unlikely(!skb))
1265 return -ENOBUFS;
1266 /* retransmit skbs might have a non zero value in skb->dev
1267 * because skb->dev is aliased with skb->rbnode.rb_left
1268 */
1269 skb->dev = NULL;
1270 }
1271
1272 inet = inet_sk(sk);
1273 tcb = TCP_SKB_CB(skb);
1274 memset(&opts, 0, sizeof(opts));
1275
1276 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
1277 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
1278 } else {
1279 tcp_options_size = tcp_established_options(sk, skb, &opts,
1280 &md5);
1281 /* Force a PSH flag on all (GSO) packets to expedite GRO flush
1282 * at receiver : This slightly improve GRO performance.
1283 * Note that we do not force the PSH flag for non GSO packets,
1284 * because they might be sent under high congestion events,
1285 * and in this case it is better to delay the delivery of 1-MSS
1286 * packets and thus the corresponding ACK packet that would
1287 * release the following packet.
1288 */
1289 if (tcp_skb_pcount(skb) > 1)
1290 tcb->tcp_flags |= TCPHDR_PSH;
1291 }
1292 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1293
1294 /* if no packet is in qdisc/device queue, then allow XPS to select
1295 * another queue. We can be called from tcp_tsq_handler()
1296 * which holds one reference to sk.
1297 *
1298 * TODO: Ideally, in-flight pure ACK packets should not matter here.
1299 * One way to get this would be to set skb->truesize = 2 on them.
1300 */
1301 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
1302
1303 /* If we had to use memory reserve to allocate this skb,
1304 * this might cause drops if packet is looped back :
1305 * Other socket might not have SOCK_MEMALLOC.
1306 * Packets not looped back do not care about pfmemalloc.
1307 */
1308 skb->pfmemalloc = 0;
1309
1310 skb_push(skb, tcp_header_size);
1311 skb_reset_transport_header(skb);
1312
1313 skb_orphan(skb);
1314 skb->sk = sk;
1315 skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
1316 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
1317
1318 skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm);
1319
1320 /* Build TCP header and checksum it. */
1321 th = (struct tcphdr *)skb->data;
1322 th->source = inet->inet_sport;
1323 th->dest = inet->inet_dport;
1324 th->seq = htonl(tcb->seq);
1325 th->ack_seq = htonl(rcv_nxt);
1326 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
1327 tcb->tcp_flags);
1328
1329 th->check = 0;
1330 th->urg_ptr = 0;
1331
1332 /* The urg_mode check is necessary during a below snd_una win probe */
1333 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1334 if (before(tp->snd_up, tcb->seq + 0x10000)) {
1335 th->urg_ptr = htons(tp->snd_up - tcb->seq);
1336 th->urg = 1;
1337 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1338 th->urg_ptr = htons(0xFFFF);
1339 th->urg = 1;
1340 }
1341 }
1342
1343 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1344 if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1345 th->window = htons(tcp_select_window(sk));
1346 tcp_ecn_send(sk, skb, th, tcp_header_size);
1347 } else {
1348 /* RFC1323: The window in SYN & SYN/ACK segments
1349 * is never scaled.
1350 */
1351 th->window = htons(min(tp->rcv_wnd, 65535U));
1352 }
1353
1354 tcp_options_write(th, tp, &opts);
1355
1356#ifdef CONFIG_TCP_MD5SIG
1357 /* Calculate the MD5 hash, as we have all we need now */
1358 if (md5) {
1359 sk_gso_disable(sk);
1360 tp->af_specific->calc_md5_hash(opts.hash_location,
1361 md5, sk, skb);
1362 }
1363#endif
1364
1365 /* BPF prog is the last one writing header option */
1366 bpf_skops_write_hdr_opt(sk, skb, NULL, NULL, 0, &opts);
1367
1368 INDIRECT_CALL_INET(icsk->icsk_af_ops->send_check,
1369 tcp_v6_send_check, tcp_v4_send_check,
1370 sk, skb);
1371
1372 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1373 tcp_event_ack_sent(sk, tcp_skb_pcount(skb), rcv_nxt);
1374
1375 if (skb->len != tcp_header_size) {
1376 tcp_event_data_sent(tp, sk);
1377 tp->data_segs_out += tcp_skb_pcount(skb);
1378 tp->bytes_sent += skb->len - tcp_header_size;
1379 }
1380
1381 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1382 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1383 tcp_skb_pcount(skb));
1384
1385 tp->segs_out += tcp_skb_pcount(skb);
1386 skb_set_hash_from_sk(skb, sk);
1387 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1388 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1389 skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1390
1391 /* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */
1392
1393 /* Cleanup our debris for IP stacks */
1394 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1395 sizeof(struct inet6_skb_parm)));
1396
1397 tcp_add_tx_delay(skb, tp);
1398
1399 err = INDIRECT_CALL_INET(icsk->icsk_af_ops->queue_xmit,
1400 inet6_csk_xmit, ip_queue_xmit,
1401 sk, skb, &inet->cork.fl);
1402
1403 if (unlikely(err > 0)) {
1404 tcp_enter_cwr(sk);
1405 err = net_xmit_eval(err);
1406 }
1407 if (!err && oskb) {
1408 tcp_update_skb_after_send(sk, oskb, prior_wstamp);
1409 tcp_rate_skb_sent(sk, oskb);
1410 }
1411 return err;
1412}
1413
1414static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
1415 gfp_t gfp_mask)
1416{
1417 return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask,
1418 tcp_sk(sk)->rcv_nxt);
1419}
1420
1421/* This routine just queues the buffer for sending.
1422 *
1423 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1424 * otherwise socket can stall.
1425 */
1426static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1427{
1428 struct tcp_sock *tp = tcp_sk(sk);
1429
1430 /* Advance write_seq and place onto the write_queue. */
1431 WRITE_ONCE(tp->write_seq, TCP_SKB_CB(skb)->end_seq);
1432 __skb_header_release(skb);
1433 tcp_add_write_queue_tail(sk, skb);
1434 sk_wmem_queued_add(sk, skb->truesize);
1435 sk_mem_charge(sk, skb->truesize);
1436}
1437
1438/* Initialize TSO segments for a packet. */
1439static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1440{
1441 if (skb->len <= mss_now) {
1442 /* Avoid the costly divide in the normal
1443 * non-TSO case.
1444 */
1445 tcp_skb_pcount_set(skb, 1);
1446 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1447 } else {
1448 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1449 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1450 }
1451}
1452
1453/* Pcount in the middle of the write queue got changed, we need to do various
1454 * tweaks to fix counters
1455 */
1456static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1457{
1458 struct tcp_sock *tp = tcp_sk(sk);
1459
1460 tp->packets_out -= decr;
1461
1462 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1463 tp->sacked_out -= decr;
1464 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1465 tp->retrans_out -= decr;
1466 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1467 tp->lost_out -= decr;
1468
1469 /* Reno case is special. Sigh... */
1470 if (tcp_is_reno(tp) && decr > 0)
1471 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1472
1473 if (tp->lost_skb_hint &&
1474 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1475 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
1476 tp->lost_cnt_hint -= decr;
1477
1478 tcp_verify_left_out(tp);
1479}
1480
1481static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1482{
1483 return TCP_SKB_CB(skb)->txstamp_ack ||
1484 (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1485}
1486
1487static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1488{
1489 struct skb_shared_info *shinfo = skb_shinfo(skb);
1490
1491 if (unlikely(tcp_has_tx_tstamp(skb)) &&
1492 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1493 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1494 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1495
1496 shinfo->tx_flags &= ~tsflags;
1497 shinfo2->tx_flags |= tsflags;
1498 swap(shinfo->tskey, shinfo2->tskey);
1499 TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1500 TCP_SKB_CB(skb)->txstamp_ack = 0;
1501 }
1502}
1503
1504static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1505{
1506 TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1507 TCP_SKB_CB(skb)->eor = 0;
1508}
1509
1510/* Insert buff after skb on the write or rtx queue of sk. */
1511static void tcp_insert_write_queue_after(struct sk_buff *skb,
1512 struct sk_buff *buff,
1513 struct sock *sk,
1514 enum tcp_queue tcp_queue)
1515{
1516 if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE)
1517 __skb_queue_after(&sk->sk_write_queue, skb, buff);
1518 else
1519 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
1520}
1521
1522/* Function to create two new TCP segments. Shrinks the given segment
1523 * to the specified size and appends a new segment with the rest of the
1524 * packet to the list. This won't be called frequently, I hope.
1525 * Remember, these are still headerless SKBs at this point.
1526 */
1527int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1528 struct sk_buff *skb, u32 len,
1529 unsigned int mss_now, gfp_t gfp)
1530{
1531 struct tcp_sock *tp = tcp_sk(sk);
1532 struct sk_buff *buff;
1533 int nsize, old_factor;
1534 long limit;
1535 int nlen;
1536 u8 flags;
1537
1538 if (WARN_ON(len > skb->len))
1539 return -EINVAL;
1540
1541 nsize = skb_headlen(skb) - len;
1542 if (nsize < 0)
1543 nsize = 0;
1544
1545 /* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb.
1546 * We need some allowance to not penalize applications setting small
1547 * SO_SNDBUF values.
1548 * Also allow first and last skb in retransmit queue to be split.
1549 */
1550 limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_LEGACY_MAX_SIZE);
1551 if (unlikely((sk->sk_wmem_queued >> 1) > limit &&
1552 tcp_queue != TCP_FRAG_IN_WRITE_QUEUE &&
1553 skb != tcp_rtx_queue_head(sk) &&
1554 skb != tcp_rtx_queue_tail(sk))) {
1555 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG);
1556 return -ENOMEM;
1557 }
1558
1559 if (skb_unclone_keeptruesize(skb, gfp))
1560 return -ENOMEM;
1561
1562 /* Get a new skb... force flag on. */
1563 buff = tcp_stream_alloc_skb(sk, nsize, gfp, true);
1564 if (!buff)
1565 return -ENOMEM; /* We'll just try again later. */
1566 skb_copy_decrypted(buff, skb);
1567 mptcp_skb_ext_copy(buff, skb);
1568
1569 sk_wmem_queued_add(sk, buff->truesize);
1570 sk_mem_charge(sk, buff->truesize);
1571 nlen = skb->len - len - nsize;
1572 buff->truesize += nlen;
1573 skb->truesize -= nlen;
1574
1575 /* Correct the sequence numbers. */
1576 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1577 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1578 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1579
1580 /* PSH and FIN should only be set in the second packet. */
1581 flags = TCP_SKB_CB(skb)->tcp_flags;
1582 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1583 TCP_SKB_CB(buff)->tcp_flags = flags;
1584 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1585 tcp_skb_fragment_eor(skb, buff);
1586
1587 skb_split(skb, buff, len);
1588
1589 skb_set_delivery_time(buff, skb->tstamp, true);
1590 tcp_fragment_tstamp(skb, buff);
1591
1592 old_factor = tcp_skb_pcount(skb);
1593
1594 /* Fix up tso_factor for both original and new SKB. */
1595 tcp_set_skb_tso_segs(skb, mss_now);
1596 tcp_set_skb_tso_segs(buff, mss_now);
1597
1598 /* Update delivered info for the new segment */
1599 TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1600
1601 /* If this packet has been sent out already, we must
1602 * adjust the various packet counters.
1603 */
1604 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1605 int diff = old_factor - tcp_skb_pcount(skb) -
1606 tcp_skb_pcount(buff);
1607
1608 if (diff)
1609 tcp_adjust_pcount(sk, skb, diff);
1610 }
1611
1612 /* Link BUFF into the send queue. */
1613 __skb_header_release(buff);
1614 tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1615 if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE)
1616 list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor);
1617
1618 return 0;
1619}
1620
1621/* This is similar to __pskb_pull_tail(). The difference is that pulled
1622 * data is not copied, but immediately discarded.
1623 */
1624static int __pskb_trim_head(struct sk_buff *skb, int len)
1625{
1626 struct skb_shared_info *shinfo;
1627 int i, k, eat;
1628
1629 eat = min_t(int, len, skb_headlen(skb));
1630 if (eat) {
1631 __skb_pull(skb, eat);
1632 len -= eat;
1633 if (!len)
1634 return 0;
1635 }
1636 eat = len;
1637 k = 0;
1638 shinfo = skb_shinfo(skb);
1639 for (i = 0; i < shinfo->nr_frags; i++) {
1640 int size = skb_frag_size(&shinfo->frags[i]);
1641
1642 if (size <= eat) {
1643 skb_frag_unref(skb, i);
1644 eat -= size;
1645 } else {
1646 shinfo->frags[k] = shinfo->frags[i];
1647 if (eat) {
1648 skb_frag_off_add(&shinfo->frags[k], eat);
1649 skb_frag_size_sub(&shinfo->frags[k], eat);
1650 eat = 0;
1651 }
1652 k++;
1653 }
1654 }
1655 shinfo->nr_frags = k;
1656
1657 skb->data_len -= len;
1658 skb->len = skb->data_len;
1659 return len;
1660}
1661
1662/* Remove acked data from a packet in the transmit queue. */
1663int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1664{
1665 u32 delta_truesize;
1666
1667 if (skb_unclone_keeptruesize(skb, GFP_ATOMIC))
1668 return -ENOMEM;
1669
1670 delta_truesize = __pskb_trim_head(skb, len);
1671
1672 TCP_SKB_CB(skb)->seq += len;
1673
1674 if (delta_truesize) {
1675 skb->truesize -= delta_truesize;
1676 sk_wmem_queued_add(sk, -delta_truesize);
1677 if (!skb_zcopy_pure(skb))
1678 sk_mem_uncharge(sk, delta_truesize);
1679 }
1680
1681 /* Any change of skb->len requires recalculation of tso factor. */
1682 if (tcp_skb_pcount(skb) > 1)
1683 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1684
1685 return 0;
1686}
1687
1688/* Calculate MSS not accounting any TCP options. */
1689static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1690{
1691 const struct tcp_sock *tp = tcp_sk(sk);
1692 const struct inet_connection_sock *icsk = inet_csk(sk);
1693 int mss_now;
1694
1695 /* Calculate base mss without TCP options:
1696 It is MMS_S - sizeof(tcphdr) of rfc1122
1697 */
1698 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1699
1700 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1701 if (icsk->icsk_af_ops->net_frag_header_len) {
1702 const struct dst_entry *dst = __sk_dst_get(sk);
1703
1704 if (dst && dst_allfrag(dst))
1705 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1706 }
1707
1708 /* Clamp it (mss_clamp does not include tcp options) */
1709 if (mss_now > tp->rx_opt.mss_clamp)
1710 mss_now = tp->rx_opt.mss_clamp;
1711
1712 /* Now subtract optional transport overhead */
1713 mss_now -= icsk->icsk_ext_hdr_len;
1714
1715 /* Then reserve room for full set of TCP options and 8 bytes of data */
1716 mss_now = max(mss_now,
1717 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss));
1718 return mss_now;
1719}
1720
1721/* Calculate MSS. Not accounting for SACKs here. */
1722int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1723{
1724 /* Subtract TCP options size, not including SACKs */
1725 return __tcp_mtu_to_mss(sk, pmtu) -
1726 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1727}
1728EXPORT_SYMBOL(tcp_mtu_to_mss);
1729
1730/* Inverse of above */
1731int tcp_mss_to_mtu(struct sock *sk, int mss)
1732{
1733 const struct tcp_sock *tp = tcp_sk(sk);
1734 const struct inet_connection_sock *icsk = inet_csk(sk);
1735 int mtu;
1736
1737 mtu = mss +
1738 tp->tcp_header_len +
1739 icsk->icsk_ext_hdr_len +
1740 icsk->icsk_af_ops->net_header_len;
1741
1742 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1743 if (icsk->icsk_af_ops->net_frag_header_len) {
1744 const struct dst_entry *dst = __sk_dst_get(sk);
1745
1746 if (dst && dst_allfrag(dst))
1747 mtu += icsk->icsk_af_ops->net_frag_header_len;
1748 }
1749 return mtu;
1750}
1751EXPORT_SYMBOL(tcp_mss_to_mtu);
1752
1753/* MTU probing init per socket */
1754void tcp_mtup_init(struct sock *sk)
1755{
1756 struct tcp_sock *tp = tcp_sk(sk);
1757 struct inet_connection_sock *icsk = inet_csk(sk);
1758 struct net *net = sock_net(sk);
1759
1760 icsk->icsk_mtup.enabled = READ_ONCE(net->ipv4.sysctl_tcp_mtu_probing) > 1;
1761 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1762 icsk->icsk_af_ops->net_header_len;
1763 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, READ_ONCE(net->ipv4.sysctl_tcp_base_mss));
1764 icsk->icsk_mtup.probe_size = 0;
1765 if (icsk->icsk_mtup.enabled)
1766 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
1767}
1768EXPORT_SYMBOL(tcp_mtup_init);
1769
1770/* This function synchronize snd mss to current pmtu/exthdr set.
1771
1772 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1773 for TCP options, but includes only bare TCP header.
1774
1775 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1776 It is minimum of user_mss and mss received with SYN.
1777 It also does not include TCP options.
1778
1779 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1780
1781 tp->mss_cache is current effective sending mss, including
1782 all tcp options except for SACKs. It is evaluated,
1783 taking into account current pmtu, but never exceeds
1784 tp->rx_opt.mss_clamp.
1785
1786 NOTE1. rfc1122 clearly states that advertised MSS
1787 DOES NOT include either tcp or ip options.
1788
1789 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1790 are READ ONLY outside this function. --ANK (980731)
1791 */
1792unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1793{
1794 struct tcp_sock *tp = tcp_sk(sk);
1795 struct inet_connection_sock *icsk = inet_csk(sk);
1796 int mss_now;
1797
1798 if (icsk->icsk_mtup.search_high > pmtu)
1799 icsk->icsk_mtup.search_high = pmtu;
1800
1801 mss_now = tcp_mtu_to_mss(sk, pmtu);
1802 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1803
1804 /* And store cached results */
1805 icsk->icsk_pmtu_cookie = pmtu;
1806 if (icsk->icsk_mtup.enabled)
1807 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1808 tp->mss_cache = mss_now;
1809
1810 return mss_now;
1811}
1812EXPORT_SYMBOL(tcp_sync_mss);
1813
1814/* Compute the current effective MSS, taking SACKs and IP options,
1815 * and even PMTU discovery events into account.
1816 */
1817unsigned int tcp_current_mss(struct sock *sk)
1818{
1819 const struct tcp_sock *tp = tcp_sk(sk);
1820 const struct dst_entry *dst = __sk_dst_get(sk);
1821 u32 mss_now;
1822 unsigned int header_len;
1823 struct tcp_out_options opts;
1824 struct tcp_md5sig_key *md5;
1825
1826 mss_now = tp->mss_cache;
1827
1828 if (dst) {
1829 u32 mtu = dst_mtu(dst);
1830 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1831 mss_now = tcp_sync_mss(sk, mtu);
1832 }
1833
1834 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1835 sizeof(struct tcphdr);
1836 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1837 * some common options. If this is an odd packet (because we have SACK
1838 * blocks etc) then our calculated header_len will be different, and
1839 * we have to adjust mss_now correspondingly */
1840 if (header_len != tp->tcp_header_len) {
1841 int delta = (int) header_len - tp->tcp_header_len;
1842 mss_now -= delta;
1843 }
1844
1845 return mss_now;
1846}
1847
1848/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1849 * As additional protections, we do not touch cwnd in retransmission phases,
1850 * and if application hit its sndbuf limit recently.
1851 */
1852static void tcp_cwnd_application_limited(struct sock *sk)
1853{
1854 struct tcp_sock *tp = tcp_sk(sk);
1855
1856 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1857 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1858 /* Limited by application or receiver window. */
1859 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1860 u32 win_used = max(tp->snd_cwnd_used, init_win);
1861 if (win_used < tcp_snd_cwnd(tp)) {
1862 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1863 tcp_snd_cwnd_set(tp, (tcp_snd_cwnd(tp) + win_used) >> 1);
1864 }
1865 tp->snd_cwnd_used = 0;
1866 }
1867 tp->snd_cwnd_stamp = tcp_jiffies32;
1868}
1869
1870static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1871{
1872 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1873 struct tcp_sock *tp = tcp_sk(sk);
1874
1875 /* Track the strongest available signal of the degree to which the cwnd
1876 * is fully utilized. If cwnd-limited then remember that fact for the
1877 * current window. If not cwnd-limited then track the maximum number of
1878 * outstanding packets in the current window. (If cwnd-limited then we
1879 * chose to not update tp->max_packets_out to avoid an extra else
1880 * clause with no functional impact.)
1881 */
1882 if (!before(tp->snd_una, tp->cwnd_usage_seq) ||
1883 is_cwnd_limited ||
1884 (!tp->is_cwnd_limited &&
1885 tp->packets_out > tp->max_packets_out)) {
1886 tp->is_cwnd_limited = is_cwnd_limited;
1887 tp->max_packets_out = tp->packets_out;
1888 tp->cwnd_usage_seq = tp->snd_nxt;
1889 }
1890
1891 if (tcp_is_cwnd_limited(sk)) {
1892 /* Network is feed fully. */
1893 tp->snd_cwnd_used = 0;
1894 tp->snd_cwnd_stamp = tcp_jiffies32;
1895 } else {
1896 /* Network starves. */
1897 if (tp->packets_out > tp->snd_cwnd_used)
1898 tp->snd_cwnd_used = tp->packets_out;
1899
1900 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle) &&
1901 (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto &&
1902 !ca_ops->cong_control)
1903 tcp_cwnd_application_limited(sk);
1904
1905 /* The following conditions together indicate the starvation
1906 * is caused by insufficient sender buffer:
1907 * 1) just sent some data (see tcp_write_xmit)
1908 * 2) not cwnd limited (this else condition)
1909 * 3) no more data to send (tcp_write_queue_empty())
1910 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1911 */
1912 if (tcp_write_queue_empty(sk) && sk->sk_socket &&
1913 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
1914 (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1915 tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED);
1916 }
1917}
1918
1919/* Minshall's variant of the Nagle send check. */
1920static bool tcp_minshall_check(const struct tcp_sock *tp)
1921{
1922 return after(tp->snd_sml, tp->snd_una) &&
1923 !after(tp->snd_sml, tp->snd_nxt);
1924}
1925
1926/* Update snd_sml if this skb is under mss
1927 * Note that a TSO packet might end with a sub-mss segment
1928 * The test is really :
1929 * if ((skb->len % mss) != 0)
1930 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1931 * But we can avoid doing the divide again given we already have
1932 * skb_pcount = skb->len / mss_now
1933 */
1934static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1935 const struct sk_buff *skb)
1936{
1937 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1938 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1939}
1940
1941/* Return false, if packet can be sent now without violation Nagle's rules:
1942 * 1. It is full sized. (provided by caller in %partial bool)
1943 * 2. Or it contains FIN. (already checked by caller)
1944 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1945 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1946 * With Minshall's modification: all sent small packets are ACKed.
1947 */
1948static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1949 int nonagle)
1950{
1951 return partial &&
1952 ((nonagle & TCP_NAGLE_CORK) ||
1953 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1954}
1955
1956/* Return how many segs we'd like on a TSO packet,
1957 * depending on current pacing rate, and how close the peer is.
1958 *
1959 * Rationale is:
1960 * - For close peers, we rather send bigger packets to reduce
1961 * cpu costs, because occasional losses will be repaired fast.
1962 * - For long distance/rtt flows, we would like to get ACK clocking
1963 * with 1 ACK per ms.
1964 *
1965 * Use min_rtt to help adapt TSO burst size, with smaller min_rtt resulting
1966 * in bigger TSO bursts. We we cut the RTT-based allowance in half
1967 * for every 2^9 usec (aka 512 us) of RTT, so that the RTT-based allowance
1968 * is below 1500 bytes after 6 * ~500 usec = 3ms.
1969 */
1970static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
1971 int min_tso_segs)
1972{
1973 unsigned long bytes;
1974 u32 r;
1975
1976 bytes = sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift);
1977
1978 r = tcp_min_rtt(tcp_sk(sk)) >> READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_rtt_log);
1979 if (r < BITS_PER_TYPE(sk->sk_gso_max_size))
1980 bytes += sk->sk_gso_max_size >> r;
1981
1982 bytes = min_t(unsigned long, bytes, sk->sk_gso_max_size);
1983
1984 return max_t(u32, bytes / mss_now, min_tso_segs);
1985}
1986
1987/* Return the number of segments we want in the skb we are transmitting.
1988 * See if congestion control module wants to decide; otherwise, autosize.
1989 */
1990static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
1991{
1992 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1993 u32 min_tso, tso_segs;
1994
1995 min_tso = ca_ops->min_tso_segs ?
1996 ca_ops->min_tso_segs(sk) :
1997 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs);
1998
1999 tso_segs = tcp_tso_autosize(sk, mss_now, min_tso);
2000 return min_t(u32, tso_segs, sk->sk_gso_max_segs);
2001}
2002
2003/* Returns the portion of skb which can be sent right away */
2004static unsigned int tcp_mss_split_point(const struct sock *sk,
2005 const struct sk_buff *skb,
2006 unsigned int mss_now,
2007 unsigned int max_segs,
2008 int nonagle)
2009{
2010 const struct tcp_sock *tp = tcp_sk(sk);
2011 u32 partial, needed, window, max_len;
2012
2013 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2014 max_len = mss_now * max_segs;
2015
2016 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
2017 return max_len;
2018
2019 needed = min(skb->len, window);
2020
2021 if (max_len <= needed)
2022 return max_len;
2023
2024 partial = needed % mss_now;
2025 /* If last segment is not a full MSS, check if Nagle rules allow us
2026 * to include this last segment in this skb.
2027 * Otherwise, we'll split the skb at last MSS boundary
2028 */
2029 if (tcp_nagle_check(partial != 0, tp, nonagle))
2030 return needed - partial;
2031
2032 return needed;
2033}
2034
2035/* Can at least one segment of SKB be sent right now, according to the
2036 * congestion window rules? If so, return how many segments are allowed.
2037 */
2038static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
2039 const struct sk_buff *skb)
2040{
2041 u32 in_flight, cwnd, halfcwnd;
2042
2043 /* Don't be strict about the congestion window for the final FIN. */
2044 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
2045 tcp_skb_pcount(skb) == 1)
2046 return 1;
2047
2048 in_flight = tcp_packets_in_flight(tp);
2049 cwnd = tcp_snd_cwnd(tp);
2050 if (in_flight >= cwnd)
2051 return 0;
2052
2053 /* For better scheduling, ensure we have at least
2054 * 2 GSO packets in flight.
2055 */
2056 halfcwnd = max(cwnd >> 1, 1U);
2057 return min(halfcwnd, cwnd - in_flight);
2058}
2059
2060/* Initialize TSO state of a skb.
2061 * This must be invoked the first time we consider transmitting
2062 * SKB onto the wire.
2063 */
2064static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
2065{
2066 int tso_segs = tcp_skb_pcount(skb);
2067
2068 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
2069 tcp_set_skb_tso_segs(skb, mss_now);
2070 tso_segs = tcp_skb_pcount(skb);
2071 }
2072 return tso_segs;
2073}
2074
2075
2076/* Return true if the Nagle test allows this packet to be
2077 * sent now.
2078 */
2079static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
2080 unsigned int cur_mss, int nonagle)
2081{
2082 /* Nagle rule does not apply to frames, which sit in the middle of the
2083 * write_queue (they have no chances to get new data).
2084 *
2085 * This is implemented in the callers, where they modify the 'nonagle'
2086 * argument based upon the location of SKB in the send queue.
2087 */
2088 if (nonagle & TCP_NAGLE_PUSH)
2089 return true;
2090
2091 /* Don't use the nagle rule for urgent data (or for the final FIN). */
2092 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
2093 return true;
2094
2095 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
2096 return true;
2097
2098 return false;
2099}
2100
2101/* Does at least the first segment of SKB fit into the send window? */
2102static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
2103 const struct sk_buff *skb,
2104 unsigned int cur_mss)
2105{
2106 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
2107
2108 if (skb->len > cur_mss)
2109 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
2110
2111 return !after(end_seq, tcp_wnd_end(tp));
2112}
2113
2114/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
2115 * which is put after SKB on the list. It is very much like
2116 * tcp_fragment() except that it may make several kinds of assumptions
2117 * in order to speed up the splitting operation. In particular, we
2118 * know that all the data is in scatter-gather pages, and that the
2119 * packet has never been sent out before (and thus is not cloned).
2120 */
2121static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
2122 unsigned int mss_now, gfp_t gfp)
2123{
2124 int nlen = skb->len - len;
2125 struct sk_buff *buff;
2126 u8 flags;
2127
2128 /* All of a TSO frame must be composed of paged data. */
2129 if (skb->len != skb->data_len)
2130 return tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
2131 skb, len, mss_now, gfp);
2132
2133 buff = tcp_stream_alloc_skb(sk, 0, gfp, true);
2134 if (unlikely(!buff))
2135 return -ENOMEM;
2136 skb_copy_decrypted(buff, skb);
2137 mptcp_skb_ext_copy(buff, skb);
2138
2139 sk_wmem_queued_add(sk, buff->truesize);
2140 sk_mem_charge(sk, buff->truesize);
2141 buff->truesize += nlen;
2142 skb->truesize -= nlen;
2143
2144 /* Correct the sequence numbers. */
2145 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
2146 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
2147 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
2148
2149 /* PSH and FIN should only be set in the second packet. */
2150 flags = TCP_SKB_CB(skb)->tcp_flags;
2151 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
2152 TCP_SKB_CB(buff)->tcp_flags = flags;
2153
2154 tcp_skb_fragment_eor(skb, buff);
2155
2156 skb_split(skb, buff, len);
2157 tcp_fragment_tstamp(skb, buff);
2158
2159 /* Fix up tso_factor for both original and new SKB. */
2160 tcp_set_skb_tso_segs(skb, mss_now);
2161 tcp_set_skb_tso_segs(buff, mss_now);
2162
2163 /* Link BUFF into the send queue. */
2164 __skb_header_release(buff);
2165 tcp_insert_write_queue_after(skb, buff, sk, TCP_FRAG_IN_WRITE_QUEUE);
2166
2167 return 0;
2168}
2169
2170/* Try to defer sending, if possible, in order to minimize the amount
2171 * of TSO splitting we do. View it as a kind of TSO Nagle test.
2172 *
2173 * This algorithm is from John Heffner.
2174 */
2175static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
2176 bool *is_cwnd_limited,
2177 bool *is_rwnd_limited,
2178 u32 max_segs)
2179{
2180 const struct inet_connection_sock *icsk = inet_csk(sk);
2181 u32 send_win, cong_win, limit, in_flight;
2182 struct tcp_sock *tp = tcp_sk(sk);
2183 struct sk_buff *head;
2184 int win_divisor;
2185 s64 delta;
2186
2187 if (icsk->icsk_ca_state >= TCP_CA_Recovery)
2188 goto send_now;
2189
2190 /* Avoid bursty behavior by allowing defer
2191 * only if the last write was recent (1 ms).
2192 * Note that tp->tcp_wstamp_ns can be in the future if we have
2193 * packets waiting in a qdisc or device for EDT delivery.
2194 */
2195 delta = tp->tcp_clock_cache - tp->tcp_wstamp_ns - NSEC_PER_MSEC;
2196 if (delta > 0)
2197 goto send_now;
2198
2199 in_flight = tcp_packets_in_flight(tp);
2200
2201 BUG_ON(tcp_skb_pcount(skb) <= 1);
2202 BUG_ON(tcp_snd_cwnd(tp) <= in_flight);
2203
2204 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2205
2206 /* From in_flight test above, we know that cwnd > in_flight. */
2207 cong_win = (tcp_snd_cwnd(tp) - in_flight) * tp->mss_cache;
2208
2209 limit = min(send_win, cong_win);
2210
2211 /* If a full-sized TSO skb can be sent, do it. */
2212 if (limit >= max_segs * tp->mss_cache)
2213 goto send_now;
2214
2215 /* Middle in queue won't get any more data, full sendable already? */
2216 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
2217 goto send_now;
2218
2219 win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor);
2220 if (win_divisor) {
2221 u32 chunk = min(tp->snd_wnd, tcp_snd_cwnd(tp) * tp->mss_cache);
2222
2223 /* If at least some fraction of a window is available,
2224 * just use it.
2225 */
2226 chunk /= win_divisor;
2227 if (limit >= chunk)
2228 goto send_now;
2229 } else {
2230 /* Different approach, try not to defer past a single
2231 * ACK. Receiver should ACK every other full sized
2232 * frame, so if we have space for more than 3 frames
2233 * then send now.
2234 */
2235 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
2236 goto send_now;
2237 }
2238
2239 /* TODO : use tsorted_sent_queue ? */
2240 head = tcp_rtx_queue_head(sk);
2241 if (!head)
2242 goto send_now;
2243 delta = tp->tcp_clock_cache - head->tstamp;
2244 /* If next ACK is likely to come too late (half srtt), do not defer */
2245 if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0)
2246 goto send_now;
2247
2248 /* Ok, it looks like it is advisable to defer.
2249 * Three cases are tracked :
2250 * 1) We are cwnd-limited
2251 * 2) We are rwnd-limited
2252 * 3) We are application limited.
2253 */
2254 if (cong_win < send_win) {
2255 if (cong_win <= skb->len) {
2256 *is_cwnd_limited = true;
2257 return true;
2258 }
2259 } else {
2260 if (send_win <= skb->len) {
2261 *is_rwnd_limited = true;
2262 return true;
2263 }
2264 }
2265
2266 /* If this packet won't get more data, do not wait. */
2267 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) ||
2268 TCP_SKB_CB(skb)->eor)
2269 goto send_now;
2270
2271 return true;
2272
2273send_now:
2274 return false;
2275}
2276
2277static inline void tcp_mtu_check_reprobe(struct sock *sk)
2278{
2279 struct inet_connection_sock *icsk = inet_csk(sk);
2280 struct tcp_sock *tp = tcp_sk(sk);
2281 struct net *net = sock_net(sk);
2282 u32 interval;
2283 s32 delta;
2284
2285 interval = READ_ONCE(net->ipv4.sysctl_tcp_probe_interval);
2286 delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp;
2287 if (unlikely(delta >= interval * HZ)) {
2288 int mss = tcp_current_mss(sk);
2289
2290 /* Update current search range */
2291 icsk->icsk_mtup.probe_size = 0;
2292 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
2293 sizeof(struct tcphdr) +
2294 icsk->icsk_af_ops->net_header_len;
2295 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
2296
2297 /* Update probe time stamp */
2298 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
2299 }
2300}
2301
2302static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len)
2303{
2304 struct sk_buff *skb, *next;
2305
2306 skb = tcp_send_head(sk);
2307 tcp_for_write_queue_from_safe(skb, next, sk) {
2308 if (len <= skb->len)
2309 break;
2310
2311 if (unlikely(TCP_SKB_CB(skb)->eor) ||
2312 tcp_has_tx_tstamp(skb) ||
2313 !skb_pure_zcopy_same(skb, next))
2314 return false;
2315
2316 len -= skb->len;
2317 }
2318
2319 return true;
2320}
2321
2322/* Create a new MTU probe if we are ready.
2323 * MTU probe is regularly attempting to increase the path MTU by
2324 * deliberately sending larger packets. This discovers routing
2325 * changes resulting in larger path MTUs.
2326 *
2327 * Returns 0 if we should wait to probe (no cwnd available),
2328 * 1 if a probe was sent,
2329 * -1 otherwise
2330 */
2331static int tcp_mtu_probe(struct sock *sk)
2332{
2333 struct inet_connection_sock *icsk = inet_csk(sk);
2334 struct tcp_sock *tp = tcp_sk(sk);
2335 struct sk_buff *skb, *nskb, *next;
2336 struct net *net = sock_net(sk);
2337 int probe_size;
2338 int size_needed;
2339 int copy, len;
2340 int mss_now;
2341 int interval;
2342
2343 /* Not currently probing/verifying,
2344 * not in recovery,
2345 * have enough cwnd, and
2346 * not SACKing (the variable headers throw things off)
2347 */
2348 if (likely(!icsk->icsk_mtup.enabled ||
2349 icsk->icsk_mtup.probe_size ||
2350 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
2351 tcp_snd_cwnd(tp) < 11 ||
2352 tp->rx_opt.num_sacks || tp->rx_opt.dsack))
2353 return -1;
2354
2355 /* Use binary search for probe_size between tcp_mss_base,
2356 * and current mss_clamp. if (search_high - search_low)
2357 * smaller than a threshold, backoff from probing.
2358 */
2359 mss_now = tcp_current_mss(sk);
2360 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
2361 icsk->icsk_mtup.search_low) >> 1);
2362 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
2363 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
2364 /* When misfortune happens, we are reprobing actively,
2365 * and then reprobe timer has expired. We stick with current
2366 * probing process by not resetting search range to its orignal.
2367 */
2368 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
2369 interval < READ_ONCE(net->ipv4.sysctl_tcp_probe_threshold)) {
2370 /* Check whether enough time has elaplased for
2371 * another round of probing.
2372 */
2373 tcp_mtu_check_reprobe(sk);
2374 return -1;
2375 }
2376
2377 /* Have enough data in the send queue to probe? */
2378 if (tp->write_seq - tp->snd_nxt < size_needed)
2379 return -1;
2380
2381 if (tp->snd_wnd < size_needed)
2382 return -1;
2383 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
2384 return 0;
2385
2386 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2387 if (tcp_packets_in_flight(tp) + 2 > tcp_snd_cwnd(tp)) {
2388 if (!tcp_packets_in_flight(tp))
2389 return -1;
2390 else
2391 return 0;
2392 }
2393
2394 if (!tcp_can_coalesce_send_queue_head(sk, probe_size))
2395 return -1;
2396
2397 /* We're allowed to probe. Build it now. */
2398 nskb = tcp_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
2399 if (!nskb)
2400 return -1;
2401 sk_wmem_queued_add(sk, nskb->truesize);
2402 sk_mem_charge(sk, nskb->truesize);
2403
2404 skb = tcp_send_head(sk);
2405 skb_copy_decrypted(nskb, skb);
2406 mptcp_skb_ext_copy(nskb, skb);
2407
2408 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2409 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2410 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2411
2412 tcp_insert_write_queue_before(nskb, skb, sk);
2413 tcp_highest_sack_replace(sk, skb, nskb);
2414
2415 len = 0;
2416 tcp_for_write_queue_from_safe(skb, next, sk) {
2417 copy = min_t(int, skb->len, probe_size - len);
2418 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
2419
2420 if (skb->len <= copy) {
2421 /* We've eaten all the data from this skb.
2422 * Throw it away. */
2423 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2424 /* If this is the last SKB we copy and eor is set
2425 * we need to propagate it to the new skb.
2426 */
2427 TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor;
2428 tcp_skb_collapse_tstamp(nskb, skb);
2429 tcp_unlink_write_queue(skb, sk);
2430 tcp_wmem_free_skb(sk, skb);
2431 } else {
2432 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2433 ~(TCPHDR_FIN|TCPHDR_PSH);
2434 if (!skb_shinfo(skb)->nr_frags) {
2435 skb_pull(skb, copy);
2436 } else {
2437 __pskb_trim_head(skb, copy);
2438 tcp_set_skb_tso_segs(skb, mss_now);
2439 }
2440 TCP_SKB_CB(skb)->seq += copy;
2441 }
2442
2443 len += copy;
2444
2445 if (len >= probe_size)
2446 break;
2447 }
2448 tcp_init_tso_segs(nskb, nskb->len);
2449
2450 /* We're ready to send. If this fails, the probe will
2451 * be resegmented into mss-sized pieces by tcp_write_xmit().
2452 */
2453 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2454 /* Decrement cwnd here because we are sending
2455 * effectively two packets. */
2456 tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) - 1);
2457 tcp_event_new_data_sent(sk, nskb);
2458
2459 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2460 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2461 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2462
2463 return 1;
2464 }
2465
2466 return -1;
2467}
2468
2469static bool tcp_pacing_check(struct sock *sk)
2470{
2471 struct tcp_sock *tp = tcp_sk(sk);
2472
2473 if (!tcp_needs_internal_pacing(sk))
2474 return false;
2475
2476 if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache)
2477 return false;
2478
2479 if (!hrtimer_is_queued(&tp->pacing_timer)) {
2480 hrtimer_start(&tp->pacing_timer,
2481 ns_to_ktime(tp->tcp_wstamp_ns),
2482 HRTIMER_MODE_ABS_PINNED_SOFT);
2483 sock_hold(sk);
2484 }
2485 return true;
2486}
2487
2488/* TCP Small Queues :
2489 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2490 * (These limits are doubled for retransmits)
2491 * This allows for :
2492 * - better RTT estimation and ACK scheduling
2493 * - faster recovery
2494 * - high rates
2495 * Alas, some drivers / subsystems require a fair amount
2496 * of queued bytes to ensure line rate.
2497 * One example is wifi aggregation (802.11 AMPDU)
2498 */
2499static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2500 unsigned int factor)
2501{
2502 unsigned long limit;
2503
2504 limit = max_t(unsigned long,
2505 2 * skb->truesize,
2506 sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift));
2507 if (sk->sk_pacing_status == SK_PACING_NONE)
2508 limit = min_t(unsigned long, limit,
2509 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes));
2510 limit <<= factor;
2511
2512 if (static_branch_unlikely(&tcp_tx_delay_enabled) &&
2513 tcp_sk(sk)->tcp_tx_delay) {
2514 u64 extra_bytes = (u64)sk->sk_pacing_rate * tcp_sk(sk)->tcp_tx_delay;
2515
2516 /* TSQ is based on skb truesize sum (sk_wmem_alloc), so we
2517 * approximate our needs assuming an ~100% skb->truesize overhead.
2518 * USEC_PER_SEC is approximated by 2^20.
2519 * do_div(extra_bytes, USEC_PER_SEC/2) is replaced by a right shift.
2520 */
2521 extra_bytes >>= (20 - 1);
2522 limit += extra_bytes;
2523 }
2524 if (refcount_read(&sk->sk_wmem_alloc) > limit) {
2525 /* Always send skb if rtx queue is empty.
2526 * No need to wait for TX completion to call us back,
2527 * after softirq/tasklet schedule.
2528 * This helps when TX completions are delayed too much.
2529 */
2530 if (tcp_rtx_queue_empty(sk))
2531 return false;
2532
2533 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2534 /* It is possible TX completion already happened
2535 * before we set TSQ_THROTTLED, so we must
2536 * test again the condition.
2537 */
2538 smp_mb__after_atomic();
2539 if (refcount_read(&sk->sk_wmem_alloc) > limit)
2540 return true;
2541 }
2542 return false;
2543}
2544
2545static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2546{
2547 const u32 now = tcp_jiffies32;
2548 enum tcp_chrono old = tp->chrono_type;
2549
2550 if (old > TCP_CHRONO_UNSPEC)
2551 tp->chrono_stat[old - 1] += now - tp->chrono_start;
2552 tp->chrono_start = now;
2553 tp->chrono_type = new;
2554}
2555
2556void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2557{
2558 struct tcp_sock *tp = tcp_sk(sk);
2559
2560 /* If there are multiple conditions worthy of tracking in a
2561 * chronograph then the highest priority enum takes precedence
2562 * over the other conditions. So that if something "more interesting"
2563 * starts happening, stop the previous chrono and start a new one.
2564 */
2565 if (type > tp->chrono_type)
2566 tcp_chrono_set(tp, type);
2567}
2568
2569void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2570{
2571 struct tcp_sock *tp = tcp_sk(sk);
2572
2573
2574 /* There are multiple conditions worthy of tracking in a
2575 * chronograph, so that the highest priority enum takes
2576 * precedence over the other conditions (see tcp_chrono_start).
2577 * If a condition stops, we only stop chrono tracking if
2578 * it's the "most interesting" or current chrono we are
2579 * tracking and starts busy chrono if we have pending data.
2580 */
2581 if (tcp_rtx_and_write_queues_empty(sk))
2582 tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
2583 else if (type == tp->chrono_type)
2584 tcp_chrono_set(tp, TCP_CHRONO_BUSY);
2585}
2586
2587/* This routine writes packets to the network. It advances the
2588 * send_head. This happens as incoming acks open up the remote
2589 * window for us.
2590 *
2591 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2592 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2593 * account rare use of URG, this is not a big flaw.
2594 *
2595 * Send at most one packet when push_one > 0. Temporarily ignore
2596 * cwnd limit to force at most one packet out when push_one == 2.
2597
2598 * Returns true, if no segments are in flight and we have queued segments,
2599 * but cannot send anything now because of SWS or another problem.
2600 */
2601static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2602 int push_one, gfp_t gfp)
2603{
2604 struct tcp_sock *tp = tcp_sk(sk);
2605 struct sk_buff *skb;
2606 unsigned int tso_segs, sent_pkts;
2607 int cwnd_quota;
2608 int result;
2609 bool is_cwnd_limited = false, is_rwnd_limited = false;
2610 u32 max_segs;
2611
2612 sent_pkts = 0;
2613
2614 tcp_mstamp_refresh(tp);
2615 if (!push_one) {
2616 /* Do MTU probing. */
2617 result = tcp_mtu_probe(sk);
2618 if (!result) {
2619 return false;
2620 } else if (result > 0) {
2621 sent_pkts = 1;
2622 }
2623 }
2624
2625 max_segs = tcp_tso_segs(sk, mss_now);
2626 while ((skb = tcp_send_head(sk))) {
2627 unsigned int limit;
2628
2629 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2630 /* "skb_mstamp_ns" is used as a start point for the retransmit timer */
2631 tp->tcp_wstamp_ns = tp->tcp_clock_cache;
2632 skb_set_delivery_time(skb, tp->tcp_wstamp_ns, true);
2633 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
2634 tcp_init_tso_segs(skb, mss_now);
2635 goto repair; /* Skip network transmission */
2636 }
2637
2638 if (tcp_pacing_check(sk))
2639 break;
2640
2641 tso_segs = tcp_init_tso_segs(skb, mss_now);
2642 BUG_ON(!tso_segs);
2643
2644 cwnd_quota = tcp_cwnd_test(tp, skb);
2645 if (!cwnd_quota) {
2646 if (push_one == 2)
2647 /* Force out a loss probe pkt. */
2648 cwnd_quota = 1;
2649 else
2650 break;
2651 }
2652
2653 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2654 is_rwnd_limited = true;
2655 break;
2656 }
2657
2658 if (tso_segs == 1) {
2659 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2660 (tcp_skb_is_last(sk, skb) ?
2661 nonagle : TCP_NAGLE_PUSH))))
2662 break;
2663 } else {
2664 if (!push_one &&
2665 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2666 &is_rwnd_limited, max_segs))
2667 break;
2668 }
2669
2670 limit = mss_now;
2671 if (tso_segs > 1 && !tcp_urg_mode(tp))
2672 limit = tcp_mss_split_point(sk, skb, mss_now,
2673 min_t(unsigned int,
2674 cwnd_quota,
2675 max_segs),
2676 nonagle);
2677
2678 if (skb->len > limit &&
2679 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2680 break;
2681
2682 if (tcp_small_queue_check(sk, skb, 0))
2683 break;
2684
2685 /* Argh, we hit an empty skb(), presumably a thread
2686 * is sleeping in sendmsg()/sk_stream_wait_memory().
2687 * We do not want to send a pure-ack packet and have
2688 * a strange looking rtx queue with empty packet(s).
2689 */
2690 if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq)
2691 break;
2692
2693 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2694 break;
2695
2696repair:
2697 /* Advance the send_head. This one is sent out.
2698 * This call will increment packets_out.
2699 */
2700 tcp_event_new_data_sent(sk, skb);
2701
2702 tcp_minshall_update(tp, mss_now, skb);
2703 sent_pkts += tcp_skb_pcount(skb);
2704
2705 if (push_one)
2706 break;
2707 }
2708
2709 if (is_rwnd_limited)
2710 tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
2711 else
2712 tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
2713
2714 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp));
2715 if (likely(sent_pkts || is_cwnd_limited))
2716 tcp_cwnd_validate(sk, is_cwnd_limited);
2717
2718 if (likely(sent_pkts)) {
2719 if (tcp_in_cwnd_reduction(sk))
2720 tp->prr_out += sent_pkts;
2721
2722 /* Send one loss probe per tail loss episode. */
2723 if (push_one != 2)
2724 tcp_schedule_loss_probe(sk, false);
2725 return false;
2726 }
2727 return !tp->packets_out && !tcp_write_queue_empty(sk);
2728}
2729
2730bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto)
2731{
2732 struct inet_connection_sock *icsk = inet_csk(sk);
2733 struct tcp_sock *tp = tcp_sk(sk);
2734 u32 timeout, rto_delta_us;
2735 int early_retrans;
2736
2737 /* Don't do any loss probe on a Fast Open connection before 3WHS
2738 * finishes.
2739 */
2740 if (rcu_access_pointer(tp->fastopen_rsk))
2741 return false;
2742
2743 early_retrans = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_early_retrans);
2744 /* Schedule a loss probe in 2*RTT for SACK capable connections
2745 * not in loss recovery, that are either limited by cwnd or application.
2746 */
2747 if ((early_retrans != 3 && early_retrans != 4) ||
2748 !tp->packets_out || !tcp_is_sack(tp) ||
2749 (icsk->icsk_ca_state != TCP_CA_Open &&
2750 icsk->icsk_ca_state != TCP_CA_CWR))
2751 return false;
2752
2753 /* Probe timeout is 2*rtt. Add minimum RTO to account
2754 * for delayed ack when there's one outstanding packet. If no RTT
2755 * sample is available then probe after TCP_TIMEOUT_INIT.
2756 */
2757 if (tp->srtt_us) {
2758 timeout = usecs_to_jiffies(tp->srtt_us >> 2);
2759 if (tp->packets_out == 1)
2760 timeout += TCP_RTO_MIN;
2761 else
2762 timeout += TCP_TIMEOUT_MIN;
2763 } else {
2764 timeout = TCP_TIMEOUT_INIT;
2765 }
2766
2767 /* If the RTO formula yields an earlier time, then use that time. */
2768 rto_delta_us = advancing_rto ?
2769 jiffies_to_usecs(inet_csk(sk)->icsk_rto) :
2770 tcp_rto_delta_us(sk); /* How far in future is RTO? */
2771 if (rto_delta_us > 0)
2772 timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us));
2773
2774 tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, TCP_RTO_MAX);
2775 return true;
2776}
2777
2778/* Thanks to skb fast clones, we can detect if a prior transmit of
2779 * a packet is still in a qdisc or driver queue.
2780 * In this case, there is very little point doing a retransmit !
2781 */
2782static bool skb_still_in_host_queue(struct sock *sk,
2783 const struct sk_buff *skb)
2784{
2785 if (unlikely(skb_fclone_busy(sk, skb))) {
2786 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2787 smp_mb__after_atomic();
2788 if (skb_fclone_busy(sk, skb)) {
2789 NET_INC_STATS(sock_net(sk),
2790 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2791 return true;
2792 }
2793 }
2794 return false;
2795}
2796
2797/* When probe timeout (PTO) fires, try send a new segment if possible, else
2798 * retransmit the last segment.
2799 */
2800void tcp_send_loss_probe(struct sock *sk)
2801{
2802 struct tcp_sock *tp = tcp_sk(sk);
2803 struct sk_buff *skb;
2804 int pcount;
2805 int mss = tcp_current_mss(sk);
2806
2807 /* At most one outstanding TLP */
2808 if (tp->tlp_high_seq)
2809 goto rearm_timer;
2810
2811 tp->tlp_retrans = 0;
2812 skb = tcp_send_head(sk);
2813 if (skb && tcp_snd_wnd_test(tp, skb, mss)) {
2814 pcount = tp->packets_out;
2815 tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2816 if (tp->packets_out > pcount)
2817 goto probe_sent;
2818 goto rearm_timer;
2819 }
2820 skb = skb_rb_last(&sk->tcp_rtx_queue);
2821 if (unlikely(!skb)) {
2822 WARN_ONCE(tp->packets_out,
2823 "invalid inflight: %u state %u cwnd %u mss %d\n",
2824 tp->packets_out, sk->sk_state, tcp_snd_cwnd(tp), mss);
2825 inet_csk(sk)->icsk_pending = 0;
2826 return;
2827 }
2828
2829 if (skb_still_in_host_queue(sk, skb))
2830 goto rearm_timer;
2831
2832 pcount = tcp_skb_pcount(skb);
2833 if (WARN_ON(!pcount))
2834 goto rearm_timer;
2835
2836 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2837 if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2838 (pcount - 1) * mss, mss,
2839 GFP_ATOMIC)))
2840 goto rearm_timer;
2841 skb = skb_rb_next(skb);
2842 }
2843
2844 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2845 goto rearm_timer;
2846
2847 if (__tcp_retransmit_skb(sk, skb, 1))
2848 goto rearm_timer;
2849
2850 tp->tlp_retrans = 1;
2851
2852probe_sent:
2853 /* Record snd_nxt for loss detection. */
2854 tp->tlp_high_seq = tp->snd_nxt;
2855
2856 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2857 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2858 inet_csk(sk)->icsk_pending = 0;
2859rearm_timer:
2860 tcp_rearm_rto(sk);
2861}
2862
2863/* Push out any pending frames which were held back due to
2864 * TCP_CORK or attempt at coalescing tiny packets.
2865 * The socket must be locked by the caller.
2866 */
2867void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2868 int nonagle)
2869{
2870 /* If we are closed, the bytes will have to remain here.
2871 * In time closedown will finish, we empty the write queue and
2872 * all will be happy.
2873 */
2874 if (unlikely(sk->sk_state == TCP_CLOSE))
2875 return;
2876
2877 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2878 sk_gfp_mask(sk, GFP_ATOMIC)))
2879 tcp_check_probe_timer(sk);
2880}
2881
2882/* Send _single_ skb sitting at the send head. This function requires
2883 * true push pending frames to setup probe timer etc.
2884 */
2885void tcp_push_one(struct sock *sk, unsigned int mss_now)
2886{
2887 struct sk_buff *skb = tcp_send_head(sk);
2888
2889 BUG_ON(!skb || skb->len < mss_now);
2890
2891 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2892}
2893
2894/* This function returns the amount that we can raise the
2895 * usable window based on the following constraints
2896 *
2897 * 1. The window can never be shrunk once it is offered (RFC 793)
2898 * 2. We limit memory per socket
2899 *
2900 * RFC 1122:
2901 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2902 * RECV.NEXT + RCV.WIN fixed until:
2903 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2904 *
2905 * i.e. don't raise the right edge of the window until you can raise
2906 * it at least MSS bytes.
2907 *
2908 * Unfortunately, the recommended algorithm breaks header prediction,
2909 * since header prediction assumes th->window stays fixed.
2910 *
2911 * Strictly speaking, keeping th->window fixed violates the receiver
2912 * side SWS prevention criteria. The problem is that under this rule
2913 * a stream of single byte packets will cause the right side of the
2914 * window to always advance by a single byte.
2915 *
2916 * Of course, if the sender implements sender side SWS prevention
2917 * then this will not be a problem.
2918 *
2919 * BSD seems to make the following compromise:
2920 *
2921 * If the free space is less than the 1/4 of the maximum
2922 * space available and the free space is less than 1/2 mss,
2923 * then set the window to 0.
2924 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2925 * Otherwise, just prevent the window from shrinking
2926 * and from being larger than the largest representable value.
2927 *
2928 * This prevents incremental opening of the window in the regime
2929 * where TCP is limited by the speed of the reader side taking
2930 * data out of the TCP receive queue. It does nothing about
2931 * those cases where the window is constrained on the sender side
2932 * because the pipeline is full.
2933 *
2934 * BSD also seems to "accidentally" limit itself to windows that are a
2935 * multiple of MSS, at least until the free space gets quite small.
2936 * This would appear to be a side effect of the mbuf implementation.
2937 * Combining these two algorithms results in the observed behavior
2938 * of having a fixed window size at almost all times.
2939 *
2940 * Below we obtain similar behavior by forcing the offered window to
2941 * a multiple of the mss when it is feasible to do so.
2942 *
2943 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2944 * Regular options like TIMESTAMP are taken into account.
2945 */
2946u32 __tcp_select_window(struct sock *sk)
2947{
2948 struct inet_connection_sock *icsk = inet_csk(sk);
2949 struct tcp_sock *tp = tcp_sk(sk);
2950 /* MSS for the peer's data. Previous versions used mss_clamp
2951 * here. I don't know if the value based on our guesses
2952 * of peer's MSS is better for the performance. It's more correct
2953 * but may be worse for the performance because of rcv_mss
2954 * fluctuations. --SAW 1998/11/1
2955 */
2956 int mss = icsk->icsk_ack.rcv_mss;
2957 int free_space = tcp_space(sk);
2958 int allowed_space = tcp_full_space(sk);
2959 int full_space, window;
2960
2961 if (sk_is_mptcp(sk))
2962 mptcp_space(sk, &free_space, &allowed_space);
2963
2964 full_space = min_t(int, tp->window_clamp, allowed_space);
2965
2966 if (unlikely(mss > full_space)) {
2967 mss = full_space;
2968 if (mss <= 0)
2969 return 0;
2970 }
2971 if (free_space < (full_space >> 1)) {
2972 icsk->icsk_ack.quick = 0;
2973
2974 if (tcp_under_memory_pressure(sk))
2975 tcp_adjust_rcv_ssthresh(sk);
2976
2977 /* free_space might become our new window, make sure we don't
2978 * increase it due to wscale.
2979 */
2980 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2981
2982 /* if free space is less than mss estimate, or is below 1/16th
2983 * of the maximum allowed, try to move to zero-window, else
2984 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2985 * new incoming data is dropped due to memory limits.
2986 * With large window, mss test triggers way too late in order
2987 * to announce zero window in time before rmem limit kicks in.
2988 */
2989 if (free_space < (allowed_space >> 4) || free_space < mss)
2990 return 0;
2991 }
2992
2993 if (free_space > tp->rcv_ssthresh)
2994 free_space = tp->rcv_ssthresh;
2995
2996 /* Don't do rounding if we are using window scaling, since the
2997 * scaled window will not line up with the MSS boundary anyway.
2998 */
2999 if (tp->rx_opt.rcv_wscale) {
3000 window = free_space;
3001
3002 /* Advertise enough space so that it won't get scaled away.
3003 * Import case: prevent zero window announcement if
3004 * 1<<rcv_wscale > mss.
3005 */
3006 window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
3007 } else {
3008 window = tp->rcv_wnd;
3009 /* Get the largest window that is a nice multiple of mss.
3010 * Window clamp already applied above.
3011 * If our current window offering is within 1 mss of the
3012 * free space we just keep it. This prevents the divide
3013 * and multiply from happening most of the time.
3014 * We also don't do any window rounding when the free space
3015 * is too small.
3016 */
3017 if (window <= free_space - mss || window > free_space)
3018 window = rounddown(free_space, mss);
3019 else if (mss == full_space &&
3020 free_space > window + (full_space >> 1))
3021 window = free_space;
3022 }
3023
3024 return window;
3025}
3026
3027void tcp_skb_collapse_tstamp(struct sk_buff *skb,
3028 const struct sk_buff *next_skb)
3029{
3030 if (unlikely(tcp_has_tx_tstamp(next_skb))) {
3031 const struct skb_shared_info *next_shinfo =
3032 skb_shinfo(next_skb);
3033 struct skb_shared_info *shinfo = skb_shinfo(skb);
3034
3035 shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
3036 shinfo->tskey = next_shinfo->tskey;
3037 TCP_SKB_CB(skb)->txstamp_ack |=
3038 TCP_SKB_CB(next_skb)->txstamp_ack;
3039 }
3040}
3041
3042/* Collapses two adjacent SKB's during retransmission. */
3043static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
3044{
3045 struct tcp_sock *tp = tcp_sk(sk);
3046 struct sk_buff *next_skb = skb_rb_next(skb);
3047 int next_skb_size;
3048
3049 next_skb_size = next_skb->len;
3050
3051 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
3052
3053 if (next_skb_size && !tcp_skb_shift(skb, next_skb, 1, next_skb_size))
3054 return false;
3055
3056 tcp_highest_sack_replace(sk, next_skb, skb);
3057
3058 /* Update sequence range on original skb. */
3059 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
3060
3061 /* Merge over control information. This moves PSH/FIN etc. over */
3062 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
3063
3064 /* All done, get rid of second SKB and account for it so
3065 * packet counting does not break.
3066 */
3067 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
3068 TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
3069
3070 /* changed transmit queue under us so clear hints */
3071 tcp_clear_retrans_hints_partial(tp);
3072 if (next_skb == tp->retransmit_skb_hint)
3073 tp->retransmit_skb_hint = skb;
3074
3075 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
3076
3077 tcp_skb_collapse_tstamp(skb, next_skb);
3078
3079 tcp_rtx_queue_unlink_and_free(next_skb, sk);
3080 return true;
3081}
3082
3083/* Check if coalescing SKBs is legal. */
3084static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
3085{
3086 if (tcp_skb_pcount(skb) > 1)
3087 return false;
3088 if (skb_cloned(skb))
3089 return false;
3090 /* Some heuristics for collapsing over SACK'd could be invented */
3091 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
3092 return false;
3093
3094 return true;
3095}
3096
3097/* Collapse packets in the retransmit queue to make to create
3098 * less packets on the wire. This is only done on retransmission.
3099 */
3100static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
3101 int space)
3102{
3103 struct tcp_sock *tp = tcp_sk(sk);
3104 struct sk_buff *skb = to, *tmp;
3105 bool first = true;
3106
3107 if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse))
3108 return;
3109 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
3110 return;
3111
3112 skb_rbtree_walk_from_safe(skb, tmp) {
3113 if (!tcp_can_collapse(sk, skb))
3114 break;
3115
3116 if (!tcp_skb_can_collapse(to, skb))
3117 break;
3118
3119 space -= skb->len;
3120
3121 if (first) {
3122 first = false;
3123 continue;
3124 }
3125
3126 if (space < 0)
3127 break;
3128
3129 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
3130 break;
3131
3132 if (!tcp_collapse_retrans(sk, to))
3133 break;
3134 }
3135}
3136
3137/* This retransmits one SKB. Policy decisions and retransmit queue
3138 * state updates are done by the caller. Returns non-zero if an
3139 * error occurred which prevented the send.
3140 */
3141int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
3142{
3143 struct inet_connection_sock *icsk = inet_csk(sk);
3144 struct tcp_sock *tp = tcp_sk(sk);
3145 unsigned int cur_mss;
3146 int diff, len, err;
3147 int avail_wnd;
3148
3149 /* Inconclusive MTU probe */
3150 if (icsk->icsk_mtup.probe_size)
3151 icsk->icsk_mtup.probe_size = 0;
3152
3153 if (skb_still_in_host_queue(sk, skb))
3154 return -EBUSY;
3155
3156 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
3157 if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) {
3158 WARN_ON_ONCE(1);
3159 return -EINVAL;
3160 }
3161 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3162 return -ENOMEM;
3163 }
3164
3165 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3166 return -EHOSTUNREACH; /* Routing failure or similar. */
3167
3168 cur_mss = tcp_current_mss(sk);
3169 avail_wnd = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3170
3171 /* If receiver has shrunk his window, and skb is out of
3172 * new window, do not retransmit it. The exception is the
3173 * case, when window is shrunk to zero. In this case
3174 * our retransmit of one segment serves as a zero window probe.
3175 */
3176 if (avail_wnd <= 0) {
3177 if (TCP_SKB_CB(skb)->seq != tp->snd_una)
3178 return -EAGAIN;
3179 avail_wnd = cur_mss;
3180 }
3181
3182 len = cur_mss * segs;
3183 if (len > avail_wnd) {
3184 len = rounddown(avail_wnd, cur_mss);
3185 if (!len)
3186 len = avail_wnd;
3187 }
3188 if (skb->len > len) {
3189 if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len,
3190 cur_mss, GFP_ATOMIC))
3191 return -ENOMEM; /* We'll try again later. */
3192 } else {
3193 if (skb_unclone_keeptruesize(skb, GFP_ATOMIC))
3194 return -ENOMEM;
3195
3196 diff = tcp_skb_pcount(skb);
3197 tcp_set_skb_tso_segs(skb, cur_mss);
3198 diff -= tcp_skb_pcount(skb);
3199 if (diff)
3200 tcp_adjust_pcount(sk, skb, diff);
3201 avail_wnd = min_t(int, avail_wnd, cur_mss);
3202 if (skb->len < avail_wnd)
3203 tcp_retrans_try_collapse(sk, skb, avail_wnd);
3204 }
3205
3206 /* RFC3168, section 6.1.1.1. ECN fallback */
3207 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
3208 tcp_ecn_clear_syn(sk, skb);
3209
3210 /* Update global and local TCP statistics. */
3211 segs = tcp_skb_pcount(skb);
3212 TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
3213 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
3214 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3215 tp->total_retrans += segs;
3216 tp->bytes_retrans += skb->len;
3217
3218 /* make sure skb->data is aligned on arches that require it
3219 * and check if ack-trimming & collapsing extended the headroom
3220 * beyond what csum_start can cover.
3221 */
3222 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
3223 skb_headroom(skb) >= 0xFFFF)) {
3224 struct sk_buff *nskb;
3225
3226 tcp_skb_tsorted_save(skb) {
3227 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
3228 if (nskb) {
3229 nskb->dev = NULL;
3230 err = tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC);
3231 } else {
3232 err = -ENOBUFS;
3233 }
3234 } tcp_skb_tsorted_restore(skb);
3235
3236 if (!err) {
3237 tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns);
3238 tcp_rate_skb_sent(sk, skb);
3239 }
3240 } else {
3241 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3242 }
3243
3244 /* To avoid taking spuriously low RTT samples based on a timestamp
3245 * for a transmit that never happened, always mark EVER_RETRANS
3246 */
3247 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
3248
3249 if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG))
3250 tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB,
3251 TCP_SKB_CB(skb)->seq, segs, err);
3252
3253 if (likely(!err)) {
3254 trace_tcp_retransmit_skb(sk, skb);
3255 } else if (err != -EBUSY) {
3256 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs);
3257 }
3258 return err;
3259}
3260
3261int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
3262{
3263 struct tcp_sock *tp = tcp_sk(sk);
3264 int err = __tcp_retransmit_skb(sk, skb, segs);
3265
3266 if (err == 0) {
3267#if FASTRETRANS_DEBUG > 0
3268 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
3269 net_dbg_ratelimited("retrans_out leaked\n");
3270 }
3271#endif
3272 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
3273 tp->retrans_out += tcp_skb_pcount(skb);
3274 }
3275
3276 /* Save stamp of the first (attempted) retransmit. */
3277 if (!tp->retrans_stamp)
3278 tp->retrans_stamp = tcp_skb_timestamp(skb);
3279
3280 if (tp->undo_retrans < 0)
3281 tp->undo_retrans = 0;
3282 tp->undo_retrans += tcp_skb_pcount(skb);
3283 return err;
3284}
3285
3286/* This gets called after a retransmit timeout, and the initially
3287 * retransmitted data is acknowledged. It tries to continue
3288 * resending the rest of the retransmit queue, until either
3289 * we've sent it all or the congestion window limit is reached.
3290 */
3291void tcp_xmit_retransmit_queue(struct sock *sk)
3292{
3293 const struct inet_connection_sock *icsk = inet_csk(sk);
3294 struct sk_buff *skb, *rtx_head, *hole = NULL;
3295 struct tcp_sock *tp = tcp_sk(sk);
3296 bool rearm_timer = false;
3297 u32 max_segs;
3298 int mib_idx;
3299
3300 if (!tp->packets_out)
3301 return;
3302
3303 rtx_head = tcp_rtx_queue_head(sk);
3304 skb = tp->retransmit_skb_hint ?: rtx_head;
3305 max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
3306 skb_rbtree_walk_from(skb) {
3307 __u8 sacked;
3308 int segs;
3309
3310 if (tcp_pacing_check(sk))
3311 break;
3312
3313 /* we could do better than to assign each time */
3314 if (!hole)
3315 tp->retransmit_skb_hint = skb;
3316
3317 segs = tcp_snd_cwnd(tp) - tcp_packets_in_flight(tp);
3318 if (segs <= 0)
3319 break;
3320 sacked = TCP_SKB_CB(skb)->sacked;
3321 /* In case tcp_shift_skb_data() have aggregated large skbs,
3322 * we need to make sure not sending too bigs TSO packets
3323 */
3324 segs = min_t(int, segs, max_segs);
3325
3326 if (tp->retrans_out >= tp->lost_out) {
3327 break;
3328 } else if (!(sacked & TCPCB_LOST)) {
3329 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
3330 hole = skb;
3331 continue;
3332
3333 } else {
3334 if (icsk->icsk_ca_state != TCP_CA_Loss)
3335 mib_idx = LINUX_MIB_TCPFASTRETRANS;
3336 else
3337 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
3338 }
3339
3340 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
3341 continue;
3342
3343 if (tcp_small_queue_check(sk, skb, 1))
3344 break;
3345
3346 if (tcp_retransmit_skb(sk, skb, segs))
3347 break;
3348
3349 NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
3350
3351 if (tcp_in_cwnd_reduction(sk))
3352 tp->prr_out += tcp_skb_pcount(skb);
3353
3354 if (skb == rtx_head &&
3355 icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
3356 rearm_timer = true;
3357
3358 }
3359 if (rearm_timer)
3360 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3361 inet_csk(sk)->icsk_rto,
3362 TCP_RTO_MAX);
3363}
3364
3365/* We allow to exceed memory limits for FIN packets to expedite
3366 * connection tear down and (memory) recovery.
3367 * Otherwise tcp_send_fin() could be tempted to either delay FIN
3368 * or even be forced to close flow without any FIN.
3369 * In general, we want to allow one skb per socket to avoid hangs
3370 * with edge trigger epoll()
3371 */
3372void sk_forced_mem_schedule(struct sock *sk, int size)
3373{
3374 int delta, amt;
3375
3376 delta = size - sk->sk_forward_alloc;
3377 if (delta <= 0)
3378 return;
3379 amt = sk_mem_pages(delta);
3380 sk->sk_forward_alloc += amt << PAGE_SHIFT;
3381 sk_memory_allocated_add(sk, amt);
3382
3383 if (mem_cgroup_sockets_enabled && sk->sk_memcg)
3384 mem_cgroup_charge_skmem(sk->sk_memcg, amt,
3385 gfp_memcg_charge() | __GFP_NOFAIL);
3386}
3387
3388/* Send a FIN. The caller locks the socket for us.
3389 * We should try to send a FIN packet really hard, but eventually give up.
3390 */
3391void tcp_send_fin(struct sock *sk)
3392{
3393 struct sk_buff *skb, *tskb, *tail = tcp_write_queue_tail(sk);
3394 struct tcp_sock *tp = tcp_sk(sk);
3395
3396 /* Optimization, tack on the FIN if we have one skb in write queue and
3397 * this skb was not yet sent, or we are under memory pressure.
3398 * Note: in the latter case, FIN packet will be sent after a timeout,
3399 * as TCP stack thinks it has already been transmitted.
3400 */
3401 tskb = tail;
3402 if (!tskb && tcp_under_memory_pressure(sk))
3403 tskb = skb_rb_last(&sk->tcp_rtx_queue);
3404
3405 if (tskb) {
3406 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
3407 TCP_SKB_CB(tskb)->end_seq++;
3408 tp->write_seq++;
3409 if (!tail) {
3410 /* This means tskb was already sent.
3411 * Pretend we included the FIN on previous transmit.
3412 * We need to set tp->snd_nxt to the value it would have
3413 * if FIN had been sent. This is because retransmit path
3414 * does not change tp->snd_nxt.
3415 */
3416 WRITE_ONCE(tp->snd_nxt, tp->snd_nxt + 1);
3417 return;
3418 }
3419 } else {
3420 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
3421 if (unlikely(!skb))
3422 return;
3423
3424 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
3425 skb_reserve(skb, MAX_TCP_HEADER);
3426 sk_forced_mem_schedule(sk, skb->truesize);
3427 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3428 tcp_init_nondata_skb(skb, tp->write_seq,
3429 TCPHDR_ACK | TCPHDR_FIN);
3430 tcp_queue_skb(sk, skb);
3431 }
3432 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
3433}
3434
3435/* We get here when a process closes a file descriptor (either due to
3436 * an explicit close() or as a byproduct of exit()'ing) and there
3437 * was unread data in the receive queue. This behavior is recommended
3438 * by RFC 2525, section 2.17. -DaveM
3439 */
3440void tcp_send_active_reset(struct sock *sk, gfp_t priority)
3441{
3442 struct sk_buff *skb;
3443
3444 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3445
3446 /* NOTE: No TCP options attached and we never retransmit this. */
3447 skb = alloc_skb(MAX_TCP_HEADER, priority);
3448 if (!skb) {
3449 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3450 return;
3451 }
3452
3453 /* Reserve space for headers and prepare control bits. */
3454 skb_reserve(skb, MAX_TCP_HEADER);
3455 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
3456 TCPHDR_ACK | TCPHDR_RST);
3457 tcp_mstamp_refresh(tcp_sk(sk));
3458 /* Send it off. */
3459 if (tcp_transmit_skb(sk, skb, 0, priority))
3460 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3461
3462 /* skb of trace_tcp_send_reset() keeps the skb that caused RST,
3463 * skb here is different to the troublesome skb, so use NULL
3464 */
3465 trace_tcp_send_reset(sk, NULL);
3466}
3467
3468/* Send a crossed SYN-ACK during socket establishment.
3469 * WARNING: This routine must only be called when we have already sent
3470 * a SYN packet that crossed the incoming SYN that caused this routine
3471 * to get called. If this assumption fails then the initial rcv_wnd
3472 * and rcv_wscale values will not be correct.
3473 */
3474int tcp_send_synack(struct sock *sk)
3475{
3476 struct sk_buff *skb;
3477
3478 skb = tcp_rtx_queue_head(sk);
3479 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3480 pr_err("%s: wrong queue state\n", __func__);
3481 return -EFAULT;
3482 }
3483 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3484 if (skb_cloned(skb)) {
3485 struct sk_buff *nskb;
3486
3487 tcp_skb_tsorted_save(skb) {
3488 nskb = skb_copy(skb, GFP_ATOMIC);
3489 } tcp_skb_tsorted_restore(skb);
3490 if (!nskb)
3491 return -ENOMEM;
3492 INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor);
3493 tcp_highest_sack_replace(sk, skb, nskb);
3494 tcp_rtx_queue_unlink_and_free(skb, sk);
3495 __skb_header_release(nskb);
3496 tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb);
3497 sk_wmem_queued_add(sk, nskb->truesize);
3498 sk_mem_charge(sk, nskb->truesize);
3499 skb = nskb;
3500 }
3501
3502 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3503 tcp_ecn_send_synack(sk, skb);
3504 }
3505 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3506}
3507
3508/**
3509 * tcp_make_synack - Allocate one skb and build a SYNACK packet.
3510 * @sk: listener socket
3511 * @dst: dst entry attached to the SYNACK. It is consumed and caller
3512 * should not use it again.
3513 * @req: request_sock pointer
3514 * @foc: cookie for tcp fast open
3515 * @synack_type: Type of synack to prepare
3516 * @syn_skb: SYN packet just received. It could be NULL for rtx case.
3517 */
3518struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3519 struct request_sock *req,
3520 struct tcp_fastopen_cookie *foc,
3521 enum tcp_synack_type synack_type,
3522 struct sk_buff *syn_skb)
3523{
3524 struct inet_request_sock *ireq = inet_rsk(req);
3525 const struct tcp_sock *tp = tcp_sk(sk);
3526 struct tcp_md5sig_key *md5 = NULL;
3527 struct tcp_out_options opts;
3528 struct sk_buff *skb;
3529 int tcp_header_size;
3530 struct tcphdr *th;
3531 int mss;
3532 u64 now;
3533
3534 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3535 if (unlikely(!skb)) {
3536 dst_release(dst);
3537 return NULL;
3538 }
3539 /* Reserve space for headers. */
3540 skb_reserve(skb, MAX_TCP_HEADER);
3541
3542 switch (synack_type) {
3543 case TCP_SYNACK_NORMAL:
3544 skb_set_owner_w(skb, req_to_sk(req));
3545 break;
3546 case TCP_SYNACK_COOKIE:
3547 /* Under synflood, we do not attach skb to a socket,
3548 * to avoid false sharing.
3549 */
3550 break;
3551 case TCP_SYNACK_FASTOPEN:
3552 /* sk is a const pointer, because we want to express multiple
3553 * cpu might call us concurrently.
3554 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3555 */
3556 skb_set_owner_w(skb, (struct sock *)sk);
3557 break;
3558 }
3559 skb_dst_set(skb, dst);
3560
3561 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3562
3563 memset(&opts, 0, sizeof(opts));
3564 now = tcp_clock_ns();
3565#ifdef CONFIG_SYN_COOKIES
3566 if (unlikely(synack_type == TCP_SYNACK_COOKIE && ireq->tstamp_ok))
3567 skb_set_delivery_time(skb, cookie_init_timestamp(req, now),
3568 true);
3569 else
3570#endif
3571 {
3572 skb_set_delivery_time(skb, now, true);
3573 if (!tcp_rsk(req)->snt_synack) /* Timestamp first SYNACK */
3574 tcp_rsk(req)->snt_synack = tcp_skb_timestamp_us(skb);
3575 }
3576
3577#ifdef CONFIG_TCP_MD5SIG
3578 rcu_read_lock();
3579 md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3580#endif
3581 skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
3582 /* bpf program will be interested in the tcp_flags */
3583 TCP_SKB_CB(skb)->tcp_flags = TCPHDR_SYN | TCPHDR_ACK;
3584 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5,
3585 foc, synack_type,
3586 syn_skb) + sizeof(*th);
3587
3588 skb_push(skb, tcp_header_size);
3589 skb_reset_transport_header(skb);
3590
3591 th = (struct tcphdr *)skb->data;
3592 memset(th, 0, sizeof(struct tcphdr));
3593 th->syn = 1;
3594 th->ack = 1;
3595 tcp_ecn_make_synack(req, th);
3596 th->source = htons(ireq->ir_num);
3597 th->dest = ireq->ir_rmt_port;
3598 skb->mark = ireq->ir_mark;
3599 skb->ip_summed = CHECKSUM_PARTIAL;
3600 th->seq = htonl(tcp_rsk(req)->snt_isn);
3601 /* XXX data is queued and acked as is. No buffer/window check */
3602 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3603
3604 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3605 th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3606 tcp_options_write(th, NULL, &opts);
3607 th->doff = (tcp_header_size >> 2);
3608 __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3609
3610#ifdef CONFIG_TCP_MD5SIG
3611 /* Okay, we have all we need - do the md5 hash if needed */
3612 if (md5)
3613 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3614 md5, req_to_sk(req), skb);
3615 rcu_read_unlock();
3616#endif
3617
3618 bpf_skops_write_hdr_opt((struct sock *)sk, skb, req, syn_skb,
3619 synack_type, &opts);
3620
3621 skb_set_delivery_time(skb, now, true);
3622 tcp_add_tx_delay(skb, tp);
3623
3624 return skb;
3625}
3626EXPORT_SYMBOL(tcp_make_synack);
3627
3628static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3629{
3630 struct inet_connection_sock *icsk = inet_csk(sk);
3631 const struct tcp_congestion_ops *ca;
3632 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3633
3634 if (ca_key == TCP_CA_UNSPEC)
3635 return;
3636
3637 rcu_read_lock();
3638 ca = tcp_ca_find_key(ca_key);
3639 if (likely(ca && bpf_try_module_get(ca, ca->owner))) {
3640 bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner);
3641 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3642 icsk->icsk_ca_ops = ca;
3643 }
3644 rcu_read_unlock();
3645}
3646
3647/* Do all connect socket setups that can be done AF independent. */
3648static void tcp_connect_init(struct sock *sk)
3649{
3650 const struct dst_entry *dst = __sk_dst_get(sk);
3651 struct tcp_sock *tp = tcp_sk(sk);
3652 __u8 rcv_wscale;
3653 u32 rcv_wnd;
3654
3655 /* We'll fix this up when we get a response from the other end.
3656 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3657 */
3658 tp->tcp_header_len = sizeof(struct tcphdr);
3659 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_timestamps))
3660 tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED;
3661
3662#ifdef CONFIG_TCP_MD5SIG
3663 if (tp->af_specific->md5_lookup(sk, sk))
3664 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3665#endif
3666
3667 /* If user gave his TCP_MAXSEG, record it to clamp */
3668 if (tp->rx_opt.user_mss)
3669 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3670 tp->max_window = 0;
3671 tcp_mtup_init(sk);
3672 tcp_sync_mss(sk, dst_mtu(dst));
3673
3674 tcp_ca_dst_init(sk, dst);
3675
3676 if (!tp->window_clamp)
3677 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3678 tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3679
3680 tcp_initialize_rcv_mss(sk);
3681
3682 /* limit the window selection if the user enforce a smaller rx buffer */
3683 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3684 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3685 tp->window_clamp = tcp_full_space(sk);
3686
3687 rcv_wnd = tcp_rwnd_init_bpf(sk);
3688 if (rcv_wnd == 0)
3689 rcv_wnd = dst_metric(dst, RTAX_INITRWND);
3690
3691 tcp_select_initial_window(sk, tcp_full_space(sk),
3692 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3693 &tp->rcv_wnd,
3694 &tp->window_clamp,
3695 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_window_scaling),
3696 &rcv_wscale,
3697 rcv_wnd);
3698
3699 tp->rx_opt.rcv_wscale = rcv_wscale;
3700 tp->rcv_ssthresh = tp->rcv_wnd;
3701
3702 sk->sk_err = 0;
3703 sock_reset_flag(sk, SOCK_DONE);
3704 tp->snd_wnd = 0;
3705 tcp_init_wl(tp, 0);
3706 tcp_write_queue_purge(sk);
3707 tp->snd_una = tp->write_seq;
3708 tp->snd_sml = tp->write_seq;
3709 tp->snd_up = tp->write_seq;
3710 WRITE_ONCE(tp->snd_nxt, tp->write_seq);
3711
3712 if (likely(!tp->repair))
3713 tp->rcv_nxt = 0;
3714 else
3715 tp->rcv_tstamp = tcp_jiffies32;
3716 tp->rcv_wup = tp->rcv_nxt;
3717 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
3718
3719 inet_csk(sk)->icsk_rto = tcp_timeout_init(sk);
3720 inet_csk(sk)->icsk_retransmits = 0;
3721 tcp_clear_retrans(tp);
3722}
3723
3724static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3725{
3726 struct tcp_sock *tp = tcp_sk(sk);
3727 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3728
3729 tcb->end_seq += skb->len;
3730 __skb_header_release(skb);
3731 sk_wmem_queued_add(sk, skb->truesize);
3732 sk_mem_charge(sk, skb->truesize);
3733 WRITE_ONCE(tp->write_seq, tcb->end_seq);
3734 tp->packets_out += tcp_skb_pcount(skb);
3735}
3736
3737/* Build and send a SYN with data and (cached) Fast Open cookie. However,
3738 * queue a data-only packet after the regular SYN, such that regular SYNs
3739 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3740 * only the SYN sequence, the data are retransmitted in the first ACK.
3741 * If cookie is not cached or other error occurs, falls back to send a
3742 * regular SYN with Fast Open cookie request option.
3743 */
3744static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3745{
3746 struct inet_connection_sock *icsk = inet_csk(sk);
3747 struct tcp_sock *tp = tcp_sk(sk);
3748 struct tcp_fastopen_request *fo = tp->fastopen_req;
3749 int space, err = 0;
3750 struct sk_buff *syn_data;
3751
3752 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
3753 if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie))
3754 goto fallback;
3755
3756 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3757 * user-MSS. Reserve maximum option space for middleboxes that add
3758 * private TCP options. The cost is reduced data space in SYN :(
3759 */
3760 tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp);
3761 /* Sync mss_cache after updating the mss_clamp */
3762 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
3763
3764 space = __tcp_mtu_to_mss(sk, icsk->icsk_pmtu_cookie) -
3765 MAX_TCP_OPTION_SPACE;
3766
3767 space = min_t(size_t, space, fo->size);
3768
3769 /* limit to order-0 allocations */
3770 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3771
3772 syn_data = tcp_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3773 if (!syn_data)
3774 goto fallback;
3775 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3776 if (space) {
3777 int copied = copy_from_iter(skb_put(syn_data, space), space,
3778 &fo->data->msg_iter);
3779 if (unlikely(!copied)) {
3780 tcp_skb_tsorted_anchor_cleanup(syn_data);
3781 kfree_skb(syn_data);
3782 goto fallback;
3783 }
3784 if (copied != space) {
3785 skb_trim(syn_data, copied);
3786 space = copied;
3787 }
3788 skb_zcopy_set(syn_data, fo->uarg, NULL);
3789 }
3790 /* No more data pending in inet_wait_for_connect() */
3791 if (space == fo->size)
3792 fo->data = NULL;
3793 fo->copied = space;
3794
3795 tcp_connect_queue_skb(sk, syn_data);
3796 if (syn_data->len)
3797 tcp_chrono_start(sk, TCP_CHRONO_BUSY);
3798
3799 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3800
3801 skb_set_delivery_time(syn, syn_data->skb_mstamp_ns, true);
3802
3803 /* Now full SYN+DATA was cloned and sent (or not),
3804 * remove the SYN from the original skb (syn_data)
3805 * we keep in write queue in case of a retransmit, as we
3806 * also have the SYN packet (with no data) in the same queue.
3807 */
3808 TCP_SKB_CB(syn_data)->seq++;
3809 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3810 if (!err) {
3811 tp->syn_data = (fo->copied > 0);
3812 tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data);
3813 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3814 goto done;
3815 }
3816
3817 /* data was not sent, put it in write_queue */
3818 __skb_queue_tail(&sk->sk_write_queue, syn_data);
3819 tp->packets_out -= tcp_skb_pcount(syn_data);
3820
3821fallback:
3822 /* Send a regular SYN with Fast Open cookie request option */
3823 if (fo->cookie.len > 0)
3824 fo->cookie.len = 0;
3825 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3826 if (err)
3827 tp->syn_fastopen = 0;
3828done:
3829 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3830 return err;
3831}
3832
3833/* Build a SYN and send it off. */
3834int tcp_connect(struct sock *sk)
3835{
3836 struct tcp_sock *tp = tcp_sk(sk);
3837 struct sk_buff *buff;
3838 int err;
3839
3840 tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL);
3841
3842 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3843 return -EHOSTUNREACH; /* Routing failure or similar. */
3844
3845 tcp_connect_init(sk);
3846
3847 if (unlikely(tp->repair)) {
3848 tcp_finish_connect(sk, NULL);
3849 return 0;
3850 }
3851
3852 buff = tcp_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3853 if (unlikely(!buff))
3854 return -ENOBUFS;
3855
3856 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3857 tcp_mstamp_refresh(tp);
3858 tp->retrans_stamp = tcp_time_stamp(tp);
3859 tcp_connect_queue_skb(sk, buff);
3860 tcp_ecn_send_syn(sk, buff);
3861 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
3862
3863 /* Send off SYN; include data in Fast Open. */
3864 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3865 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3866 if (err == -ECONNREFUSED)
3867 return err;
3868
3869 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3870 * in order to make this packet get counted in tcpOutSegs.
3871 */
3872 WRITE_ONCE(tp->snd_nxt, tp->write_seq);
3873 tp->pushed_seq = tp->write_seq;
3874 buff = tcp_send_head(sk);
3875 if (unlikely(buff)) {
3876 WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(buff)->seq);
3877 tp->pushed_seq = TCP_SKB_CB(buff)->seq;
3878 }
3879 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3880
3881 /* Timer for repeating the SYN until an answer. */
3882 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3883 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3884 return 0;
3885}
3886EXPORT_SYMBOL(tcp_connect);
3887
3888/* Send out a delayed ack, the caller does the policy checking
3889 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3890 * for details.
3891 */
3892void tcp_send_delayed_ack(struct sock *sk)
3893{
3894 struct inet_connection_sock *icsk = inet_csk(sk);
3895 int ato = icsk->icsk_ack.ato;
3896 unsigned long timeout;
3897
3898 if (ato > TCP_DELACK_MIN) {
3899 const struct tcp_sock *tp = tcp_sk(sk);
3900 int max_ato = HZ / 2;
3901
3902 if (inet_csk_in_pingpong_mode(sk) ||
3903 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3904 max_ato = TCP_DELACK_MAX;
3905
3906 /* Slow path, intersegment interval is "high". */
3907
3908 /* If some rtt estimate is known, use it to bound delayed ack.
3909 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3910 * directly.
3911 */
3912 if (tp->srtt_us) {
3913 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3914 TCP_DELACK_MIN);
3915
3916 if (rtt < max_ato)
3917 max_ato = rtt;
3918 }
3919
3920 ato = min(ato, max_ato);
3921 }
3922
3923 ato = min_t(u32, ato, inet_csk(sk)->icsk_delack_max);
3924
3925 /* Stay within the limit we were given */
3926 timeout = jiffies + ato;
3927
3928 /* Use new timeout only if there wasn't a older one earlier. */
3929 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3930 /* If delack timer is about to expire, send ACK now. */
3931 if (time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3932 tcp_send_ack(sk);
3933 return;
3934 }
3935
3936 if (!time_before(timeout, icsk->icsk_ack.timeout))
3937 timeout = icsk->icsk_ack.timeout;
3938 }
3939 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3940 icsk->icsk_ack.timeout = timeout;
3941 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3942}
3943
3944/* This routine sends an ack and also updates the window. */
3945void __tcp_send_ack(struct sock *sk, u32 rcv_nxt)
3946{
3947 struct sk_buff *buff;
3948
3949 /* If we have been reset, we may not send again. */
3950 if (sk->sk_state == TCP_CLOSE)
3951 return;
3952
3953 /* We are not putting this on the write queue, so
3954 * tcp_transmit_skb() will set the ownership to this
3955 * sock.
3956 */
3957 buff = alloc_skb(MAX_TCP_HEADER,
3958 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3959 if (unlikely(!buff)) {
3960 struct inet_connection_sock *icsk = inet_csk(sk);
3961 unsigned long delay;
3962
3963 delay = TCP_DELACK_MAX << icsk->icsk_ack.retry;
3964 if (delay < TCP_RTO_MAX)
3965 icsk->icsk_ack.retry++;
3966 inet_csk_schedule_ack(sk);
3967 icsk->icsk_ack.ato = TCP_ATO_MIN;
3968 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, delay, TCP_RTO_MAX);
3969 return;
3970 }
3971
3972 /* Reserve space for headers and prepare control bits. */
3973 skb_reserve(buff, MAX_TCP_HEADER);
3974 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3975
3976 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3977 * too much.
3978 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3979 */
3980 skb_set_tcp_pure_ack(buff);
3981
3982 /* Send it off, this clears delayed acks for us. */
3983 __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt);
3984}
3985EXPORT_SYMBOL_GPL(__tcp_send_ack);
3986
3987void tcp_send_ack(struct sock *sk)
3988{
3989 __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt);
3990}
3991
3992/* This routine sends a packet with an out of date sequence
3993 * number. It assumes the other end will try to ack it.
3994 *
3995 * Question: what should we make while urgent mode?
3996 * 4.4BSD forces sending single byte of data. We cannot send
3997 * out of window data, because we have SND.NXT==SND.MAX...
3998 *
3999 * Current solution: to send TWO zero-length segments in urgent mode:
4000 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
4001 * out-of-date with SND.UNA-1 to probe window.
4002 */
4003static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
4004{
4005 struct tcp_sock *tp = tcp_sk(sk);
4006 struct sk_buff *skb;
4007
4008 /* We don't queue it, tcp_transmit_skb() sets ownership. */
4009 skb = alloc_skb(MAX_TCP_HEADER,
4010 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
4011 if (!skb)
4012 return -1;
4013
4014 /* Reserve space for headers and set control bits. */
4015 skb_reserve(skb, MAX_TCP_HEADER);
4016 /* Use a previous sequence. This should cause the other
4017 * end to send an ack. Don't queue or clone SKB, just
4018 * send it.
4019 */
4020 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
4021 NET_INC_STATS(sock_net(sk), mib);
4022 return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
4023}
4024
4025/* Called from setsockopt( ... TCP_REPAIR ) */
4026void tcp_send_window_probe(struct sock *sk)
4027{
4028 if (sk->sk_state == TCP_ESTABLISHED) {
4029 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
4030 tcp_mstamp_refresh(tcp_sk(sk));
4031 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
4032 }
4033}
4034
4035/* Initiate keepalive or window probe from timer. */
4036int tcp_write_wakeup(struct sock *sk, int mib)
4037{
4038 struct tcp_sock *tp = tcp_sk(sk);
4039 struct sk_buff *skb;
4040
4041 if (sk->sk_state == TCP_CLOSE)
4042 return -1;
4043
4044 skb = tcp_send_head(sk);
4045 if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
4046 int err;
4047 unsigned int mss = tcp_current_mss(sk);
4048 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
4049
4050 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
4051 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
4052
4053 /* We are probing the opening of a window
4054 * but the window size is != 0
4055 * must have been a result SWS avoidance ( sender )
4056 */
4057 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
4058 skb->len > mss) {
4059 seg_size = min(seg_size, mss);
4060 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
4061 if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
4062 skb, seg_size, mss, GFP_ATOMIC))
4063 return -1;
4064 } else if (!tcp_skb_pcount(skb))
4065 tcp_set_skb_tso_segs(skb, mss);
4066
4067 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
4068 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
4069 if (!err)
4070 tcp_event_new_data_sent(sk, skb);
4071 return err;
4072 } else {
4073 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
4074 tcp_xmit_probe_skb(sk, 1, mib);
4075 return tcp_xmit_probe_skb(sk, 0, mib);
4076 }
4077}
4078
4079/* A window probe timeout has occurred. If window is not closed send
4080 * a partial packet else a zero probe.
4081 */
4082void tcp_send_probe0(struct sock *sk)
4083{
4084 struct inet_connection_sock *icsk = inet_csk(sk);
4085 struct tcp_sock *tp = tcp_sk(sk);
4086 struct net *net = sock_net(sk);
4087 unsigned long timeout;
4088 int err;
4089
4090 err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
4091
4092 if (tp->packets_out || tcp_write_queue_empty(sk)) {
4093 /* Cancel probe timer, if it is not required. */
4094 icsk->icsk_probes_out = 0;
4095 icsk->icsk_backoff = 0;
4096 icsk->icsk_probes_tstamp = 0;
4097 return;
4098 }
4099
4100 icsk->icsk_probes_out++;
4101 if (err <= 0) {
4102 if (icsk->icsk_backoff < READ_ONCE(net->ipv4.sysctl_tcp_retries2))
4103 icsk->icsk_backoff++;
4104 timeout = tcp_probe0_when(sk, TCP_RTO_MAX);
4105 } else {
4106 /* If packet was not sent due to local congestion,
4107 * Let senders fight for local resources conservatively.
4108 */
4109 timeout = TCP_RESOURCE_PROBE_INTERVAL;
4110 }
4111
4112 timeout = tcp_clamp_probe0_to_user_timeout(sk, timeout);
4113 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, timeout, TCP_RTO_MAX);
4114}
4115
4116int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
4117{
4118 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
4119 struct flowi fl;
4120 int res;
4121
4122 /* Paired with WRITE_ONCE() in sock_setsockopt() */
4123 if (READ_ONCE(sk->sk_txrehash) == SOCK_TXREHASH_ENABLED)
4124 tcp_rsk(req)->txhash = net_tx_rndhash();
4125 res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL,
4126 NULL);
4127 if (!res) {
4128 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
4129 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
4130 if (unlikely(tcp_passive_fastopen(sk)))
4131 tcp_sk(sk)->total_retrans++;
4132 trace_tcp_retransmit_synack(sk, req);
4133 }
4134 return res;
4135}
4136EXPORT_SYMBOL(tcp_rtx_synack);
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
6 *
7 * Implementation of the Transmission Control Protocol(TCP).
8 *
9 * Authors: Ross Biro
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
20 */
21
22/*
23 * Changes: Pedro Roque : Retransmit queue handled by TCP.
24 * : Fragmentation on mtu decrease
25 * : Segment collapse on retransmit
26 * : AF independence
27 *
28 * Linus Torvalds : send_delayed_ack
29 * David S. Miller : Charge memory using the right skb
30 * during syn/ack processing.
31 * David S. Miller : Output engine completely rewritten.
32 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
33 * Cacophonix Gaul : draft-minshall-nagle-01
34 * J Hadi Salim : ECN support
35 *
36 */
37
38#define pr_fmt(fmt) "TCP: " fmt
39
40#include <net/tcp.h>
41#include <net/mptcp.h>
42
43#include <linux/compiler.h>
44#include <linux/gfp.h>
45#include <linux/module.h>
46#include <linux/static_key.h>
47
48#include <trace/events/tcp.h>
49
50/* Refresh clocks of a TCP socket,
51 * ensuring monotically increasing values.
52 */
53void tcp_mstamp_refresh(struct tcp_sock *tp)
54{
55 u64 val = tcp_clock_ns();
56
57 tp->tcp_clock_cache = val;
58 tp->tcp_mstamp = div_u64(val, NSEC_PER_USEC);
59}
60
61static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
62 int push_one, gfp_t gfp);
63
64/* Account for new data that has been sent to the network. */
65static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
66{
67 struct inet_connection_sock *icsk = inet_csk(sk);
68 struct tcp_sock *tp = tcp_sk(sk);
69 unsigned int prior_packets = tp->packets_out;
70
71 WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(skb)->end_seq);
72
73 __skb_unlink(skb, &sk->sk_write_queue);
74 tcp_rbtree_insert(&sk->tcp_rtx_queue, skb);
75
76 if (tp->highest_sack == NULL)
77 tp->highest_sack = skb;
78
79 tp->packets_out += tcp_skb_pcount(skb);
80 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
81 tcp_rearm_rto(sk);
82
83 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
84 tcp_skb_pcount(skb));
85 tcp_check_space(sk);
86}
87
88/* SND.NXT, if window was not shrunk or the amount of shrunk was less than one
89 * window scaling factor due to loss of precision.
90 * If window has been shrunk, what should we make? It is not clear at all.
91 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
92 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
93 * invalid. OK, let's make this for now:
94 */
95static inline __u32 tcp_acceptable_seq(const struct sock *sk)
96{
97 const struct tcp_sock *tp = tcp_sk(sk);
98
99 if (!before(tcp_wnd_end(tp), tp->snd_nxt) ||
100 (tp->rx_opt.wscale_ok &&
101 ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale))))
102 return tp->snd_nxt;
103 else
104 return tcp_wnd_end(tp);
105}
106
107/* Calculate mss to advertise in SYN segment.
108 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
109 *
110 * 1. It is independent of path mtu.
111 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
112 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
113 * attached devices, because some buggy hosts are confused by
114 * large MSS.
115 * 4. We do not make 3, we advertise MSS, calculated from first
116 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
117 * This may be overridden via information stored in routing table.
118 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
119 * probably even Jumbo".
120 */
121static __u16 tcp_advertise_mss(struct sock *sk)
122{
123 struct tcp_sock *tp = tcp_sk(sk);
124 const struct dst_entry *dst = __sk_dst_get(sk);
125 int mss = tp->advmss;
126
127 if (dst) {
128 unsigned int metric = dst_metric_advmss(dst);
129
130 if (metric < mss) {
131 mss = metric;
132 tp->advmss = mss;
133 }
134 }
135
136 return (__u16)mss;
137}
138
139/* RFC2861. Reset CWND after idle period longer RTO to "restart window".
140 * This is the first part of cwnd validation mechanism.
141 */
142void tcp_cwnd_restart(struct sock *sk, s32 delta)
143{
144 struct tcp_sock *tp = tcp_sk(sk);
145 u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
146 u32 cwnd = tcp_snd_cwnd(tp);
147
148 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
149
150 tp->snd_ssthresh = tcp_current_ssthresh(sk);
151 restart_cwnd = min(restart_cwnd, cwnd);
152
153 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
154 cwnd >>= 1;
155 tcp_snd_cwnd_set(tp, max(cwnd, restart_cwnd));
156 tp->snd_cwnd_stamp = tcp_jiffies32;
157 tp->snd_cwnd_used = 0;
158}
159
160/* Congestion state accounting after a packet has been sent. */
161static void tcp_event_data_sent(struct tcp_sock *tp,
162 struct sock *sk)
163{
164 struct inet_connection_sock *icsk = inet_csk(sk);
165 const u32 now = tcp_jiffies32;
166
167 if (tcp_packets_in_flight(tp) == 0)
168 tcp_ca_event(sk, CA_EVENT_TX_START);
169
170 tp->lsndtime = now;
171
172 /* If it is a reply for ato after last received
173 * packet, increase pingpong count.
174 */
175 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
176 inet_csk_inc_pingpong_cnt(sk);
177}
178
179/* Account for an ACK we sent. */
180static inline void tcp_event_ack_sent(struct sock *sk, u32 rcv_nxt)
181{
182 struct tcp_sock *tp = tcp_sk(sk);
183
184 if (unlikely(tp->compressed_ack)) {
185 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
186 tp->compressed_ack);
187 tp->compressed_ack = 0;
188 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
189 __sock_put(sk);
190 }
191
192 if (unlikely(rcv_nxt != tp->rcv_nxt))
193 return; /* Special ACK sent by DCTCP to reflect ECN */
194 tcp_dec_quickack_mode(sk);
195 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
196}
197
198/* Determine a window scaling and initial window to offer.
199 * Based on the assumption that the given amount of space
200 * will be offered. Store the results in the tp structure.
201 * NOTE: for smooth operation initial space offering should
202 * be a multiple of mss if possible. We assume here that mss >= 1.
203 * This MUST be enforced by all callers.
204 */
205void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss,
206 __u32 *rcv_wnd, __u32 *window_clamp,
207 int wscale_ok, __u8 *rcv_wscale,
208 __u32 init_rcv_wnd)
209{
210 unsigned int space = (__space < 0 ? 0 : __space);
211
212 /* If no clamp set the clamp to the max possible scaled window */
213 if (*window_clamp == 0)
214 (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE);
215 space = min(*window_clamp, space);
216
217 /* Quantize space offering to a multiple of mss if possible. */
218 if (space > mss)
219 space = rounddown(space, mss);
220
221 /* NOTE: offering an initial window larger than 32767
222 * will break some buggy TCP stacks. If the admin tells us
223 * it is likely we could be speaking with such a buggy stack
224 * we will truncate our initial window offering to 32K-1
225 * unless the remote has sent us a window scaling option,
226 * which we interpret as a sign the remote TCP is not
227 * misinterpreting the window field as a signed quantity.
228 */
229 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows))
230 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
231 else
232 (*rcv_wnd) = min_t(u32, space, U16_MAX);
233
234 if (init_rcv_wnd)
235 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
236
237 *rcv_wscale = 0;
238 if (wscale_ok) {
239 /* Set window scaling on max possible window */
240 space = max_t(u32, space, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
241 space = max_t(u32, space, READ_ONCE(sysctl_rmem_max));
242 space = min_t(u32, space, *window_clamp);
243 *rcv_wscale = clamp_t(int, ilog2(space) - 15,
244 0, TCP_MAX_WSCALE);
245 }
246 /* Set the clamp no higher than max representable value */
247 (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp);
248}
249EXPORT_SYMBOL(tcp_select_initial_window);
250
251/* Chose a new window to advertise, update state in tcp_sock for the
252 * socket, and return result with RFC1323 scaling applied. The return
253 * value can be stuffed directly into th->window for an outgoing
254 * frame.
255 */
256static u16 tcp_select_window(struct sock *sk)
257{
258 struct tcp_sock *tp = tcp_sk(sk);
259 struct net *net = sock_net(sk);
260 u32 old_win = tp->rcv_wnd;
261 u32 cur_win, new_win;
262
263 /* Make the window 0 if we failed to queue the data because we
264 * are out of memory. The window is temporary, so we don't store
265 * it on the socket.
266 */
267 if (unlikely(inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOMEM))
268 return 0;
269
270 cur_win = tcp_receive_window(tp);
271 new_win = __tcp_select_window(sk);
272 if (new_win < cur_win) {
273 /* Danger Will Robinson!
274 * Don't update rcv_wup/rcv_wnd here or else
275 * we will not be able to advertise a zero
276 * window in time. --DaveM
277 *
278 * Relax Will Robinson.
279 */
280 if (!READ_ONCE(net->ipv4.sysctl_tcp_shrink_window) || !tp->rx_opt.rcv_wscale) {
281 /* Never shrink the offered window */
282 if (new_win == 0)
283 NET_INC_STATS(net, LINUX_MIB_TCPWANTZEROWINDOWADV);
284 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
285 }
286 }
287
288 tp->rcv_wnd = new_win;
289 tp->rcv_wup = tp->rcv_nxt;
290
291 /* Make sure we do not exceed the maximum possible
292 * scaled window.
293 */
294 if (!tp->rx_opt.rcv_wscale &&
295 READ_ONCE(net->ipv4.sysctl_tcp_workaround_signed_windows))
296 new_win = min(new_win, MAX_TCP_WINDOW);
297 else
298 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
299
300 /* RFC1323 scaling applied */
301 new_win >>= tp->rx_opt.rcv_wscale;
302
303 /* If we advertise zero window, disable fast path. */
304 if (new_win == 0) {
305 tp->pred_flags = 0;
306 if (old_win)
307 NET_INC_STATS(net, LINUX_MIB_TCPTOZEROWINDOWADV);
308 } else if (old_win == 0) {
309 NET_INC_STATS(net, LINUX_MIB_TCPFROMZEROWINDOWADV);
310 }
311
312 return new_win;
313}
314
315/* Packet ECN state for a SYN-ACK */
316static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
317{
318 const struct tcp_sock *tp = tcp_sk(sk);
319
320 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
321 if (!(tp->ecn_flags & TCP_ECN_OK))
322 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
323 else if (tcp_ca_needs_ecn(sk) ||
324 tcp_bpf_ca_needs_ecn(sk))
325 INET_ECN_xmit(sk);
326}
327
328/* Packet ECN state for a SYN. */
329static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
330{
331 struct tcp_sock *tp = tcp_sk(sk);
332 bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk);
333 bool use_ecn = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_ecn) == 1 ||
334 tcp_ca_needs_ecn(sk) || bpf_needs_ecn;
335
336 if (!use_ecn) {
337 const struct dst_entry *dst = __sk_dst_get(sk);
338
339 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
340 use_ecn = true;
341 }
342
343 tp->ecn_flags = 0;
344
345 if (use_ecn) {
346 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
347 tp->ecn_flags = TCP_ECN_OK;
348 if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn)
349 INET_ECN_xmit(sk);
350 }
351}
352
353static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
354{
355 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback))
356 /* tp->ecn_flags are cleared at a later point in time when
357 * SYN ACK is ultimatively being received.
358 */
359 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
360}
361
362static void
363tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
364{
365 if (inet_rsk(req)->ecn_ok)
366 th->ece = 1;
367}
368
369/* Set up ECN state for a packet on a ESTABLISHED socket that is about to
370 * be sent.
371 */
372static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
373 struct tcphdr *th, int tcp_header_len)
374{
375 struct tcp_sock *tp = tcp_sk(sk);
376
377 if (tp->ecn_flags & TCP_ECN_OK) {
378 /* Not-retransmitted data segment: set ECT and inject CWR. */
379 if (skb->len != tcp_header_len &&
380 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
381 INET_ECN_xmit(sk);
382 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
383 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
384 th->cwr = 1;
385 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
386 }
387 } else if (!tcp_ca_needs_ecn(sk)) {
388 /* ACK or retransmitted segment: clear ECT|CE */
389 INET_ECN_dontxmit(sk);
390 }
391 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
392 th->ece = 1;
393 }
394}
395
396/* Constructs common control bits of non-data skb. If SYN/FIN is present,
397 * auto increment end seqno.
398 */
399static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
400{
401 skb->ip_summed = CHECKSUM_PARTIAL;
402
403 TCP_SKB_CB(skb)->tcp_flags = flags;
404
405 tcp_skb_pcount_set(skb, 1);
406
407 TCP_SKB_CB(skb)->seq = seq;
408 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
409 seq++;
410 TCP_SKB_CB(skb)->end_seq = seq;
411}
412
413static inline bool tcp_urg_mode(const struct tcp_sock *tp)
414{
415 return tp->snd_una != tp->snd_up;
416}
417
418#define OPTION_SACK_ADVERTISE BIT(0)
419#define OPTION_TS BIT(1)
420#define OPTION_MD5 BIT(2)
421#define OPTION_WSCALE BIT(3)
422#define OPTION_FAST_OPEN_COOKIE BIT(8)
423#define OPTION_SMC BIT(9)
424#define OPTION_MPTCP BIT(10)
425#define OPTION_AO BIT(11)
426
427static void smc_options_write(__be32 *ptr, u16 *options)
428{
429#if IS_ENABLED(CONFIG_SMC)
430 if (static_branch_unlikely(&tcp_have_smc)) {
431 if (unlikely(OPTION_SMC & *options)) {
432 *ptr++ = htonl((TCPOPT_NOP << 24) |
433 (TCPOPT_NOP << 16) |
434 (TCPOPT_EXP << 8) |
435 (TCPOLEN_EXP_SMC_BASE));
436 *ptr++ = htonl(TCPOPT_SMC_MAGIC);
437 }
438 }
439#endif
440}
441
442struct tcp_out_options {
443 u16 options; /* bit field of OPTION_* */
444 u16 mss; /* 0 to disable */
445 u8 ws; /* window scale, 0 to disable */
446 u8 num_sack_blocks; /* number of SACK blocks to include */
447 u8 hash_size; /* bytes in hash_location */
448 u8 bpf_opt_len; /* length of BPF hdr option */
449 __u8 *hash_location; /* temporary pointer, overloaded */
450 __u32 tsval, tsecr; /* need to include OPTION_TS */
451 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
452 struct mptcp_out_options mptcp;
453};
454
455static void mptcp_options_write(struct tcphdr *th, __be32 *ptr,
456 struct tcp_sock *tp,
457 struct tcp_out_options *opts)
458{
459#if IS_ENABLED(CONFIG_MPTCP)
460 if (unlikely(OPTION_MPTCP & opts->options))
461 mptcp_write_options(th, ptr, tp, &opts->mptcp);
462#endif
463}
464
465#ifdef CONFIG_CGROUP_BPF
466static int bpf_skops_write_hdr_opt_arg0(struct sk_buff *skb,
467 enum tcp_synack_type synack_type)
468{
469 if (unlikely(!skb))
470 return BPF_WRITE_HDR_TCP_CURRENT_MSS;
471
472 if (unlikely(synack_type == TCP_SYNACK_COOKIE))
473 return BPF_WRITE_HDR_TCP_SYNACK_COOKIE;
474
475 return 0;
476}
477
478/* req, syn_skb and synack_type are used when writing synack */
479static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb,
480 struct request_sock *req,
481 struct sk_buff *syn_skb,
482 enum tcp_synack_type synack_type,
483 struct tcp_out_options *opts,
484 unsigned int *remaining)
485{
486 struct bpf_sock_ops_kern sock_ops;
487 int err;
488
489 if (likely(!BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk),
490 BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG)) ||
491 !*remaining)
492 return;
493
494 /* *remaining has already been aligned to 4 bytes, so *remaining >= 4 */
495
496 /* init sock_ops */
497 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
498
499 sock_ops.op = BPF_SOCK_OPS_HDR_OPT_LEN_CB;
500
501 if (req) {
502 /* The listen "sk" cannot be passed here because
503 * it is not locked. It would not make too much
504 * sense to do bpf_setsockopt(listen_sk) based
505 * on individual connection request also.
506 *
507 * Thus, "req" is passed here and the cgroup-bpf-progs
508 * of the listen "sk" will be run.
509 *
510 * "req" is also used here for fastopen even the "sk" here is
511 * a fullsock "child" sk. It is to keep the behavior
512 * consistent between fastopen and non-fastopen on
513 * the bpf programming side.
514 */
515 sock_ops.sk = (struct sock *)req;
516 sock_ops.syn_skb = syn_skb;
517 } else {
518 sock_owned_by_me(sk);
519
520 sock_ops.is_fullsock = 1;
521 sock_ops.sk = sk;
522 }
523
524 sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type);
525 sock_ops.remaining_opt_len = *remaining;
526 /* tcp_current_mss() does not pass a skb */
527 if (skb)
528 bpf_skops_init_skb(&sock_ops, skb, 0);
529
530 err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk);
531
532 if (err || sock_ops.remaining_opt_len == *remaining)
533 return;
534
535 opts->bpf_opt_len = *remaining - sock_ops.remaining_opt_len;
536 /* round up to 4 bytes */
537 opts->bpf_opt_len = (opts->bpf_opt_len + 3) & ~3;
538
539 *remaining -= opts->bpf_opt_len;
540}
541
542static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb,
543 struct request_sock *req,
544 struct sk_buff *syn_skb,
545 enum tcp_synack_type synack_type,
546 struct tcp_out_options *opts)
547{
548 u8 first_opt_off, nr_written, max_opt_len = opts->bpf_opt_len;
549 struct bpf_sock_ops_kern sock_ops;
550 int err;
551
552 if (likely(!max_opt_len))
553 return;
554
555 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
556
557 sock_ops.op = BPF_SOCK_OPS_WRITE_HDR_OPT_CB;
558
559 if (req) {
560 sock_ops.sk = (struct sock *)req;
561 sock_ops.syn_skb = syn_skb;
562 } else {
563 sock_owned_by_me(sk);
564
565 sock_ops.is_fullsock = 1;
566 sock_ops.sk = sk;
567 }
568
569 sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type);
570 sock_ops.remaining_opt_len = max_opt_len;
571 first_opt_off = tcp_hdrlen(skb) - max_opt_len;
572 bpf_skops_init_skb(&sock_ops, skb, first_opt_off);
573
574 err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk);
575
576 if (err)
577 nr_written = 0;
578 else
579 nr_written = max_opt_len - sock_ops.remaining_opt_len;
580
581 if (nr_written < max_opt_len)
582 memset(skb->data + first_opt_off + nr_written, TCPOPT_NOP,
583 max_opt_len - nr_written);
584}
585#else
586static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb,
587 struct request_sock *req,
588 struct sk_buff *syn_skb,
589 enum tcp_synack_type synack_type,
590 struct tcp_out_options *opts,
591 unsigned int *remaining)
592{
593}
594
595static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb,
596 struct request_sock *req,
597 struct sk_buff *syn_skb,
598 enum tcp_synack_type synack_type,
599 struct tcp_out_options *opts)
600{
601}
602#endif
603
604static __be32 *process_tcp_ao_options(struct tcp_sock *tp,
605 const struct tcp_request_sock *tcprsk,
606 struct tcp_out_options *opts,
607 struct tcp_key *key, __be32 *ptr)
608{
609#ifdef CONFIG_TCP_AO
610 u8 maclen = tcp_ao_maclen(key->ao_key);
611
612 if (tcprsk) {
613 u8 aolen = maclen + sizeof(struct tcp_ao_hdr);
614
615 *ptr++ = htonl((TCPOPT_AO << 24) | (aolen << 16) |
616 (tcprsk->ao_keyid << 8) |
617 (tcprsk->ao_rcv_next));
618 } else {
619 struct tcp_ao_key *rnext_key;
620 struct tcp_ao_info *ao_info;
621
622 ao_info = rcu_dereference_check(tp->ao_info,
623 lockdep_sock_is_held(&tp->inet_conn.icsk_inet.sk));
624 rnext_key = READ_ONCE(ao_info->rnext_key);
625 if (WARN_ON_ONCE(!rnext_key))
626 return ptr;
627 *ptr++ = htonl((TCPOPT_AO << 24) |
628 (tcp_ao_len(key->ao_key) << 16) |
629 (key->ao_key->sndid << 8) |
630 (rnext_key->rcvid));
631 }
632 opts->hash_location = (__u8 *)ptr;
633 ptr += maclen / sizeof(*ptr);
634 if (unlikely(maclen % sizeof(*ptr))) {
635 memset(ptr, TCPOPT_NOP, sizeof(*ptr));
636 ptr++;
637 }
638#endif
639 return ptr;
640}
641
642/* Write previously computed TCP options to the packet.
643 *
644 * Beware: Something in the Internet is very sensitive to the ordering of
645 * TCP options, we learned this through the hard way, so be careful here.
646 * Luckily we can at least blame others for their non-compliance but from
647 * inter-operability perspective it seems that we're somewhat stuck with
648 * the ordering which we have been using if we want to keep working with
649 * those broken things (not that it currently hurts anybody as there isn't
650 * particular reason why the ordering would need to be changed).
651 *
652 * At least SACK_PERM as the first option is known to lead to a disaster
653 * (but it may well be that other scenarios fail similarly).
654 */
655static void tcp_options_write(struct tcphdr *th, struct tcp_sock *tp,
656 const struct tcp_request_sock *tcprsk,
657 struct tcp_out_options *opts,
658 struct tcp_key *key)
659{
660 __be32 *ptr = (__be32 *)(th + 1);
661 u16 options = opts->options; /* mungable copy */
662
663 if (tcp_key_is_md5(key)) {
664 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
665 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
666 /* overload cookie hash location */
667 opts->hash_location = (__u8 *)ptr;
668 ptr += 4;
669 } else if (tcp_key_is_ao(key)) {
670 ptr = process_tcp_ao_options(tp, tcprsk, opts, key, ptr);
671 }
672 if (unlikely(opts->mss)) {
673 *ptr++ = htonl((TCPOPT_MSS << 24) |
674 (TCPOLEN_MSS << 16) |
675 opts->mss);
676 }
677
678 if (likely(OPTION_TS & options)) {
679 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
680 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
681 (TCPOLEN_SACK_PERM << 16) |
682 (TCPOPT_TIMESTAMP << 8) |
683 TCPOLEN_TIMESTAMP);
684 options &= ~OPTION_SACK_ADVERTISE;
685 } else {
686 *ptr++ = htonl((TCPOPT_NOP << 24) |
687 (TCPOPT_NOP << 16) |
688 (TCPOPT_TIMESTAMP << 8) |
689 TCPOLEN_TIMESTAMP);
690 }
691 *ptr++ = htonl(opts->tsval);
692 *ptr++ = htonl(opts->tsecr);
693 }
694
695 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
696 *ptr++ = htonl((TCPOPT_NOP << 24) |
697 (TCPOPT_NOP << 16) |
698 (TCPOPT_SACK_PERM << 8) |
699 TCPOLEN_SACK_PERM);
700 }
701
702 if (unlikely(OPTION_WSCALE & options)) {
703 *ptr++ = htonl((TCPOPT_NOP << 24) |
704 (TCPOPT_WINDOW << 16) |
705 (TCPOLEN_WINDOW << 8) |
706 opts->ws);
707 }
708
709 if (unlikely(opts->num_sack_blocks)) {
710 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
711 tp->duplicate_sack : tp->selective_acks;
712 int this_sack;
713
714 *ptr++ = htonl((TCPOPT_NOP << 24) |
715 (TCPOPT_NOP << 16) |
716 (TCPOPT_SACK << 8) |
717 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
718 TCPOLEN_SACK_PERBLOCK)));
719
720 for (this_sack = 0; this_sack < opts->num_sack_blocks;
721 ++this_sack) {
722 *ptr++ = htonl(sp[this_sack].start_seq);
723 *ptr++ = htonl(sp[this_sack].end_seq);
724 }
725
726 tp->rx_opt.dsack = 0;
727 }
728
729 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
730 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
731 u8 *p = (u8 *)ptr;
732 u32 len; /* Fast Open option length */
733
734 if (foc->exp) {
735 len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
736 *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
737 TCPOPT_FASTOPEN_MAGIC);
738 p += TCPOLEN_EXP_FASTOPEN_BASE;
739 } else {
740 len = TCPOLEN_FASTOPEN_BASE + foc->len;
741 *p++ = TCPOPT_FASTOPEN;
742 *p++ = len;
743 }
744
745 memcpy(p, foc->val, foc->len);
746 if ((len & 3) == 2) {
747 p[foc->len] = TCPOPT_NOP;
748 p[foc->len + 1] = TCPOPT_NOP;
749 }
750 ptr += (len + 3) >> 2;
751 }
752
753 smc_options_write(ptr, &options);
754
755 mptcp_options_write(th, ptr, tp, opts);
756}
757
758static void smc_set_option(const struct tcp_sock *tp,
759 struct tcp_out_options *opts,
760 unsigned int *remaining)
761{
762#if IS_ENABLED(CONFIG_SMC)
763 if (static_branch_unlikely(&tcp_have_smc)) {
764 if (tp->syn_smc) {
765 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
766 opts->options |= OPTION_SMC;
767 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
768 }
769 }
770 }
771#endif
772}
773
774static void smc_set_option_cond(const struct tcp_sock *tp,
775 const struct inet_request_sock *ireq,
776 struct tcp_out_options *opts,
777 unsigned int *remaining)
778{
779#if IS_ENABLED(CONFIG_SMC)
780 if (static_branch_unlikely(&tcp_have_smc)) {
781 if (tp->syn_smc && ireq->smc_ok) {
782 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
783 opts->options |= OPTION_SMC;
784 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
785 }
786 }
787 }
788#endif
789}
790
791static void mptcp_set_option_cond(const struct request_sock *req,
792 struct tcp_out_options *opts,
793 unsigned int *remaining)
794{
795 if (rsk_is_mptcp(req)) {
796 unsigned int size;
797
798 if (mptcp_synack_options(req, &size, &opts->mptcp)) {
799 if (*remaining >= size) {
800 opts->options |= OPTION_MPTCP;
801 *remaining -= size;
802 }
803 }
804 }
805}
806
807/* Compute TCP options for SYN packets. This is not the final
808 * network wire format yet.
809 */
810static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
811 struct tcp_out_options *opts,
812 struct tcp_key *key)
813{
814 struct tcp_sock *tp = tcp_sk(sk);
815 unsigned int remaining = MAX_TCP_OPTION_SPACE;
816 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
817 bool timestamps;
818
819 /* Better than switch (key.type) as it has static branches */
820 if (tcp_key_is_md5(key)) {
821 timestamps = false;
822 opts->options |= OPTION_MD5;
823 remaining -= TCPOLEN_MD5SIG_ALIGNED;
824 } else {
825 timestamps = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_timestamps);
826 if (tcp_key_is_ao(key)) {
827 opts->options |= OPTION_AO;
828 remaining -= tcp_ao_len_aligned(key->ao_key);
829 }
830 }
831
832 /* We always get an MSS option. The option bytes which will be seen in
833 * normal data packets should timestamps be used, must be in the MSS
834 * advertised. But we subtract them from tp->mss_cache so that
835 * calculations in tcp_sendmsg are simpler etc. So account for this
836 * fact here if necessary. If we don't do this correctly, as a
837 * receiver we won't recognize data packets as being full sized when we
838 * should, and thus we won't abide by the delayed ACK rules correctly.
839 * SACKs don't matter, we never delay an ACK when we have any of those
840 * going out. */
841 opts->mss = tcp_advertise_mss(sk);
842 remaining -= TCPOLEN_MSS_ALIGNED;
843
844 if (likely(timestamps)) {
845 opts->options |= OPTION_TS;
846 opts->tsval = tcp_skb_timestamp_ts(tp->tcp_usec_ts, skb) + tp->tsoffset;
847 opts->tsecr = tp->rx_opt.ts_recent;
848 remaining -= TCPOLEN_TSTAMP_ALIGNED;
849 }
850 if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_window_scaling))) {
851 opts->ws = tp->rx_opt.rcv_wscale;
852 opts->options |= OPTION_WSCALE;
853 remaining -= TCPOLEN_WSCALE_ALIGNED;
854 }
855 if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_sack))) {
856 opts->options |= OPTION_SACK_ADVERTISE;
857 if (unlikely(!(OPTION_TS & opts->options)))
858 remaining -= TCPOLEN_SACKPERM_ALIGNED;
859 }
860
861 if (fastopen && fastopen->cookie.len >= 0) {
862 u32 need = fastopen->cookie.len;
863
864 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
865 TCPOLEN_FASTOPEN_BASE;
866 need = (need + 3) & ~3U; /* Align to 32 bits */
867 if (remaining >= need) {
868 opts->options |= OPTION_FAST_OPEN_COOKIE;
869 opts->fastopen_cookie = &fastopen->cookie;
870 remaining -= need;
871 tp->syn_fastopen = 1;
872 tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
873 }
874 }
875
876 smc_set_option(tp, opts, &remaining);
877
878 if (sk_is_mptcp(sk)) {
879 unsigned int size;
880
881 if (mptcp_syn_options(sk, skb, &size, &opts->mptcp)) {
882 opts->options |= OPTION_MPTCP;
883 remaining -= size;
884 }
885 }
886
887 bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining);
888
889 return MAX_TCP_OPTION_SPACE - remaining;
890}
891
892/* Set up TCP options for SYN-ACKs. */
893static unsigned int tcp_synack_options(const struct sock *sk,
894 struct request_sock *req,
895 unsigned int mss, struct sk_buff *skb,
896 struct tcp_out_options *opts,
897 const struct tcp_key *key,
898 struct tcp_fastopen_cookie *foc,
899 enum tcp_synack_type synack_type,
900 struct sk_buff *syn_skb)
901{
902 struct inet_request_sock *ireq = inet_rsk(req);
903 unsigned int remaining = MAX_TCP_OPTION_SPACE;
904
905 if (tcp_key_is_md5(key)) {
906 opts->options |= OPTION_MD5;
907 remaining -= TCPOLEN_MD5SIG_ALIGNED;
908
909 /* We can't fit any SACK blocks in a packet with MD5 + TS
910 * options. There was discussion about disabling SACK
911 * rather than TS in order to fit in better with old,
912 * buggy kernels, but that was deemed to be unnecessary.
913 */
914 if (synack_type != TCP_SYNACK_COOKIE)
915 ireq->tstamp_ok &= !ireq->sack_ok;
916 } else if (tcp_key_is_ao(key)) {
917 opts->options |= OPTION_AO;
918 remaining -= tcp_ao_len_aligned(key->ao_key);
919 ireq->tstamp_ok &= !ireq->sack_ok;
920 }
921
922 /* We always send an MSS option. */
923 opts->mss = mss;
924 remaining -= TCPOLEN_MSS_ALIGNED;
925
926 if (likely(ireq->wscale_ok)) {
927 opts->ws = ireq->rcv_wscale;
928 opts->options |= OPTION_WSCALE;
929 remaining -= TCPOLEN_WSCALE_ALIGNED;
930 }
931 if (likely(ireq->tstamp_ok)) {
932 opts->options |= OPTION_TS;
933 opts->tsval = tcp_skb_timestamp_ts(tcp_rsk(req)->req_usec_ts, skb) +
934 tcp_rsk(req)->ts_off;
935 opts->tsecr = READ_ONCE(req->ts_recent);
936 remaining -= TCPOLEN_TSTAMP_ALIGNED;
937 }
938 if (likely(ireq->sack_ok)) {
939 opts->options |= OPTION_SACK_ADVERTISE;
940 if (unlikely(!ireq->tstamp_ok))
941 remaining -= TCPOLEN_SACKPERM_ALIGNED;
942 }
943 if (foc != NULL && foc->len >= 0) {
944 u32 need = foc->len;
945
946 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
947 TCPOLEN_FASTOPEN_BASE;
948 need = (need + 3) & ~3U; /* Align to 32 bits */
949 if (remaining >= need) {
950 opts->options |= OPTION_FAST_OPEN_COOKIE;
951 opts->fastopen_cookie = foc;
952 remaining -= need;
953 }
954 }
955
956 mptcp_set_option_cond(req, opts, &remaining);
957
958 smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining);
959
960 bpf_skops_hdr_opt_len((struct sock *)sk, skb, req, syn_skb,
961 synack_type, opts, &remaining);
962
963 return MAX_TCP_OPTION_SPACE - remaining;
964}
965
966/* Compute TCP options for ESTABLISHED sockets. This is not the
967 * final wire format yet.
968 */
969static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
970 struct tcp_out_options *opts,
971 struct tcp_key *key)
972{
973 struct tcp_sock *tp = tcp_sk(sk);
974 unsigned int size = 0;
975 unsigned int eff_sacks;
976
977 opts->options = 0;
978
979 /* Better than switch (key.type) as it has static branches */
980 if (tcp_key_is_md5(key)) {
981 opts->options |= OPTION_MD5;
982 size += TCPOLEN_MD5SIG_ALIGNED;
983 } else if (tcp_key_is_ao(key)) {
984 opts->options |= OPTION_AO;
985 size += tcp_ao_len_aligned(key->ao_key);
986 }
987
988 if (likely(tp->rx_opt.tstamp_ok)) {
989 opts->options |= OPTION_TS;
990 opts->tsval = skb ? tcp_skb_timestamp_ts(tp->tcp_usec_ts, skb) +
991 tp->tsoffset : 0;
992 opts->tsecr = tp->rx_opt.ts_recent;
993 size += TCPOLEN_TSTAMP_ALIGNED;
994 }
995
996 /* MPTCP options have precedence over SACK for the limited TCP
997 * option space because a MPTCP connection would be forced to
998 * fall back to regular TCP if a required multipath option is
999 * missing. SACK still gets a chance to use whatever space is
1000 * left.
1001 */
1002 if (sk_is_mptcp(sk)) {
1003 unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
1004 unsigned int opt_size = 0;
1005
1006 if (mptcp_established_options(sk, skb, &opt_size, remaining,
1007 &opts->mptcp)) {
1008 opts->options |= OPTION_MPTCP;
1009 size += opt_size;
1010 }
1011 }
1012
1013 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
1014 if (unlikely(eff_sacks)) {
1015 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
1016 if (unlikely(remaining < TCPOLEN_SACK_BASE_ALIGNED +
1017 TCPOLEN_SACK_PERBLOCK))
1018 return size;
1019
1020 opts->num_sack_blocks =
1021 min_t(unsigned int, eff_sacks,
1022 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
1023 TCPOLEN_SACK_PERBLOCK);
1024
1025 size += TCPOLEN_SACK_BASE_ALIGNED +
1026 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
1027 }
1028
1029 if (unlikely(BPF_SOCK_OPS_TEST_FLAG(tp,
1030 BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG))) {
1031 unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
1032
1033 bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining);
1034
1035 size = MAX_TCP_OPTION_SPACE - remaining;
1036 }
1037
1038 return size;
1039}
1040
1041
1042/* TCP SMALL QUEUES (TSQ)
1043 *
1044 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
1045 * to reduce RTT and bufferbloat.
1046 * We do this using a special skb destructor (tcp_wfree).
1047 *
1048 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
1049 * needs to be reallocated in a driver.
1050 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
1051 *
1052 * Since transmit from skb destructor is forbidden, we use a tasklet
1053 * to process all sockets that eventually need to send more skbs.
1054 * We use one tasklet per cpu, with its own queue of sockets.
1055 */
1056struct tsq_tasklet {
1057 struct tasklet_struct tasklet;
1058 struct list_head head; /* queue of tcp sockets */
1059};
1060static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
1061
1062static void tcp_tsq_write(struct sock *sk)
1063{
1064 if ((1 << sk->sk_state) &
1065 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
1066 TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) {
1067 struct tcp_sock *tp = tcp_sk(sk);
1068
1069 if (tp->lost_out > tp->retrans_out &&
1070 tcp_snd_cwnd(tp) > tcp_packets_in_flight(tp)) {
1071 tcp_mstamp_refresh(tp);
1072 tcp_xmit_retransmit_queue(sk);
1073 }
1074
1075 tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
1076 0, GFP_ATOMIC);
1077 }
1078}
1079
1080static void tcp_tsq_handler(struct sock *sk)
1081{
1082 bh_lock_sock(sk);
1083 if (!sock_owned_by_user(sk))
1084 tcp_tsq_write(sk);
1085 else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
1086 sock_hold(sk);
1087 bh_unlock_sock(sk);
1088}
1089/*
1090 * One tasklet per cpu tries to send more skbs.
1091 * We run in tasklet context but need to disable irqs when
1092 * transferring tsq->head because tcp_wfree() might
1093 * interrupt us (non NAPI drivers)
1094 */
1095static void tcp_tasklet_func(struct tasklet_struct *t)
1096{
1097 struct tsq_tasklet *tsq = from_tasklet(tsq, t, tasklet);
1098 LIST_HEAD(list);
1099 unsigned long flags;
1100 struct list_head *q, *n;
1101 struct tcp_sock *tp;
1102 struct sock *sk;
1103
1104 local_irq_save(flags);
1105 list_splice_init(&tsq->head, &list);
1106 local_irq_restore(flags);
1107
1108 list_for_each_safe(q, n, &list) {
1109 tp = list_entry(q, struct tcp_sock, tsq_node);
1110 list_del(&tp->tsq_node);
1111
1112 sk = (struct sock *)tp;
1113 smp_mb__before_atomic();
1114 clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
1115
1116 tcp_tsq_handler(sk);
1117 sk_free(sk);
1118 }
1119}
1120
1121#define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
1122 TCPF_WRITE_TIMER_DEFERRED | \
1123 TCPF_DELACK_TIMER_DEFERRED | \
1124 TCPF_MTU_REDUCED_DEFERRED | \
1125 TCPF_ACK_DEFERRED)
1126/**
1127 * tcp_release_cb - tcp release_sock() callback
1128 * @sk: socket
1129 *
1130 * called from release_sock() to perform protocol dependent
1131 * actions before socket release.
1132 */
1133void tcp_release_cb(struct sock *sk)
1134{
1135 unsigned long flags = smp_load_acquire(&sk->sk_tsq_flags);
1136 unsigned long nflags;
1137
1138 /* perform an atomic operation only if at least one flag is set */
1139 do {
1140 if (!(flags & TCP_DEFERRED_ALL))
1141 return;
1142 nflags = flags & ~TCP_DEFERRED_ALL;
1143 } while (!try_cmpxchg(&sk->sk_tsq_flags, &flags, nflags));
1144
1145 if (flags & TCPF_TSQ_DEFERRED) {
1146 tcp_tsq_write(sk);
1147 __sock_put(sk);
1148 }
1149
1150 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
1151 tcp_write_timer_handler(sk);
1152 __sock_put(sk);
1153 }
1154 if (flags & TCPF_DELACK_TIMER_DEFERRED) {
1155 tcp_delack_timer_handler(sk);
1156 __sock_put(sk);
1157 }
1158 if (flags & TCPF_MTU_REDUCED_DEFERRED) {
1159 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
1160 __sock_put(sk);
1161 }
1162 if ((flags & TCPF_ACK_DEFERRED) && inet_csk_ack_scheduled(sk))
1163 tcp_send_ack(sk);
1164}
1165EXPORT_SYMBOL(tcp_release_cb);
1166
1167void __init tcp_tasklet_init(void)
1168{
1169 int i;
1170
1171 for_each_possible_cpu(i) {
1172 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
1173
1174 INIT_LIST_HEAD(&tsq->head);
1175 tasklet_setup(&tsq->tasklet, tcp_tasklet_func);
1176 }
1177}
1178
1179/*
1180 * Write buffer destructor automatically called from kfree_skb.
1181 * We can't xmit new skbs from this context, as we might already
1182 * hold qdisc lock.
1183 */
1184void tcp_wfree(struct sk_buff *skb)
1185{
1186 struct sock *sk = skb->sk;
1187 struct tcp_sock *tp = tcp_sk(sk);
1188 unsigned long flags, nval, oval;
1189 struct tsq_tasklet *tsq;
1190 bool empty;
1191
1192 /* Keep one reference on sk_wmem_alloc.
1193 * Will be released by sk_free() from here or tcp_tasklet_func()
1194 */
1195 WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc));
1196
1197 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
1198 * Wait until our queues (qdisc + devices) are drained.
1199 * This gives :
1200 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
1201 * - chance for incoming ACK (processed by another cpu maybe)
1202 * to migrate this flow (skb->ooo_okay will be eventually set)
1203 */
1204 if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
1205 goto out;
1206
1207 oval = smp_load_acquire(&sk->sk_tsq_flags);
1208 do {
1209 if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
1210 goto out;
1211
1212 nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED;
1213 } while (!try_cmpxchg(&sk->sk_tsq_flags, &oval, nval));
1214
1215 /* queue this socket to tasklet queue */
1216 local_irq_save(flags);
1217 tsq = this_cpu_ptr(&tsq_tasklet);
1218 empty = list_empty(&tsq->head);
1219 list_add(&tp->tsq_node, &tsq->head);
1220 if (empty)
1221 tasklet_schedule(&tsq->tasklet);
1222 local_irq_restore(flags);
1223 return;
1224out:
1225 sk_free(sk);
1226}
1227
1228/* Note: Called under soft irq.
1229 * We can call TCP stack right away, unless socket is owned by user.
1230 */
1231enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer)
1232{
1233 struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer);
1234 struct sock *sk = (struct sock *)tp;
1235
1236 tcp_tsq_handler(sk);
1237 sock_put(sk);
1238
1239 return HRTIMER_NORESTART;
1240}
1241
1242static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb,
1243 u64 prior_wstamp)
1244{
1245 struct tcp_sock *tp = tcp_sk(sk);
1246
1247 if (sk->sk_pacing_status != SK_PACING_NONE) {
1248 unsigned long rate = READ_ONCE(sk->sk_pacing_rate);
1249
1250 /* Original sch_fq does not pace first 10 MSS
1251 * Note that tp->data_segs_out overflows after 2^32 packets,
1252 * this is a minor annoyance.
1253 */
1254 if (rate != ~0UL && rate && tp->data_segs_out >= 10) {
1255 u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate);
1256 u64 credit = tp->tcp_wstamp_ns - prior_wstamp;
1257
1258 /* take into account OS jitter */
1259 len_ns -= min_t(u64, len_ns / 2, credit);
1260 tp->tcp_wstamp_ns += len_ns;
1261 }
1262 }
1263 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
1264}
1265
1266INDIRECT_CALLABLE_DECLARE(int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl));
1267INDIRECT_CALLABLE_DECLARE(int inet6_csk_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl));
1268INDIRECT_CALLABLE_DECLARE(void tcp_v4_send_check(struct sock *sk, struct sk_buff *skb));
1269
1270/* This routine actually transmits TCP packets queued in by
1271 * tcp_do_sendmsg(). This is used by both the initial
1272 * transmission and possible later retransmissions.
1273 * All SKB's seen here are completely headerless. It is our
1274 * job to build the TCP header, and pass the packet down to
1275 * IP so it can do the same plus pass the packet off to the
1276 * device.
1277 *
1278 * We are working here with either a clone of the original
1279 * SKB, or a fresh unique copy made by the retransmit engine.
1280 */
1281static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb,
1282 int clone_it, gfp_t gfp_mask, u32 rcv_nxt)
1283{
1284 const struct inet_connection_sock *icsk = inet_csk(sk);
1285 struct inet_sock *inet;
1286 struct tcp_sock *tp;
1287 struct tcp_skb_cb *tcb;
1288 struct tcp_out_options opts;
1289 unsigned int tcp_options_size, tcp_header_size;
1290 struct sk_buff *oskb = NULL;
1291 struct tcp_key key;
1292 struct tcphdr *th;
1293 u64 prior_wstamp;
1294 int err;
1295
1296 BUG_ON(!skb || !tcp_skb_pcount(skb));
1297 tp = tcp_sk(sk);
1298 prior_wstamp = tp->tcp_wstamp_ns;
1299 tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache);
1300 skb_set_delivery_time(skb, tp->tcp_wstamp_ns, true);
1301 if (clone_it) {
1302 oskb = skb;
1303
1304 tcp_skb_tsorted_save(oskb) {
1305 if (unlikely(skb_cloned(oskb)))
1306 skb = pskb_copy(oskb, gfp_mask);
1307 else
1308 skb = skb_clone(oskb, gfp_mask);
1309 } tcp_skb_tsorted_restore(oskb);
1310
1311 if (unlikely(!skb))
1312 return -ENOBUFS;
1313 /* retransmit skbs might have a non zero value in skb->dev
1314 * because skb->dev is aliased with skb->rbnode.rb_left
1315 */
1316 skb->dev = NULL;
1317 }
1318
1319 inet = inet_sk(sk);
1320 tcb = TCP_SKB_CB(skb);
1321 memset(&opts, 0, sizeof(opts));
1322
1323 tcp_get_current_key(sk, &key);
1324 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
1325 tcp_options_size = tcp_syn_options(sk, skb, &opts, &key);
1326 } else {
1327 tcp_options_size = tcp_established_options(sk, skb, &opts, &key);
1328 /* Force a PSH flag on all (GSO) packets to expedite GRO flush
1329 * at receiver : This slightly improve GRO performance.
1330 * Note that we do not force the PSH flag for non GSO packets,
1331 * because they might be sent under high congestion events,
1332 * and in this case it is better to delay the delivery of 1-MSS
1333 * packets and thus the corresponding ACK packet that would
1334 * release the following packet.
1335 */
1336 if (tcp_skb_pcount(skb) > 1)
1337 tcb->tcp_flags |= TCPHDR_PSH;
1338 }
1339 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1340
1341 /* We set skb->ooo_okay to one if this packet can select
1342 * a different TX queue than prior packets of this flow,
1343 * to avoid self inflicted reorders.
1344 * The 'other' queue decision is based on current cpu number
1345 * if XPS is enabled, or sk->sk_txhash otherwise.
1346 * We can switch to another (and better) queue if:
1347 * 1) No packet with payload is in qdisc/device queues.
1348 * Delays in TX completion can defeat the test
1349 * even if packets were already sent.
1350 * 2) Or rtx queue is empty.
1351 * This mitigates above case if ACK packets for
1352 * all prior packets were already processed.
1353 */
1354 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1) ||
1355 tcp_rtx_queue_empty(sk);
1356
1357 /* If we had to use memory reserve to allocate this skb,
1358 * this might cause drops if packet is looped back :
1359 * Other socket might not have SOCK_MEMALLOC.
1360 * Packets not looped back do not care about pfmemalloc.
1361 */
1362 skb->pfmemalloc = 0;
1363
1364 skb_push(skb, tcp_header_size);
1365 skb_reset_transport_header(skb);
1366
1367 skb_orphan(skb);
1368 skb->sk = sk;
1369 skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
1370 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
1371
1372 skb_set_dst_pending_confirm(skb, READ_ONCE(sk->sk_dst_pending_confirm));
1373
1374 /* Build TCP header and checksum it. */
1375 th = (struct tcphdr *)skb->data;
1376 th->source = inet->inet_sport;
1377 th->dest = inet->inet_dport;
1378 th->seq = htonl(tcb->seq);
1379 th->ack_seq = htonl(rcv_nxt);
1380 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
1381 tcb->tcp_flags);
1382
1383 th->check = 0;
1384 th->urg_ptr = 0;
1385
1386 /* The urg_mode check is necessary during a below snd_una win probe */
1387 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1388 if (before(tp->snd_up, tcb->seq + 0x10000)) {
1389 th->urg_ptr = htons(tp->snd_up - tcb->seq);
1390 th->urg = 1;
1391 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1392 th->urg_ptr = htons(0xFFFF);
1393 th->urg = 1;
1394 }
1395 }
1396
1397 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1398 if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1399 th->window = htons(tcp_select_window(sk));
1400 tcp_ecn_send(sk, skb, th, tcp_header_size);
1401 } else {
1402 /* RFC1323: The window in SYN & SYN/ACK segments
1403 * is never scaled.
1404 */
1405 th->window = htons(min(tp->rcv_wnd, 65535U));
1406 }
1407
1408 tcp_options_write(th, tp, NULL, &opts, &key);
1409
1410 if (tcp_key_is_md5(&key)) {
1411#ifdef CONFIG_TCP_MD5SIG
1412 /* Calculate the MD5 hash, as we have all we need now */
1413 sk_gso_disable(sk);
1414 tp->af_specific->calc_md5_hash(opts.hash_location,
1415 key.md5_key, sk, skb);
1416#endif
1417 } else if (tcp_key_is_ao(&key)) {
1418 int err;
1419
1420 err = tcp_ao_transmit_skb(sk, skb, key.ao_key, th,
1421 opts.hash_location);
1422 if (err) {
1423 kfree_skb_reason(skb, SKB_DROP_REASON_NOT_SPECIFIED);
1424 return -ENOMEM;
1425 }
1426 }
1427
1428 /* BPF prog is the last one writing header option */
1429 bpf_skops_write_hdr_opt(sk, skb, NULL, NULL, 0, &opts);
1430
1431 INDIRECT_CALL_INET(icsk->icsk_af_ops->send_check,
1432 tcp_v6_send_check, tcp_v4_send_check,
1433 sk, skb);
1434
1435 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1436 tcp_event_ack_sent(sk, rcv_nxt);
1437
1438 if (skb->len != tcp_header_size) {
1439 tcp_event_data_sent(tp, sk);
1440 tp->data_segs_out += tcp_skb_pcount(skb);
1441 tp->bytes_sent += skb->len - tcp_header_size;
1442 }
1443
1444 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1445 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1446 tcp_skb_pcount(skb));
1447
1448 tp->segs_out += tcp_skb_pcount(skb);
1449 skb_set_hash_from_sk(skb, sk);
1450 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1451 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1452 skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1453
1454 /* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */
1455
1456 /* Cleanup our debris for IP stacks */
1457 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1458 sizeof(struct inet6_skb_parm)));
1459
1460 tcp_add_tx_delay(skb, tp);
1461
1462 err = INDIRECT_CALL_INET(icsk->icsk_af_ops->queue_xmit,
1463 inet6_csk_xmit, ip_queue_xmit,
1464 sk, skb, &inet->cork.fl);
1465
1466 if (unlikely(err > 0)) {
1467 tcp_enter_cwr(sk);
1468 err = net_xmit_eval(err);
1469 }
1470 if (!err && oskb) {
1471 tcp_update_skb_after_send(sk, oskb, prior_wstamp);
1472 tcp_rate_skb_sent(sk, oskb);
1473 }
1474 return err;
1475}
1476
1477static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
1478 gfp_t gfp_mask)
1479{
1480 return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask,
1481 tcp_sk(sk)->rcv_nxt);
1482}
1483
1484/* This routine just queues the buffer for sending.
1485 *
1486 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1487 * otherwise socket can stall.
1488 */
1489static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1490{
1491 struct tcp_sock *tp = tcp_sk(sk);
1492
1493 /* Advance write_seq and place onto the write_queue. */
1494 WRITE_ONCE(tp->write_seq, TCP_SKB_CB(skb)->end_seq);
1495 __skb_header_release(skb);
1496 tcp_add_write_queue_tail(sk, skb);
1497 sk_wmem_queued_add(sk, skb->truesize);
1498 sk_mem_charge(sk, skb->truesize);
1499}
1500
1501/* Initialize TSO segments for a packet. */
1502static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1503{
1504 if (skb->len <= mss_now) {
1505 /* Avoid the costly divide in the normal
1506 * non-TSO case.
1507 */
1508 tcp_skb_pcount_set(skb, 1);
1509 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1510 } else {
1511 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1512 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1513 }
1514}
1515
1516/* Pcount in the middle of the write queue got changed, we need to do various
1517 * tweaks to fix counters
1518 */
1519static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1520{
1521 struct tcp_sock *tp = tcp_sk(sk);
1522
1523 tp->packets_out -= decr;
1524
1525 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1526 tp->sacked_out -= decr;
1527 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1528 tp->retrans_out -= decr;
1529 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1530 tp->lost_out -= decr;
1531
1532 /* Reno case is special. Sigh... */
1533 if (tcp_is_reno(tp) && decr > 0)
1534 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1535
1536 if (tp->lost_skb_hint &&
1537 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1538 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
1539 tp->lost_cnt_hint -= decr;
1540
1541 tcp_verify_left_out(tp);
1542}
1543
1544static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1545{
1546 return TCP_SKB_CB(skb)->txstamp_ack ||
1547 (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1548}
1549
1550static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1551{
1552 struct skb_shared_info *shinfo = skb_shinfo(skb);
1553
1554 if (unlikely(tcp_has_tx_tstamp(skb)) &&
1555 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1556 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1557 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1558
1559 shinfo->tx_flags &= ~tsflags;
1560 shinfo2->tx_flags |= tsflags;
1561 swap(shinfo->tskey, shinfo2->tskey);
1562 TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1563 TCP_SKB_CB(skb)->txstamp_ack = 0;
1564 }
1565}
1566
1567static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1568{
1569 TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1570 TCP_SKB_CB(skb)->eor = 0;
1571}
1572
1573/* Insert buff after skb on the write or rtx queue of sk. */
1574static void tcp_insert_write_queue_after(struct sk_buff *skb,
1575 struct sk_buff *buff,
1576 struct sock *sk,
1577 enum tcp_queue tcp_queue)
1578{
1579 if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE)
1580 __skb_queue_after(&sk->sk_write_queue, skb, buff);
1581 else
1582 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
1583}
1584
1585/* Function to create two new TCP segments. Shrinks the given segment
1586 * to the specified size and appends a new segment with the rest of the
1587 * packet to the list. This won't be called frequently, I hope.
1588 * Remember, these are still headerless SKBs at this point.
1589 */
1590int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1591 struct sk_buff *skb, u32 len,
1592 unsigned int mss_now, gfp_t gfp)
1593{
1594 struct tcp_sock *tp = tcp_sk(sk);
1595 struct sk_buff *buff;
1596 int old_factor;
1597 long limit;
1598 int nlen;
1599 u8 flags;
1600
1601 if (WARN_ON(len > skb->len))
1602 return -EINVAL;
1603
1604 DEBUG_NET_WARN_ON_ONCE(skb_headlen(skb));
1605
1606 /* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb.
1607 * We need some allowance to not penalize applications setting small
1608 * SO_SNDBUF values.
1609 * Also allow first and last skb in retransmit queue to be split.
1610 */
1611 limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_LEGACY_MAX_SIZE);
1612 if (unlikely((sk->sk_wmem_queued >> 1) > limit &&
1613 tcp_queue != TCP_FRAG_IN_WRITE_QUEUE &&
1614 skb != tcp_rtx_queue_head(sk) &&
1615 skb != tcp_rtx_queue_tail(sk))) {
1616 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG);
1617 return -ENOMEM;
1618 }
1619
1620 if (skb_unclone_keeptruesize(skb, gfp))
1621 return -ENOMEM;
1622
1623 /* Get a new skb... force flag on. */
1624 buff = tcp_stream_alloc_skb(sk, gfp, true);
1625 if (!buff)
1626 return -ENOMEM; /* We'll just try again later. */
1627 skb_copy_decrypted(buff, skb);
1628 mptcp_skb_ext_copy(buff, skb);
1629
1630 sk_wmem_queued_add(sk, buff->truesize);
1631 sk_mem_charge(sk, buff->truesize);
1632 nlen = skb->len - len;
1633 buff->truesize += nlen;
1634 skb->truesize -= nlen;
1635
1636 /* Correct the sequence numbers. */
1637 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1638 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1639 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1640
1641 /* PSH and FIN should only be set in the second packet. */
1642 flags = TCP_SKB_CB(skb)->tcp_flags;
1643 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1644 TCP_SKB_CB(buff)->tcp_flags = flags;
1645 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1646 tcp_skb_fragment_eor(skb, buff);
1647
1648 skb_split(skb, buff, len);
1649
1650 skb_set_delivery_time(buff, skb->tstamp, true);
1651 tcp_fragment_tstamp(skb, buff);
1652
1653 old_factor = tcp_skb_pcount(skb);
1654
1655 /* Fix up tso_factor for both original and new SKB. */
1656 tcp_set_skb_tso_segs(skb, mss_now);
1657 tcp_set_skb_tso_segs(buff, mss_now);
1658
1659 /* Update delivered info for the new segment */
1660 TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1661
1662 /* If this packet has been sent out already, we must
1663 * adjust the various packet counters.
1664 */
1665 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1666 int diff = old_factor - tcp_skb_pcount(skb) -
1667 tcp_skb_pcount(buff);
1668
1669 if (diff)
1670 tcp_adjust_pcount(sk, skb, diff);
1671 }
1672
1673 /* Link BUFF into the send queue. */
1674 __skb_header_release(buff);
1675 tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1676 if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE)
1677 list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor);
1678
1679 return 0;
1680}
1681
1682/* This is similar to __pskb_pull_tail(). The difference is that pulled
1683 * data is not copied, but immediately discarded.
1684 */
1685static int __pskb_trim_head(struct sk_buff *skb, int len)
1686{
1687 struct skb_shared_info *shinfo;
1688 int i, k, eat;
1689
1690 DEBUG_NET_WARN_ON_ONCE(skb_headlen(skb));
1691 eat = len;
1692 k = 0;
1693 shinfo = skb_shinfo(skb);
1694 for (i = 0; i < shinfo->nr_frags; i++) {
1695 int size = skb_frag_size(&shinfo->frags[i]);
1696
1697 if (size <= eat) {
1698 skb_frag_unref(skb, i);
1699 eat -= size;
1700 } else {
1701 shinfo->frags[k] = shinfo->frags[i];
1702 if (eat) {
1703 skb_frag_off_add(&shinfo->frags[k], eat);
1704 skb_frag_size_sub(&shinfo->frags[k], eat);
1705 eat = 0;
1706 }
1707 k++;
1708 }
1709 }
1710 shinfo->nr_frags = k;
1711
1712 skb->data_len -= len;
1713 skb->len = skb->data_len;
1714 return len;
1715}
1716
1717/* Remove acked data from a packet in the transmit queue. */
1718int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1719{
1720 u32 delta_truesize;
1721
1722 if (skb_unclone_keeptruesize(skb, GFP_ATOMIC))
1723 return -ENOMEM;
1724
1725 delta_truesize = __pskb_trim_head(skb, len);
1726
1727 TCP_SKB_CB(skb)->seq += len;
1728
1729 skb->truesize -= delta_truesize;
1730 sk_wmem_queued_add(sk, -delta_truesize);
1731 if (!skb_zcopy_pure(skb))
1732 sk_mem_uncharge(sk, delta_truesize);
1733
1734 /* Any change of skb->len requires recalculation of tso factor. */
1735 if (tcp_skb_pcount(skb) > 1)
1736 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1737
1738 return 0;
1739}
1740
1741/* Calculate MSS not accounting any TCP options. */
1742static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1743{
1744 const struct tcp_sock *tp = tcp_sk(sk);
1745 const struct inet_connection_sock *icsk = inet_csk(sk);
1746 int mss_now;
1747
1748 /* Calculate base mss without TCP options:
1749 It is MMS_S - sizeof(tcphdr) of rfc1122
1750 */
1751 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1752
1753 /* Clamp it (mss_clamp does not include tcp options) */
1754 if (mss_now > tp->rx_opt.mss_clamp)
1755 mss_now = tp->rx_opt.mss_clamp;
1756
1757 /* Now subtract optional transport overhead */
1758 mss_now -= icsk->icsk_ext_hdr_len;
1759
1760 /* Then reserve room for full set of TCP options and 8 bytes of data */
1761 mss_now = max(mss_now,
1762 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss));
1763 return mss_now;
1764}
1765
1766/* Calculate MSS. Not accounting for SACKs here. */
1767int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1768{
1769 /* Subtract TCP options size, not including SACKs */
1770 return __tcp_mtu_to_mss(sk, pmtu) -
1771 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1772}
1773EXPORT_SYMBOL(tcp_mtu_to_mss);
1774
1775/* Inverse of above */
1776int tcp_mss_to_mtu(struct sock *sk, int mss)
1777{
1778 const struct tcp_sock *tp = tcp_sk(sk);
1779 const struct inet_connection_sock *icsk = inet_csk(sk);
1780
1781 return mss +
1782 tp->tcp_header_len +
1783 icsk->icsk_ext_hdr_len +
1784 icsk->icsk_af_ops->net_header_len;
1785}
1786EXPORT_SYMBOL(tcp_mss_to_mtu);
1787
1788/* MTU probing init per socket */
1789void tcp_mtup_init(struct sock *sk)
1790{
1791 struct tcp_sock *tp = tcp_sk(sk);
1792 struct inet_connection_sock *icsk = inet_csk(sk);
1793 struct net *net = sock_net(sk);
1794
1795 icsk->icsk_mtup.enabled = READ_ONCE(net->ipv4.sysctl_tcp_mtu_probing) > 1;
1796 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1797 icsk->icsk_af_ops->net_header_len;
1798 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, READ_ONCE(net->ipv4.sysctl_tcp_base_mss));
1799 icsk->icsk_mtup.probe_size = 0;
1800 if (icsk->icsk_mtup.enabled)
1801 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
1802}
1803EXPORT_SYMBOL(tcp_mtup_init);
1804
1805/* This function synchronize snd mss to current pmtu/exthdr set.
1806
1807 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1808 for TCP options, but includes only bare TCP header.
1809
1810 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1811 It is minimum of user_mss and mss received with SYN.
1812 It also does not include TCP options.
1813
1814 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1815
1816 tp->mss_cache is current effective sending mss, including
1817 all tcp options except for SACKs. It is evaluated,
1818 taking into account current pmtu, but never exceeds
1819 tp->rx_opt.mss_clamp.
1820
1821 NOTE1. rfc1122 clearly states that advertised MSS
1822 DOES NOT include either tcp or ip options.
1823
1824 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1825 are READ ONLY outside this function. --ANK (980731)
1826 */
1827unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1828{
1829 struct tcp_sock *tp = tcp_sk(sk);
1830 struct inet_connection_sock *icsk = inet_csk(sk);
1831 int mss_now;
1832
1833 if (icsk->icsk_mtup.search_high > pmtu)
1834 icsk->icsk_mtup.search_high = pmtu;
1835
1836 mss_now = tcp_mtu_to_mss(sk, pmtu);
1837 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1838
1839 /* And store cached results */
1840 icsk->icsk_pmtu_cookie = pmtu;
1841 if (icsk->icsk_mtup.enabled)
1842 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1843 tp->mss_cache = mss_now;
1844
1845 return mss_now;
1846}
1847EXPORT_SYMBOL(tcp_sync_mss);
1848
1849/* Compute the current effective MSS, taking SACKs and IP options,
1850 * and even PMTU discovery events into account.
1851 */
1852unsigned int tcp_current_mss(struct sock *sk)
1853{
1854 const struct tcp_sock *tp = tcp_sk(sk);
1855 const struct dst_entry *dst = __sk_dst_get(sk);
1856 u32 mss_now;
1857 unsigned int header_len;
1858 struct tcp_out_options opts;
1859 struct tcp_key key;
1860
1861 mss_now = tp->mss_cache;
1862
1863 if (dst) {
1864 u32 mtu = dst_mtu(dst);
1865 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1866 mss_now = tcp_sync_mss(sk, mtu);
1867 }
1868 tcp_get_current_key(sk, &key);
1869 header_len = tcp_established_options(sk, NULL, &opts, &key) +
1870 sizeof(struct tcphdr);
1871 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1872 * some common options. If this is an odd packet (because we have SACK
1873 * blocks etc) then our calculated header_len will be different, and
1874 * we have to adjust mss_now correspondingly */
1875 if (header_len != tp->tcp_header_len) {
1876 int delta = (int) header_len - tp->tcp_header_len;
1877 mss_now -= delta;
1878 }
1879
1880 return mss_now;
1881}
1882
1883/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1884 * As additional protections, we do not touch cwnd in retransmission phases,
1885 * and if application hit its sndbuf limit recently.
1886 */
1887static void tcp_cwnd_application_limited(struct sock *sk)
1888{
1889 struct tcp_sock *tp = tcp_sk(sk);
1890
1891 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1892 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1893 /* Limited by application or receiver window. */
1894 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1895 u32 win_used = max(tp->snd_cwnd_used, init_win);
1896 if (win_used < tcp_snd_cwnd(tp)) {
1897 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1898 tcp_snd_cwnd_set(tp, (tcp_snd_cwnd(tp) + win_used) >> 1);
1899 }
1900 tp->snd_cwnd_used = 0;
1901 }
1902 tp->snd_cwnd_stamp = tcp_jiffies32;
1903}
1904
1905static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1906{
1907 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1908 struct tcp_sock *tp = tcp_sk(sk);
1909
1910 /* Track the strongest available signal of the degree to which the cwnd
1911 * is fully utilized. If cwnd-limited then remember that fact for the
1912 * current window. If not cwnd-limited then track the maximum number of
1913 * outstanding packets in the current window. (If cwnd-limited then we
1914 * chose to not update tp->max_packets_out to avoid an extra else
1915 * clause with no functional impact.)
1916 */
1917 if (!before(tp->snd_una, tp->cwnd_usage_seq) ||
1918 is_cwnd_limited ||
1919 (!tp->is_cwnd_limited &&
1920 tp->packets_out > tp->max_packets_out)) {
1921 tp->is_cwnd_limited = is_cwnd_limited;
1922 tp->max_packets_out = tp->packets_out;
1923 tp->cwnd_usage_seq = tp->snd_nxt;
1924 }
1925
1926 if (tcp_is_cwnd_limited(sk)) {
1927 /* Network is feed fully. */
1928 tp->snd_cwnd_used = 0;
1929 tp->snd_cwnd_stamp = tcp_jiffies32;
1930 } else {
1931 /* Network starves. */
1932 if (tp->packets_out > tp->snd_cwnd_used)
1933 tp->snd_cwnd_used = tp->packets_out;
1934
1935 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle) &&
1936 (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto &&
1937 !ca_ops->cong_control)
1938 tcp_cwnd_application_limited(sk);
1939
1940 /* The following conditions together indicate the starvation
1941 * is caused by insufficient sender buffer:
1942 * 1) just sent some data (see tcp_write_xmit)
1943 * 2) not cwnd limited (this else condition)
1944 * 3) no more data to send (tcp_write_queue_empty())
1945 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1946 */
1947 if (tcp_write_queue_empty(sk) && sk->sk_socket &&
1948 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
1949 (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1950 tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED);
1951 }
1952}
1953
1954/* Minshall's variant of the Nagle send check. */
1955static bool tcp_minshall_check(const struct tcp_sock *tp)
1956{
1957 return after(tp->snd_sml, tp->snd_una) &&
1958 !after(tp->snd_sml, tp->snd_nxt);
1959}
1960
1961/* Update snd_sml if this skb is under mss
1962 * Note that a TSO packet might end with a sub-mss segment
1963 * The test is really :
1964 * if ((skb->len % mss) != 0)
1965 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1966 * But we can avoid doing the divide again given we already have
1967 * skb_pcount = skb->len / mss_now
1968 */
1969static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1970 const struct sk_buff *skb)
1971{
1972 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1973 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1974}
1975
1976/* Return false, if packet can be sent now without violation Nagle's rules:
1977 * 1. It is full sized. (provided by caller in %partial bool)
1978 * 2. Or it contains FIN. (already checked by caller)
1979 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1980 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1981 * With Minshall's modification: all sent small packets are ACKed.
1982 */
1983static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1984 int nonagle)
1985{
1986 return partial &&
1987 ((nonagle & TCP_NAGLE_CORK) ||
1988 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1989}
1990
1991/* Return how many segs we'd like on a TSO packet,
1992 * depending on current pacing rate, and how close the peer is.
1993 *
1994 * Rationale is:
1995 * - For close peers, we rather send bigger packets to reduce
1996 * cpu costs, because occasional losses will be repaired fast.
1997 * - For long distance/rtt flows, we would like to get ACK clocking
1998 * with 1 ACK per ms.
1999 *
2000 * Use min_rtt to help adapt TSO burst size, with smaller min_rtt resulting
2001 * in bigger TSO bursts. We we cut the RTT-based allowance in half
2002 * for every 2^9 usec (aka 512 us) of RTT, so that the RTT-based allowance
2003 * is below 1500 bytes after 6 * ~500 usec = 3ms.
2004 */
2005static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
2006 int min_tso_segs)
2007{
2008 unsigned long bytes;
2009 u32 r;
2010
2011 bytes = READ_ONCE(sk->sk_pacing_rate) >> READ_ONCE(sk->sk_pacing_shift);
2012
2013 r = tcp_min_rtt(tcp_sk(sk)) >> READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_rtt_log);
2014 if (r < BITS_PER_TYPE(sk->sk_gso_max_size))
2015 bytes += sk->sk_gso_max_size >> r;
2016
2017 bytes = min_t(unsigned long, bytes, sk->sk_gso_max_size);
2018
2019 return max_t(u32, bytes / mss_now, min_tso_segs);
2020}
2021
2022/* Return the number of segments we want in the skb we are transmitting.
2023 * See if congestion control module wants to decide; otherwise, autosize.
2024 */
2025static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
2026{
2027 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2028 u32 min_tso, tso_segs;
2029
2030 min_tso = ca_ops->min_tso_segs ?
2031 ca_ops->min_tso_segs(sk) :
2032 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs);
2033
2034 tso_segs = tcp_tso_autosize(sk, mss_now, min_tso);
2035 return min_t(u32, tso_segs, sk->sk_gso_max_segs);
2036}
2037
2038/* Returns the portion of skb which can be sent right away */
2039static unsigned int tcp_mss_split_point(const struct sock *sk,
2040 const struct sk_buff *skb,
2041 unsigned int mss_now,
2042 unsigned int max_segs,
2043 int nonagle)
2044{
2045 const struct tcp_sock *tp = tcp_sk(sk);
2046 u32 partial, needed, window, max_len;
2047
2048 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2049 max_len = mss_now * max_segs;
2050
2051 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
2052 return max_len;
2053
2054 needed = min(skb->len, window);
2055
2056 if (max_len <= needed)
2057 return max_len;
2058
2059 partial = needed % mss_now;
2060 /* If last segment is not a full MSS, check if Nagle rules allow us
2061 * to include this last segment in this skb.
2062 * Otherwise, we'll split the skb at last MSS boundary
2063 */
2064 if (tcp_nagle_check(partial != 0, tp, nonagle))
2065 return needed - partial;
2066
2067 return needed;
2068}
2069
2070/* Can at least one segment of SKB be sent right now, according to the
2071 * congestion window rules? If so, return how many segments are allowed.
2072 */
2073static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
2074 const struct sk_buff *skb)
2075{
2076 u32 in_flight, cwnd, halfcwnd;
2077
2078 /* Don't be strict about the congestion window for the final FIN. */
2079 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
2080 tcp_skb_pcount(skb) == 1)
2081 return 1;
2082
2083 in_flight = tcp_packets_in_flight(tp);
2084 cwnd = tcp_snd_cwnd(tp);
2085 if (in_flight >= cwnd)
2086 return 0;
2087
2088 /* For better scheduling, ensure we have at least
2089 * 2 GSO packets in flight.
2090 */
2091 halfcwnd = max(cwnd >> 1, 1U);
2092 return min(halfcwnd, cwnd - in_flight);
2093}
2094
2095/* Initialize TSO state of a skb.
2096 * This must be invoked the first time we consider transmitting
2097 * SKB onto the wire.
2098 */
2099static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
2100{
2101 int tso_segs = tcp_skb_pcount(skb);
2102
2103 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
2104 tcp_set_skb_tso_segs(skb, mss_now);
2105 tso_segs = tcp_skb_pcount(skb);
2106 }
2107 return tso_segs;
2108}
2109
2110
2111/* Return true if the Nagle test allows this packet to be
2112 * sent now.
2113 */
2114static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
2115 unsigned int cur_mss, int nonagle)
2116{
2117 /* Nagle rule does not apply to frames, which sit in the middle of the
2118 * write_queue (they have no chances to get new data).
2119 *
2120 * This is implemented in the callers, where they modify the 'nonagle'
2121 * argument based upon the location of SKB in the send queue.
2122 */
2123 if (nonagle & TCP_NAGLE_PUSH)
2124 return true;
2125
2126 /* Don't use the nagle rule for urgent data (or for the final FIN). */
2127 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
2128 return true;
2129
2130 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
2131 return true;
2132
2133 return false;
2134}
2135
2136/* Does at least the first segment of SKB fit into the send window? */
2137static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
2138 const struct sk_buff *skb,
2139 unsigned int cur_mss)
2140{
2141 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
2142
2143 if (skb->len > cur_mss)
2144 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
2145
2146 return !after(end_seq, tcp_wnd_end(tp));
2147}
2148
2149/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
2150 * which is put after SKB on the list. It is very much like
2151 * tcp_fragment() except that it may make several kinds of assumptions
2152 * in order to speed up the splitting operation. In particular, we
2153 * know that all the data is in scatter-gather pages, and that the
2154 * packet has never been sent out before (and thus is not cloned).
2155 */
2156static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
2157 unsigned int mss_now, gfp_t gfp)
2158{
2159 int nlen = skb->len - len;
2160 struct sk_buff *buff;
2161 u8 flags;
2162
2163 /* All of a TSO frame must be composed of paged data. */
2164 DEBUG_NET_WARN_ON_ONCE(skb->len != skb->data_len);
2165
2166 buff = tcp_stream_alloc_skb(sk, gfp, true);
2167 if (unlikely(!buff))
2168 return -ENOMEM;
2169 skb_copy_decrypted(buff, skb);
2170 mptcp_skb_ext_copy(buff, skb);
2171
2172 sk_wmem_queued_add(sk, buff->truesize);
2173 sk_mem_charge(sk, buff->truesize);
2174 buff->truesize += nlen;
2175 skb->truesize -= nlen;
2176
2177 /* Correct the sequence numbers. */
2178 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
2179 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
2180 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
2181
2182 /* PSH and FIN should only be set in the second packet. */
2183 flags = TCP_SKB_CB(skb)->tcp_flags;
2184 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
2185 TCP_SKB_CB(buff)->tcp_flags = flags;
2186
2187 tcp_skb_fragment_eor(skb, buff);
2188
2189 skb_split(skb, buff, len);
2190 tcp_fragment_tstamp(skb, buff);
2191
2192 /* Fix up tso_factor for both original and new SKB. */
2193 tcp_set_skb_tso_segs(skb, mss_now);
2194 tcp_set_skb_tso_segs(buff, mss_now);
2195
2196 /* Link BUFF into the send queue. */
2197 __skb_header_release(buff);
2198 tcp_insert_write_queue_after(skb, buff, sk, TCP_FRAG_IN_WRITE_QUEUE);
2199
2200 return 0;
2201}
2202
2203/* Try to defer sending, if possible, in order to minimize the amount
2204 * of TSO splitting we do. View it as a kind of TSO Nagle test.
2205 *
2206 * This algorithm is from John Heffner.
2207 */
2208static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
2209 bool *is_cwnd_limited,
2210 bool *is_rwnd_limited,
2211 u32 max_segs)
2212{
2213 const struct inet_connection_sock *icsk = inet_csk(sk);
2214 u32 send_win, cong_win, limit, in_flight;
2215 struct tcp_sock *tp = tcp_sk(sk);
2216 struct sk_buff *head;
2217 int win_divisor;
2218 s64 delta;
2219
2220 if (icsk->icsk_ca_state >= TCP_CA_Recovery)
2221 goto send_now;
2222
2223 /* Avoid bursty behavior by allowing defer
2224 * only if the last write was recent (1 ms).
2225 * Note that tp->tcp_wstamp_ns can be in the future if we have
2226 * packets waiting in a qdisc or device for EDT delivery.
2227 */
2228 delta = tp->tcp_clock_cache - tp->tcp_wstamp_ns - NSEC_PER_MSEC;
2229 if (delta > 0)
2230 goto send_now;
2231
2232 in_flight = tcp_packets_in_flight(tp);
2233
2234 BUG_ON(tcp_skb_pcount(skb) <= 1);
2235 BUG_ON(tcp_snd_cwnd(tp) <= in_flight);
2236
2237 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2238
2239 /* From in_flight test above, we know that cwnd > in_flight. */
2240 cong_win = (tcp_snd_cwnd(tp) - in_flight) * tp->mss_cache;
2241
2242 limit = min(send_win, cong_win);
2243
2244 /* If a full-sized TSO skb can be sent, do it. */
2245 if (limit >= max_segs * tp->mss_cache)
2246 goto send_now;
2247
2248 /* Middle in queue won't get any more data, full sendable already? */
2249 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
2250 goto send_now;
2251
2252 win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor);
2253 if (win_divisor) {
2254 u32 chunk = min(tp->snd_wnd, tcp_snd_cwnd(tp) * tp->mss_cache);
2255
2256 /* If at least some fraction of a window is available,
2257 * just use it.
2258 */
2259 chunk /= win_divisor;
2260 if (limit >= chunk)
2261 goto send_now;
2262 } else {
2263 /* Different approach, try not to defer past a single
2264 * ACK. Receiver should ACK every other full sized
2265 * frame, so if we have space for more than 3 frames
2266 * then send now.
2267 */
2268 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
2269 goto send_now;
2270 }
2271
2272 /* TODO : use tsorted_sent_queue ? */
2273 head = tcp_rtx_queue_head(sk);
2274 if (!head)
2275 goto send_now;
2276 delta = tp->tcp_clock_cache - head->tstamp;
2277 /* If next ACK is likely to come too late (half srtt), do not defer */
2278 if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0)
2279 goto send_now;
2280
2281 /* Ok, it looks like it is advisable to defer.
2282 * Three cases are tracked :
2283 * 1) We are cwnd-limited
2284 * 2) We are rwnd-limited
2285 * 3) We are application limited.
2286 */
2287 if (cong_win < send_win) {
2288 if (cong_win <= skb->len) {
2289 *is_cwnd_limited = true;
2290 return true;
2291 }
2292 } else {
2293 if (send_win <= skb->len) {
2294 *is_rwnd_limited = true;
2295 return true;
2296 }
2297 }
2298
2299 /* If this packet won't get more data, do not wait. */
2300 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) ||
2301 TCP_SKB_CB(skb)->eor)
2302 goto send_now;
2303
2304 return true;
2305
2306send_now:
2307 return false;
2308}
2309
2310static inline void tcp_mtu_check_reprobe(struct sock *sk)
2311{
2312 struct inet_connection_sock *icsk = inet_csk(sk);
2313 struct tcp_sock *tp = tcp_sk(sk);
2314 struct net *net = sock_net(sk);
2315 u32 interval;
2316 s32 delta;
2317
2318 interval = READ_ONCE(net->ipv4.sysctl_tcp_probe_interval);
2319 delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp;
2320 if (unlikely(delta >= interval * HZ)) {
2321 int mss = tcp_current_mss(sk);
2322
2323 /* Update current search range */
2324 icsk->icsk_mtup.probe_size = 0;
2325 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
2326 sizeof(struct tcphdr) +
2327 icsk->icsk_af_ops->net_header_len;
2328 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
2329
2330 /* Update probe time stamp */
2331 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
2332 }
2333}
2334
2335static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len)
2336{
2337 struct sk_buff *skb, *next;
2338
2339 skb = tcp_send_head(sk);
2340 tcp_for_write_queue_from_safe(skb, next, sk) {
2341 if (len <= skb->len)
2342 break;
2343
2344 if (unlikely(TCP_SKB_CB(skb)->eor) ||
2345 tcp_has_tx_tstamp(skb) ||
2346 !skb_pure_zcopy_same(skb, next))
2347 return false;
2348
2349 len -= skb->len;
2350 }
2351
2352 return true;
2353}
2354
2355static int tcp_clone_payload(struct sock *sk, struct sk_buff *to,
2356 int probe_size)
2357{
2358 skb_frag_t *lastfrag = NULL, *fragto = skb_shinfo(to)->frags;
2359 int i, todo, len = 0, nr_frags = 0;
2360 const struct sk_buff *skb;
2361
2362 if (!sk_wmem_schedule(sk, to->truesize + probe_size))
2363 return -ENOMEM;
2364
2365 skb_queue_walk(&sk->sk_write_queue, skb) {
2366 const skb_frag_t *fragfrom = skb_shinfo(skb)->frags;
2367
2368 if (skb_headlen(skb))
2369 return -EINVAL;
2370
2371 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++, fragfrom++) {
2372 if (len >= probe_size)
2373 goto commit;
2374 todo = min_t(int, skb_frag_size(fragfrom),
2375 probe_size - len);
2376 len += todo;
2377 if (lastfrag &&
2378 skb_frag_page(fragfrom) == skb_frag_page(lastfrag) &&
2379 skb_frag_off(fragfrom) == skb_frag_off(lastfrag) +
2380 skb_frag_size(lastfrag)) {
2381 skb_frag_size_add(lastfrag, todo);
2382 continue;
2383 }
2384 if (unlikely(nr_frags == MAX_SKB_FRAGS))
2385 return -E2BIG;
2386 skb_frag_page_copy(fragto, fragfrom);
2387 skb_frag_off_copy(fragto, fragfrom);
2388 skb_frag_size_set(fragto, todo);
2389 nr_frags++;
2390 lastfrag = fragto++;
2391 }
2392 }
2393commit:
2394 WARN_ON_ONCE(len != probe_size);
2395 for (i = 0; i < nr_frags; i++)
2396 skb_frag_ref(to, i);
2397
2398 skb_shinfo(to)->nr_frags = nr_frags;
2399 to->truesize += probe_size;
2400 to->len += probe_size;
2401 to->data_len += probe_size;
2402 __skb_header_release(to);
2403 return 0;
2404}
2405
2406/* Create a new MTU probe if we are ready.
2407 * MTU probe is regularly attempting to increase the path MTU by
2408 * deliberately sending larger packets. This discovers routing
2409 * changes resulting in larger path MTUs.
2410 *
2411 * Returns 0 if we should wait to probe (no cwnd available),
2412 * 1 if a probe was sent,
2413 * -1 otherwise
2414 */
2415static int tcp_mtu_probe(struct sock *sk)
2416{
2417 struct inet_connection_sock *icsk = inet_csk(sk);
2418 struct tcp_sock *tp = tcp_sk(sk);
2419 struct sk_buff *skb, *nskb, *next;
2420 struct net *net = sock_net(sk);
2421 int probe_size;
2422 int size_needed;
2423 int copy, len;
2424 int mss_now;
2425 int interval;
2426
2427 /* Not currently probing/verifying,
2428 * not in recovery,
2429 * have enough cwnd, and
2430 * not SACKing (the variable headers throw things off)
2431 */
2432 if (likely(!icsk->icsk_mtup.enabled ||
2433 icsk->icsk_mtup.probe_size ||
2434 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
2435 tcp_snd_cwnd(tp) < 11 ||
2436 tp->rx_opt.num_sacks || tp->rx_opt.dsack))
2437 return -1;
2438
2439 /* Use binary search for probe_size between tcp_mss_base,
2440 * and current mss_clamp. if (search_high - search_low)
2441 * smaller than a threshold, backoff from probing.
2442 */
2443 mss_now = tcp_current_mss(sk);
2444 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
2445 icsk->icsk_mtup.search_low) >> 1);
2446 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
2447 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
2448 /* When misfortune happens, we are reprobing actively,
2449 * and then reprobe timer has expired. We stick with current
2450 * probing process by not resetting search range to its orignal.
2451 */
2452 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
2453 interval < READ_ONCE(net->ipv4.sysctl_tcp_probe_threshold)) {
2454 /* Check whether enough time has elaplased for
2455 * another round of probing.
2456 */
2457 tcp_mtu_check_reprobe(sk);
2458 return -1;
2459 }
2460
2461 /* Have enough data in the send queue to probe? */
2462 if (tp->write_seq - tp->snd_nxt < size_needed)
2463 return -1;
2464
2465 if (tp->snd_wnd < size_needed)
2466 return -1;
2467 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
2468 return 0;
2469
2470 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2471 if (tcp_packets_in_flight(tp) + 2 > tcp_snd_cwnd(tp)) {
2472 if (!tcp_packets_in_flight(tp))
2473 return -1;
2474 else
2475 return 0;
2476 }
2477
2478 if (!tcp_can_coalesce_send_queue_head(sk, probe_size))
2479 return -1;
2480
2481 /* We're allowed to probe. Build it now. */
2482 nskb = tcp_stream_alloc_skb(sk, GFP_ATOMIC, false);
2483 if (!nskb)
2484 return -1;
2485
2486 /* build the payload, and be prepared to abort if this fails. */
2487 if (tcp_clone_payload(sk, nskb, probe_size)) {
2488 tcp_skb_tsorted_anchor_cleanup(nskb);
2489 consume_skb(nskb);
2490 return -1;
2491 }
2492 sk_wmem_queued_add(sk, nskb->truesize);
2493 sk_mem_charge(sk, nskb->truesize);
2494
2495 skb = tcp_send_head(sk);
2496 skb_copy_decrypted(nskb, skb);
2497 mptcp_skb_ext_copy(nskb, skb);
2498
2499 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2500 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2501 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2502
2503 tcp_insert_write_queue_before(nskb, skb, sk);
2504 tcp_highest_sack_replace(sk, skb, nskb);
2505
2506 len = 0;
2507 tcp_for_write_queue_from_safe(skb, next, sk) {
2508 copy = min_t(int, skb->len, probe_size - len);
2509
2510 if (skb->len <= copy) {
2511 /* We've eaten all the data from this skb.
2512 * Throw it away. */
2513 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2514 /* If this is the last SKB we copy and eor is set
2515 * we need to propagate it to the new skb.
2516 */
2517 TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor;
2518 tcp_skb_collapse_tstamp(nskb, skb);
2519 tcp_unlink_write_queue(skb, sk);
2520 tcp_wmem_free_skb(sk, skb);
2521 } else {
2522 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2523 ~(TCPHDR_FIN|TCPHDR_PSH);
2524 __pskb_trim_head(skb, copy);
2525 tcp_set_skb_tso_segs(skb, mss_now);
2526 TCP_SKB_CB(skb)->seq += copy;
2527 }
2528
2529 len += copy;
2530
2531 if (len >= probe_size)
2532 break;
2533 }
2534 tcp_init_tso_segs(nskb, nskb->len);
2535
2536 /* We're ready to send. If this fails, the probe will
2537 * be resegmented into mss-sized pieces by tcp_write_xmit().
2538 */
2539 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2540 /* Decrement cwnd here because we are sending
2541 * effectively two packets. */
2542 tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) - 1);
2543 tcp_event_new_data_sent(sk, nskb);
2544
2545 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2546 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2547 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2548
2549 return 1;
2550 }
2551
2552 return -1;
2553}
2554
2555static bool tcp_pacing_check(struct sock *sk)
2556{
2557 struct tcp_sock *tp = tcp_sk(sk);
2558
2559 if (!tcp_needs_internal_pacing(sk))
2560 return false;
2561
2562 if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache)
2563 return false;
2564
2565 if (!hrtimer_is_queued(&tp->pacing_timer)) {
2566 hrtimer_start(&tp->pacing_timer,
2567 ns_to_ktime(tp->tcp_wstamp_ns),
2568 HRTIMER_MODE_ABS_PINNED_SOFT);
2569 sock_hold(sk);
2570 }
2571 return true;
2572}
2573
2574static bool tcp_rtx_queue_empty_or_single_skb(const struct sock *sk)
2575{
2576 const struct rb_node *node = sk->tcp_rtx_queue.rb_node;
2577
2578 /* No skb in the rtx queue. */
2579 if (!node)
2580 return true;
2581
2582 /* Only one skb in rtx queue. */
2583 return !node->rb_left && !node->rb_right;
2584}
2585
2586/* TCP Small Queues :
2587 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2588 * (These limits are doubled for retransmits)
2589 * This allows for :
2590 * - better RTT estimation and ACK scheduling
2591 * - faster recovery
2592 * - high rates
2593 * Alas, some drivers / subsystems require a fair amount
2594 * of queued bytes to ensure line rate.
2595 * One example is wifi aggregation (802.11 AMPDU)
2596 */
2597static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2598 unsigned int factor)
2599{
2600 unsigned long limit;
2601
2602 limit = max_t(unsigned long,
2603 2 * skb->truesize,
2604 READ_ONCE(sk->sk_pacing_rate) >> READ_ONCE(sk->sk_pacing_shift));
2605 if (sk->sk_pacing_status == SK_PACING_NONE)
2606 limit = min_t(unsigned long, limit,
2607 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes));
2608 limit <<= factor;
2609
2610 if (static_branch_unlikely(&tcp_tx_delay_enabled) &&
2611 tcp_sk(sk)->tcp_tx_delay) {
2612 u64 extra_bytes = (u64)READ_ONCE(sk->sk_pacing_rate) *
2613 tcp_sk(sk)->tcp_tx_delay;
2614
2615 /* TSQ is based on skb truesize sum (sk_wmem_alloc), so we
2616 * approximate our needs assuming an ~100% skb->truesize overhead.
2617 * USEC_PER_SEC is approximated by 2^20.
2618 * do_div(extra_bytes, USEC_PER_SEC/2) is replaced by a right shift.
2619 */
2620 extra_bytes >>= (20 - 1);
2621 limit += extra_bytes;
2622 }
2623 if (refcount_read(&sk->sk_wmem_alloc) > limit) {
2624 /* Always send skb if rtx queue is empty or has one skb.
2625 * No need to wait for TX completion to call us back,
2626 * after softirq/tasklet schedule.
2627 * This helps when TX completions are delayed too much.
2628 */
2629 if (tcp_rtx_queue_empty_or_single_skb(sk))
2630 return false;
2631
2632 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2633 /* It is possible TX completion already happened
2634 * before we set TSQ_THROTTLED, so we must
2635 * test again the condition.
2636 */
2637 smp_mb__after_atomic();
2638 if (refcount_read(&sk->sk_wmem_alloc) > limit)
2639 return true;
2640 }
2641 return false;
2642}
2643
2644static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2645{
2646 const u32 now = tcp_jiffies32;
2647 enum tcp_chrono old = tp->chrono_type;
2648
2649 if (old > TCP_CHRONO_UNSPEC)
2650 tp->chrono_stat[old - 1] += now - tp->chrono_start;
2651 tp->chrono_start = now;
2652 tp->chrono_type = new;
2653}
2654
2655void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2656{
2657 struct tcp_sock *tp = tcp_sk(sk);
2658
2659 /* If there are multiple conditions worthy of tracking in a
2660 * chronograph then the highest priority enum takes precedence
2661 * over the other conditions. So that if something "more interesting"
2662 * starts happening, stop the previous chrono and start a new one.
2663 */
2664 if (type > tp->chrono_type)
2665 tcp_chrono_set(tp, type);
2666}
2667
2668void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2669{
2670 struct tcp_sock *tp = tcp_sk(sk);
2671
2672
2673 /* There are multiple conditions worthy of tracking in a
2674 * chronograph, so that the highest priority enum takes
2675 * precedence over the other conditions (see tcp_chrono_start).
2676 * If a condition stops, we only stop chrono tracking if
2677 * it's the "most interesting" or current chrono we are
2678 * tracking and starts busy chrono if we have pending data.
2679 */
2680 if (tcp_rtx_and_write_queues_empty(sk))
2681 tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
2682 else if (type == tp->chrono_type)
2683 tcp_chrono_set(tp, TCP_CHRONO_BUSY);
2684}
2685
2686/* This routine writes packets to the network. It advances the
2687 * send_head. This happens as incoming acks open up the remote
2688 * window for us.
2689 *
2690 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2691 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2692 * account rare use of URG, this is not a big flaw.
2693 *
2694 * Send at most one packet when push_one > 0. Temporarily ignore
2695 * cwnd limit to force at most one packet out when push_one == 2.
2696
2697 * Returns true, if no segments are in flight and we have queued segments,
2698 * but cannot send anything now because of SWS or another problem.
2699 */
2700static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2701 int push_one, gfp_t gfp)
2702{
2703 struct tcp_sock *tp = tcp_sk(sk);
2704 struct sk_buff *skb;
2705 unsigned int tso_segs, sent_pkts;
2706 int cwnd_quota;
2707 int result;
2708 bool is_cwnd_limited = false, is_rwnd_limited = false;
2709 u32 max_segs;
2710
2711 sent_pkts = 0;
2712
2713 tcp_mstamp_refresh(tp);
2714 if (!push_one) {
2715 /* Do MTU probing. */
2716 result = tcp_mtu_probe(sk);
2717 if (!result) {
2718 return false;
2719 } else if (result > 0) {
2720 sent_pkts = 1;
2721 }
2722 }
2723
2724 max_segs = tcp_tso_segs(sk, mss_now);
2725 while ((skb = tcp_send_head(sk))) {
2726 unsigned int limit;
2727
2728 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2729 /* "skb_mstamp_ns" is used as a start point for the retransmit timer */
2730 tp->tcp_wstamp_ns = tp->tcp_clock_cache;
2731 skb_set_delivery_time(skb, tp->tcp_wstamp_ns, true);
2732 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
2733 tcp_init_tso_segs(skb, mss_now);
2734 goto repair; /* Skip network transmission */
2735 }
2736
2737 if (tcp_pacing_check(sk))
2738 break;
2739
2740 tso_segs = tcp_init_tso_segs(skb, mss_now);
2741 BUG_ON(!tso_segs);
2742
2743 cwnd_quota = tcp_cwnd_test(tp, skb);
2744 if (!cwnd_quota) {
2745 if (push_one == 2)
2746 /* Force out a loss probe pkt. */
2747 cwnd_quota = 1;
2748 else
2749 break;
2750 }
2751
2752 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2753 is_rwnd_limited = true;
2754 break;
2755 }
2756
2757 if (tso_segs == 1) {
2758 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2759 (tcp_skb_is_last(sk, skb) ?
2760 nonagle : TCP_NAGLE_PUSH))))
2761 break;
2762 } else {
2763 if (!push_one &&
2764 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2765 &is_rwnd_limited, max_segs))
2766 break;
2767 }
2768
2769 limit = mss_now;
2770 if (tso_segs > 1 && !tcp_urg_mode(tp))
2771 limit = tcp_mss_split_point(sk, skb, mss_now,
2772 min_t(unsigned int,
2773 cwnd_quota,
2774 max_segs),
2775 nonagle);
2776
2777 if (skb->len > limit &&
2778 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2779 break;
2780
2781 if (tcp_small_queue_check(sk, skb, 0))
2782 break;
2783
2784 /* Argh, we hit an empty skb(), presumably a thread
2785 * is sleeping in sendmsg()/sk_stream_wait_memory().
2786 * We do not want to send a pure-ack packet and have
2787 * a strange looking rtx queue with empty packet(s).
2788 */
2789 if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq)
2790 break;
2791
2792 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2793 break;
2794
2795repair:
2796 /* Advance the send_head. This one is sent out.
2797 * This call will increment packets_out.
2798 */
2799 tcp_event_new_data_sent(sk, skb);
2800
2801 tcp_minshall_update(tp, mss_now, skb);
2802 sent_pkts += tcp_skb_pcount(skb);
2803
2804 if (push_one)
2805 break;
2806 }
2807
2808 if (is_rwnd_limited)
2809 tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
2810 else
2811 tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
2812
2813 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp));
2814 if (likely(sent_pkts || is_cwnd_limited))
2815 tcp_cwnd_validate(sk, is_cwnd_limited);
2816
2817 if (likely(sent_pkts)) {
2818 if (tcp_in_cwnd_reduction(sk))
2819 tp->prr_out += sent_pkts;
2820
2821 /* Send one loss probe per tail loss episode. */
2822 if (push_one != 2)
2823 tcp_schedule_loss_probe(sk, false);
2824 return false;
2825 }
2826 return !tp->packets_out && !tcp_write_queue_empty(sk);
2827}
2828
2829bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto)
2830{
2831 struct inet_connection_sock *icsk = inet_csk(sk);
2832 struct tcp_sock *tp = tcp_sk(sk);
2833 u32 timeout, timeout_us, rto_delta_us;
2834 int early_retrans;
2835
2836 /* Don't do any loss probe on a Fast Open connection before 3WHS
2837 * finishes.
2838 */
2839 if (rcu_access_pointer(tp->fastopen_rsk))
2840 return false;
2841
2842 early_retrans = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_early_retrans);
2843 /* Schedule a loss probe in 2*RTT for SACK capable connections
2844 * not in loss recovery, that are either limited by cwnd or application.
2845 */
2846 if ((early_retrans != 3 && early_retrans != 4) ||
2847 !tp->packets_out || !tcp_is_sack(tp) ||
2848 (icsk->icsk_ca_state != TCP_CA_Open &&
2849 icsk->icsk_ca_state != TCP_CA_CWR))
2850 return false;
2851
2852 /* Probe timeout is 2*rtt. Add minimum RTO to account
2853 * for delayed ack when there's one outstanding packet. If no RTT
2854 * sample is available then probe after TCP_TIMEOUT_INIT.
2855 */
2856 if (tp->srtt_us) {
2857 timeout_us = tp->srtt_us >> 2;
2858 if (tp->packets_out == 1)
2859 timeout_us += tcp_rto_min_us(sk);
2860 else
2861 timeout_us += TCP_TIMEOUT_MIN_US;
2862 timeout = usecs_to_jiffies(timeout_us);
2863 } else {
2864 timeout = TCP_TIMEOUT_INIT;
2865 }
2866
2867 /* If the RTO formula yields an earlier time, then use that time. */
2868 rto_delta_us = advancing_rto ?
2869 jiffies_to_usecs(inet_csk(sk)->icsk_rto) :
2870 tcp_rto_delta_us(sk); /* How far in future is RTO? */
2871 if (rto_delta_us > 0)
2872 timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us));
2873
2874 tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, TCP_RTO_MAX);
2875 return true;
2876}
2877
2878/* Thanks to skb fast clones, we can detect if a prior transmit of
2879 * a packet is still in a qdisc or driver queue.
2880 * In this case, there is very little point doing a retransmit !
2881 */
2882static bool skb_still_in_host_queue(struct sock *sk,
2883 const struct sk_buff *skb)
2884{
2885 if (unlikely(skb_fclone_busy(sk, skb))) {
2886 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2887 smp_mb__after_atomic();
2888 if (skb_fclone_busy(sk, skb)) {
2889 NET_INC_STATS(sock_net(sk),
2890 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2891 return true;
2892 }
2893 }
2894 return false;
2895}
2896
2897/* When probe timeout (PTO) fires, try send a new segment if possible, else
2898 * retransmit the last segment.
2899 */
2900void tcp_send_loss_probe(struct sock *sk)
2901{
2902 struct tcp_sock *tp = tcp_sk(sk);
2903 struct sk_buff *skb;
2904 int pcount;
2905 int mss = tcp_current_mss(sk);
2906
2907 /* At most one outstanding TLP */
2908 if (tp->tlp_high_seq)
2909 goto rearm_timer;
2910
2911 tp->tlp_retrans = 0;
2912 skb = tcp_send_head(sk);
2913 if (skb && tcp_snd_wnd_test(tp, skb, mss)) {
2914 pcount = tp->packets_out;
2915 tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2916 if (tp->packets_out > pcount)
2917 goto probe_sent;
2918 goto rearm_timer;
2919 }
2920 skb = skb_rb_last(&sk->tcp_rtx_queue);
2921 if (unlikely(!skb)) {
2922 WARN_ONCE(tp->packets_out,
2923 "invalid inflight: %u state %u cwnd %u mss %d\n",
2924 tp->packets_out, sk->sk_state, tcp_snd_cwnd(tp), mss);
2925 inet_csk(sk)->icsk_pending = 0;
2926 return;
2927 }
2928
2929 if (skb_still_in_host_queue(sk, skb))
2930 goto rearm_timer;
2931
2932 pcount = tcp_skb_pcount(skb);
2933 if (WARN_ON(!pcount))
2934 goto rearm_timer;
2935
2936 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2937 if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2938 (pcount - 1) * mss, mss,
2939 GFP_ATOMIC)))
2940 goto rearm_timer;
2941 skb = skb_rb_next(skb);
2942 }
2943
2944 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2945 goto rearm_timer;
2946
2947 if (__tcp_retransmit_skb(sk, skb, 1))
2948 goto rearm_timer;
2949
2950 tp->tlp_retrans = 1;
2951
2952probe_sent:
2953 /* Record snd_nxt for loss detection. */
2954 tp->tlp_high_seq = tp->snd_nxt;
2955
2956 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2957 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2958 inet_csk(sk)->icsk_pending = 0;
2959rearm_timer:
2960 tcp_rearm_rto(sk);
2961}
2962
2963/* Push out any pending frames which were held back due to
2964 * TCP_CORK or attempt at coalescing tiny packets.
2965 * The socket must be locked by the caller.
2966 */
2967void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2968 int nonagle)
2969{
2970 /* If we are closed, the bytes will have to remain here.
2971 * In time closedown will finish, we empty the write queue and
2972 * all will be happy.
2973 */
2974 if (unlikely(sk->sk_state == TCP_CLOSE))
2975 return;
2976
2977 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2978 sk_gfp_mask(sk, GFP_ATOMIC)))
2979 tcp_check_probe_timer(sk);
2980}
2981
2982/* Send _single_ skb sitting at the send head. This function requires
2983 * true push pending frames to setup probe timer etc.
2984 */
2985void tcp_push_one(struct sock *sk, unsigned int mss_now)
2986{
2987 struct sk_buff *skb = tcp_send_head(sk);
2988
2989 BUG_ON(!skb || skb->len < mss_now);
2990
2991 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2992}
2993
2994/* This function returns the amount that we can raise the
2995 * usable window based on the following constraints
2996 *
2997 * 1. The window can never be shrunk once it is offered (RFC 793)
2998 * 2. We limit memory per socket
2999 *
3000 * RFC 1122:
3001 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
3002 * RECV.NEXT + RCV.WIN fixed until:
3003 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
3004 *
3005 * i.e. don't raise the right edge of the window until you can raise
3006 * it at least MSS bytes.
3007 *
3008 * Unfortunately, the recommended algorithm breaks header prediction,
3009 * since header prediction assumes th->window stays fixed.
3010 *
3011 * Strictly speaking, keeping th->window fixed violates the receiver
3012 * side SWS prevention criteria. The problem is that under this rule
3013 * a stream of single byte packets will cause the right side of the
3014 * window to always advance by a single byte.
3015 *
3016 * Of course, if the sender implements sender side SWS prevention
3017 * then this will not be a problem.
3018 *
3019 * BSD seems to make the following compromise:
3020 *
3021 * If the free space is less than the 1/4 of the maximum
3022 * space available and the free space is less than 1/2 mss,
3023 * then set the window to 0.
3024 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
3025 * Otherwise, just prevent the window from shrinking
3026 * and from being larger than the largest representable value.
3027 *
3028 * This prevents incremental opening of the window in the regime
3029 * where TCP is limited by the speed of the reader side taking
3030 * data out of the TCP receive queue. It does nothing about
3031 * those cases where the window is constrained on the sender side
3032 * because the pipeline is full.
3033 *
3034 * BSD also seems to "accidentally" limit itself to windows that are a
3035 * multiple of MSS, at least until the free space gets quite small.
3036 * This would appear to be a side effect of the mbuf implementation.
3037 * Combining these two algorithms results in the observed behavior
3038 * of having a fixed window size at almost all times.
3039 *
3040 * Below we obtain similar behavior by forcing the offered window to
3041 * a multiple of the mss when it is feasible to do so.
3042 *
3043 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
3044 * Regular options like TIMESTAMP are taken into account.
3045 */
3046u32 __tcp_select_window(struct sock *sk)
3047{
3048 struct inet_connection_sock *icsk = inet_csk(sk);
3049 struct tcp_sock *tp = tcp_sk(sk);
3050 struct net *net = sock_net(sk);
3051 /* MSS for the peer's data. Previous versions used mss_clamp
3052 * here. I don't know if the value based on our guesses
3053 * of peer's MSS is better for the performance. It's more correct
3054 * but may be worse for the performance because of rcv_mss
3055 * fluctuations. --SAW 1998/11/1
3056 */
3057 int mss = icsk->icsk_ack.rcv_mss;
3058 int free_space = tcp_space(sk);
3059 int allowed_space = tcp_full_space(sk);
3060 int full_space, window;
3061
3062 if (sk_is_mptcp(sk))
3063 mptcp_space(sk, &free_space, &allowed_space);
3064
3065 full_space = min_t(int, tp->window_clamp, allowed_space);
3066
3067 if (unlikely(mss > full_space)) {
3068 mss = full_space;
3069 if (mss <= 0)
3070 return 0;
3071 }
3072
3073 /* Only allow window shrink if the sysctl is enabled and we have
3074 * a non-zero scaling factor in effect.
3075 */
3076 if (READ_ONCE(net->ipv4.sysctl_tcp_shrink_window) && tp->rx_opt.rcv_wscale)
3077 goto shrink_window_allowed;
3078
3079 /* do not allow window to shrink */
3080
3081 if (free_space < (full_space >> 1)) {
3082 icsk->icsk_ack.quick = 0;
3083
3084 if (tcp_under_memory_pressure(sk))
3085 tcp_adjust_rcv_ssthresh(sk);
3086
3087 /* free_space might become our new window, make sure we don't
3088 * increase it due to wscale.
3089 */
3090 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
3091
3092 /* if free space is less than mss estimate, or is below 1/16th
3093 * of the maximum allowed, try to move to zero-window, else
3094 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
3095 * new incoming data is dropped due to memory limits.
3096 * With large window, mss test triggers way too late in order
3097 * to announce zero window in time before rmem limit kicks in.
3098 */
3099 if (free_space < (allowed_space >> 4) || free_space < mss)
3100 return 0;
3101 }
3102
3103 if (free_space > tp->rcv_ssthresh)
3104 free_space = tp->rcv_ssthresh;
3105
3106 /* Don't do rounding if we are using window scaling, since the
3107 * scaled window will not line up with the MSS boundary anyway.
3108 */
3109 if (tp->rx_opt.rcv_wscale) {
3110 window = free_space;
3111
3112 /* Advertise enough space so that it won't get scaled away.
3113 * Import case: prevent zero window announcement if
3114 * 1<<rcv_wscale > mss.
3115 */
3116 window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
3117 } else {
3118 window = tp->rcv_wnd;
3119 /* Get the largest window that is a nice multiple of mss.
3120 * Window clamp already applied above.
3121 * If our current window offering is within 1 mss of the
3122 * free space we just keep it. This prevents the divide
3123 * and multiply from happening most of the time.
3124 * We also don't do any window rounding when the free space
3125 * is too small.
3126 */
3127 if (window <= free_space - mss || window > free_space)
3128 window = rounddown(free_space, mss);
3129 else if (mss == full_space &&
3130 free_space > window + (full_space >> 1))
3131 window = free_space;
3132 }
3133
3134 return window;
3135
3136shrink_window_allowed:
3137 /* new window should always be an exact multiple of scaling factor */
3138 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
3139
3140 if (free_space < (full_space >> 1)) {
3141 icsk->icsk_ack.quick = 0;
3142
3143 if (tcp_under_memory_pressure(sk))
3144 tcp_adjust_rcv_ssthresh(sk);
3145
3146 /* if free space is too low, return a zero window */
3147 if (free_space < (allowed_space >> 4) || free_space < mss ||
3148 free_space < (1 << tp->rx_opt.rcv_wscale))
3149 return 0;
3150 }
3151
3152 if (free_space > tp->rcv_ssthresh) {
3153 free_space = tp->rcv_ssthresh;
3154 /* new window should always be an exact multiple of scaling factor
3155 *
3156 * For this case, we ALIGN "up" (increase free_space) because
3157 * we know free_space is not zero here, it has been reduced from
3158 * the memory-based limit, and rcv_ssthresh is not a hard limit
3159 * (unlike sk_rcvbuf).
3160 */
3161 free_space = ALIGN(free_space, (1 << tp->rx_opt.rcv_wscale));
3162 }
3163
3164 return free_space;
3165}
3166
3167void tcp_skb_collapse_tstamp(struct sk_buff *skb,
3168 const struct sk_buff *next_skb)
3169{
3170 if (unlikely(tcp_has_tx_tstamp(next_skb))) {
3171 const struct skb_shared_info *next_shinfo =
3172 skb_shinfo(next_skb);
3173 struct skb_shared_info *shinfo = skb_shinfo(skb);
3174
3175 shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
3176 shinfo->tskey = next_shinfo->tskey;
3177 TCP_SKB_CB(skb)->txstamp_ack |=
3178 TCP_SKB_CB(next_skb)->txstamp_ack;
3179 }
3180}
3181
3182/* Collapses two adjacent SKB's during retransmission. */
3183static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
3184{
3185 struct tcp_sock *tp = tcp_sk(sk);
3186 struct sk_buff *next_skb = skb_rb_next(skb);
3187 int next_skb_size;
3188
3189 next_skb_size = next_skb->len;
3190
3191 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
3192
3193 if (next_skb_size && !tcp_skb_shift(skb, next_skb, 1, next_skb_size))
3194 return false;
3195
3196 tcp_highest_sack_replace(sk, next_skb, skb);
3197
3198 /* Update sequence range on original skb. */
3199 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
3200
3201 /* Merge over control information. This moves PSH/FIN etc. over */
3202 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
3203
3204 /* All done, get rid of second SKB and account for it so
3205 * packet counting does not break.
3206 */
3207 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
3208 TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
3209
3210 /* changed transmit queue under us so clear hints */
3211 tcp_clear_retrans_hints_partial(tp);
3212 if (next_skb == tp->retransmit_skb_hint)
3213 tp->retransmit_skb_hint = skb;
3214
3215 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
3216
3217 tcp_skb_collapse_tstamp(skb, next_skb);
3218
3219 tcp_rtx_queue_unlink_and_free(next_skb, sk);
3220 return true;
3221}
3222
3223/* Check if coalescing SKBs is legal. */
3224static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
3225{
3226 if (tcp_skb_pcount(skb) > 1)
3227 return false;
3228 if (skb_cloned(skb))
3229 return false;
3230 /* Some heuristics for collapsing over SACK'd could be invented */
3231 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
3232 return false;
3233
3234 return true;
3235}
3236
3237/* Collapse packets in the retransmit queue to make to create
3238 * less packets on the wire. This is only done on retransmission.
3239 */
3240static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
3241 int space)
3242{
3243 struct tcp_sock *tp = tcp_sk(sk);
3244 struct sk_buff *skb = to, *tmp;
3245 bool first = true;
3246
3247 if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse))
3248 return;
3249 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
3250 return;
3251
3252 skb_rbtree_walk_from_safe(skb, tmp) {
3253 if (!tcp_can_collapse(sk, skb))
3254 break;
3255
3256 if (!tcp_skb_can_collapse(to, skb))
3257 break;
3258
3259 space -= skb->len;
3260
3261 if (first) {
3262 first = false;
3263 continue;
3264 }
3265
3266 if (space < 0)
3267 break;
3268
3269 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
3270 break;
3271
3272 if (!tcp_collapse_retrans(sk, to))
3273 break;
3274 }
3275}
3276
3277/* This retransmits one SKB. Policy decisions and retransmit queue
3278 * state updates are done by the caller. Returns non-zero if an
3279 * error occurred which prevented the send.
3280 */
3281int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
3282{
3283 struct inet_connection_sock *icsk = inet_csk(sk);
3284 struct tcp_sock *tp = tcp_sk(sk);
3285 unsigned int cur_mss;
3286 int diff, len, err;
3287 int avail_wnd;
3288
3289 /* Inconclusive MTU probe */
3290 if (icsk->icsk_mtup.probe_size)
3291 icsk->icsk_mtup.probe_size = 0;
3292
3293 if (skb_still_in_host_queue(sk, skb))
3294 return -EBUSY;
3295
3296start:
3297 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
3298 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3299 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_SYN;
3300 TCP_SKB_CB(skb)->seq++;
3301 goto start;
3302 }
3303 if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) {
3304 WARN_ON_ONCE(1);
3305 return -EINVAL;
3306 }
3307 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3308 return -ENOMEM;
3309 }
3310
3311 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3312 return -EHOSTUNREACH; /* Routing failure or similar. */
3313
3314 cur_mss = tcp_current_mss(sk);
3315 avail_wnd = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3316
3317 /* If receiver has shrunk his window, and skb is out of
3318 * new window, do not retransmit it. The exception is the
3319 * case, when window is shrunk to zero. In this case
3320 * our retransmit of one segment serves as a zero window probe.
3321 */
3322 if (avail_wnd <= 0) {
3323 if (TCP_SKB_CB(skb)->seq != tp->snd_una)
3324 return -EAGAIN;
3325 avail_wnd = cur_mss;
3326 }
3327
3328 len = cur_mss * segs;
3329 if (len > avail_wnd) {
3330 len = rounddown(avail_wnd, cur_mss);
3331 if (!len)
3332 len = avail_wnd;
3333 }
3334 if (skb->len > len) {
3335 if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len,
3336 cur_mss, GFP_ATOMIC))
3337 return -ENOMEM; /* We'll try again later. */
3338 } else {
3339 if (skb_unclone_keeptruesize(skb, GFP_ATOMIC))
3340 return -ENOMEM;
3341
3342 diff = tcp_skb_pcount(skb);
3343 tcp_set_skb_tso_segs(skb, cur_mss);
3344 diff -= tcp_skb_pcount(skb);
3345 if (diff)
3346 tcp_adjust_pcount(sk, skb, diff);
3347 avail_wnd = min_t(int, avail_wnd, cur_mss);
3348 if (skb->len < avail_wnd)
3349 tcp_retrans_try_collapse(sk, skb, avail_wnd);
3350 }
3351
3352 /* RFC3168, section 6.1.1.1. ECN fallback */
3353 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
3354 tcp_ecn_clear_syn(sk, skb);
3355
3356 /* Update global and local TCP statistics. */
3357 segs = tcp_skb_pcount(skb);
3358 TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
3359 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
3360 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3361 tp->total_retrans += segs;
3362 tp->bytes_retrans += skb->len;
3363
3364 /* make sure skb->data is aligned on arches that require it
3365 * and check if ack-trimming & collapsing extended the headroom
3366 * beyond what csum_start can cover.
3367 */
3368 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
3369 skb_headroom(skb) >= 0xFFFF)) {
3370 struct sk_buff *nskb;
3371
3372 tcp_skb_tsorted_save(skb) {
3373 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
3374 if (nskb) {
3375 nskb->dev = NULL;
3376 err = tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC);
3377 } else {
3378 err = -ENOBUFS;
3379 }
3380 } tcp_skb_tsorted_restore(skb);
3381
3382 if (!err) {
3383 tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns);
3384 tcp_rate_skb_sent(sk, skb);
3385 }
3386 } else {
3387 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3388 }
3389
3390 /* To avoid taking spuriously low RTT samples based on a timestamp
3391 * for a transmit that never happened, always mark EVER_RETRANS
3392 */
3393 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
3394
3395 if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG))
3396 tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB,
3397 TCP_SKB_CB(skb)->seq, segs, err);
3398
3399 if (likely(!err)) {
3400 trace_tcp_retransmit_skb(sk, skb);
3401 } else if (err != -EBUSY) {
3402 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs);
3403 }
3404 return err;
3405}
3406
3407int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
3408{
3409 struct tcp_sock *tp = tcp_sk(sk);
3410 int err = __tcp_retransmit_skb(sk, skb, segs);
3411
3412 if (err == 0) {
3413#if FASTRETRANS_DEBUG > 0
3414 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
3415 net_dbg_ratelimited("retrans_out leaked\n");
3416 }
3417#endif
3418 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
3419 tp->retrans_out += tcp_skb_pcount(skb);
3420 }
3421
3422 /* Save stamp of the first (attempted) retransmit. */
3423 if (!tp->retrans_stamp)
3424 tp->retrans_stamp = tcp_skb_timestamp_ts(tp->tcp_usec_ts, skb);
3425
3426 if (tp->undo_retrans < 0)
3427 tp->undo_retrans = 0;
3428 tp->undo_retrans += tcp_skb_pcount(skb);
3429 return err;
3430}
3431
3432/* This gets called after a retransmit timeout, and the initially
3433 * retransmitted data is acknowledged. It tries to continue
3434 * resending the rest of the retransmit queue, until either
3435 * we've sent it all or the congestion window limit is reached.
3436 */
3437void tcp_xmit_retransmit_queue(struct sock *sk)
3438{
3439 const struct inet_connection_sock *icsk = inet_csk(sk);
3440 struct sk_buff *skb, *rtx_head, *hole = NULL;
3441 struct tcp_sock *tp = tcp_sk(sk);
3442 bool rearm_timer = false;
3443 u32 max_segs;
3444 int mib_idx;
3445
3446 if (!tp->packets_out)
3447 return;
3448
3449 rtx_head = tcp_rtx_queue_head(sk);
3450 skb = tp->retransmit_skb_hint ?: rtx_head;
3451 max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
3452 skb_rbtree_walk_from(skb) {
3453 __u8 sacked;
3454 int segs;
3455
3456 if (tcp_pacing_check(sk))
3457 break;
3458
3459 /* we could do better than to assign each time */
3460 if (!hole)
3461 tp->retransmit_skb_hint = skb;
3462
3463 segs = tcp_snd_cwnd(tp) - tcp_packets_in_flight(tp);
3464 if (segs <= 0)
3465 break;
3466 sacked = TCP_SKB_CB(skb)->sacked;
3467 /* In case tcp_shift_skb_data() have aggregated large skbs,
3468 * we need to make sure not sending too bigs TSO packets
3469 */
3470 segs = min_t(int, segs, max_segs);
3471
3472 if (tp->retrans_out >= tp->lost_out) {
3473 break;
3474 } else if (!(sacked & TCPCB_LOST)) {
3475 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
3476 hole = skb;
3477 continue;
3478
3479 } else {
3480 if (icsk->icsk_ca_state != TCP_CA_Loss)
3481 mib_idx = LINUX_MIB_TCPFASTRETRANS;
3482 else
3483 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
3484 }
3485
3486 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
3487 continue;
3488
3489 if (tcp_small_queue_check(sk, skb, 1))
3490 break;
3491
3492 if (tcp_retransmit_skb(sk, skb, segs))
3493 break;
3494
3495 NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
3496
3497 if (tcp_in_cwnd_reduction(sk))
3498 tp->prr_out += tcp_skb_pcount(skb);
3499
3500 if (skb == rtx_head &&
3501 icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
3502 rearm_timer = true;
3503
3504 }
3505 if (rearm_timer)
3506 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3507 inet_csk(sk)->icsk_rto,
3508 TCP_RTO_MAX);
3509}
3510
3511/* We allow to exceed memory limits for FIN packets to expedite
3512 * connection tear down and (memory) recovery.
3513 * Otherwise tcp_send_fin() could be tempted to either delay FIN
3514 * or even be forced to close flow without any FIN.
3515 * In general, we want to allow one skb per socket to avoid hangs
3516 * with edge trigger epoll()
3517 */
3518void sk_forced_mem_schedule(struct sock *sk, int size)
3519{
3520 int delta, amt;
3521
3522 delta = size - sk->sk_forward_alloc;
3523 if (delta <= 0)
3524 return;
3525 amt = sk_mem_pages(delta);
3526 sk_forward_alloc_add(sk, amt << PAGE_SHIFT);
3527 sk_memory_allocated_add(sk, amt);
3528
3529 if (mem_cgroup_sockets_enabled && sk->sk_memcg)
3530 mem_cgroup_charge_skmem(sk->sk_memcg, amt,
3531 gfp_memcg_charge() | __GFP_NOFAIL);
3532}
3533
3534/* Send a FIN. The caller locks the socket for us.
3535 * We should try to send a FIN packet really hard, but eventually give up.
3536 */
3537void tcp_send_fin(struct sock *sk)
3538{
3539 struct sk_buff *skb, *tskb, *tail = tcp_write_queue_tail(sk);
3540 struct tcp_sock *tp = tcp_sk(sk);
3541
3542 /* Optimization, tack on the FIN if we have one skb in write queue and
3543 * this skb was not yet sent, or we are under memory pressure.
3544 * Note: in the latter case, FIN packet will be sent after a timeout,
3545 * as TCP stack thinks it has already been transmitted.
3546 */
3547 tskb = tail;
3548 if (!tskb && tcp_under_memory_pressure(sk))
3549 tskb = skb_rb_last(&sk->tcp_rtx_queue);
3550
3551 if (tskb) {
3552 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
3553 TCP_SKB_CB(tskb)->end_seq++;
3554 tp->write_seq++;
3555 if (!tail) {
3556 /* This means tskb was already sent.
3557 * Pretend we included the FIN on previous transmit.
3558 * We need to set tp->snd_nxt to the value it would have
3559 * if FIN had been sent. This is because retransmit path
3560 * does not change tp->snd_nxt.
3561 */
3562 WRITE_ONCE(tp->snd_nxt, tp->snd_nxt + 1);
3563 return;
3564 }
3565 } else {
3566 skb = alloc_skb_fclone(MAX_TCP_HEADER,
3567 sk_gfp_mask(sk, GFP_ATOMIC |
3568 __GFP_NOWARN));
3569 if (unlikely(!skb))
3570 return;
3571
3572 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
3573 skb_reserve(skb, MAX_TCP_HEADER);
3574 sk_forced_mem_schedule(sk, skb->truesize);
3575 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3576 tcp_init_nondata_skb(skb, tp->write_seq,
3577 TCPHDR_ACK | TCPHDR_FIN);
3578 tcp_queue_skb(sk, skb);
3579 }
3580 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
3581}
3582
3583/* We get here when a process closes a file descriptor (either due to
3584 * an explicit close() or as a byproduct of exit()'ing) and there
3585 * was unread data in the receive queue. This behavior is recommended
3586 * by RFC 2525, section 2.17. -DaveM
3587 */
3588void tcp_send_active_reset(struct sock *sk, gfp_t priority)
3589{
3590 struct sk_buff *skb;
3591
3592 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3593
3594 /* NOTE: No TCP options attached and we never retransmit this. */
3595 skb = alloc_skb(MAX_TCP_HEADER, priority);
3596 if (!skb) {
3597 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3598 return;
3599 }
3600
3601 /* Reserve space for headers and prepare control bits. */
3602 skb_reserve(skb, MAX_TCP_HEADER);
3603 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
3604 TCPHDR_ACK | TCPHDR_RST);
3605 tcp_mstamp_refresh(tcp_sk(sk));
3606 /* Send it off. */
3607 if (tcp_transmit_skb(sk, skb, 0, priority))
3608 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3609
3610 /* skb of trace_tcp_send_reset() keeps the skb that caused RST,
3611 * skb here is different to the troublesome skb, so use NULL
3612 */
3613 trace_tcp_send_reset(sk, NULL);
3614}
3615
3616/* Send a crossed SYN-ACK during socket establishment.
3617 * WARNING: This routine must only be called when we have already sent
3618 * a SYN packet that crossed the incoming SYN that caused this routine
3619 * to get called. If this assumption fails then the initial rcv_wnd
3620 * and rcv_wscale values will not be correct.
3621 */
3622int tcp_send_synack(struct sock *sk)
3623{
3624 struct sk_buff *skb;
3625
3626 skb = tcp_rtx_queue_head(sk);
3627 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3628 pr_err("%s: wrong queue state\n", __func__);
3629 return -EFAULT;
3630 }
3631 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3632 if (skb_cloned(skb)) {
3633 struct sk_buff *nskb;
3634
3635 tcp_skb_tsorted_save(skb) {
3636 nskb = skb_copy(skb, GFP_ATOMIC);
3637 } tcp_skb_tsorted_restore(skb);
3638 if (!nskb)
3639 return -ENOMEM;
3640 INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor);
3641 tcp_highest_sack_replace(sk, skb, nskb);
3642 tcp_rtx_queue_unlink_and_free(skb, sk);
3643 __skb_header_release(nskb);
3644 tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb);
3645 sk_wmem_queued_add(sk, nskb->truesize);
3646 sk_mem_charge(sk, nskb->truesize);
3647 skb = nskb;
3648 }
3649
3650 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3651 tcp_ecn_send_synack(sk, skb);
3652 }
3653 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3654}
3655
3656/**
3657 * tcp_make_synack - Allocate one skb and build a SYNACK packet.
3658 * @sk: listener socket
3659 * @dst: dst entry attached to the SYNACK. It is consumed and caller
3660 * should not use it again.
3661 * @req: request_sock pointer
3662 * @foc: cookie for tcp fast open
3663 * @synack_type: Type of synack to prepare
3664 * @syn_skb: SYN packet just received. It could be NULL for rtx case.
3665 */
3666struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3667 struct request_sock *req,
3668 struct tcp_fastopen_cookie *foc,
3669 enum tcp_synack_type synack_type,
3670 struct sk_buff *syn_skb)
3671{
3672 struct inet_request_sock *ireq = inet_rsk(req);
3673 const struct tcp_sock *tp = tcp_sk(sk);
3674 struct tcp_out_options opts;
3675 struct tcp_key key = {};
3676 struct sk_buff *skb;
3677 int tcp_header_size;
3678 struct tcphdr *th;
3679 int mss;
3680 u64 now;
3681
3682 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3683 if (unlikely(!skb)) {
3684 dst_release(dst);
3685 return NULL;
3686 }
3687 /* Reserve space for headers. */
3688 skb_reserve(skb, MAX_TCP_HEADER);
3689
3690 switch (synack_type) {
3691 case TCP_SYNACK_NORMAL:
3692 skb_set_owner_w(skb, req_to_sk(req));
3693 break;
3694 case TCP_SYNACK_COOKIE:
3695 /* Under synflood, we do not attach skb to a socket,
3696 * to avoid false sharing.
3697 */
3698 break;
3699 case TCP_SYNACK_FASTOPEN:
3700 /* sk is a const pointer, because we want to express multiple
3701 * cpu might call us concurrently.
3702 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3703 */
3704 skb_set_owner_w(skb, (struct sock *)sk);
3705 break;
3706 }
3707 skb_dst_set(skb, dst);
3708
3709 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3710
3711 memset(&opts, 0, sizeof(opts));
3712 now = tcp_clock_ns();
3713#ifdef CONFIG_SYN_COOKIES
3714 if (unlikely(synack_type == TCP_SYNACK_COOKIE && ireq->tstamp_ok))
3715 skb_set_delivery_time(skb, cookie_init_timestamp(req, now),
3716 true);
3717 else
3718#endif
3719 {
3720 skb_set_delivery_time(skb, now, true);
3721 if (!tcp_rsk(req)->snt_synack) /* Timestamp first SYNACK */
3722 tcp_rsk(req)->snt_synack = tcp_skb_timestamp_us(skb);
3723 }
3724
3725#if defined(CONFIG_TCP_MD5SIG) || defined(CONFIG_TCP_AO)
3726 rcu_read_lock();
3727#endif
3728 if (tcp_rsk_used_ao(req)) {
3729#ifdef CONFIG_TCP_AO
3730 struct tcp_ao_key *ao_key = NULL;
3731 u8 keyid = tcp_rsk(req)->ao_keyid;
3732
3733 ao_key = tcp_sk(sk)->af_specific->ao_lookup(sk, req_to_sk(req),
3734 keyid, -1);
3735 /* If there is no matching key - avoid sending anything,
3736 * especially usigned segments. It could try harder and lookup
3737 * for another peer-matching key, but the peer has requested
3738 * ao_keyid (RFC5925 RNextKeyID), so let's keep it simple here.
3739 */
3740 if (unlikely(!ao_key)) {
3741 rcu_read_unlock();
3742 kfree_skb(skb);
3743 net_warn_ratelimited("TCP-AO: the keyid %u from SYN packet is not present - not sending SYNACK\n",
3744 keyid);
3745 return NULL;
3746 }
3747 key.ao_key = ao_key;
3748 key.type = TCP_KEY_AO;
3749#endif
3750 } else {
3751#ifdef CONFIG_TCP_MD5SIG
3752 key.md5_key = tcp_rsk(req)->af_specific->req_md5_lookup(sk,
3753 req_to_sk(req));
3754 if (key.md5_key)
3755 key.type = TCP_KEY_MD5;
3756#endif
3757 }
3758 skb_set_hash(skb, READ_ONCE(tcp_rsk(req)->txhash), PKT_HASH_TYPE_L4);
3759 /* bpf program will be interested in the tcp_flags */
3760 TCP_SKB_CB(skb)->tcp_flags = TCPHDR_SYN | TCPHDR_ACK;
3761 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts,
3762 &key, foc, synack_type, syn_skb)
3763 + sizeof(*th);
3764
3765 skb_push(skb, tcp_header_size);
3766 skb_reset_transport_header(skb);
3767
3768 th = (struct tcphdr *)skb->data;
3769 memset(th, 0, sizeof(struct tcphdr));
3770 th->syn = 1;
3771 th->ack = 1;
3772 tcp_ecn_make_synack(req, th);
3773 th->source = htons(ireq->ir_num);
3774 th->dest = ireq->ir_rmt_port;
3775 skb->mark = ireq->ir_mark;
3776 skb->ip_summed = CHECKSUM_PARTIAL;
3777 th->seq = htonl(tcp_rsk(req)->snt_isn);
3778 /* XXX data is queued and acked as is. No buffer/window check */
3779 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3780
3781 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3782 th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3783 tcp_options_write(th, NULL, tcp_rsk(req), &opts, &key);
3784 th->doff = (tcp_header_size >> 2);
3785 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3786
3787 /* Okay, we have all we need - do the md5 hash if needed */
3788 if (tcp_key_is_md5(&key)) {
3789#ifdef CONFIG_TCP_MD5SIG
3790 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3791 key.md5_key, req_to_sk(req), skb);
3792#endif
3793 } else if (tcp_key_is_ao(&key)) {
3794#ifdef CONFIG_TCP_AO
3795 tcp_rsk(req)->af_specific->ao_synack_hash(opts.hash_location,
3796 key.ao_key, req, skb,
3797 opts.hash_location - (u8 *)th, 0);
3798#endif
3799 }
3800#if defined(CONFIG_TCP_MD5SIG) || defined(CONFIG_TCP_AO)
3801 rcu_read_unlock();
3802#endif
3803
3804 bpf_skops_write_hdr_opt((struct sock *)sk, skb, req, syn_skb,
3805 synack_type, &opts);
3806
3807 skb_set_delivery_time(skb, now, true);
3808 tcp_add_tx_delay(skb, tp);
3809
3810 return skb;
3811}
3812EXPORT_SYMBOL(tcp_make_synack);
3813
3814static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3815{
3816 struct inet_connection_sock *icsk = inet_csk(sk);
3817 const struct tcp_congestion_ops *ca;
3818 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3819
3820 if (ca_key == TCP_CA_UNSPEC)
3821 return;
3822
3823 rcu_read_lock();
3824 ca = tcp_ca_find_key(ca_key);
3825 if (likely(ca && bpf_try_module_get(ca, ca->owner))) {
3826 bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner);
3827 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3828 icsk->icsk_ca_ops = ca;
3829 }
3830 rcu_read_unlock();
3831}
3832
3833/* Do all connect socket setups that can be done AF independent. */
3834static void tcp_connect_init(struct sock *sk)
3835{
3836 const struct dst_entry *dst = __sk_dst_get(sk);
3837 struct tcp_sock *tp = tcp_sk(sk);
3838 __u8 rcv_wscale;
3839 u32 rcv_wnd;
3840
3841 /* We'll fix this up when we get a response from the other end.
3842 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3843 */
3844 tp->tcp_header_len = sizeof(struct tcphdr);
3845 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_timestamps))
3846 tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED;
3847
3848 tcp_ao_connect_init(sk);
3849
3850 /* If user gave his TCP_MAXSEG, record it to clamp */
3851 if (tp->rx_opt.user_mss)
3852 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3853 tp->max_window = 0;
3854 tcp_mtup_init(sk);
3855 tcp_sync_mss(sk, dst_mtu(dst));
3856
3857 tcp_ca_dst_init(sk, dst);
3858
3859 if (!tp->window_clamp)
3860 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3861 tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3862
3863 tcp_initialize_rcv_mss(sk);
3864
3865 /* limit the window selection if the user enforce a smaller rx buffer */
3866 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3867 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3868 tp->window_clamp = tcp_full_space(sk);
3869
3870 rcv_wnd = tcp_rwnd_init_bpf(sk);
3871 if (rcv_wnd == 0)
3872 rcv_wnd = dst_metric(dst, RTAX_INITRWND);
3873
3874 tcp_select_initial_window(sk, tcp_full_space(sk),
3875 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3876 &tp->rcv_wnd,
3877 &tp->window_clamp,
3878 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_window_scaling),
3879 &rcv_wscale,
3880 rcv_wnd);
3881
3882 tp->rx_opt.rcv_wscale = rcv_wscale;
3883 tp->rcv_ssthresh = tp->rcv_wnd;
3884
3885 WRITE_ONCE(sk->sk_err, 0);
3886 sock_reset_flag(sk, SOCK_DONE);
3887 tp->snd_wnd = 0;
3888 tcp_init_wl(tp, 0);
3889 tcp_write_queue_purge(sk);
3890 tp->snd_una = tp->write_seq;
3891 tp->snd_sml = tp->write_seq;
3892 tp->snd_up = tp->write_seq;
3893 WRITE_ONCE(tp->snd_nxt, tp->write_seq);
3894
3895 if (likely(!tp->repair))
3896 tp->rcv_nxt = 0;
3897 else
3898 tp->rcv_tstamp = tcp_jiffies32;
3899 tp->rcv_wup = tp->rcv_nxt;
3900 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
3901
3902 inet_csk(sk)->icsk_rto = tcp_timeout_init(sk);
3903 inet_csk(sk)->icsk_retransmits = 0;
3904 tcp_clear_retrans(tp);
3905}
3906
3907static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3908{
3909 struct tcp_sock *tp = tcp_sk(sk);
3910 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3911
3912 tcb->end_seq += skb->len;
3913 __skb_header_release(skb);
3914 sk_wmem_queued_add(sk, skb->truesize);
3915 sk_mem_charge(sk, skb->truesize);
3916 WRITE_ONCE(tp->write_seq, tcb->end_seq);
3917 tp->packets_out += tcp_skb_pcount(skb);
3918}
3919
3920/* Build and send a SYN with data and (cached) Fast Open cookie. However,
3921 * queue a data-only packet after the regular SYN, such that regular SYNs
3922 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3923 * only the SYN sequence, the data are retransmitted in the first ACK.
3924 * If cookie is not cached or other error occurs, falls back to send a
3925 * regular SYN with Fast Open cookie request option.
3926 */
3927static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3928{
3929 struct inet_connection_sock *icsk = inet_csk(sk);
3930 struct tcp_sock *tp = tcp_sk(sk);
3931 struct tcp_fastopen_request *fo = tp->fastopen_req;
3932 struct page_frag *pfrag = sk_page_frag(sk);
3933 struct sk_buff *syn_data;
3934 int space, err = 0;
3935
3936 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
3937 if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie))
3938 goto fallback;
3939
3940 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3941 * user-MSS. Reserve maximum option space for middleboxes that add
3942 * private TCP options. The cost is reduced data space in SYN :(
3943 */
3944 tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp);
3945 /* Sync mss_cache after updating the mss_clamp */
3946 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
3947
3948 space = __tcp_mtu_to_mss(sk, icsk->icsk_pmtu_cookie) -
3949 MAX_TCP_OPTION_SPACE;
3950
3951 space = min_t(size_t, space, fo->size);
3952
3953 if (space &&
3954 !skb_page_frag_refill(min_t(size_t, space, PAGE_SIZE),
3955 pfrag, sk->sk_allocation))
3956 goto fallback;
3957 syn_data = tcp_stream_alloc_skb(sk, sk->sk_allocation, false);
3958 if (!syn_data)
3959 goto fallback;
3960 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3961 if (space) {
3962 space = min_t(size_t, space, pfrag->size - pfrag->offset);
3963 space = tcp_wmem_schedule(sk, space);
3964 }
3965 if (space) {
3966 space = copy_page_from_iter(pfrag->page, pfrag->offset,
3967 space, &fo->data->msg_iter);
3968 if (unlikely(!space)) {
3969 tcp_skb_tsorted_anchor_cleanup(syn_data);
3970 kfree_skb(syn_data);
3971 goto fallback;
3972 }
3973 skb_fill_page_desc(syn_data, 0, pfrag->page,
3974 pfrag->offset, space);
3975 page_ref_inc(pfrag->page);
3976 pfrag->offset += space;
3977 skb_len_add(syn_data, space);
3978 skb_zcopy_set(syn_data, fo->uarg, NULL);
3979 }
3980 /* No more data pending in inet_wait_for_connect() */
3981 if (space == fo->size)
3982 fo->data = NULL;
3983 fo->copied = space;
3984
3985 tcp_connect_queue_skb(sk, syn_data);
3986 if (syn_data->len)
3987 tcp_chrono_start(sk, TCP_CHRONO_BUSY);
3988
3989 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3990
3991 skb_set_delivery_time(syn, syn_data->skb_mstamp_ns, true);
3992
3993 /* Now full SYN+DATA was cloned and sent (or not),
3994 * remove the SYN from the original skb (syn_data)
3995 * we keep in write queue in case of a retransmit, as we
3996 * also have the SYN packet (with no data) in the same queue.
3997 */
3998 TCP_SKB_CB(syn_data)->seq++;
3999 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
4000 if (!err) {
4001 tp->syn_data = (fo->copied > 0);
4002 tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data);
4003 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
4004 goto done;
4005 }
4006
4007 /* data was not sent, put it in write_queue */
4008 __skb_queue_tail(&sk->sk_write_queue, syn_data);
4009 tp->packets_out -= tcp_skb_pcount(syn_data);
4010
4011fallback:
4012 /* Send a regular SYN with Fast Open cookie request option */
4013 if (fo->cookie.len > 0)
4014 fo->cookie.len = 0;
4015 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
4016 if (err)
4017 tp->syn_fastopen = 0;
4018done:
4019 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
4020 return err;
4021}
4022
4023/* Build a SYN and send it off. */
4024int tcp_connect(struct sock *sk)
4025{
4026 struct tcp_sock *tp = tcp_sk(sk);
4027 struct sk_buff *buff;
4028 int err;
4029
4030 tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL);
4031
4032#if defined(CONFIG_TCP_MD5SIG) && defined(CONFIG_TCP_AO)
4033 /* Has to be checked late, after setting daddr/saddr/ops.
4034 * Return error if the peer has both a md5 and a tcp-ao key
4035 * configured as this is ambiguous.
4036 */
4037 if (unlikely(rcu_dereference_protected(tp->md5sig_info,
4038 lockdep_sock_is_held(sk)))) {
4039 bool needs_ao = !!tp->af_specific->ao_lookup(sk, sk, -1, -1);
4040 bool needs_md5 = !!tp->af_specific->md5_lookup(sk, sk);
4041 struct tcp_ao_info *ao_info;
4042
4043 ao_info = rcu_dereference_check(tp->ao_info,
4044 lockdep_sock_is_held(sk));
4045 if (ao_info) {
4046 /* This is an extra check: tcp_ao_required() in
4047 * tcp_v{4,6}_parse_md5_keys() should prevent adding
4048 * md5 keys on ao_required socket.
4049 */
4050 needs_ao |= ao_info->ao_required;
4051 WARN_ON_ONCE(ao_info->ao_required && needs_md5);
4052 }
4053 if (needs_md5 && needs_ao)
4054 return -EKEYREJECTED;
4055
4056 /* If we have a matching md5 key and no matching tcp-ao key
4057 * then free up ao_info if allocated.
4058 */
4059 if (needs_md5) {
4060 tcp_ao_destroy_sock(sk, false);
4061 } else if (needs_ao) {
4062 tcp_clear_md5_list(sk);
4063 kfree(rcu_replace_pointer(tp->md5sig_info, NULL,
4064 lockdep_sock_is_held(sk)));
4065 }
4066 }
4067#endif
4068#ifdef CONFIG_TCP_AO
4069 if (unlikely(rcu_dereference_protected(tp->ao_info,
4070 lockdep_sock_is_held(sk)))) {
4071 /* Don't allow connecting if ao is configured but no
4072 * matching key is found.
4073 */
4074 if (!tp->af_specific->ao_lookup(sk, sk, -1, -1))
4075 return -EKEYREJECTED;
4076 }
4077#endif
4078
4079 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
4080 return -EHOSTUNREACH; /* Routing failure or similar. */
4081
4082 tcp_connect_init(sk);
4083
4084 if (unlikely(tp->repair)) {
4085 tcp_finish_connect(sk, NULL);
4086 return 0;
4087 }
4088
4089 buff = tcp_stream_alloc_skb(sk, sk->sk_allocation, true);
4090 if (unlikely(!buff))
4091 return -ENOBUFS;
4092
4093 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
4094 tcp_mstamp_refresh(tp);
4095 tp->retrans_stamp = tcp_time_stamp_ts(tp);
4096 tcp_connect_queue_skb(sk, buff);
4097 tcp_ecn_send_syn(sk, buff);
4098 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
4099
4100 /* Send off SYN; include data in Fast Open. */
4101 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
4102 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
4103 if (err == -ECONNREFUSED)
4104 return err;
4105
4106 /* We change tp->snd_nxt after the tcp_transmit_skb() call
4107 * in order to make this packet get counted in tcpOutSegs.
4108 */
4109 WRITE_ONCE(tp->snd_nxt, tp->write_seq);
4110 tp->pushed_seq = tp->write_seq;
4111 buff = tcp_send_head(sk);
4112 if (unlikely(buff)) {
4113 WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(buff)->seq);
4114 tp->pushed_seq = TCP_SKB_CB(buff)->seq;
4115 }
4116 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
4117
4118 /* Timer for repeating the SYN until an answer. */
4119 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
4120 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
4121 return 0;
4122}
4123EXPORT_SYMBOL(tcp_connect);
4124
4125u32 tcp_delack_max(const struct sock *sk)
4126{
4127 const struct dst_entry *dst = __sk_dst_get(sk);
4128 u32 delack_max = inet_csk(sk)->icsk_delack_max;
4129
4130 if (dst && dst_metric_locked(dst, RTAX_RTO_MIN)) {
4131 u32 rto_min = dst_metric_rtt(dst, RTAX_RTO_MIN);
4132 u32 delack_from_rto_min = max_t(int, 1, rto_min - 1);
4133
4134 delack_max = min_t(u32, delack_max, delack_from_rto_min);
4135 }
4136 return delack_max;
4137}
4138
4139/* Send out a delayed ack, the caller does the policy checking
4140 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
4141 * for details.
4142 */
4143void tcp_send_delayed_ack(struct sock *sk)
4144{
4145 struct inet_connection_sock *icsk = inet_csk(sk);
4146 int ato = icsk->icsk_ack.ato;
4147 unsigned long timeout;
4148
4149 if (ato > TCP_DELACK_MIN) {
4150 const struct tcp_sock *tp = tcp_sk(sk);
4151 int max_ato = HZ / 2;
4152
4153 if (inet_csk_in_pingpong_mode(sk) ||
4154 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
4155 max_ato = TCP_DELACK_MAX;
4156
4157 /* Slow path, intersegment interval is "high". */
4158
4159 /* If some rtt estimate is known, use it to bound delayed ack.
4160 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
4161 * directly.
4162 */
4163 if (tp->srtt_us) {
4164 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
4165 TCP_DELACK_MIN);
4166
4167 if (rtt < max_ato)
4168 max_ato = rtt;
4169 }
4170
4171 ato = min(ato, max_ato);
4172 }
4173
4174 ato = min_t(u32, ato, tcp_delack_max(sk));
4175
4176 /* Stay within the limit we were given */
4177 timeout = jiffies + ato;
4178
4179 /* Use new timeout only if there wasn't a older one earlier. */
4180 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
4181 /* If delack timer is about to expire, send ACK now. */
4182 if (time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
4183 tcp_send_ack(sk);
4184 return;
4185 }
4186
4187 if (!time_before(timeout, icsk->icsk_ack.timeout))
4188 timeout = icsk->icsk_ack.timeout;
4189 }
4190 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
4191 icsk->icsk_ack.timeout = timeout;
4192 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
4193}
4194
4195/* This routine sends an ack and also updates the window. */
4196void __tcp_send_ack(struct sock *sk, u32 rcv_nxt)
4197{
4198 struct sk_buff *buff;
4199
4200 /* If we have been reset, we may not send again. */
4201 if (sk->sk_state == TCP_CLOSE)
4202 return;
4203
4204 /* We are not putting this on the write queue, so
4205 * tcp_transmit_skb() will set the ownership to this
4206 * sock.
4207 */
4208 buff = alloc_skb(MAX_TCP_HEADER,
4209 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
4210 if (unlikely(!buff)) {
4211 struct inet_connection_sock *icsk = inet_csk(sk);
4212 unsigned long delay;
4213
4214 delay = TCP_DELACK_MAX << icsk->icsk_ack.retry;
4215 if (delay < TCP_RTO_MAX)
4216 icsk->icsk_ack.retry++;
4217 inet_csk_schedule_ack(sk);
4218 icsk->icsk_ack.ato = TCP_ATO_MIN;
4219 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, delay, TCP_RTO_MAX);
4220 return;
4221 }
4222
4223 /* Reserve space for headers and prepare control bits. */
4224 skb_reserve(buff, MAX_TCP_HEADER);
4225 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
4226
4227 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
4228 * too much.
4229 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
4230 */
4231 skb_set_tcp_pure_ack(buff);
4232
4233 /* Send it off, this clears delayed acks for us. */
4234 __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt);
4235}
4236EXPORT_SYMBOL_GPL(__tcp_send_ack);
4237
4238void tcp_send_ack(struct sock *sk)
4239{
4240 __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt);
4241}
4242
4243/* This routine sends a packet with an out of date sequence
4244 * number. It assumes the other end will try to ack it.
4245 *
4246 * Question: what should we make while urgent mode?
4247 * 4.4BSD forces sending single byte of data. We cannot send
4248 * out of window data, because we have SND.NXT==SND.MAX...
4249 *
4250 * Current solution: to send TWO zero-length segments in urgent mode:
4251 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
4252 * out-of-date with SND.UNA-1 to probe window.
4253 */
4254static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
4255{
4256 struct tcp_sock *tp = tcp_sk(sk);
4257 struct sk_buff *skb;
4258
4259 /* We don't queue it, tcp_transmit_skb() sets ownership. */
4260 skb = alloc_skb(MAX_TCP_HEADER,
4261 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
4262 if (!skb)
4263 return -1;
4264
4265 /* Reserve space for headers and set control bits. */
4266 skb_reserve(skb, MAX_TCP_HEADER);
4267 /* Use a previous sequence. This should cause the other
4268 * end to send an ack. Don't queue or clone SKB, just
4269 * send it.
4270 */
4271 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
4272 NET_INC_STATS(sock_net(sk), mib);
4273 return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
4274}
4275
4276/* Called from setsockopt( ... TCP_REPAIR ) */
4277void tcp_send_window_probe(struct sock *sk)
4278{
4279 if (sk->sk_state == TCP_ESTABLISHED) {
4280 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
4281 tcp_mstamp_refresh(tcp_sk(sk));
4282 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
4283 }
4284}
4285
4286/* Initiate keepalive or window probe from timer. */
4287int tcp_write_wakeup(struct sock *sk, int mib)
4288{
4289 struct tcp_sock *tp = tcp_sk(sk);
4290 struct sk_buff *skb;
4291
4292 if (sk->sk_state == TCP_CLOSE)
4293 return -1;
4294
4295 skb = tcp_send_head(sk);
4296 if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
4297 int err;
4298 unsigned int mss = tcp_current_mss(sk);
4299 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
4300
4301 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
4302 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
4303
4304 /* We are probing the opening of a window
4305 * but the window size is != 0
4306 * must have been a result SWS avoidance ( sender )
4307 */
4308 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
4309 skb->len > mss) {
4310 seg_size = min(seg_size, mss);
4311 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
4312 if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
4313 skb, seg_size, mss, GFP_ATOMIC))
4314 return -1;
4315 } else if (!tcp_skb_pcount(skb))
4316 tcp_set_skb_tso_segs(skb, mss);
4317
4318 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
4319 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
4320 if (!err)
4321 tcp_event_new_data_sent(sk, skb);
4322 return err;
4323 } else {
4324 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
4325 tcp_xmit_probe_skb(sk, 1, mib);
4326 return tcp_xmit_probe_skb(sk, 0, mib);
4327 }
4328}
4329
4330/* A window probe timeout has occurred. If window is not closed send
4331 * a partial packet else a zero probe.
4332 */
4333void tcp_send_probe0(struct sock *sk)
4334{
4335 struct inet_connection_sock *icsk = inet_csk(sk);
4336 struct tcp_sock *tp = tcp_sk(sk);
4337 struct net *net = sock_net(sk);
4338 unsigned long timeout;
4339 int err;
4340
4341 err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
4342
4343 if (tp->packets_out || tcp_write_queue_empty(sk)) {
4344 /* Cancel probe timer, if it is not required. */
4345 icsk->icsk_probes_out = 0;
4346 icsk->icsk_backoff = 0;
4347 icsk->icsk_probes_tstamp = 0;
4348 return;
4349 }
4350
4351 icsk->icsk_probes_out++;
4352 if (err <= 0) {
4353 if (icsk->icsk_backoff < READ_ONCE(net->ipv4.sysctl_tcp_retries2))
4354 icsk->icsk_backoff++;
4355 timeout = tcp_probe0_when(sk, TCP_RTO_MAX);
4356 } else {
4357 /* If packet was not sent due to local congestion,
4358 * Let senders fight for local resources conservatively.
4359 */
4360 timeout = TCP_RESOURCE_PROBE_INTERVAL;
4361 }
4362
4363 timeout = tcp_clamp_probe0_to_user_timeout(sk, timeout);
4364 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, timeout, TCP_RTO_MAX);
4365}
4366
4367int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
4368{
4369 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
4370 struct flowi fl;
4371 int res;
4372
4373 /* Paired with WRITE_ONCE() in sock_setsockopt() */
4374 if (READ_ONCE(sk->sk_txrehash) == SOCK_TXREHASH_ENABLED)
4375 WRITE_ONCE(tcp_rsk(req)->txhash, net_tx_rndhash());
4376 res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL,
4377 NULL);
4378 if (!res) {
4379 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
4380 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
4381 if (unlikely(tcp_passive_fastopen(sk))) {
4382 /* sk has const attribute because listeners are lockless.
4383 * However in this case, we are dealing with a passive fastopen
4384 * socket thus we can change total_retrans value.
4385 */
4386 tcp_sk_rw(sk)->total_retrans++;
4387 }
4388 trace_tcp_retransmit_synack(sk, req);
4389 }
4390 return res;
4391}
4392EXPORT_SYMBOL(tcp_rtx_synack);