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