1/*
2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
5 *
6 * Implementation of the Transmission Control Protocol(TCP).
7 *
8 * Authors: Ross Biro
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
19 */
20
21/*
22 * Changes: Pedro Roque : Retransmit queue handled by TCP.
23 * : Fragmentation on mtu decrease
24 * : Segment collapse on retransmit
25 * : AF independence
26 *
27 * Linus Torvalds : send_delayed_ack
28 * David S. Miller : Charge memory using the right skb
29 * during syn/ack processing.
30 * David S. Miller : Output engine completely rewritten.
31 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
32 * Cacophonix Gaul : draft-minshall-nagle-01
33 * J Hadi Salim : ECN support
34 *
35 */
36
37#define pr_fmt(fmt) "TCP: " fmt
38
39#include <net/tcp.h>
40
41#include <linux/compiler.h>
42#include <linux/gfp.h>
43#include <linux/module.h>
44#include <linux/static_key.h>
45
46#include <trace/events/tcp.h>
47
48/* Refresh clocks of a TCP socket,
49 * ensuring monotically increasing values.
50 */
51void tcp_mstamp_refresh(struct tcp_sock *tp)
52{
53 u64 val = tcp_clock_ns();
54
55 if (val > tp->tcp_clock_cache)
56 tp->tcp_clock_cache = val;
57
58 val = div_u64(val, NSEC_PER_USEC);
59 if (val > tp->tcp_mstamp)
60 tp->tcp_mstamp = val;
61}
62
63static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
64 int push_one, gfp_t gfp);
65
66/* Account for new data that has been sent to the network. */
67static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
68{
69 struct inet_connection_sock *icsk = inet_csk(sk);
70 struct tcp_sock *tp = tcp_sk(sk);
71 unsigned int prior_packets = tp->packets_out;
72
73 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
74
75 __skb_unlink(skb, &sk->sk_write_queue);
76 tcp_rbtree_insert(&sk->tcp_rtx_queue, skb);
77
78 tp->packets_out += tcp_skb_pcount(skb);
79 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
80 tcp_rearm_rto(sk);
81
82 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
83 tcp_skb_pcount(skb));
84}
85
86/* SND.NXT, if window was not shrunk or the amount of shrunk was less than one
87 * window scaling factor due to loss of precision.
88 * If window has been shrunk, what should we make? It is not clear at all.
89 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
90 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
91 * invalid. OK, let's make this for now:
92 */
93static inline __u32 tcp_acceptable_seq(const struct sock *sk)
94{
95 const struct tcp_sock *tp = tcp_sk(sk);
96
97 if (!before(tcp_wnd_end(tp), tp->snd_nxt) ||
98 (tp->rx_opt.wscale_ok &&
99 ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale))))
100 return tp->snd_nxt;
101 else
102 return tcp_wnd_end(tp);
103}
104
105/* Calculate mss to advertise in SYN segment.
106 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
107 *
108 * 1. It is independent of path mtu.
109 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
110 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
111 * attached devices, because some buggy hosts are confused by
112 * large MSS.
113 * 4. We do not make 3, we advertise MSS, calculated from first
114 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
115 * This may be overridden via information stored in routing table.
116 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
117 * probably even Jumbo".
118 */
119static __u16 tcp_advertise_mss(struct sock *sk)
120{
121 struct tcp_sock *tp = tcp_sk(sk);
122 const struct dst_entry *dst = __sk_dst_get(sk);
123 int mss = tp->advmss;
124
125 if (dst) {
126 unsigned int metric = dst_metric_advmss(dst);
127
128 if (metric < mss) {
129 mss = metric;
130 tp->advmss = mss;
131 }
132 }
133
134 return (__u16)mss;
135}
136
137/* RFC2861. Reset CWND after idle period longer RTO to "restart window".
138 * This is the first part of cwnd validation mechanism.
139 */
140void tcp_cwnd_restart(struct sock *sk, s32 delta)
141{
142 struct tcp_sock *tp = tcp_sk(sk);
143 u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
144 u32 cwnd = tp->snd_cwnd;
145
146 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
147
148 tp->snd_ssthresh = tcp_current_ssthresh(sk);
149 restart_cwnd = min(restart_cwnd, cwnd);
150
151 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
152 cwnd >>= 1;
153 tp->snd_cwnd = max(cwnd, restart_cwnd);
154 tp->snd_cwnd_stamp = tcp_jiffies32;
155 tp->snd_cwnd_used = 0;
156}
157
158/* Congestion state accounting after a packet has been sent. */
159static void tcp_event_data_sent(struct tcp_sock *tp,
160 struct sock *sk)
161{
162 struct inet_connection_sock *icsk = inet_csk(sk);
163 const u32 now = tcp_jiffies32;
164
165 if (tcp_packets_in_flight(tp) == 0)
166 tcp_ca_event(sk, CA_EVENT_TX_START);
167
168 /* If this is the first data packet sent in response to the
169 * previous received data,
170 * and it is a reply for ato after last received packet,
171 * increase pingpong count.
172 */
173 if (before(tp->lsndtime, icsk->icsk_ack.lrcvtime) &&
174 (u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
175 inet_csk_inc_pingpong_cnt(sk);
176
177 tp->lsndtime = now;
178}
179
180/* Account for an ACK we sent. */
181static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts,
182 u32 rcv_nxt)
183{
184 struct tcp_sock *tp = tcp_sk(sk);
185
186 if (unlikely(tp->compressed_ack > TCP_FASTRETRANS_THRESH)) {
187 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
188 tp->compressed_ack - TCP_FASTRETRANS_THRESH);
189 tp->compressed_ack = TCP_FASTRETRANS_THRESH;
190 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
191 __sock_put(sk);
192 }
193
194 if (unlikely(rcv_nxt != tp->rcv_nxt))
195 return; /* Special ACK sent by DCTCP to reflect ECN */
196 tcp_dec_quickack_mode(sk, pkts);
197 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
198}
199
200/* Determine a window scaling and initial window to offer.
201 * Based on the assumption that the given amount of space
202 * will be offered. Store the results in the tp structure.
203 * NOTE: for smooth operation initial space offering should
204 * be a multiple of mss if possible. We assume here that mss >= 1.
205 * This MUST be enforced by all callers.
206 */
207void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss,
208 __u32 *rcv_wnd, __u32 *window_clamp,
209 int wscale_ok, __u8 *rcv_wscale,
210 __u32 init_rcv_wnd)
211{
212 unsigned int space = (__space < 0 ? 0 : __space);
213
214 /* If no clamp set the clamp to the max possible scaled window */
215 if (*window_clamp == 0)
216 (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE);
217 space = min(*window_clamp, space);
218
219 /* Quantize space offering to a multiple of mss if possible. */
220 if (space > mss)
221 space = rounddown(space, mss);
222
223 /* NOTE: offering an initial window larger than 32767
224 * will break some buggy TCP stacks. If the admin tells us
225 * it is likely we could be speaking with such a buggy stack
226 * we will truncate our initial window offering to 32K-1
227 * unless the remote has sent us a window scaling option,
228 * which we interpret as a sign the remote TCP is not
229 * misinterpreting the window field as a signed quantity.
230 */
231 if (sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
232 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
233 else
234 (*rcv_wnd) = min_t(u32, space, U16_MAX);
235
236 if (init_rcv_wnd)
237 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
238
239 *rcv_wscale = 0;
240 if (wscale_ok) {
241 /* Set window scaling on max possible window */
242 space = max_t(u32, space, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
243 space = max_t(u32, space, sysctl_rmem_max);
244 space = min_t(u32, space, *window_clamp);
245 *rcv_wscale = clamp_t(int, ilog2(space) - 15,
246 0, TCP_MAX_WSCALE);
247 }
248 /* Set the clamp no higher than max representable value */
249 (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp);
250}
251EXPORT_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 */
258static u16 tcp_select_window(struct sock *sk)
259{
260 struct tcp_sock *tp = tcp_sk(sk);
261 u32 old_win = tp->rcv_wnd;
262 u32 cur_win = tcp_receive_window(tp);
263 u32 new_win = __tcp_select_window(sk);
264
265 /* Never shrink the offered window */
266 if (new_win < cur_win) {
267 /* Danger Will Robinson!
268 * Don't update rcv_wup/rcv_wnd here or else
269 * we will not be able to advertise a zero
270 * window in time. --DaveM
271 *
272 * Relax Will Robinson.
273 */
274 if (new_win == 0)
275 NET_INC_STATS(sock_net(sk),
276 LINUX_MIB_TCPWANTZEROWINDOWADV);
277 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
278 }
279 tp->rcv_wnd = new_win;
280 tp->rcv_wup = tp->rcv_nxt;
281
282 /* Make sure we do not exceed the maximum possible
283 * scaled window.
284 */
285 if (!tp->rx_opt.rcv_wscale &&
286 sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
287 new_win = min(new_win, MAX_TCP_WINDOW);
288 else
289 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
290
291 /* RFC1323 scaling applied */
292 new_win >>= tp->rx_opt.rcv_wscale;
293
294 /* If we advertise zero window, disable fast path. */
295 if (new_win == 0) {
296 tp->pred_flags = 0;
297 if (old_win)
298 NET_INC_STATS(sock_net(sk),
299 LINUX_MIB_TCPTOZEROWINDOWADV);
300 } else if (old_win == 0) {
301 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
302 }
303
304 return new_win;
305}
306
307/* Packet ECN state for a SYN-ACK */
308static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
309{
310 const struct tcp_sock *tp = tcp_sk(sk);
311
312 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
313 if (!(tp->ecn_flags & TCP_ECN_OK))
314 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
315 else if (tcp_ca_needs_ecn(sk) ||
316 tcp_bpf_ca_needs_ecn(sk))
317 INET_ECN_xmit(sk);
318}
319
320/* Packet ECN state for a SYN. */
321static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
322{
323 struct tcp_sock *tp = tcp_sk(sk);
324 bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk);
325 bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
326 tcp_ca_needs_ecn(sk) || bpf_needs_ecn;
327
328 if (!use_ecn) {
329 const struct dst_entry *dst = __sk_dst_get(sk);
330
331 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
332 use_ecn = true;
333 }
334
335 tp->ecn_flags = 0;
336
337 if (use_ecn) {
338 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
339 tp->ecn_flags = TCP_ECN_OK;
340 if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn)
341 INET_ECN_xmit(sk);
342 }
343}
344
345static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
346{
347 if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
348 /* tp->ecn_flags are cleared at a later point in time when
349 * SYN ACK is ultimatively being received.
350 */
351 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
352}
353
354static void
355tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
356{
357 if (inet_rsk(req)->ecn_ok)
358 th->ece = 1;
359}
360
361/* Set up ECN state for a packet on a ESTABLISHED socket that is about to
362 * be sent.
363 */
364static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
365 struct tcphdr *th, int tcp_header_len)
366{
367 struct tcp_sock *tp = tcp_sk(sk);
368
369 if (tp->ecn_flags & TCP_ECN_OK) {
370 /* Not-retransmitted data segment: set ECT and inject CWR. */
371 if (skb->len != tcp_header_len &&
372 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
373 INET_ECN_xmit(sk);
374 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
375 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
376 th->cwr = 1;
377 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
378 }
379 } else if (!tcp_ca_needs_ecn(sk)) {
380 /* ACK or retransmitted segment: clear ECT|CE */
381 INET_ECN_dontxmit(sk);
382 }
383 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
384 th->ece = 1;
385 }
386}
387
388/* Constructs common control bits of non-data skb. If SYN/FIN is present,
389 * auto increment end seqno.
390 */
391static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
392{
393 skb->ip_summed = CHECKSUM_PARTIAL;
394
395 TCP_SKB_CB(skb)->tcp_flags = flags;
396 TCP_SKB_CB(skb)->sacked = 0;
397
398 tcp_skb_pcount_set(skb, 1);
399
400 TCP_SKB_CB(skb)->seq = seq;
401 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
402 seq++;
403 TCP_SKB_CB(skb)->end_seq = seq;
404}
405
406static inline bool tcp_urg_mode(const struct tcp_sock *tp)
407{
408 return tp->snd_una != tp->snd_up;
409}
410
411#define OPTION_SACK_ADVERTISE (1 << 0)
412#define OPTION_TS (1 << 1)
413#define OPTION_MD5 (1 << 2)
414#define OPTION_WSCALE (1 << 3)
415#define OPTION_FAST_OPEN_COOKIE (1 << 8)
416#define OPTION_SMC (1 << 9)
417
418static void smc_options_write(__be32 *ptr, u16 *options)
419{
420#if IS_ENABLED(CONFIG_SMC)
421 if (static_branch_unlikely(&tcp_have_smc)) {
422 if (unlikely(OPTION_SMC & *options)) {
423 *ptr++ = htonl((TCPOPT_NOP << 24) |
424 (TCPOPT_NOP << 16) |
425 (TCPOPT_EXP << 8) |
426 (TCPOLEN_EXP_SMC_BASE));
427 *ptr++ = htonl(TCPOPT_SMC_MAGIC);
428 }
429 }
430#endif
431}
432
433struct tcp_out_options {
434 u16 options; /* bit field of OPTION_* */
435 u16 mss; /* 0 to disable */
436 u8 ws; /* window scale, 0 to disable */
437 u8 num_sack_blocks; /* number of SACK blocks to include */
438 u8 hash_size; /* bytes in hash_location */
439 __u8 *hash_location; /* temporary pointer, overloaded */
440 __u32 tsval, tsecr; /* need to include OPTION_TS */
441 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
442};
443
444/* Write previously computed TCP options to the packet.
445 *
446 * Beware: Something in the Internet is very sensitive to the ordering of
447 * TCP options, we learned this through the hard way, so be careful here.
448 * Luckily we can at least blame others for their non-compliance but from
449 * inter-operability perspective it seems that we're somewhat stuck with
450 * the ordering which we have been using if we want to keep working with
451 * those broken things (not that it currently hurts anybody as there isn't
452 * particular reason why the ordering would need to be changed).
453 *
454 * At least SACK_PERM as the first option is known to lead to a disaster
455 * (but it may well be that other scenarios fail similarly).
456 */
457static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
458 struct tcp_out_options *opts)
459{
460 u16 options = opts->options; /* mungable copy */
461
462 if (unlikely(OPTION_MD5 & options)) {
463 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
464 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
465 /* overload cookie hash location */
466 opts->hash_location = (__u8 *)ptr;
467 ptr += 4;
468 }
469
470 if (unlikely(opts->mss)) {
471 *ptr++ = htonl((TCPOPT_MSS << 24) |
472 (TCPOLEN_MSS << 16) |
473 opts->mss);
474 }
475
476 if (likely(OPTION_TS & options)) {
477 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
478 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
479 (TCPOLEN_SACK_PERM << 16) |
480 (TCPOPT_TIMESTAMP << 8) |
481 TCPOLEN_TIMESTAMP);
482 options &= ~OPTION_SACK_ADVERTISE;
483 } else {
484 *ptr++ = htonl((TCPOPT_NOP << 24) |
485 (TCPOPT_NOP << 16) |
486 (TCPOPT_TIMESTAMP << 8) |
487 TCPOLEN_TIMESTAMP);
488 }
489 *ptr++ = htonl(opts->tsval);
490 *ptr++ = htonl(opts->tsecr);
491 }
492
493 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
494 *ptr++ = htonl((TCPOPT_NOP << 24) |
495 (TCPOPT_NOP << 16) |
496 (TCPOPT_SACK_PERM << 8) |
497 TCPOLEN_SACK_PERM);
498 }
499
500 if (unlikely(OPTION_WSCALE & options)) {
501 *ptr++ = htonl((TCPOPT_NOP << 24) |
502 (TCPOPT_WINDOW << 16) |
503 (TCPOLEN_WINDOW << 8) |
504 opts->ws);
505 }
506
507 if (unlikely(opts->num_sack_blocks)) {
508 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
509 tp->duplicate_sack : tp->selective_acks;
510 int this_sack;
511
512 *ptr++ = htonl((TCPOPT_NOP << 24) |
513 (TCPOPT_NOP << 16) |
514 (TCPOPT_SACK << 8) |
515 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
516 TCPOLEN_SACK_PERBLOCK)));
517
518 for (this_sack = 0; this_sack < opts->num_sack_blocks;
519 ++this_sack) {
520 *ptr++ = htonl(sp[this_sack].start_seq);
521 *ptr++ = htonl(sp[this_sack].end_seq);
522 }
523
524 tp->rx_opt.dsack = 0;
525 }
526
527 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
528 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
529 u8 *p = (u8 *)ptr;
530 u32 len; /* Fast Open option length */
531
532 if (foc->exp) {
533 len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
534 *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
535 TCPOPT_FASTOPEN_MAGIC);
536 p += TCPOLEN_EXP_FASTOPEN_BASE;
537 } else {
538 len = TCPOLEN_FASTOPEN_BASE + foc->len;
539 *p++ = TCPOPT_FASTOPEN;
540 *p++ = len;
541 }
542
543 memcpy(p, foc->val, foc->len);
544 if ((len & 3) == 2) {
545 p[foc->len] = TCPOPT_NOP;
546 p[foc->len + 1] = TCPOPT_NOP;
547 }
548 ptr += (len + 3) >> 2;
549 }
550
551 smc_options_write(ptr, &options);
552}
553
554static void smc_set_option(const struct tcp_sock *tp,
555 struct tcp_out_options *opts,
556 unsigned int *remaining)
557{
558#if IS_ENABLED(CONFIG_SMC)
559 if (static_branch_unlikely(&tcp_have_smc)) {
560 if (tp->syn_smc) {
561 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
562 opts->options |= OPTION_SMC;
563 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
564 }
565 }
566 }
567#endif
568}
569
570static void smc_set_option_cond(const struct tcp_sock *tp,
571 const struct inet_request_sock *ireq,
572 struct tcp_out_options *opts,
573 unsigned int *remaining)
574{
575#if IS_ENABLED(CONFIG_SMC)
576 if (static_branch_unlikely(&tcp_have_smc)) {
577 if (tp->syn_smc && ireq->smc_ok) {
578 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
579 opts->options |= OPTION_SMC;
580 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
581 }
582 }
583 }
584#endif
585}
586
587/* Compute TCP options for SYN packets. This is not the final
588 * network wire format yet.
589 */
590static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
591 struct tcp_out_options *opts,
592 struct tcp_md5sig_key **md5)
593{
594 struct tcp_sock *tp = tcp_sk(sk);
595 unsigned int remaining = MAX_TCP_OPTION_SPACE;
596 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
597
598 *md5 = NULL;
599#ifdef CONFIG_TCP_MD5SIG
600 if (static_branch_unlikely(&tcp_md5_needed) &&
601 rcu_access_pointer(tp->md5sig_info)) {
602 *md5 = tp->af_specific->md5_lookup(sk, sk);
603 if (*md5) {
604 opts->options |= OPTION_MD5;
605 remaining -= TCPOLEN_MD5SIG_ALIGNED;
606 }
607 }
608#endif
609
610 /* We always get an MSS option. The option bytes which will be seen in
611 * normal data packets should timestamps be used, must be in the MSS
612 * advertised. But we subtract them from tp->mss_cache so that
613 * calculations in tcp_sendmsg are simpler etc. So account for this
614 * fact here if necessary. If we don't do this correctly, as a
615 * receiver we won't recognize data packets as being full sized when we
616 * should, and thus we won't abide by the delayed ACK rules correctly.
617 * SACKs don't matter, we never delay an ACK when we have any of those
618 * going out. */
619 opts->mss = tcp_advertise_mss(sk);
620 remaining -= TCPOLEN_MSS_ALIGNED;
621
622 if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) {
623 opts->options |= OPTION_TS;
624 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
625 opts->tsecr = tp->rx_opt.ts_recent;
626 remaining -= TCPOLEN_TSTAMP_ALIGNED;
627 }
628 if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) {
629 opts->ws = tp->rx_opt.rcv_wscale;
630 opts->options |= OPTION_WSCALE;
631 remaining -= TCPOLEN_WSCALE_ALIGNED;
632 }
633 if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) {
634 opts->options |= OPTION_SACK_ADVERTISE;
635 if (unlikely(!(OPTION_TS & opts->options)))
636 remaining -= TCPOLEN_SACKPERM_ALIGNED;
637 }
638
639 if (fastopen && fastopen->cookie.len >= 0) {
640 u32 need = fastopen->cookie.len;
641
642 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
643 TCPOLEN_FASTOPEN_BASE;
644 need = (need + 3) & ~3U; /* Align to 32 bits */
645 if (remaining >= need) {
646 opts->options |= OPTION_FAST_OPEN_COOKIE;
647 opts->fastopen_cookie = &fastopen->cookie;
648 remaining -= need;
649 tp->syn_fastopen = 1;
650 tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
651 }
652 }
653
654 smc_set_option(tp, opts, &remaining);
655
656 return MAX_TCP_OPTION_SPACE - remaining;
657}
658
659/* Set up TCP options for SYN-ACKs. */
660static unsigned int tcp_synack_options(const struct sock *sk,
661 struct request_sock *req,
662 unsigned int mss, struct sk_buff *skb,
663 struct tcp_out_options *opts,
664 const struct tcp_md5sig_key *md5,
665 struct tcp_fastopen_cookie *foc)
666{
667 struct inet_request_sock *ireq = inet_rsk(req);
668 unsigned int remaining = MAX_TCP_OPTION_SPACE;
669
670#ifdef CONFIG_TCP_MD5SIG
671 if (md5) {
672 opts->options |= OPTION_MD5;
673 remaining -= TCPOLEN_MD5SIG_ALIGNED;
674
675 /* We can't fit any SACK blocks in a packet with MD5 + TS
676 * options. There was discussion about disabling SACK
677 * rather than TS in order to fit in better with old,
678 * buggy kernels, but that was deemed to be unnecessary.
679 */
680 ireq->tstamp_ok &= !ireq->sack_ok;
681 }
682#endif
683
684 /* We always send an MSS option. */
685 opts->mss = mss;
686 remaining -= TCPOLEN_MSS_ALIGNED;
687
688 if (likely(ireq->wscale_ok)) {
689 opts->ws = ireq->rcv_wscale;
690 opts->options |= OPTION_WSCALE;
691 remaining -= TCPOLEN_WSCALE_ALIGNED;
692 }
693 if (likely(ireq->tstamp_ok)) {
694 opts->options |= OPTION_TS;
695 opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off;
696 opts->tsecr = req->ts_recent;
697 remaining -= TCPOLEN_TSTAMP_ALIGNED;
698 }
699 if (likely(ireq->sack_ok)) {
700 opts->options |= OPTION_SACK_ADVERTISE;
701 if (unlikely(!ireq->tstamp_ok))
702 remaining -= TCPOLEN_SACKPERM_ALIGNED;
703 }
704 if (foc != NULL && foc->len >= 0) {
705 u32 need = foc->len;
706
707 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
708 TCPOLEN_FASTOPEN_BASE;
709 need = (need + 3) & ~3U; /* Align to 32 bits */
710 if (remaining >= need) {
711 opts->options |= OPTION_FAST_OPEN_COOKIE;
712 opts->fastopen_cookie = foc;
713 remaining -= need;
714 }
715 }
716
717 smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining);
718
719 return MAX_TCP_OPTION_SPACE - remaining;
720}
721
722/* Compute TCP options for ESTABLISHED sockets. This is not the
723 * final wire format yet.
724 */
725static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
726 struct tcp_out_options *opts,
727 struct tcp_md5sig_key **md5)
728{
729 struct tcp_sock *tp = tcp_sk(sk);
730 unsigned int size = 0;
731 unsigned int eff_sacks;
732
733 opts->options = 0;
734
735 *md5 = NULL;
736#ifdef CONFIG_TCP_MD5SIG
737 if (static_branch_unlikely(&tcp_md5_needed) &&
738 rcu_access_pointer(tp->md5sig_info)) {
739 *md5 = tp->af_specific->md5_lookup(sk, sk);
740 if (*md5) {
741 opts->options |= OPTION_MD5;
742 size += TCPOLEN_MD5SIG_ALIGNED;
743 }
744 }
745#endif
746
747 if (likely(tp->rx_opt.tstamp_ok)) {
748 opts->options |= OPTION_TS;
749 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
750 opts->tsecr = tp->rx_opt.ts_recent;
751 size += TCPOLEN_TSTAMP_ALIGNED;
752 }
753
754 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
755 if (unlikely(eff_sacks)) {
756 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
757 opts->num_sack_blocks =
758 min_t(unsigned int, eff_sacks,
759 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
760 TCPOLEN_SACK_PERBLOCK);
761 size += TCPOLEN_SACK_BASE_ALIGNED +
762 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
763 }
764
765 return size;
766}
767
768
769/* TCP SMALL QUEUES (TSQ)
770 *
771 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
772 * to reduce RTT and bufferbloat.
773 * We do this using a special skb destructor (tcp_wfree).
774 *
775 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
776 * needs to be reallocated in a driver.
777 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
778 *
779 * Since transmit from skb destructor is forbidden, we use a tasklet
780 * to process all sockets that eventually need to send more skbs.
781 * We use one tasklet per cpu, with its own queue of sockets.
782 */
783struct tsq_tasklet {
784 struct tasklet_struct tasklet;
785 struct list_head head; /* queue of tcp sockets */
786};
787static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
788
789static void tcp_tsq_write(struct sock *sk)
790{
791 if ((1 << sk->sk_state) &
792 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
793 TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) {
794 struct tcp_sock *tp = tcp_sk(sk);
795
796 if (tp->lost_out > tp->retrans_out &&
797 tp->snd_cwnd > tcp_packets_in_flight(tp)) {
798 tcp_mstamp_refresh(tp);
799 tcp_xmit_retransmit_queue(sk);
800 }
801
802 tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
803 0, GFP_ATOMIC);
804 }
805}
806
807static void tcp_tsq_handler(struct sock *sk)
808{
809 bh_lock_sock(sk);
810 if (!sock_owned_by_user(sk))
811 tcp_tsq_write(sk);
812 else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
813 sock_hold(sk);
814 bh_unlock_sock(sk);
815}
816/*
817 * One tasklet per cpu tries to send more skbs.
818 * We run in tasklet context but need to disable irqs when
819 * transferring tsq->head because tcp_wfree() might
820 * interrupt us (non NAPI drivers)
821 */
822static void tcp_tasklet_func(unsigned long data)
823{
824 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
825 LIST_HEAD(list);
826 unsigned long flags;
827 struct list_head *q, *n;
828 struct tcp_sock *tp;
829 struct sock *sk;
830
831 local_irq_save(flags);
832 list_splice_init(&tsq->head, &list);
833 local_irq_restore(flags);
834
835 list_for_each_safe(q, n, &list) {
836 tp = list_entry(q, struct tcp_sock, tsq_node);
837 list_del(&tp->tsq_node);
838
839 sk = (struct sock *)tp;
840 smp_mb__before_atomic();
841 clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
842
843 tcp_tsq_handler(sk);
844 sk_free(sk);
845 }
846}
847
848#define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
849 TCPF_WRITE_TIMER_DEFERRED | \
850 TCPF_DELACK_TIMER_DEFERRED | \
851 TCPF_MTU_REDUCED_DEFERRED)
852/**
853 * tcp_release_cb - tcp release_sock() callback
854 * @sk: socket
855 *
856 * called from release_sock() to perform protocol dependent
857 * actions before socket release.
858 */
859void tcp_release_cb(struct sock *sk)
860{
861 unsigned long flags, nflags;
862
863 /* perform an atomic operation only if at least one flag is set */
864 do {
865 flags = sk->sk_tsq_flags;
866 if (!(flags & TCP_DEFERRED_ALL))
867 return;
868 nflags = flags & ~TCP_DEFERRED_ALL;
869 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
870
871 if (flags & TCPF_TSQ_DEFERRED) {
872 tcp_tsq_write(sk);
873 __sock_put(sk);
874 }
875 /* Here begins the tricky part :
876 * We are called from release_sock() with :
877 * 1) BH disabled
878 * 2) sk_lock.slock spinlock held
879 * 3) socket owned by us (sk->sk_lock.owned == 1)
880 *
881 * But following code is meant to be called from BH handlers,
882 * so we should keep BH disabled, but early release socket ownership
883 */
884 sock_release_ownership(sk);
885
886 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
887 tcp_write_timer_handler(sk);
888 __sock_put(sk);
889 }
890 if (flags & TCPF_DELACK_TIMER_DEFERRED) {
891 tcp_delack_timer_handler(sk);
892 __sock_put(sk);
893 }
894 if (flags & TCPF_MTU_REDUCED_DEFERRED) {
895 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
896 __sock_put(sk);
897 }
898}
899EXPORT_SYMBOL(tcp_release_cb);
900
901void __init tcp_tasklet_init(void)
902{
903 int i;
904
905 for_each_possible_cpu(i) {
906 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
907
908 INIT_LIST_HEAD(&tsq->head);
909 tasklet_init(&tsq->tasklet,
910 tcp_tasklet_func,
911 (unsigned long)tsq);
912 }
913}
914
915/*
916 * Write buffer destructor automatically called from kfree_skb.
917 * We can't xmit new skbs from this context, as we might already
918 * hold qdisc lock.
919 */
920void tcp_wfree(struct sk_buff *skb)
921{
922 struct sock *sk = skb->sk;
923 struct tcp_sock *tp = tcp_sk(sk);
924 unsigned long flags, nval, oval;
925
926 /* Keep one reference on sk_wmem_alloc.
927 * Will be released by sk_free() from here or tcp_tasklet_func()
928 */
929 WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc));
930
931 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
932 * Wait until our queues (qdisc + devices) are drained.
933 * This gives :
934 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
935 * - chance for incoming ACK (processed by another cpu maybe)
936 * to migrate this flow (skb->ooo_okay will be eventually set)
937 */
938 if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
939 goto out;
940
941 for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
942 struct tsq_tasklet *tsq;
943 bool empty;
944
945 if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
946 goto out;
947
948 nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED;
949 nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
950 if (nval != oval)
951 continue;
952
953 /* queue this socket to tasklet queue */
954 local_irq_save(flags);
955 tsq = this_cpu_ptr(&tsq_tasklet);
956 empty = list_empty(&tsq->head);
957 list_add(&tp->tsq_node, &tsq->head);
958 if (empty)
959 tasklet_schedule(&tsq->tasklet);
960 local_irq_restore(flags);
961 return;
962 }
963out:
964 sk_free(sk);
965}
966
967/* Note: Called under soft irq.
968 * We can call TCP stack right away, unless socket is owned by user.
969 */
970enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer)
971{
972 struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer);
973 struct sock *sk = (struct sock *)tp;
974
975 tcp_tsq_handler(sk);
976 sock_put(sk);
977
978 return HRTIMER_NORESTART;
979}
980
981static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb,
982 u64 prior_wstamp)
983{
984 struct tcp_sock *tp = tcp_sk(sk);
985
986 if (sk->sk_pacing_status != SK_PACING_NONE) {
987 unsigned long rate = sk->sk_pacing_rate;
988
989 /* Original sch_fq does not pace first 10 MSS
990 * Note that tp->data_segs_out overflows after 2^32 packets,
991 * this is a minor annoyance.
992 */
993 if (rate != ~0UL && rate && tp->data_segs_out >= 10) {
994 u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate);
995 u64 credit = tp->tcp_wstamp_ns - prior_wstamp;
996
997 /* take into account OS jitter */
998 len_ns -= min_t(u64, len_ns / 2, credit);
999 tp->tcp_wstamp_ns += len_ns;
1000 }
1001 }
1002 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
1003}
1004
1005/* This routine actually transmits TCP packets queued in by
1006 * tcp_do_sendmsg(). This is used by both the initial
1007 * transmission and possible later retransmissions.
1008 * All SKB's seen here are completely headerless. It is our
1009 * job to build the TCP header, and pass the packet down to
1010 * IP so it can do the same plus pass the packet off to the
1011 * device.
1012 *
1013 * We are working here with either a clone of the original
1014 * SKB, or a fresh unique copy made by the retransmit engine.
1015 */
1016static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb,
1017 int clone_it, gfp_t gfp_mask, u32 rcv_nxt)
1018{
1019 const struct inet_connection_sock *icsk = inet_csk(sk);
1020 struct inet_sock *inet;
1021 struct tcp_sock *tp;
1022 struct tcp_skb_cb *tcb;
1023 struct tcp_out_options opts;
1024 unsigned int tcp_options_size, tcp_header_size;
1025 struct sk_buff *oskb = NULL;
1026 struct tcp_md5sig_key *md5;
1027 struct tcphdr *th;
1028 u64 prior_wstamp;
1029 int err;
1030
1031 BUG_ON(!skb || !tcp_skb_pcount(skb));
1032 tp = tcp_sk(sk);
1033 prior_wstamp = tp->tcp_wstamp_ns;
1034 tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache);
1035 skb->skb_mstamp_ns = tp->tcp_wstamp_ns;
1036 if (clone_it) {
1037 TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq
1038 - tp->snd_una;
1039 oskb = skb;
1040
1041 tcp_skb_tsorted_save(oskb) {
1042 if (unlikely(skb_cloned(oskb)))
1043 skb = pskb_copy(oskb, gfp_mask);
1044 else
1045 skb = skb_clone(oskb, gfp_mask);
1046 } tcp_skb_tsorted_restore(oskb);
1047
1048 if (unlikely(!skb))
1049 return -ENOBUFS;
1050 }
1051
1052 inet = inet_sk(sk);
1053 tcb = TCP_SKB_CB(skb);
1054 memset(&opts, 0, sizeof(opts));
1055
1056 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
1057 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
1058 else
1059 tcp_options_size = tcp_established_options(sk, skb, &opts,
1060 &md5);
1061 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1062
1063 /* if no packet is in qdisc/device queue, then allow XPS to select
1064 * another queue. We can be called from tcp_tsq_handler()
1065 * which holds one reference to sk.
1066 *
1067 * TODO: Ideally, in-flight pure ACK packets should not matter here.
1068 * One way to get this would be to set skb->truesize = 2 on them.
1069 */
1070 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
1071
1072 /* If we had to use memory reserve to allocate this skb,
1073 * this might cause drops if packet is looped back :
1074 * Other socket might not have SOCK_MEMALLOC.
1075 * Packets not looped back do not care about pfmemalloc.
1076 */
1077 skb->pfmemalloc = 0;
1078
1079 skb_push(skb, tcp_header_size);
1080 skb_reset_transport_header(skb);
1081
1082 skb_orphan(skb);
1083 skb->sk = sk;
1084 skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
1085 skb_set_hash_from_sk(skb, sk);
1086 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
1087
1088 skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm);
1089
1090 /* Build TCP header and checksum it. */
1091 th = (struct tcphdr *)skb->data;
1092 th->source = inet->inet_sport;
1093 th->dest = inet->inet_dport;
1094 th->seq = htonl(tcb->seq);
1095 th->ack_seq = htonl(rcv_nxt);
1096 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
1097 tcb->tcp_flags);
1098
1099 th->check = 0;
1100 th->urg_ptr = 0;
1101
1102 /* The urg_mode check is necessary during a below snd_una win probe */
1103 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1104 if (before(tp->snd_up, tcb->seq + 0x10000)) {
1105 th->urg_ptr = htons(tp->snd_up - tcb->seq);
1106 th->urg = 1;
1107 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1108 th->urg_ptr = htons(0xFFFF);
1109 th->urg = 1;
1110 }
1111 }
1112
1113 tcp_options_write((__be32 *)(th + 1), tp, &opts);
1114 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1115 if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1116 th->window = htons(tcp_select_window(sk));
1117 tcp_ecn_send(sk, skb, th, tcp_header_size);
1118 } else {
1119 /* RFC1323: The window in SYN & SYN/ACK segments
1120 * is never scaled.
1121 */
1122 th->window = htons(min(tp->rcv_wnd, 65535U));
1123 }
1124#ifdef CONFIG_TCP_MD5SIG
1125 /* Calculate the MD5 hash, as we have all we need now */
1126 if (md5) {
1127 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1128 tp->af_specific->calc_md5_hash(opts.hash_location,
1129 md5, sk, skb);
1130 }
1131#endif
1132
1133 icsk->icsk_af_ops->send_check(sk, skb);
1134
1135 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1136 tcp_event_ack_sent(sk, tcp_skb_pcount(skb), rcv_nxt);
1137
1138 if (skb->len != tcp_header_size) {
1139 tcp_event_data_sent(tp, sk);
1140 tp->data_segs_out += tcp_skb_pcount(skb);
1141 tp->bytes_sent += skb->len - tcp_header_size;
1142 }
1143
1144 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1145 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1146 tcp_skb_pcount(skb));
1147
1148 tp->segs_out += tcp_skb_pcount(skb);
1149 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1150 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1151 skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1152
1153 /* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */
1154
1155 /* Cleanup our debris for IP stacks */
1156 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1157 sizeof(struct inet6_skb_parm)));
1158
1159 err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
1160
1161 if (unlikely(err > 0)) {
1162 tcp_enter_cwr(sk);
1163 err = net_xmit_eval(err);
1164 }
1165 if (!err && oskb) {
1166 tcp_update_skb_after_send(sk, oskb, prior_wstamp);
1167 tcp_rate_skb_sent(sk, oskb);
1168 }
1169 return err;
1170}
1171
1172static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
1173 gfp_t gfp_mask)
1174{
1175 return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask,
1176 tcp_sk(sk)->rcv_nxt);
1177}
1178
1179/* This routine just queues the buffer for sending.
1180 *
1181 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1182 * otherwise socket can stall.
1183 */
1184static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1185{
1186 struct tcp_sock *tp = tcp_sk(sk);
1187
1188 /* Advance write_seq and place onto the write_queue. */
1189 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
1190 __skb_header_release(skb);
1191 tcp_add_write_queue_tail(sk, skb);
1192 sk->sk_wmem_queued += skb->truesize;
1193 sk_mem_charge(sk, skb->truesize);
1194}
1195
1196/* Initialize TSO segments for a packet. */
1197static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1198{
1199 if (skb->len <= mss_now) {
1200 /* Avoid the costly divide in the normal
1201 * non-TSO case.
1202 */
1203 tcp_skb_pcount_set(skb, 1);
1204 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1205 } else {
1206 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1207 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1208 }
1209}
1210
1211/* Pcount in the middle of the write queue got changed, we need to do various
1212 * tweaks to fix counters
1213 */
1214static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1215{
1216 struct tcp_sock *tp = tcp_sk(sk);
1217
1218 tp->packets_out -= decr;
1219
1220 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1221 tp->sacked_out -= decr;
1222 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1223 tp->retrans_out -= decr;
1224 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1225 tp->lost_out -= decr;
1226
1227 /* Reno case is special. Sigh... */
1228 if (tcp_is_reno(tp) && decr > 0)
1229 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1230
1231 if (tp->lost_skb_hint &&
1232 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1233 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
1234 tp->lost_cnt_hint -= decr;
1235
1236 tcp_verify_left_out(tp);
1237}
1238
1239static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1240{
1241 return TCP_SKB_CB(skb)->txstamp_ack ||
1242 (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1243}
1244
1245static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1246{
1247 struct skb_shared_info *shinfo = skb_shinfo(skb);
1248
1249 if (unlikely(tcp_has_tx_tstamp(skb)) &&
1250 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1251 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1252 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1253
1254 shinfo->tx_flags &= ~tsflags;
1255 shinfo2->tx_flags |= tsflags;
1256 swap(shinfo->tskey, shinfo2->tskey);
1257 TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1258 TCP_SKB_CB(skb)->txstamp_ack = 0;
1259 }
1260}
1261
1262static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1263{
1264 TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1265 TCP_SKB_CB(skb)->eor = 0;
1266}
1267
1268/* Insert buff after skb on the write or rtx queue of sk. */
1269static void tcp_insert_write_queue_after(struct sk_buff *skb,
1270 struct sk_buff *buff,
1271 struct sock *sk,
1272 enum tcp_queue tcp_queue)
1273{
1274 if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE)
1275 __skb_queue_after(&sk->sk_write_queue, skb, buff);
1276 else
1277 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
1278}
1279
1280/* Function to create two new TCP segments. Shrinks the given segment
1281 * to the specified size and appends a new segment with the rest of the
1282 * packet to the list. This won't be called frequently, I hope.
1283 * Remember, these are still headerless SKBs at this point.
1284 */
1285int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1286 struct sk_buff *skb, u32 len,
1287 unsigned int mss_now, gfp_t gfp)
1288{
1289 struct tcp_sock *tp = tcp_sk(sk);
1290 struct sk_buff *buff;
1291 int nsize, old_factor;
1292 int nlen;
1293 u8 flags;
1294
1295 if (WARN_ON(len > skb->len))
1296 return -EINVAL;
1297
1298 nsize = skb_headlen(skb) - len;
1299 if (nsize < 0)
1300 nsize = 0;
1301
1302 if (skb_unclone(skb, gfp))
1303 return -ENOMEM;
1304
1305 /* Get a new skb... force flag on. */
1306 buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
1307 if (!buff)
1308 return -ENOMEM; /* We'll just try again later. */
1309
1310 sk->sk_wmem_queued += buff->truesize;
1311 sk_mem_charge(sk, buff->truesize);
1312 nlen = skb->len - len - nsize;
1313 buff->truesize += nlen;
1314 skb->truesize -= nlen;
1315
1316 /* Correct the sequence numbers. */
1317 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1318 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1319 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1320
1321 /* PSH and FIN should only be set in the second packet. */
1322 flags = TCP_SKB_CB(skb)->tcp_flags;
1323 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1324 TCP_SKB_CB(buff)->tcp_flags = flags;
1325 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1326 tcp_skb_fragment_eor(skb, buff);
1327
1328 skb_split(skb, buff, len);
1329
1330 buff->ip_summed = CHECKSUM_PARTIAL;
1331
1332 buff->tstamp = skb->tstamp;
1333 tcp_fragment_tstamp(skb, buff);
1334
1335 old_factor = tcp_skb_pcount(skb);
1336
1337 /* Fix up tso_factor for both original and new SKB. */
1338 tcp_set_skb_tso_segs(skb, mss_now);
1339 tcp_set_skb_tso_segs(buff, mss_now);
1340
1341 /* Update delivered info for the new segment */
1342 TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1343
1344 /* If this packet has been sent out already, we must
1345 * adjust the various packet counters.
1346 */
1347 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1348 int diff = old_factor - tcp_skb_pcount(skb) -
1349 tcp_skb_pcount(buff);
1350
1351 if (diff)
1352 tcp_adjust_pcount(sk, skb, diff);
1353 }
1354
1355 /* Link BUFF into the send queue. */
1356 __skb_header_release(buff);
1357 tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1358 if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE)
1359 list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor);
1360
1361 return 0;
1362}
1363
1364/* This is similar to __pskb_pull_tail(). The difference is that pulled
1365 * data is not copied, but immediately discarded.
1366 */
1367static int __pskb_trim_head(struct sk_buff *skb, int len)
1368{
1369 struct skb_shared_info *shinfo;
1370 int i, k, eat;
1371
1372 eat = min_t(int, len, skb_headlen(skb));
1373 if (eat) {
1374 __skb_pull(skb, eat);
1375 len -= eat;
1376 if (!len)
1377 return 0;
1378 }
1379 eat = len;
1380 k = 0;
1381 shinfo = skb_shinfo(skb);
1382 for (i = 0; i < shinfo->nr_frags; i++) {
1383 int size = skb_frag_size(&shinfo->frags[i]);
1384
1385 if (size <= eat) {
1386 skb_frag_unref(skb, i);
1387 eat -= size;
1388 } else {
1389 shinfo->frags[k] = shinfo->frags[i];
1390 if (eat) {
1391 shinfo->frags[k].page_offset += eat;
1392 skb_frag_size_sub(&shinfo->frags[k], eat);
1393 eat = 0;
1394 }
1395 k++;
1396 }
1397 }
1398 shinfo->nr_frags = k;
1399
1400 skb->data_len -= len;
1401 skb->len = skb->data_len;
1402 return len;
1403}
1404
1405/* Remove acked data from a packet in the transmit queue. */
1406int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1407{
1408 u32 delta_truesize;
1409
1410 if (skb_unclone(skb, GFP_ATOMIC))
1411 return -ENOMEM;
1412
1413 delta_truesize = __pskb_trim_head(skb, len);
1414
1415 TCP_SKB_CB(skb)->seq += len;
1416 skb->ip_summed = CHECKSUM_PARTIAL;
1417
1418 if (delta_truesize) {
1419 skb->truesize -= delta_truesize;
1420 sk->sk_wmem_queued -= delta_truesize;
1421 sk_mem_uncharge(sk, delta_truesize);
1422 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1423 }
1424
1425 /* Any change of skb->len requires recalculation of tso factor. */
1426 if (tcp_skb_pcount(skb) > 1)
1427 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1428
1429 return 0;
1430}
1431
1432/* Calculate MSS not accounting any TCP options. */
1433static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1434{
1435 const struct tcp_sock *tp = tcp_sk(sk);
1436 const struct inet_connection_sock *icsk = inet_csk(sk);
1437 int mss_now;
1438
1439 /* Calculate base mss without TCP options:
1440 It is MMS_S - sizeof(tcphdr) of rfc1122
1441 */
1442 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1443
1444 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1445 if (icsk->icsk_af_ops->net_frag_header_len) {
1446 const struct dst_entry *dst = __sk_dst_get(sk);
1447
1448 if (dst && dst_allfrag(dst))
1449 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1450 }
1451
1452 /* Clamp it (mss_clamp does not include tcp options) */
1453 if (mss_now > tp->rx_opt.mss_clamp)
1454 mss_now = tp->rx_opt.mss_clamp;
1455
1456 /* Now subtract optional transport overhead */
1457 mss_now -= icsk->icsk_ext_hdr_len;
1458
1459 /* Then reserve room for full set of TCP options and 8 bytes of data */
1460 if (mss_now < 48)
1461 mss_now = 48;
1462 return mss_now;
1463}
1464
1465/* Calculate MSS. Not accounting for SACKs here. */
1466int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1467{
1468 /* Subtract TCP options size, not including SACKs */
1469 return __tcp_mtu_to_mss(sk, pmtu) -
1470 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1471}
1472
1473/* Inverse of above */
1474int tcp_mss_to_mtu(struct sock *sk, int mss)
1475{
1476 const struct tcp_sock *tp = tcp_sk(sk);
1477 const struct inet_connection_sock *icsk = inet_csk(sk);
1478 int mtu;
1479
1480 mtu = mss +
1481 tp->tcp_header_len +
1482 icsk->icsk_ext_hdr_len +
1483 icsk->icsk_af_ops->net_header_len;
1484
1485 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1486 if (icsk->icsk_af_ops->net_frag_header_len) {
1487 const struct dst_entry *dst = __sk_dst_get(sk);
1488
1489 if (dst && dst_allfrag(dst))
1490 mtu += icsk->icsk_af_ops->net_frag_header_len;
1491 }
1492 return mtu;
1493}
1494EXPORT_SYMBOL(tcp_mss_to_mtu);
1495
1496/* MTU probing init per socket */
1497void tcp_mtup_init(struct sock *sk)
1498{
1499 struct tcp_sock *tp = tcp_sk(sk);
1500 struct inet_connection_sock *icsk = inet_csk(sk);
1501 struct net *net = sock_net(sk);
1502
1503 icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
1504 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1505 icsk->icsk_af_ops->net_header_len;
1506 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
1507 icsk->icsk_mtup.probe_size = 0;
1508 if (icsk->icsk_mtup.enabled)
1509 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
1510}
1511EXPORT_SYMBOL(tcp_mtup_init);
1512
1513/* This function synchronize snd mss to current pmtu/exthdr set.
1514
1515 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1516 for TCP options, but includes only bare TCP header.
1517
1518 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1519 It is minimum of user_mss and mss received with SYN.
1520 It also does not include TCP options.
1521
1522 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1523
1524 tp->mss_cache is current effective sending mss, including
1525 all tcp options except for SACKs. It is evaluated,
1526 taking into account current pmtu, but never exceeds
1527 tp->rx_opt.mss_clamp.
1528
1529 NOTE1. rfc1122 clearly states that advertised MSS
1530 DOES NOT include either tcp or ip options.
1531
1532 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1533 are READ ONLY outside this function. --ANK (980731)
1534 */
1535unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1536{
1537 struct tcp_sock *tp = tcp_sk(sk);
1538 struct inet_connection_sock *icsk = inet_csk(sk);
1539 int mss_now;
1540
1541 if (icsk->icsk_mtup.search_high > pmtu)
1542 icsk->icsk_mtup.search_high = pmtu;
1543
1544 mss_now = tcp_mtu_to_mss(sk, pmtu);
1545 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1546
1547 /* And store cached results */
1548 icsk->icsk_pmtu_cookie = pmtu;
1549 if (icsk->icsk_mtup.enabled)
1550 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1551 tp->mss_cache = mss_now;
1552
1553 return mss_now;
1554}
1555EXPORT_SYMBOL(tcp_sync_mss);
1556
1557/* Compute the current effective MSS, taking SACKs and IP options,
1558 * and even PMTU discovery events into account.
1559 */
1560unsigned int tcp_current_mss(struct sock *sk)
1561{
1562 const struct tcp_sock *tp = tcp_sk(sk);
1563 const struct dst_entry *dst = __sk_dst_get(sk);
1564 u32 mss_now;
1565 unsigned int header_len;
1566 struct tcp_out_options opts;
1567 struct tcp_md5sig_key *md5;
1568
1569 mss_now = tp->mss_cache;
1570
1571 if (dst) {
1572 u32 mtu = dst_mtu(dst);
1573 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1574 mss_now = tcp_sync_mss(sk, mtu);
1575 }
1576
1577 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1578 sizeof(struct tcphdr);
1579 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1580 * some common options. If this is an odd packet (because we have SACK
1581 * blocks etc) then our calculated header_len will be different, and
1582 * we have to adjust mss_now correspondingly */
1583 if (header_len != tp->tcp_header_len) {
1584 int delta = (int) header_len - tp->tcp_header_len;
1585 mss_now -= delta;
1586 }
1587
1588 return mss_now;
1589}
1590
1591/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1592 * As additional protections, we do not touch cwnd in retransmission phases,
1593 * and if application hit its sndbuf limit recently.
1594 */
1595static void tcp_cwnd_application_limited(struct sock *sk)
1596{
1597 struct tcp_sock *tp = tcp_sk(sk);
1598
1599 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1600 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1601 /* Limited by application or receiver window. */
1602 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1603 u32 win_used = max(tp->snd_cwnd_used, init_win);
1604 if (win_used < tp->snd_cwnd) {
1605 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1606 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1607 }
1608 tp->snd_cwnd_used = 0;
1609 }
1610 tp->snd_cwnd_stamp = tcp_jiffies32;
1611}
1612
1613static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1614{
1615 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1616 struct tcp_sock *tp = tcp_sk(sk);
1617
1618 /* Track the maximum number of outstanding packets in each
1619 * window, and remember whether we were cwnd-limited then.
1620 */
1621 if (!before(tp->snd_una, tp->max_packets_seq) ||
1622 tp->packets_out > tp->max_packets_out) {
1623 tp->max_packets_out = tp->packets_out;
1624 tp->max_packets_seq = tp->snd_nxt;
1625 tp->is_cwnd_limited = is_cwnd_limited;
1626 }
1627
1628 if (tcp_is_cwnd_limited(sk)) {
1629 /* Network is feed fully. */
1630 tp->snd_cwnd_used = 0;
1631 tp->snd_cwnd_stamp = tcp_jiffies32;
1632 } else {
1633 /* Network starves. */
1634 if (tp->packets_out > tp->snd_cwnd_used)
1635 tp->snd_cwnd_used = tp->packets_out;
1636
1637 if (sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle &&
1638 (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto &&
1639 !ca_ops->cong_control)
1640 tcp_cwnd_application_limited(sk);
1641
1642 /* The following conditions together indicate the starvation
1643 * is caused by insufficient sender buffer:
1644 * 1) just sent some data (see tcp_write_xmit)
1645 * 2) not cwnd limited (this else condition)
1646 * 3) no more data to send (tcp_write_queue_empty())
1647 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1648 */
1649 if (tcp_write_queue_empty(sk) && sk->sk_socket &&
1650 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
1651 (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1652 tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED);
1653 }
1654}
1655
1656/* Minshall's variant of the Nagle send check. */
1657static bool tcp_minshall_check(const struct tcp_sock *tp)
1658{
1659 return after(tp->snd_sml, tp->snd_una) &&
1660 !after(tp->snd_sml, tp->snd_nxt);
1661}
1662
1663/* Update snd_sml if this skb is under mss
1664 * Note that a TSO packet might end with a sub-mss segment
1665 * The test is really :
1666 * if ((skb->len % mss) != 0)
1667 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1668 * But we can avoid doing the divide again given we already have
1669 * skb_pcount = skb->len / mss_now
1670 */
1671static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1672 const struct sk_buff *skb)
1673{
1674 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1675 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1676}
1677
1678/* Return false, if packet can be sent now without violation Nagle's rules:
1679 * 1. It is full sized. (provided by caller in %partial bool)
1680 * 2. Or it contains FIN. (already checked by caller)
1681 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1682 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1683 * With Minshall's modification: all sent small packets are ACKed.
1684 */
1685static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1686 int nonagle)
1687{
1688 return partial &&
1689 ((nonagle & TCP_NAGLE_CORK) ||
1690 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1691}
1692
1693/* Return how many segs we'd like on a TSO packet,
1694 * to send one TSO packet per ms
1695 */
1696static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
1697 int min_tso_segs)
1698{
1699 u32 bytes, segs;
1700
1701 bytes = min_t(unsigned long,
1702 sk->sk_pacing_rate >> sk->sk_pacing_shift,
1703 sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1704
1705 /* Goal is to send at least one packet per ms,
1706 * not one big TSO packet every 100 ms.
1707 * This preserves ACK clocking and is consistent
1708 * with tcp_tso_should_defer() heuristic.
1709 */
1710 segs = max_t(u32, bytes / mss_now, min_tso_segs);
1711
1712 return segs;
1713}
1714
1715/* Return the number of segments we want in the skb we are transmitting.
1716 * See if congestion control module wants to decide; otherwise, autosize.
1717 */
1718static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
1719{
1720 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1721 u32 min_tso, tso_segs;
1722
1723 min_tso = ca_ops->min_tso_segs ?
1724 ca_ops->min_tso_segs(sk) :
1725 sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs;
1726
1727 tso_segs = tcp_tso_autosize(sk, mss_now, min_tso);
1728 return min_t(u32, tso_segs, sk->sk_gso_max_segs);
1729}
1730
1731/* Returns the portion of skb which can be sent right away */
1732static unsigned int tcp_mss_split_point(const struct sock *sk,
1733 const struct sk_buff *skb,
1734 unsigned int mss_now,
1735 unsigned int max_segs,
1736 int nonagle)
1737{
1738 const struct tcp_sock *tp = tcp_sk(sk);
1739 u32 partial, needed, window, max_len;
1740
1741 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1742 max_len = mss_now * max_segs;
1743
1744 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1745 return max_len;
1746
1747 needed = min(skb->len, window);
1748
1749 if (max_len <= needed)
1750 return max_len;
1751
1752 partial = needed % mss_now;
1753 /* If last segment is not a full MSS, check if Nagle rules allow us
1754 * to include this last segment in this skb.
1755 * Otherwise, we'll split the skb at last MSS boundary
1756 */
1757 if (tcp_nagle_check(partial != 0, tp, nonagle))
1758 return needed - partial;
1759
1760 return needed;
1761}
1762
1763/* Can at least one segment of SKB be sent right now, according to the
1764 * congestion window rules? If so, return how many segments are allowed.
1765 */
1766static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1767 const struct sk_buff *skb)
1768{
1769 u32 in_flight, cwnd, halfcwnd;
1770
1771 /* Don't be strict about the congestion window for the final FIN. */
1772 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1773 tcp_skb_pcount(skb) == 1)
1774 return 1;
1775
1776 in_flight = tcp_packets_in_flight(tp);
1777 cwnd = tp->snd_cwnd;
1778 if (in_flight >= cwnd)
1779 return 0;
1780
1781 /* For better scheduling, ensure we have at least
1782 * 2 GSO packets in flight.
1783 */
1784 halfcwnd = max(cwnd >> 1, 1U);
1785 return min(halfcwnd, cwnd - in_flight);
1786}
1787
1788/* Initialize TSO state of a skb.
1789 * This must be invoked the first time we consider transmitting
1790 * SKB onto the wire.
1791 */
1792static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1793{
1794 int tso_segs = tcp_skb_pcount(skb);
1795
1796 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1797 tcp_set_skb_tso_segs(skb, mss_now);
1798 tso_segs = tcp_skb_pcount(skb);
1799 }
1800 return tso_segs;
1801}
1802
1803
1804/* Return true if the Nagle test allows this packet to be
1805 * sent now.
1806 */
1807static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1808 unsigned int cur_mss, int nonagle)
1809{
1810 /* Nagle rule does not apply to frames, which sit in the middle of the
1811 * write_queue (they have no chances to get new data).
1812 *
1813 * This is implemented in the callers, where they modify the 'nonagle'
1814 * argument based upon the location of SKB in the send queue.
1815 */
1816 if (nonagle & TCP_NAGLE_PUSH)
1817 return true;
1818
1819 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1820 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1821 return true;
1822
1823 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1824 return true;
1825
1826 return false;
1827}
1828
1829/* Does at least the first segment of SKB fit into the send window? */
1830static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1831 const struct sk_buff *skb,
1832 unsigned int cur_mss)
1833{
1834 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1835
1836 if (skb->len > cur_mss)
1837 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1838
1839 return !after(end_seq, tcp_wnd_end(tp));
1840}
1841
1842/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1843 * which is put after SKB on the list. It is very much like
1844 * tcp_fragment() except that it may make several kinds of assumptions
1845 * in order to speed up the splitting operation. In particular, we
1846 * know that all the data is in scatter-gather pages, and that the
1847 * packet has never been sent out before (and thus is not cloned).
1848 */
1849static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1850 unsigned int mss_now, gfp_t gfp)
1851{
1852 int nlen = skb->len - len;
1853 struct sk_buff *buff;
1854 u8 flags;
1855
1856 /* All of a TSO frame must be composed of paged data. */
1857 if (skb->len != skb->data_len)
1858 return tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
1859 skb, len, mss_now, gfp);
1860
1861 buff = sk_stream_alloc_skb(sk, 0, gfp, true);
1862 if (unlikely(!buff))
1863 return -ENOMEM;
1864
1865 sk->sk_wmem_queued += buff->truesize;
1866 sk_mem_charge(sk, buff->truesize);
1867 buff->truesize += nlen;
1868 skb->truesize -= nlen;
1869
1870 /* Correct the sequence numbers. */
1871 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1872 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1873 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1874
1875 /* PSH and FIN should only be set in the second packet. */
1876 flags = TCP_SKB_CB(skb)->tcp_flags;
1877 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1878 TCP_SKB_CB(buff)->tcp_flags = flags;
1879
1880 /* This packet was never sent out yet, so no SACK bits. */
1881 TCP_SKB_CB(buff)->sacked = 0;
1882
1883 tcp_skb_fragment_eor(skb, buff);
1884
1885 buff->ip_summed = CHECKSUM_PARTIAL;
1886 skb_split(skb, buff, len);
1887 tcp_fragment_tstamp(skb, buff);
1888
1889 /* Fix up tso_factor for both original and new SKB. */
1890 tcp_set_skb_tso_segs(skb, mss_now);
1891 tcp_set_skb_tso_segs(buff, mss_now);
1892
1893 /* Link BUFF into the send queue. */
1894 __skb_header_release(buff);
1895 tcp_insert_write_queue_after(skb, buff, sk, TCP_FRAG_IN_WRITE_QUEUE);
1896
1897 return 0;
1898}
1899
1900/* Try to defer sending, if possible, in order to minimize the amount
1901 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1902 *
1903 * This algorithm is from John Heffner.
1904 */
1905static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
1906 bool *is_cwnd_limited,
1907 bool *is_rwnd_limited,
1908 u32 max_segs)
1909{
1910 const struct inet_connection_sock *icsk = inet_csk(sk);
1911 u32 send_win, cong_win, limit, in_flight;
1912 struct tcp_sock *tp = tcp_sk(sk);
1913 struct sk_buff *head;
1914 int win_divisor;
1915 s64 delta;
1916
1917 if (icsk->icsk_ca_state >= TCP_CA_Recovery)
1918 goto send_now;
1919
1920 /* Avoid bursty behavior by allowing defer
1921 * only if the last write was recent (1 ms).
1922 * Note that tp->tcp_wstamp_ns can be in the future if we have
1923 * packets waiting in a qdisc or device for EDT delivery.
1924 */
1925 delta = tp->tcp_clock_cache - tp->tcp_wstamp_ns - NSEC_PER_MSEC;
1926 if (delta > 0)
1927 goto send_now;
1928
1929 in_flight = tcp_packets_in_flight(tp);
1930
1931 BUG_ON(tcp_skb_pcount(skb) <= 1);
1932 BUG_ON(tp->snd_cwnd <= in_flight);
1933
1934 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1935
1936 /* From in_flight test above, we know that cwnd > in_flight. */
1937 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1938
1939 limit = min(send_win, cong_win);
1940
1941 /* If a full-sized TSO skb can be sent, do it. */
1942 if (limit >= max_segs * tp->mss_cache)
1943 goto send_now;
1944
1945 /* Middle in queue won't get any more data, full sendable already? */
1946 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1947 goto send_now;
1948
1949 win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor);
1950 if (win_divisor) {
1951 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1952
1953 /* If at least some fraction of a window is available,
1954 * just use it.
1955 */
1956 chunk /= win_divisor;
1957 if (limit >= chunk)
1958 goto send_now;
1959 } else {
1960 /* Different approach, try not to defer past a single
1961 * ACK. Receiver should ACK every other full sized
1962 * frame, so if we have space for more than 3 frames
1963 * then send now.
1964 */
1965 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1966 goto send_now;
1967 }
1968
1969 /* TODO : use tsorted_sent_queue ? */
1970 head = tcp_rtx_queue_head(sk);
1971 if (!head)
1972 goto send_now;
1973 delta = tp->tcp_clock_cache - head->tstamp;
1974 /* If next ACK is likely to come too late (half srtt), do not defer */
1975 if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0)
1976 goto send_now;
1977
1978 /* Ok, it looks like it is advisable to defer.
1979 * Three cases are tracked :
1980 * 1) We are cwnd-limited
1981 * 2) We are rwnd-limited
1982 * 3) We are application limited.
1983 */
1984 if (cong_win < send_win) {
1985 if (cong_win <= skb->len) {
1986 *is_cwnd_limited = true;
1987 return true;
1988 }
1989 } else {
1990 if (send_win <= skb->len) {
1991 *is_rwnd_limited = true;
1992 return true;
1993 }
1994 }
1995
1996 /* If this packet won't get more data, do not wait. */
1997 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) ||
1998 TCP_SKB_CB(skb)->eor)
1999 goto send_now;
2000
2001 return true;
2002
2003send_now:
2004 return false;
2005}
2006
2007static inline void tcp_mtu_check_reprobe(struct sock *sk)
2008{
2009 struct inet_connection_sock *icsk = inet_csk(sk);
2010 struct tcp_sock *tp = tcp_sk(sk);
2011 struct net *net = sock_net(sk);
2012 u32 interval;
2013 s32 delta;
2014
2015 interval = net->ipv4.sysctl_tcp_probe_interval;
2016 delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp;
2017 if (unlikely(delta >= interval * HZ)) {
2018 int mss = tcp_current_mss(sk);
2019
2020 /* Update current search range */
2021 icsk->icsk_mtup.probe_size = 0;
2022 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
2023 sizeof(struct tcphdr) +
2024 icsk->icsk_af_ops->net_header_len;
2025 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
2026
2027 /* Update probe time stamp */
2028 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
2029 }
2030}
2031
2032static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len)
2033{
2034 struct sk_buff *skb, *next;
2035
2036 skb = tcp_send_head(sk);
2037 tcp_for_write_queue_from_safe(skb, next, sk) {
2038 if (len <= skb->len)
2039 break;
2040
2041 if (unlikely(TCP_SKB_CB(skb)->eor))
2042 return false;
2043
2044 len -= skb->len;
2045 }
2046
2047 return true;
2048}
2049
2050/* Create a new MTU probe if we are ready.
2051 * MTU probe is regularly attempting to increase the path MTU by
2052 * deliberately sending larger packets. This discovers routing
2053 * changes resulting in larger path MTUs.
2054 *
2055 * Returns 0 if we should wait to probe (no cwnd available),
2056 * 1 if a probe was sent,
2057 * -1 otherwise
2058 */
2059static int tcp_mtu_probe(struct sock *sk)
2060{
2061 struct inet_connection_sock *icsk = inet_csk(sk);
2062 struct tcp_sock *tp = tcp_sk(sk);
2063 struct sk_buff *skb, *nskb, *next;
2064 struct net *net = sock_net(sk);
2065 int probe_size;
2066 int size_needed;
2067 int copy, len;
2068 int mss_now;
2069 int interval;
2070
2071 /* Not currently probing/verifying,
2072 * not in recovery,
2073 * have enough cwnd, and
2074 * not SACKing (the variable headers throw things off)
2075 */
2076 if (likely(!icsk->icsk_mtup.enabled ||
2077 icsk->icsk_mtup.probe_size ||
2078 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
2079 tp->snd_cwnd < 11 ||
2080 tp->rx_opt.num_sacks || tp->rx_opt.dsack))
2081 return -1;
2082
2083 /* Use binary search for probe_size between tcp_mss_base,
2084 * and current mss_clamp. if (search_high - search_low)
2085 * smaller than a threshold, backoff from probing.
2086 */
2087 mss_now = tcp_current_mss(sk);
2088 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
2089 icsk->icsk_mtup.search_low) >> 1);
2090 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
2091 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
2092 /* When misfortune happens, we are reprobing actively,
2093 * and then reprobe timer has expired. We stick with current
2094 * probing process by not resetting search range to its orignal.
2095 */
2096 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
2097 interval < net->ipv4.sysctl_tcp_probe_threshold) {
2098 /* Check whether enough time has elaplased for
2099 * another round of probing.
2100 */
2101 tcp_mtu_check_reprobe(sk);
2102 return -1;
2103 }
2104
2105 /* Have enough data in the send queue to probe? */
2106 if (tp->write_seq - tp->snd_nxt < size_needed)
2107 return -1;
2108
2109 if (tp->snd_wnd < size_needed)
2110 return -1;
2111 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
2112 return 0;
2113
2114 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2115 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
2116 if (!tcp_packets_in_flight(tp))
2117 return -1;
2118 else
2119 return 0;
2120 }
2121
2122 if (!tcp_can_coalesce_send_queue_head(sk, probe_size))
2123 return -1;
2124
2125 /* We're allowed to probe. Build it now. */
2126 nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
2127 if (!nskb)
2128 return -1;
2129 sk->sk_wmem_queued += nskb->truesize;
2130 sk_mem_charge(sk, nskb->truesize);
2131
2132 skb = tcp_send_head(sk);
2133
2134 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2135 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2136 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2137 TCP_SKB_CB(nskb)->sacked = 0;
2138 nskb->csum = 0;
2139 nskb->ip_summed = CHECKSUM_PARTIAL;
2140
2141 tcp_insert_write_queue_before(nskb, skb, sk);
2142 tcp_highest_sack_replace(sk, skb, nskb);
2143
2144 len = 0;
2145 tcp_for_write_queue_from_safe(skb, next, sk) {
2146 copy = min_t(int, skb->len, probe_size - len);
2147 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
2148
2149 if (skb->len <= copy) {
2150 /* We've eaten all the data from this skb.
2151 * Throw it away. */
2152 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2153 /* If this is the last SKB we copy and eor is set
2154 * we need to propagate it to the new skb.
2155 */
2156 TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor;
2157 tcp_unlink_write_queue(skb, sk);
2158 sk_wmem_free_skb(sk, skb);
2159 } else {
2160 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2161 ~(TCPHDR_FIN|TCPHDR_PSH);
2162 if (!skb_shinfo(skb)->nr_frags) {
2163 skb_pull(skb, copy);
2164 } else {
2165 __pskb_trim_head(skb, copy);
2166 tcp_set_skb_tso_segs(skb, mss_now);
2167 }
2168 TCP_SKB_CB(skb)->seq += copy;
2169 }
2170
2171 len += copy;
2172
2173 if (len >= probe_size)
2174 break;
2175 }
2176 tcp_init_tso_segs(nskb, nskb->len);
2177
2178 /* We're ready to send. If this fails, the probe will
2179 * be resegmented into mss-sized pieces by tcp_write_xmit().
2180 */
2181 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2182 /* Decrement cwnd here because we are sending
2183 * effectively two packets. */
2184 tp->snd_cwnd--;
2185 tcp_event_new_data_sent(sk, nskb);
2186
2187 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2188 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2189 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2190
2191 return 1;
2192 }
2193
2194 return -1;
2195}
2196
2197static bool tcp_pacing_check(struct sock *sk)
2198{
2199 struct tcp_sock *tp = tcp_sk(sk);
2200
2201 if (!tcp_needs_internal_pacing(sk))
2202 return false;
2203
2204 if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache)
2205 return false;
2206
2207 if (!hrtimer_is_queued(&tp->pacing_timer)) {
2208 hrtimer_start(&tp->pacing_timer,
2209 ns_to_ktime(tp->tcp_wstamp_ns),
2210 HRTIMER_MODE_ABS_PINNED_SOFT);
2211 sock_hold(sk);
2212 }
2213 return true;
2214}
2215
2216/* TCP Small Queues :
2217 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2218 * (These limits are doubled for retransmits)
2219 * This allows for :
2220 * - better RTT estimation and ACK scheduling
2221 * - faster recovery
2222 * - high rates
2223 * Alas, some drivers / subsystems require a fair amount
2224 * of queued bytes to ensure line rate.
2225 * One example is wifi aggregation (802.11 AMPDU)
2226 */
2227static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2228 unsigned int factor)
2229{
2230 unsigned long limit;
2231
2232 limit = max_t(unsigned long,
2233 2 * skb->truesize,
2234 sk->sk_pacing_rate >> sk->sk_pacing_shift);
2235 if (sk->sk_pacing_status == SK_PACING_NONE)
2236 limit = min_t(unsigned long, limit,
2237 sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes);
2238 limit <<= factor;
2239
2240 if (refcount_read(&sk->sk_wmem_alloc) > limit) {
2241 /* Always send skb if rtx queue is empty.
2242 * No need to wait for TX completion to call us back,
2243 * after softirq/tasklet schedule.
2244 * This helps when TX completions are delayed too much.
2245 */
2246 if (tcp_rtx_queue_empty(sk))
2247 return false;
2248
2249 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2250 /* It is possible TX completion already happened
2251 * before we set TSQ_THROTTLED, so we must
2252 * test again the condition.
2253 */
2254 smp_mb__after_atomic();
2255 if (refcount_read(&sk->sk_wmem_alloc) > limit)
2256 return true;
2257 }
2258 return false;
2259}
2260
2261static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2262{
2263 const u32 now = tcp_jiffies32;
2264 enum tcp_chrono old = tp->chrono_type;
2265
2266 if (old > TCP_CHRONO_UNSPEC)
2267 tp->chrono_stat[old - 1] += now - tp->chrono_start;
2268 tp->chrono_start = now;
2269 tp->chrono_type = new;
2270}
2271
2272void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2273{
2274 struct tcp_sock *tp = tcp_sk(sk);
2275
2276 /* If there are multiple conditions worthy of tracking in a
2277 * chronograph then the highest priority enum takes precedence
2278 * over the other conditions. So that if something "more interesting"
2279 * starts happening, stop the previous chrono and start a new one.
2280 */
2281 if (type > tp->chrono_type)
2282 tcp_chrono_set(tp, type);
2283}
2284
2285void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2286{
2287 struct tcp_sock *tp = tcp_sk(sk);
2288
2289
2290 /* There are multiple conditions worthy of tracking in a
2291 * chronograph, so that the highest priority enum takes
2292 * precedence over the other conditions (see tcp_chrono_start).
2293 * If a condition stops, we only stop chrono tracking if
2294 * it's the "most interesting" or current chrono we are
2295 * tracking and starts busy chrono if we have pending data.
2296 */
2297 if (tcp_rtx_and_write_queues_empty(sk))
2298 tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
2299 else if (type == tp->chrono_type)
2300 tcp_chrono_set(tp, TCP_CHRONO_BUSY);
2301}
2302
2303/* This routine writes packets to the network. It advances the
2304 * send_head. This happens as incoming acks open up the remote
2305 * window for us.
2306 *
2307 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2308 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2309 * account rare use of URG, this is not a big flaw.
2310 *
2311 * Send at most one packet when push_one > 0. Temporarily ignore
2312 * cwnd limit to force at most one packet out when push_one == 2.
2313
2314 * Returns true, if no segments are in flight and we have queued segments,
2315 * but cannot send anything now because of SWS or another problem.
2316 */
2317static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2318 int push_one, gfp_t gfp)
2319{
2320 struct tcp_sock *tp = tcp_sk(sk);
2321 struct sk_buff *skb;
2322 unsigned int tso_segs, sent_pkts;
2323 int cwnd_quota;
2324 int result;
2325 bool is_cwnd_limited = false, is_rwnd_limited = false;
2326 u32 max_segs;
2327
2328 sent_pkts = 0;
2329
2330 tcp_mstamp_refresh(tp);
2331 if (!push_one) {
2332 /* Do MTU probing. */
2333 result = tcp_mtu_probe(sk);
2334 if (!result) {
2335 return false;
2336 } else if (result > 0) {
2337 sent_pkts = 1;
2338 }
2339 }
2340
2341 max_segs = tcp_tso_segs(sk, mss_now);
2342 while ((skb = tcp_send_head(sk))) {
2343 unsigned int limit;
2344
2345 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2346 /* "skb_mstamp_ns" is used as a start point for the retransmit timer */
2347 skb->skb_mstamp_ns = tp->tcp_wstamp_ns = tp->tcp_clock_cache;
2348 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
2349 tcp_init_tso_segs(skb, mss_now);
2350 goto repair; /* Skip network transmission */
2351 }
2352
2353 if (tcp_pacing_check(sk))
2354 break;
2355
2356 tso_segs = tcp_init_tso_segs(skb, mss_now);
2357 BUG_ON(!tso_segs);
2358
2359 cwnd_quota = tcp_cwnd_test(tp, skb);
2360 if (!cwnd_quota) {
2361 if (push_one == 2)
2362 /* Force out a loss probe pkt. */
2363 cwnd_quota = 1;
2364 else
2365 break;
2366 }
2367
2368 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2369 is_rwnd_limited = true;
2370 break;
2371 }
2372
2373 if (tso_segs == 1) {
2374 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2375 (tcp_skb_is_last(sk, skb) ?
2376 nonagle : TCP_NAGLE_PUSH))))
2377 break;
2378 } else {
2379 if (!push_one &&
2380 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2381 &is_rwnd_limited, max_segs))
2382 break;
2383 }
2384
2385 limit = mss_now;
2386 if (tso_segs > 1 && !tcp_urg_mode(tp))
2387 limit = tcp_mss_split_point(sk, skb, mss_now,
2388 min_t(unsigned int,
2389 cwnd_quota,
2390 max_segs),
2391 nonagle);
2392
2393 if (skb->len > limit &&
2394 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2395 break;
2396
2397 if (tcp_small_queue_check(sk, skb, 0))
2398 break;
2399
2400 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2401 break;
2402
2403repair:
2404 /* Advance the send_head. This one is sent out.
2405 * This call will increment packets_out.
2406 */
2407 tcp_event_new_data_sent(sk, skb);
2408
2409 tcp_minshall_update(tp, mss_now, skb);
2410 sent_pkts += tcp_skb_pcount(skb);
2411
2412 if (push_one)
2413 break;
2414 }
2415
2416 if (is_rwnd_limited)
2417 tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
2418 else
2419 tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
2420
2421 if (likely(sent_pkts)) {
2422 if (tcp_in_cwnd_reduction(sk))
2423 tp->prr_out += sent_pkts;
2424
2425 /* Send one loss probe per tail loss episode. */
2426 if (push_one != 2)
2427 tcp_schedule_loss_probe(sk, false);
2428 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
2429 tcp_cwnd_validate(sk, is_cwnd_limited);
2430 return false;
2431 }
2432 return !tp->packets_out && !tcp_write_queue_empty(sk);
2433}
2434
2435bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto)
2436{
2437 struct inet_connection_sock *icsk = inet_csk(sk);
2438 struct tcp_sock *tp = tcp_sk(sk);
2439 u32 timeout, rto_delta_us;
2440 int early_retrans;
2441
2442 /* Don't do any loss probe on a Fast Open connection before 3WHS
2443 * finishes.
2444 */
2445 if (tp->fastopen_rsk)
2446 return false;
2447
2448 early_retrans = sock_net(sk)->ipv4.sysctl_tcp_early_retrans;
2449 /* Schedule a loss probe in 2*RTT for SACK capable connections
2450 * not in loss recovery, that are either limited by cwnd or application.
2451 */
2452 if ((early_retrans != 3 && early_retrans != 4) ||
2453 !tp->packets_out || !tcp_is_sack(tp) ||
2454 (icsk->icsk_ca_state != TCP_CA_Open &&
2455 icsk->icsk_ca_state != TCP_CA_CWR))
2456 return false;
2457
2458 /* Probe timeout is 2*rtt. Add minimum RTO to account
2459 * for delayed ack when there's one outstanding packet. If no RTT
2460 * sample is available then probe after TCP_TIMEOUT_INIT.
2461 */
2462 if (tp->srtt_us) {
2463 timeout = usecs_to_jiffies(tp->srtt_us >> 2);
2464 if (tp->packets_out == 1)
2465 timeout += TCP_RTO_MIN;
2466 else
2467 timeout += TCP_TIMEOUT_MIN;
2468 } else {
2469 timeout = TCP_TIMEOUT_INIT;
2470 }
2471
2472 /* If the RTO formula yields an earlier time, then use that time. */
2473 rto_delta_us = advancing_rto ?
2474 jiffies_to_usecs(inet_csk(sk)->icsk_rto) :
2475 tcp_rto_delta_us(sk); /* How far in future is RTO? */
2476 if (rto_delta_us > 0)
2477 timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us));
2478
2479 tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2480 TCP_RTO_MAX, NULL);
2481 return true;
2482}
2483
2484/* Thanks to skb fast clones, we can detect if a prior transmit of
2485 * a packet is still in a qdisc or driver queue.
2486 * In this case, there is very little point doing a retransmit !
2487 */
2488static bool skb_still_in_host_queue(const struct sock *sk,
2489 const struct sk_buff *skb)
2490{
2491 if (unlikely(skb_fclone_busy(sk, skb))) {
2492 NET_INC_STATS(sock_net(sk),
2493 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2494 return true;
2495 }
2496 return false;
2497}
2498
2499/* When probe timeout (PTO) fires, try send a new segment if possible, else
2500 * retransmit the last segment.
2501 */
2502void tcp_send_loss_probe(struct sock *sk)
2503{
2504 struct tcp_sock *tp = tcp_sk(sk);
2505 struct sk_buff *skb;
2506 int pcount;
2507 int mss = tcp_current_mss(sk);
2508
2509 skb = tcp_send_head(sk);
2510 if (skb && tcp_snd_wnd_test(tp, skb, mss)) {
2511 pcount = tp->packets_out;
2512 tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2513 if (tp->packets_out > pcount)
2514 goto probe_sent;
2515 goto rearm_timer;
2516 }
2517 skb = skb_rb_last(&sk->tcp_rtx_queue);
2518 if (unlikely(!skb)) {
2519 WARN_ONCE(tp->packets_out,
2520 "invalid inflight: %u state %u cwnd %u mss %d\n",
2521 tp->packets_out, sk->sk_state, tp->snd_cwnd, mss);
2522 inet_csk(sk)->icsk_pending = 0;
2523 return;
2524 }
2525
2526 /* At most one outstanding TLP retransmission. */
2527 if (tp->tlp_high_seq)
2528 goto rearm_timer;
2529
2530 if (skb_still_in_host_queue(sk, skb))
2531 goto rearm_timer;
2532
2533 pcount = tcp_skb_pcount(skb);
2534 if (WARN_ON(!pcount))
2535 goto rearm_timer;
2536
2537 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2538 if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2539 (pcount - 1) * mss, mss,
2540 GFP_ATOMIC)))
2541 goto rearm_timer;
2542 skb = skb_rb_next(skb);
2543 }
2544
2545 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2546 goto rearm_timer;
2547
2548 if (__tcp_retransmit_skb(sk, skb, 1))
2549 goto rearm_timer;
2550
2551 /* Record snd_nxt for loss detection. */
2552 tp->tlp_high_seq = tp->snd_nxt;
2553
2554probe_sent:
2555 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2556 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2557 inet_csk(sk)->icsk_pending = 0;
2558rearm_timer:
2559 tcp_rearm_rto(sk);
2560}
2561
2562/* Push out any pending frames which were held back due to
2563 * TCP_CORK or attempt at coalescing tiny packets.
2564 * The socket must be locked by the caller.
2565 */
2566void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2567 int nonagle)
2568{
2569 /* If we are closed, the bytes will have to remain here.
2570 * In time closedown will finish, we empty the write queue and
2571 * all will be happy.
2572 */
2573 if (unlikely(sk->sk_state == TCP_CLOSE))
2574 return;
2575
2576 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2577 sk_gfp_mask(sk, GFP_ATOMIC)))
2578 tcp_check_probe_timer(sk);
2579}
2580
2581/* Send _single_ skb sitting at the send head. This function requires
2582 * true push pending frames to setup probe timer etc.
2583 */
2584void tcp_push_one(struct sock *sk, unsigned int mss_now)
2585{
2586 struct sk_buff *skb = tcp_send_head(sk);
2587
2588 BUG_ON(!skb || skb->len < mss_now);
2589
2590 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2591}
2592
2593/* This function returns the amount that we can raise the
2594 * usable window based on the following constraints
2595 *
2596 * 1. The window can never be shrunk once it is offered (RFC 793)
2597 * 2. We limit memory per socket
2598 *
2599 * RFC 1122:
2600 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2601 * RECV.NEXT + RCV.WIN fixed until:
2602 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2603 *
2604 * i.e. don't raise the right edge of the window until you can raise
2605 * it at least MSS bytes.
2606 *
2607 * Unfortunately, the recommended algorithm breaks header prediction,
2608 * since header prediction assumes th->window stays fixed.
2609 *
2610 * Strictly speaking, keeping th->window fixed violates the receiver
2611 * side SWS prevention criteria. The problem is that under this rule
2612 * a stream of single byte packets will cause the right side of the
2613 * window to always advance by a single byte.
2614 *
2615 * Of course, if the sender implements sender side SWS prevention
2616 * then this will not be a problem.
2617 *
2618 * BSD seems to make the following compromise:
2619 *
2620 * If the free space is less than the 1/4 of the maximum
2621 * space available and the free space is less than 1/2 mss,
2622 * then set the window to 0.
2623 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2624 * Otherwise, just prevent the window from shrinking
2625 * and from being larger than the largest representable value.
2626 *
2627 * This prevents incremental opening of the window in the regime
2628 * where TCP is limited by the speed of the reader side taking
2629 * data out of the TCP receive queue. It does nothing about
2630 * those cases where the window is constrained on the sender side
2631 * because the pipeline is full.
2632 *
2633 * BSD also seems to "accidentally" limit itself to windows that are a
2634 * multiple of MSS, at least until the free space gets quite small.
2635 * This would appear to be a side effect of the mbuf implementation.
2636 * Combining these two algorithms results in the observed behavior
2637 * of having a fixed window size at almost all times.
2638 *
2639 * Below we obtain similar behavior by forcing the offered window to
2640 * a multiple of the mss when it is feasible to do so.
2641 *
2642 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2643 * Regular options like TIMESTAMP are taken into account.
2644 */
2645u32 __tcp_select_window(struct sock *sk)
2646{
2647 struct inet_connection_sock *icsk = inet_csk(sk);
2648 struct tcp_sock *tp = tcp_sk(sk);
2649 /* MSS for the peer's data. Previous versions used mss_clamp
2650 * here. I don't know if the value based on our guesses
2651 * of peer's MSS is better for the performance. It's more correct
2652 * but may be worse for the performance because of rcv_mss
2653 * fluctuations. --SAW 1998/11/1
2654 */
2655 int mss = icsk->icsk_ack.rcv_mss;
2656 int free_space = tcp_space(sk);
2657 int allowed_space = tcp_full_space(sk);
2658 int full_space = min_t(int, tp->window_clamp, allowed_space);
2659 int window;
2660
2661 if (unlikely(mss > full_space)) {
2662 mss = full_space;
2663 if (mss <= 0)
2664 return 0;
2665 }
2666 if (free_space < (full_space >> 1)) {
2667 icsk->icsk_ack.quick = 0;
2668
2669 if (tcp_under_memory_pressure(sk))
2670 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2671 4U * tp->advmss);
2672
2673 /* free_space might become our new window, make sure we don't
2674 * increase it due to wscale.
2675 */
2676 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2677
2678 /* if free space is less than mss estimate, or is below 1/16th
2679 * of the maximum allowed, try to move to zero-window, else
2680 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2681 * new incoming data is dropped due to memory limits.
2682 * With large window, mss test triggers way too late in order
2683 * to announce zero window in time before rmem limit kicks in.
2684 */
2685 if (free_space < (allowed_space >> 4) || free_space < mss)
2686 return 0;
2687 }
2688
2689 if (free_space > tp->rcv_ssthresh)
2690 free_space = tp->rcv_ssthresh;
2691
2692 /* Don't do rounding if we are using window scaling, since the
2693 * scaled window will not line up with the MSS boundary anyway.
2694 */
2695 if (tp->rx_opt.rcv_wscale) {
2696 window = free_space;
2697
2698 /* Advertise enough space so that it won't get scaled away.
2699 * Import case: prevent zero window announcement if
2700 * 1<<rcv_wscale > mss.
2701 */
2702 window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
2703 } else {
2704 window = tp->rcv_wnd;
2705 /* Get the largest window that is a nice multiple of mss.
2706 * Window clamp already applied above.
2707 * If our current window offering is within 1 mss of the
2708 * free space we just keep it. This prevents the divide
2709 * and multiply from happening most of the time.
2710 * We also don't do any window rounding when the free space
2711 * is too small.
2712 */
2713 if (window <= free_space - mss || window > free_space)
2714 window = rounddown(free_space, mss);
2715 else if (mss == full_space &&
2716 free_space > window + (full_space >> 1))
2717 window = free_space;
2718 }
2719
2720 return window;
2721}
2722
2723void tcp_skb_collapse_tstamp(struct sk_buff *skb,
2724 const struct sk_buff *next_skb)
2725{
2726 if (unlikely(tcp_has_tx_tstamp(next_skb))) {
2727 const struct skb_shared_info *next_shinfo =
2728 skb_shinfo(next_skb);
2729 struct skb_shared_info *shinfo = skb_shinfo(skb);
2730
2731 shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
2732 shinfo->tskey = next_shinfo->tskey;
2733 TCP_SKB_CB(skb)->txstamp_ack |=
2734 TCP_SKB_CB(next_skb)->txstamp_ack;
2735 }
2736}
2737
2738/* Collapses two adjacent SKB's during retransmission. */
2739static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2740{
2741 struct tcp_sock *tp = tcp_sk(sk);
2742 struct sk_buff *next_skb = skb_rb_next(skb);
2743 int next_skb_size;
2744
2745 next_skb_size = next_skb->len;
2746
2747 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2748
2749 if (next_skb_size) {
2750 if (next_skb_size <= skb_availroom(skb))
2751 skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size),
2752 next_skb_size);
2753 else if (!skb_shift(skb, next_skb, next_skb_size))
2754 return false;
2755 }
2756 tcp_highest_sack_replace(sk, next_skb, skb);
2757
2758 /* Update sequence range on original skb. */
2759 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2760
2761 /* Merge over control information. This moves PSH/FIN etc. over */
2762 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2763
2764 /* All done, get rid of second SKB and account for it so
2765 * packet counting does not break.
2766 */
2767 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2768 TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
2769
2770 /* changed transmit queue under us so clear hints */
2771 tcp_clear_retrans_hints_partial(tp);
2772 if (next_skb == tp->retransmit_skb_hint)
2773 tp->retransmit_skb_hint = skb;
2774
2775 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2776
2777 tcp_skb_collapse_tstamp(skb, next_skb);
2778
2779 tcp_rtx_queue_unlink_and_free(next_skb, sk);
2780 return true;
2781}
2782
2783/* Check if coalescing SKBs is legal. */
2784static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2785{
2786 if (tcp_skb_pcount(skb) > 1)
2787 return false;
2788 if (skb_cloned(skb))
2789 return false;
2790 /* Some heuristics for collapsing over SACK'd could be invented */
2791 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2792 return false;
2793
2794 return true;
2795}
2796
2797/* Collapse packets in the retransmit queue to make to create
2798 * less packets on the wire. This is only done on retransmission.
2799 */
2800static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2801 int space)
2802{
2803 struct tcp_sock *tp = tcp_sk(sk);
2804 struct sk_buff *skb = to, *tmp;
2805 bool first = true;
2806
2807 if (!sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse)
2808 return;
2809 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2810 return;
2811
2812 skb_rbtree_walk_from_safe(skb, tmp) {
2813 if (!tcp_can_collapse(sk, skb))
2814 break;
2815
2816 if (!tcp_skb_can_collapse_to(to))
2817 break;
2818
2819 space -= skb->len;
2820
2821 if (first) {
2822 first = false;
2823 continue;
2824 }
2825
2826 if (space < 0)
2827 break;
2828
2829 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2830 break;
2831
2832 if (!tcp_collapse_retrans(sk, to))
2833 break;
2834 }
2835}
2836
2837/* This retransmits one SKB. Policy decisions and retransmit queue
2838 * state updates are done by the caller. Returns non-zero if an
2839 * error occurred which prevented the send.
2840 */
2841int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2842{
2843 struct inet_connection_sock *icsk = inet_csk(sk);
2844 struct tcp_sock *tp = tcp_sk(sk);
2845 unsigned int cur_mss;
2846 int diff, len, err;
2847
2848
2849 /* Inconclusive MTU probe */
2850 if (icsk->icsk_mtup.probe_size)
2851 icsk->icsk_mtup.probe_size = 0;
2852
2853 /* Do not sent more than we queued. 1/4 is reserved for possible
2854 * copying overhead: fragmentation, t