Loading...
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:
23 * Pedro Roque : Fast Retransmit/Recovery.
24 * Two receive queues.
25 * Retransmit queue handled by TCP.
26 * Better retransmit timer handling.
27 * New congestion avoidance.
28 * Header prediction.
29 * Variable renaming.
30 *
31 * Eric : Fast Retransmit.
32 * Randy Scott : MSS option defines.
33 * Eric Schenk : Fixes to slow start algorithm.
34 * Eric Schenk : Yet another double ACK bug.
35 * Eric Schenk : Delayed ACK bug fixes.
36 * Eric Schenk : Floyd style fast retrans war avoidance.
37 * David S. Miller : Don't allow zero congestion window.
38 * Eric Schenk : Fix retransmitter so that it sends
39 * next packet on ack of previous packet.
40 * Andi Kleen : Moved open_request checking here
41 * and process RSTs for open_requests.
42 * Andi Kleen : Better prune_queue, and other fixes.
43 * Andrey Savochkin: Fix RTT measurements in the presence of
44 * timestamps.
45 * Andrey Savochkin: Check sequence numbers correctly when
46 * removing SACKs due to in sequence incoming
47 * data segments.
48 * Andi Kleen: Make sure we never ack data there is not
49 * enough room for. Also make this condition
50 * a fatal error if it might still happen.
51 * Andi Kleen: Add tcp_measure_rcv_mss to make
52 * connections with MSS<min(MTU,ann. MSS)
53 * work without delayed acks.
54 * Andi Kleen: Process packets with PSH set in the
55 * fast path.
56 * J Hadi Salim: ECN support
57 * Andrei Gurtov,
58 * Pasi Sarolahti,
59 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
60 * engine. Lots of bugs are found.
61 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
62 */
63
64#define pr_fmt(fmt) "TCP: " fmt
65
66#include <linux/mm.h>
67#include <linux/slab.h>
68#include <linux/module.h>
69#include <linux/sysctl.h>
70#include <linux/kernel.h>
71#include <net/dst.h>
72#include <net/tcp.h>
73#include <net/inet_common.h>
74#include <linux/ipsec.h>
75#include <asm/unaligned.h>
76#include <net/netdma.h>
77
78int sysctl_tcp_timestamps __read_mostly = 1;
79int sysctl_tcp_window_scaling __read_mostly = 1;
80int sysctl_tcp_sack __read_mostly = 1;
81int sysctl_tcp_fack __read_mostly = 1;
82int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
83EXPORT_SYMBOL(sysctl_tcp_reordering);
84int sysctl_tcp_dsack __read_mostly = 1;
85int sysctl_tcp_app_win __read_mostly = 31;
86int sysctl_tcp_adv_win_scale __read_mostly = 1;
87EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
88
89/* rfc5961 challenge ack rate limiting */
90int sysctl_tcp_challenge_ack_limit = 100;
91
92int sysctl_tcp_stdurg __read_mostly;
93int sysctl_tcp_rfc1337 __read_mostly;
94int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
95int sysctl_tcp_frto __read_mostly = 2;
96
97int sysctl_tcp_thin_dupack __read_mostly;
98
99int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
100int sysctl_tcp_early_retrans __read_mostly = 3;
101
102#define FLAG_DATA 0x01 /* Incoming frame contained data. */
103#define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
104#define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
105#define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
106#define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
107#define FLAG_DATA_SACKED 0x20 /* New SACK. */
108#define FLAG_ECE 0x40 /* ECE in this ACK */
109#define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
110#define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
111#define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
112#define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
113#define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
114#define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
115
116#define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
117#define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
118#define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
119#define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
120
121#define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
122#define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
123
124/* Adapt the MSS value used to make delayed ack decision to the
125 * real world.
126 */
127static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
128{
129 struct inet_connection_sock *icsk = inet_csk(sk);
130 const unsigned int lss = icsk->icsk_ack.last_seg_size;
131 unsigned int len;
132
133 icsk->icsk_ack.last_seg_size = 0;
134
135 /* skb->len may jitter because of SACKs, even if peer
136 * sends good full-sized frames.
137 */
138 len = skb_shinfo(skb)->gso_size ? : skb->len;
139 if (len >= icsk->icsk_ack.rcv_mss) {
140 icsk->icsk_ack.rcv_mss = len;
141 } else {
142 /* Otherwise, we make more careful check taking into account,
143 * that SACKs block is variable.
144 *
145 * "len" is invariant segment length, including TCP header.
146 */
147 len += skb->data - skb_transport_header(skb);
148 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
149 /* If PSH is not set, packet should be
150 * full sized, provided peer TCP is not badly broken.
151 * This observation (if it is correct 8)) allows
152 * to handle super-low mtu links fairly.
153 */
154 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
155 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
156 /* Subtract also invariant (if peer is RFC compliant),
157 * tcp header plus fixed timestamp option length.
158 * Resulting "len" is MSS free of SACK jitter.
159 */
160 len -= tcp_sk(sk)->tcp_header_len;
161 icsk->icsk_ack.last_seg_size = len;
162 if (len == lss) {
163 icsk->icsk_ack.rcv_mss = len;
164 return;
165 }
166 }
167 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
168 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
169 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
170 }
171}
172
173static void tcp_incr_quickack(struct sock *sk)
174{
175 struct inet_connection_sock *icsk = inet_csk(sk);
176 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
177
178 if (quickacks == 0)
179 quickacks = 2;
180 if (quickacks > icsk->icsk_ack.quick)
181 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
182}
183
184static void tcp_enter_quickack_mode(struct sock *sk)
185{
186 struct inet_connection_sock *icsk = inet_csk(sk);
187 tcp_incr_quickack(sk);
188 icsk->icsk_ack.pingpong = 0;
189 icsk->icsk_ack.ato = TCP_ATO_MIN;
190}
191
192/* Send ACKs quickly, if "quick" count is not exhausted
193 * and the session is not interactive.
194 */
195
196static inline bool tcp_in_quickack_mode(const struct sock *sk)
197{
198 const struct inet_connection_sock *icsk = inet_csk(sk);
199
200 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
201}
202
203static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
204{
205 if (tp->ecn_flags & TCP_ECN_OK)
206 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
207}
208
209static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
210{
211 if (tcp_hdr(skb)->cwr)
212 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
213}
214
215static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
216{
217 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
218}
219
220static inline void TCP_ECN_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
221{
222 if (!(tp->ecn_flags & TCP_ECN_OK))
223 return;
224
225 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
226 case INET_ECN_NOT_ECT:
227 /* Funny extension: if ECT is not set on a segment,
228 * and we already seen ECT on a previous segment,
229 * it is probably a retransmit.
230 */
231 if (tp->ecn_flags & TCP_ECN_SEEN)
232 tcp_enter_quickack_mode((struct sock *)tp);
233 break;
234 case INET_ECN_CE:
235 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
236 /* Better not delay acks, sender can have a very low cwnd */
237 tcp_enter_quickack_mode((struct sock *)tp);
238 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
239 }
240 /* fallinto */
241 default:
242 tp->ecn_flags |= TCP_ECN_SEEN;
243 }
244}
245
246static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
247{
248 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
249 tp->ecn_flags &= ~TCP_ECN_OK;
250}
251
252static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
253{
254 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
255 tp->ecn_flags &= ~TCP_ECN_OK;
256}
257
258static bool TCP_ECN_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
259{
260 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
261 return true;
262 return false;
263}
264
265/* Buffer size and advertised window tuning.
266 *
267 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
268 */
269
270static void tcp_sndbuf_expand(struct sock *sk)
271{
272 const struct tcp_sock *tp = tcp_sk(sk);
273 int sndmem, per_mss;
274 u32 nr_segs;
275
276 /* Worst case is non GSO/TSO : each frame consumes one skb
277 * and skb->head is kmalloced using power of two area of memory
278 */
279 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
280 MAX_TCP_HEADER +
281 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
282
283 per_mss = roundup_pow_of_two(per_mss) +
284 SKB_DATA_ALIGN(sizeof(struct sk_buff));
285
286 nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
287 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
288
289 /* Fast Recovery (RFC 5681 3.2) :
290 * Cubic needs 1.7 factor, rounded to 2 to include
291 * extra cushion (application might react slowly to POLLOUT)
292 */
293 sndmem = 2 * nr_segs * per_mss;
294
295 if (sk->sk_sndbuf < sndmem)
296 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
297}
298
299/* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
300 *
301 * All tcp_full_space() is split to two parts: "network" buffer, allocated
302 * forward and advertised in receiver window (tp->rcv_wnd) and
303 * "application buffer", required to isolate scheduling/application
304 * latencies from network.
305 * window_clamp is maximal advertised window. It can be less than
306 * tcp_full_space(), in this case tcp_full_space() - window_clamp
307 * is reserved for "application" buffer. The less window_clamp is
308 * the smoother our behaviour from viewpoint of network, but the lower
309 * throughput and the higher sensitivity of the connection to losses. 8)
310 *
311 * rcv_ssthresh is more strict window_clamp used at "slow start"
312 * phase to predict further behaviour of this connection.
313 * It is used for two goals:
314 * - to enforce header prediction at sender, even when application
315 * requires some significant "application buffer". It is check #1.
316 * - to prevent pruning of receive queue because of misprediction
317 * of receiver window. Check #2.
318 *
319 * The scheme does not work when sender sends good segments opening
320 * window and then starts to feed us spaghetti. But it should work
321 * in common situations. Otherwise, we have to rely on queue collapsing.
322 */
323
324/* Slow part of check#2. */
325static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
326{
327 struct tcp_sock *tp = tcp_sk(sk);
328 /* Optimize this! */
329 int truesize = tcp_win_from_space(skb->truesize) >> 1;
330 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
331
332 while (tp->rcv_ssthresh <= window) {
333 if (truesize <= skb->len)
334 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
335
336 truesize >>= 1;
337 window >>= 1;
338 }
339 return 0;
340}
341
342static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
343{
344 struct tcp_sock *tp = tcp_sk(sk);
345
346 /* Check #1 */
347 if (tp->rcv_ssthresh < tp->window_clamp &&
348 (int)tp->rcv_ssthresh < tcp_space(sk) &&
349 !sk_under_memory_pressure(sk)) {
350 int incr;
351
352 /* Check #2. Increase window, if skb with such overhead
353 * will fit to rcvbuf in future.
354 */
355 if (tcp_win_from_space(skb->truesize) <= skb->len)
356 incr = 2 * tp->advmss;
357 else
358 incr = __tcp_grow_window(sk, skb);
359
360 if (incr) {
361 incr = max_t(int, incr, 2 * skb->len);
362 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
363 tp->window_clamp);
364 inet_csk(sk)->icsk_ack.quick |= 1;
365 }
366 }
367}
368
369/* 3. Tuning rcvbuf, when connection enters established state. */
370static void tcp_fixup_rcvbuf(struct sock *sk)
371{
372 u32 mss = tcp_sk(sk)->advmss;
373 int rcvmem;
374
375 rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) *
376 tcp_default_init_rwnd(mss);
377
378 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
379 * Allow enough cushion so that sender is not limited by our window
380 */
381 if (sysctl_tcp_moderate_rcvbuf)
382 rcvmem <<= 2;
383
384 if (sk->sk_rcvbuf < rcvmem)
385 sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
386}
387
388/* 4. Try to fixup all. It is made immediately after connection enters
389 * established state.
390 */
391void tcp_init_buffer_space(struct sock *sk)
392{
393 struct tcp_sock *tp = tcp_sk(sk);
394 int maxwin;
395
396 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
397 tcp_fixup_rcvbuf(sk);
398 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
399 tcp_sndbuf_expand(sk);
400
401 tp->rcvq_space.space = tp->rcv_wnd;
402 tp->rcvq_space.time = tcp_time_stamp;
403 tp->rcvq_space.seq = tp->copied_seq;
404
405 maxwin = tcp_full_space(sk);
406
407 if (tp->window_clamp >= maxwin) {
408 tp->window_clamp = maxwin;
409
410 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
411 tp->window_clamp = max(maxwin -
412 (maxwin >> sysctl_tcp_app_win),
413 4 * tp->advmss);
414 }
415
416 /* Force reservation of one segment. */
417 if (sysctl_tcp_app_win &&
418 tp->window_clamp > 2 * tp->advmss &&
419 tp->window_clamp + tp->advmss > maxwin)
420 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
421
422 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
423 tp->snd_cwnd_stamp = tcp_time_stamp;
424}
425
426/* 5. Recalculate window clamp after socket hit its memory bounds. */
427static void tcp_clamp_window(struct sock *sk)
428{
429 struct tcp_sock *tp = tcp_sk(sk);
430 struct inet_connection_sock *icsk = inet_csk(sk);
431
432 icsk->icsk_ack.quick = 0;
433
434 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
435 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
436 !sk_under_memory_pressure(sk) &&
437 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
438 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
439 sysctl_tcp_rmem[2]);
440 }
441 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
442 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
443}
444
445/* Initialize RCV_MSS value.
446 * RCV_MSS is an our guess about MSS used by the peer.
447 * We haven't any direct information about the MSS.
448 * It's better to underestimate the RCV_MSS rather than overestimate.
449 * Overestimations make us ACKing less frequently than needed.
450 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
451 */
452void tcp_initialize_rcv_mss(struct sock *sk)
453{
454 const struct tcp_sock *tp = tcp_sk(sk);
455 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
456
457 hint = min(hint, tp->rcv_wnd / 2);
458 hint = min(hint, TCP_MSS_DEFAULT);
459 hint = max(hint, TCP_MIN_MSS);
460
461 inet_csk(sk)->icsk_ack.rcv_mss = hint;
462}
463EXPORT_SYMBOL(tcp_initialize_rcv_mss);
464
465/* Receiver "autotuning" code.
466 *
467 * The algorithm for RTT estimation w/o timestamps is based on
468 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
469 * <http://public.lanl.gov/radiant/pubs.html#DRS>
470 *
471 * More detail on this code can be found at
472 * <http://staff.psc.edu/jheffner/>,
473 * though this reference is out of date. A new paper
474 * is pending.
475 */
476static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
477{
478 u32 new_sample = tp->rcv_rtt_est.rtt;
479 long m = sample;
480
481 if (m == 0)
482 m = 1;
483
484 if (new_sample != 0) {
485 /* If we sample in larger samples in the non-timestamp
486 * case, we could grossly overestimate the RTT especially
487 * with chatty applications or bulk transfer apps which
488 * are stalled on filesystem I/O.
489 *
490 * Also, since we are only going for a minimum in the
491 * non-timestamp case, we do not smooth things out
492 * else with timestamps disabled convergence takes too
493 * long.
494 */
495 if (!win_dep) {
496 m -= (new_sample >> 3);
497 new_sample += m;
498 } else {
499 m <<= 3;
500 if (m < new_sample)
501 new_sample = m;
502 }
503 } else {
504 /* No previous measure. */
505 new_sample = m << 3;
506 }
507
508 if (tp->rcv_rtt_est.rtt != new_sample)
509 tp->rcv_rtt_est.rtt = new_sample;
510}
511
512static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
513{
514 if (tp->rcv_rtt_est.time == 0)
515 goto new_measure;
516 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
517 return;
518 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1);
519
520new_measure:
521 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
522 tp->rcv_rtt_est.time = tcp_time_stamp;
523}
524
525static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
526 const struct sk_buff *skb)
527{
528 struct tcp_sock *tp = tcp_sk(sk);
529 if (tp->rx_opt.rcv_tsecr &&
530 (TCP_SKB_CB(skb)->end_seq -
531 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
532 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
533}
534
535/*
536 * This function should be called every time data is copied to user space.
537 * It calculates the appropriate TCP receive buffer space.
538 */
539void tcp_rcv_space_adjust(struct sock *sk)
540{
541 struct tcp_sock *tp = tcp_sk(sk);
542 int time;
543 int copied;
544
545 time = tcp_time_stamp - tp->rcvq_space.time;
546 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
547 return;
548
549 /* Number of bytes copied to user in last RTT */
550 copied = tp->copied_seq - tp->rcvq_space.seq;
551 if (copied <= tp->rcvq_space.space)
552 goto new_measure;
553
554 /* A bit of theory :
555 * copied = bytes received in previous RTT, our base window
556 * To cope with packet losses, we need a 2x factor
557 * To cope with slow start, and sender growing its cwin by 100 %
558 * every RTT, we need a 4x factor, because the ACK we are sending
559 * now is for the next RTT, not the current one :
560 * <prev RTT . ><current RTT .. ><next RTT .... >
561 */
562
563 if (sysctl_tcp_moderate_rcvbuf &&
564 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
565 int rcvwin, rcvmem, rcvbuf;
566
567 /* minimal window to cope with packet losses, assuming
568 * steady state. Add some cushion because of small variations.
569 */
570 rcvwin = (copied << 1) + 16 * tp->advmss;
571
572 /* If rate increased by 25%,
573 * assume slow start, rcvwin = 3 * copied
574 * If rate increased by 50%,
575 * assume sender can use 2x growth, rcvwin = 4 * copied
576 */
577 if (copied >=
578 tp->rcvq_space.space + (tp->rcvq_space.space >> 2)) {
579 if (copied >=
580 tp->rcvq_space.space + (tp->rcvq_space.space >> 1))
581 rcvwin <<= 1;
582 else
583 rcvwin += (rcvwin >> 1);
584 }
585
586 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
587 while (tcp_win_from_space(rcvmem) < tp->advmss)
588 rcvmem += 128;
589
590 rcvbuf = min(rcvwin / tp->advmss * rcvmem, sysctl_tcp_rmem[2]);
591 if (rcvbuf > sk->sk_rcvbuf) {
592 sk->sk_rcvbuf = rcvbuf;
593
594 /* Make the window clamp follow along. */
595 tp->window_clamp = rcvwin;
596 }
597 }
598 tp->rcvq_space.space = copied;
599
600new_measure:
601 tp->rcvq_space.seq = tp->copied_seq;
602 tp->rcvq_space.time = tcp_time_stamp;
603}
604
605/* There is something which you must keep in mind when you analyze the
606 * behavior of the tp->ato delayed ack timeout interval. When a
607 * connection starts up, we want to ack as quickly as possible. The
608 * problem is that "good" TCP's do slow start at the beginning of data
609 * transmission. The means that until we send the first few ACK's the
610 * sender will sit on his end and only queue most of his data, because
611 * he can only send snd_cwnd unacked packets at any given time. For
612 * each ACK we send, he increments snd_cwnd and transmits more of his
613 * queue. -DaveM
614 */
615static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
616{
617 struct tcp_sock *tp = tcp_sk(sk);
618 struct inet_connection_sock *icsk = inet_csk(sk);
619 u32 now;
620
621 inet_csk_schedule_ack(sk);
622
623 tcp_measure_rcv_mss(sk, skb);
624
625 tcp_rcv_rtt_measure(tp);
626
627 now = tcp_time_stamp;
628
629 if (!icsk->icsk_ack.ato) {
630 /* The _first_ data packet received, initialize
631 * delayed ACK engine.
632 */
633 tcp_incr_quickack(sk);
634 icsk->icsk_ack.ato = TCP_ATO_MIN;
635 } else {
636 int m = now - icsk->icsk_ack.lrcvtime;
637
638 if (m <= TCP_ATO_MIN / 2) {
639 /* The fastest case is the first. */
640 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
641 } else if (m < icsk->icsk_ack.ato) {
642 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
643 if (icsk->icsk_ack.ato > icsk->icsk_rto)
644 icsk->icsk_ack.ato = icsk->icsk_rto;
645 } else if (m > icsk->icsk_rto) {
646 /* Too long gap. Apparently sender failed to
647 * restart window, so that we send ACKs quickly.
648 */
649 tcp_incr_quickack(sk);
650 sk_mem_reclaim(sk);
651 }
652 }
653 icsk->icsk_ack.lrcvtime = now;
654
655 TCP_ECN_check_ce(tp, skb);
656
657 if (skb->len >= 128)
658 tcp_grow_window(sk, skb);
659}
660
661/* Called to compute a smoothed rtt estimate. The data fed to this
662 * routine either comes from timestamps, or from segments that were
663 * known _not_ to have been retransmitted [see Karn/Partridge
664 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
665 * piece by Van Jacobson.
666 * NOTE: the next three routines used to be one big routine.
667 * To save cycles in the RFC 1323 implementation it was better to break
668 * it up into three procedures. -- erics
669 */
670static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
671{
672 struct tcp_sock *tp = tcp_sk(sk);
673 long m = mrtt_us; /* RTT */
674 u32 srtt = tp->srtt_us;
675
676 /* The following amusing code comes from Jacobson's
677 * article in SIGCOMM '88. Note that rtt and mdev
678 * are scaled versions of rtt and mean deviation.
679 * This is designed to be as fast as possible
680 * m stands for "measurement".
681 *
682 * On a 1990 paper the rto value is changed to:
683 * RTO = rtt + 4 * mdev
684 *
685 * Funny. This algorithm seems to be very broken.
686 * These formulae increase RTO, when it should be decreased, increase
687 * too slowly, when it should be increased quickly, decrease too quickly
688 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
689 * does not matter how to _calculate_ it. Seems, it was trap
690 * that VJ failed to avoid. 8)
691 */
692 if (srtt != 0) {
693 m -= (srtt >> 3); /* m is now error in rtt est */
694 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
695 if (m < 0) {
696 m = -m; /* m is now abs(error) */
697 m -= (tp->mdev_us >> 2); /* similar update on mdev */
698 /* This is similar to one of Eifel findings.
699 * Eifel blocks mdev updates when rtt decreases.
700 * This solution is a bit different: we use finer gain
701 * for mdev in this case (alpha*beta).
702 * Like Eifel it also prevents growth of rto,
703 * but also it limits too fast rto decreases,
704 * happening in pure Eifel.
705 */
706 if (m > 0)
707 m >>= 3;
708 } else {
709 m -= (tp->mdev_us >> 2); /* similar update on mdev */
710 }
711 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
712 if (tp->mdev_us > tp->mdev_max_us) {
713 tp->mdev_max_us = tp->mdev_us;
714 if (tp->mdev_max_us > tp->rttvar_us)
715 tp->rttvar_us = tp->mdev_max_us;
716 }
717 if (after(tp->snd_una, tp->rtt_seq)) {
718 if (tp->mdev_max_us < tp->rttvar_us)
719 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
720 tp->rtt_seq = tp->snd_nxt;
721 tp->mdev_max_us = tcp_rto_min_us(sk);
722 }
723 } else {
724 /* no previous measure. */
725 srtt = m << 3; /* take the measured time to be rtt */
726 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
727 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
728 tp->mdev_max_us = tp->rttvar_us;
729 tp->rtt_seq = tp->snd_nxt;
730 }
731 tp->srtt_us = max(1U, srtt);
732}
733
734/* Set the sk_pacing_rate to allow proper sizing of TSO packets.
735 * Note: TCP stack does not yet implement pacing.
736 * FQ packet scheduler can be used to implement cheap but effective
737 * TCP pacing, to smooth the burst on large writes when packets
738 * in flight is significantly lower than cwnd (or rwin)
739 */
740static void tcp_update_pacing_rate(struct sock *sk)
741{
742 const struct tcp_sock *tp = tcp_sk(sk);
743 u64 rate;
744
745 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
746 rate = (u64)tp->mss_cache * 2 * (USEC_PER_SEC << 3);
747
748 rate *= max(tp->snd_cwnd, tp->packets_out);
749
750 if (likely(tp->srtt_us))
751 do_div(rate, tp->srtt_us);
752
753 /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate
754 * without any lock. We want to make sure compiler wont store
755 * intermediate values in this location.
756 */
757 ACCESS_ONCE(sk->sk_pacing_rate) = min_t(u64, rate,
758 sk->sk_max_pacing_rate);
759}
760
761/* Calculate rto without backoff. This is the second half of Van Jacobson's
762 * routine referred to above.
763 */
764static void tcp_set_rto(struct sock *sk)
765{
766 const struct tcp_sock *tp = tcp_sk(sk);
767 /* Old crap is replaced with new one. 8)
768 *
769 * More seriously:
770 * 1. If rtt variance happened to be less 50msec, it is hallucination.
771 * It cannot be less due to utterly erratic ACK generation made
772 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
773 * to do with delayed acks, because at cwnd>2 true delack timeout
774 * is invisible. Actually, Linux-2.4 also generates erratic
775 * ACKs in some circumstances.
776 */
777 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
778
779 /* 2. Fixups made earlier cannot be right.
780 * If we do not estimate RTO correctly without them,
781 * all the algo is pure shit and should be replaced
782 * with correct one. It is exactly, which we pretend to do.
783 */
784
785 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
786 * guarantees that rto is higher.
787 */
788 tcp_bound_rto(sk);
789}
790
791__u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
792{
793 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
794
795 if (!cwnd)
796 cwnd = TCP_INIT_CWND;
797 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
798}
799
800/*
801 * Packet counting of FACK is based on in-order assumptions, therefore TCP
802 * disables it when reordering is detected
803 */
804void tcp_disable_fack(struct tcp_sock *tp)
805{
806 /* RFC3517 uses different metric in lost marker => reset on change */
807 if (tcp_is_fack(tp))
808 tp->lost_skb_hint = NULL;
809 tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
810}
811
812/* Take a notice that peer is sending D-SACKs */
813static void tcp_dsack_seen(struct tcp_sock *tp)
814{
815 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
816}
817
818static void tcp_update_reordering(struct sock *sk, const int metric,
819 const int ts)
820{
821 struct tcp_sock *tp = tcp_sk(sk);
822 if (metric > tp->reordering) {
823 int mib_idx;
824
825 tp->reordering = min(TCP_MAX_REORDERING, metric);
826
827 /* This exciting event is worth to be remembered. 8) */
828 if (ts)
829 mib_idx = LINUX_MIB_TCPTSREORDER;
830 else if (tcp_is_reno(tp))
831 mib_idx = LINUX_MIB_TCPRENOREORDER;
832 else if (tcp_is_fack(tp))
833 mib_idx = LINUX_MIB_TCPFACKREORDER;
834 else
835 mib_idx = LINUX_MIB_TCPSACKREORDER;
836
837 NET_INC_STATS_BH(sock_net(sk), mib_idx);
838#if FASTRETRANS_DEBUG > 1
839 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
840 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
841 tp->reordering,
842 tp->fackets_out,
843 tp->sacked_out,
844 tp->undo_marker ? tp->undo_retrans : 0);
845#endif
846 tcp_disable_fack(tp);
847 }
848
849 if (metric > 0)
850 tcp_disable_early_retrans(tp);
851}
852
853/* This must be called before lost_out is incremented */
854static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
855{
856 if ((tp->retransmit_skb_hint == NULL) ||
857 before(TCP_SKB_CB(skb)->seq,
858 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
859 tp->retransmit_skb_hint = skb;
860
861 if (!tp->lost_out ||
862 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
863 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
864}
865
866static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
867{
868 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
869 tcp_verify_retransmit_hint(tp, skb);
870
871 tp->lost_out += tcp_skb_pcount(skb);
872 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
873 }
874}
875
876static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
877 struct sk_buff *skb)
878{
879 tcp_verify_retransmit_hint(tp, skb);
880
881 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
882 tp->lost_out += tcp_skb_pcount(skb);
883 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
884 }
885}
886
887/* This procedure tags the retransmission queue when SACKs arrive.
888 *
889 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
890 * Packets in queue with these bits set are counted in variables
891 * sacked_out, retrans_out and lost_out, correspondingly.
892 *
893 * Valid combinations are:
894 * Tag InFlight Description
895 * 0 1 - orig segment is in flight.
896 * S 0 - nothing flies, orig reached receiver.
897 * L 0 - nothing flies, orig lost by net.
898 * R 2 - both orig and retransmit are in flight.
899 * L|R 1 - orig is lost, retransmit is in flight.
900 * S|R 1 - orig reached receiver, retrans is still in flight.
901 * (L|S|R is logically valid, it could occur when L|R is sacked,
902 * but it is equivalent to plain S and code short-curcuits it to S.
903 * L|S is logically invalid, it would mean -1 packet in flight 8))
904 *
905 * These 6 states form finite state machine, controlled by the following events:
906 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
907 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
908 * 3. Loss detection event of two flavors:
909 * A. Scoreboard estimator decided the packet is lost.
910 * A'. Reno "three dupacks" marks head of queue lost.
911 * A''. Its FACK modification, head until snd.fack is lost.
912 * B. SACK arrives sacking SND.NXT at the moment, when the
913 * segment was retransmitted.
914 * 4. D-SACK added new rule: D-SACK changes any tag to S.
915 *
916 * It is pleasant to note, that state diagram turns out to be commutative,
917 * so that we are allowed not to be bothered by order of our actions,
918 * when multiple events arrive simultaneously. (see the function below).
919 *
920 * Reordering detection.
921 * --------------------
922 * Reordering metric is maximal distance, which a packet can be displaced
923 * in packet stream. With SACKs we can estimate it:
924 *
925 * 1. SACK fills old hole and the corresponding segment was not
926 * ever retransmitted -> reordering. Alas, we cannot use it
927 * when segment was retransmitted.
928 * 2. The last flaw is solved with D-SACK. D-SACK arrives
929 * for retransmitted and already SACKed segment -> reordering..
930 * Both of these heuristics are not used in Loss state, when we cannot
931 * account for retransmits accurately.
932 *
933 * SACK block validation.
934 * ----------------------
935 *
936 * SACK block range validation checks that the received SACK block fits to
937 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
938 * Note that SND.UNA is not included to the range though being valid because
939 * it means that the receiver is rather inconsistent with itself reporting
940 * SACK reneging when it should advance SND.UNA. Such SACK block this is
941 * perfectly valid, however, in light of RFC2018 which explicitly states
942 * that "SACK block MUST reflect the newest segment. Even if the newest
943 * segment is going to be discarded ...", not that it looks very clever
944 * in case of head skb. Due to potentional receiver driven attacks, we
945 * choose to avoid immediate execution of a walk in write queue due to
946 * reneging and defer head skb's loss recovery to standard loss recovery
947 * procedure that will eventually trigger (nothing forbids us doing this).
948 *
949 * Implements also blockage to start_seq wrap-around. Problem lies in the
950 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
951 * there's no guarantee that it will be before snd_nxt (n). The problem
952 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
953 * wrap (s_w):
954 *
955 * <- outs wnd -> <- wrapzone ->
956 * u e n u_w e_w s n_w
957 * | | | | | | |
958 * |<------------+------+----- TCP seqno space --------------+---------->|
959 * ...-- <2^31 ->| |<--------...
960 * ...---- >2^31 ------>| |<--------...
961 *
962 * Current code wouldn't be vulnerable but it's better still to discard such
963 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
964 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
965 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
966 * equal to the ideal case (infinite seqno space without wrap caused issues).
967 *
968 * With D-SACK the lower bound is extended to cover sequence space below
969 * SND.UNA down to undo_marker, which is the last point of interest. Yet
970 * again, D-SACK block must not to go across snd_una (for the same reason as
971 * for the normal SACK blocks, explained above). But there all simplicity
972 * ends, TCP might receive valid D-SACKs below that. As long as they reside
973 * fully below undo_marker they do not affect behavior in anyway and can
974 * therefore be safely ignored. In rare cases (which are more or less
975 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
976 * fragmentation and packet reordering past skb's retransmission. To consider
977 * them correctly, the acceptable range must be extended even more though
978 * the exact amount is rather hard to quantify. However, tp->max_window can
979 * be used as an exaggerated estimate.
980 */
981static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
982 u32 start_seq, u32 end_seq)
983{
984 /* Too far in future, or reversed (interpretation is ambiguous) */
985 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
986 return false;
987
988 /* Nasty start_seq wrap-around check (see comments above) */
989 if (!before(start_seq, tp->snd_nxt))
990 return false;
991
992 /* In outstanding window? ...This is valid exit for D-SACKs too.
993 * start_seq == snd_una is non-sensical (see comments above)
994 */
995 if (after(start_seq, tp->snd_una))
996 return true;
997
998 if (!is_dsack || !tp->undo_marker)
999 return false;
1000
1001 /* ...Then it's D-SACK, and must reside below snd_una completely */
1002 if (after(end_seq, tp->snd_una))
1003 return false;
1004
1005 if (!before(start_seq, tp->undo_marker))
1006 return true;
1007
1008 /* Too old */
1009 if (!after(end_seq, tp->undo_marker))
1010 return false;
1011
1012 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1013 * start_seq < undo_marker and end_seq >= undo_marker.
1014 */
1015 return !before(start_seq, end_seq - tp->max_window);
1016}
1017
1018/* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1019 * Event "B". Later note: FACK people cheated me again 8), we have to account
1020 * for reordering! Ugly, but should help.
1021 *
1022 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1023 * less than what is now known to be received by the other end (derived from
1024 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1025 * retransmitted skbs to avoid some costly processing per ACKs.
1026 */
1027static void tcp_mark_lost_retrans(struct sock *sk)
1028{
1029 const struct inet_connection_sock *icsk = inet_csk(sk);
1030 struct tcp_sock *tp = tcp_sk(sk);
1031 struct sk_buff *skb;
1032 int cnt = 0;
1033 u32 new_low_seq = tp->snd_nxt;
1034 u32 received_upto = tcp_highest_sack_seq(tp);
1035
1036 if (!tcp_is_fack(tp) || !tp->retrans_out ||
1037 !after(received_upto, tp->lost_retrans_low) ||
1038 icsk->icsk_ca_state != TCP_CA_Recovery)
1039 return;
1040
1041 tcp_for_write_queue(skb, sk) {
1042 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1043
1044 if (skb == tcp_send_head(sk))
1045 break;
1046 if (cnt == tp->retrans_out)
1047 break;
1048 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1049 continue;
1050
1051 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1052 continue;
1053
1054 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1055 * constraint here (see above) but figuring out that at
1056 * least tp->reordering SACK blocks reside between ack_seq
1057 * and received_upto is not easy task to do cheaply with
1058 * the available datastructures.
1059 *
1060 * Whether FACK should check here for tp->reordering segs
1061 * in-between one could argue for either way (it would be
1062 * rather simple to implement as we could count fack_count
1063 * during the walk and do tp->fackets_out - fack_count).
1064 */
1065 if (after(received_upto, ack_seq)) {
1066 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1067 tp->retrans_out -= tcp_skb_pcount(skb);
1068
1069 tcp_skb_mark_lost_uncond_verify(tp, skb);
1070 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1071 } else {
1072 if (before(ack_seq, new_low_seq))
1073 new_low_seq = ack_seq;
1074 cnt += tcp_skb_pcount(skb);
1075 }
1076 }
1077
1078 if (tp->retrans_out)
1079 tp->lost_retrans_low = new_low_seq;
1080}
1081
1082static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1083 struct tcp_sack_block_wire *sp, int num_sacks,
1084 u32 prior_snd_una)
1085{
1086 struct tcp_sock *tp = tcp_sk(sk);
1087 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1088 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1089 bool dup_sack = false;
1090
1091 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1092 dup_sack = true;
1093 tcp_dsack_seen(tp);
1094 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1095 } else if (num_sacks > 1) {
1096 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1097 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1098
1099 if (!after(end_seq_0, end_seq_1) &&
1100 !before(start_seq_0, start_seq_1)) {
1101 dup_sack = true;
1102 tcp_dsack_seen(tp);
1103 NET_INC_STATS_BH(sock_net(sk),
1104 LINUX_MIB_TCPDSACKOFORECV);
1105 }
1106 }
1107
1108 /* D-SACK for already forgotten data... Do dumb counting. */
1109 if (dup_sack && tp->undo_marker && tp->undo_retrans &&
1110 !after(end_seq_0, prior_snd_una) &&
1111 after(end_seq_0, tp->undo_marker))
1112 tp->undo_retrans--;
1113
1114 return dup_sack;
1115}
1116
1117struct tcp_sacktag_state {
1118 int reord;
1119 int fack_count;
1120 long rtt_us; /* RTT measured by SACKing never-retransmitted data */
1121 int flag;
1122};
1123
1124/* Check if skb is fully within the SACK block. In presence of GSO skbs,
1125 * the incoming SACK may not exactly match but we can find smaller MSS
1126 * aligned portion of it that matches. Therefore we might need to fragment
1127 * which may fail and creates some hassle (caller must handle error case
1128 * returns).
1129 *
1130 * FIXME: this could be merged to shift decision code
1131 */
1132static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1133 u32 start_seq, u32 end_seq)
1134{
1135 int err;
1136 bool in_sack;
1137 unsigned int pkt_len;
1138 unsigned int mss;
1139
1140 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1141 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1142
1143 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1144 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1145 mss = tcp_skb_mss(skb);
1146 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1147
1148 if (!in_sack) {
1149 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1150 if (pkt_len < mss)
1151 pkt_len = mss;
1152 } else {
1153 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1154 if (pkt_len < mss)
1155 return -EINVAL;
1156 }
1157
1158 /* Round if necessary so that SACKs cover only full MSSes
1159 * and/or the remaining small portion (if present)
1160 */
1161 if (pkt_len > mss) {
1162 unsigned int new_len = (pkt_len / mss) * mss;
1163 if (!in_sack && new_len < pkt_len) {
1164 new_len += mss;
1165 if (new_len > skb->len)
1166 return 0;
1167 }
1168 pkt_len = new_len;
1169 }
1170 err = tcp_fragment(sk, skb, pkt_len, mss);
1171 if (err < 0)
1172 return err;
1173 }
1174
1175 return in_sack;
1176}
1177
1178/* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1179static u8 tcp_sacktag_one(struct sock *sk,
1180 struct tcp_sacktag_state *state, u8 sacked,
1181 u32 start_seq, u32 end_seq,
1182 int dup_sack, int pcount,
1183 const struct skb_mstamp *xmit_time)
1184{
1185 struct tcp_sock *tp = tcp_sk(sk);
1186 int fack_count = state->fack_count;
1187
1188 /* Account D-SACK for retransmitted packet. */
1189 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1190 if (tp->undo_marker && tp->undo_retrans &&
1191 after(end_seq, tp->undo_marker))
1192 tp->undo_retrans--;
1193 if (sacked & TCPCB_SACKED_ACKED)
1194 state->reord = min(fack_count, state->reord);
1195 }
1196
1197 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1198 if (!after(end_seq, tp->snd_una))
1199 return sacked;
1200
1201 if (!(sacked & TCPCB_SACKED_ACKED)) {
1202 if (sacked & TCPCB_SACKED_RETRANS) {
1203 /* If the segment is not tagged as lost,
1204 * we do not clear RETRANS, believing
1205 * that retransmission is still in flight.
1206 */
1207 if (sacked & TCPCB_LOST) {
1208 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1209 tp->lost_out -= pcount;
1210 tp->retrans_out -= pcount;
1211 }
1212 } else {
1213 if (!(sacked & TCPCB_RETRANS)) {
1214 /* New sack for not retransmitted frame,
1215 * which was in hole. It is reordering.
1216 */
1217 if (before(start_seq,
1218 tcp_highest_sack_seq(tp)))
1219 state->reord = min(fack_count,
1220 state->reord);
1221 if (!after(end_seq, tp->high_seq))
1222 state->flag |= FLAG_ORIG_SACK_ACKED;
1223 /* Pick the earliest sequence sacked for RTT */
1224 if (state->rtt_us < 0) {
1225 struct skb_mstamp now;
1226
1227 skb_mstamp_get(&now);
1228 state->rtt_us = skb_mstamp_us_delta(&now,
1229 xmit_time);
1230 }
1231 }
1232
1233 if (sacked & TCPCB_LOST) {
1234 sacked &= ~TCPCB_LOST;
1235 tp->lost_out -= pcount;
1236 }
1237 }
1238
1239 sacked |= TCPCB_SACKED_ACKED;
1240 state->flag |= FLAG_DATA_SACKED;
1241 tp->sacked_out += pcount;
1242
1243 fack_count += pcount;
1244
1245 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1246 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1247 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1248 tp->lost_cnt_hint += pcount;
1249
1250 if (fack_count > tp->fackets_out)
1251 tp->fackets_out = fack_count;
1252 }
1253
1254 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1255 * frames and clear it. undo_retrans is decreased above, L|R frames
1256 * are accounted above as well.
1257 */
1258 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1259 sacked &= ~TCPCB_SACKED_RETRANS;
1260 tp->retrans_out -= pcount;
1261 }
1262
1263 return sacked;
1264}
1265
1266/* Shift newly-SACKed bytes from this skb to the immediately previous
1267 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1268 */
1269static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1270 struct tcp_sacktag_state *state,
1271 unsigned int pcount, int shifted, int mss,
1272 bool dup_sack)
1273{
1274 struct tcp_sock *tp = tcp_sk(sk);
1275 struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1276 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1277 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1278
1279 BUG_ON(!pcount);
1280
1281 /* Adjust counters and hints for the newly sacked sequence
1282 * range but discard the return value since prev is already
1283 * marked. We must tag the range first because the seq
1284 * advancement below implicitly advances
1285 * tcp_highest_sack_seq() when skb is highest_sack.
1286 */
1287 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1288 start_seq, end_seq, dup_sack, pcount,
1289 &skb->skb_mstamp);
1290
1291 if (skb == tp->lost_skb_hint)
1292 tp->lost_cnt_hint += pcount;
1293
1294 TCP_SKB_CB(prev)->end_seq += shifted;
1295 TCP_SKB_CB(skb)->seq += shifted;
1296
1297 skb_shinfo(prev)->gso_segs += pcount;
1298 BUG_ON(skb_shinfo(skb)->gso_segs < pcount);
1299 skb_shinfo(skb)->gso_segs -= pcount;
1300
1301 /* When we're adding to gso_segs == 1, gso_size will be zero,
1302 * in theory this shouldn't be necessary but as long as DSACK
1303 * code can come after this skb later on it's better to keep
1304 * setting gso_size to something.
1305 */
1306 if (!skb_shinfo(prev)->gso_size) {
1307 skb_shinfo(prev)->gso_size = mss;
1308 skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1309 }
1310
1311 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1312 if (skb_shinfo(skb)->gso_segs <= 1) {
1313 skb_shinfo(skb)->gso_size = 0;
1314 skb_shinfo(skb)->gso_type = 0;
1315 }
1316
1317 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1318 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1319
1320 if (skb->len > 0) {
1321 BUG_ON(!tcp_skb_pcount(skb));
1322 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1323 return false;
1324 }
1325
1326 /* Whole SKB was eaten :-) */
1327
1328 if (skb == tp->retransmit_skb_hint)
1329 tp->retransmit_skb_hint = prev;
1330 if (skb == tp->lost_skb_hint) {
1331 tp->lost_skb_hint = prev;
1332 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1333 }
1334
1335 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1336 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1337 TCP_SKB_CB(prev)->end_seq++;
1338
1339 if (skb == tcp_highest_sack(sk))
1340 tcp_advance_highest_sack(sk, skb);
1341
1342 tcp_unlink_write_queue(skb, sk);
1343 sk_wmem_free_skb(sk, skb);
1344
1345 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1346
1347 return true;
1348}
1349
1350/* I wish gso_size would have a bit more sane initialization than
1351 * something-or-zero which complicates things
1352 */
1353static int tcp_skb_seglen(const struct sk_buff *skb)
1354{
1355 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1356}
1357
1358/* Shifting pages past head area doesn't work */
1359static int skb_can_shift(const struct sk_buff *skb)
1360{
1361 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1362}
1363
1364/* Try collapsing SACK blocks spanning across multiple skbs to a single
1365 * skb.
1366 */
1367static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1368 struct tcp_sacktag_state *state,
1369 u32 start_seq, u32 end_seq,
1370 bool dup_sack)
1371{
1372 struct tcp_sock *tp = tcp_sk(sk);
1373 struct sk_buff *prev;
1374 int mss;
1375 int pcount = 0;
1376 int len;
1377 int in_sack;
1378
1379 if (!sk_can_gso(sk))
1380 goto fallback;
1381
1382 /* Normally R but no L won't result in plain S */
1383 if (!dup_sack &&
1384 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1385 goto fallback;
1386 if (!skb_can_shift(skb))
1387 goto fallback;
1388 /* This frame is about to be dropped (was ACKed). */
1389 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1390 goto fallback;
1391
1392 /* Can only happen with delayed DSACK + discard craziness */
1393 if (unlikely(skb == tcp_write_queue_head(sk)))
1394 goto fallback;
1395 prev = tcp_write_queue_prev(sk, skb);
1396
1397 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1398 goto fallback;
1399
1400 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1401 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1402
1403 if (in_sack) {
1404 len = skb->len;
1405 pcount = tcp_skb_pcount(skb);
1406 mss = tcp_skb_seglen(skb);
1407
1408 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1409 * drop this restriction as unnecessary
1410 */
1411 if (mss != tcp_skb_seglen(prev))
1412 goto fallback;
1413 } else {
1414 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1415 goto noop;
1416 /* CHECKME: This is non-MSS split case only?, this will
1417 * cause skipped skbs due to advancing loop btw, original
1418 * has that feature too
1419 */
1420 if (tcp_skb_pcount(skb) <= 1)
1421 goto noop;
1422
1423 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1424 if (!in_sack) {
1425 /* TODO: head merge to next could be attempted here
1426 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1427 * though it might not be worth of the additional hassle
1428 *
1429 * ...we can probably just fallback to what was done
1430 * previously. We could try merging non-SACKed ones
1431 * as well but it probably isn't going to buy off
1432 * because later SACKs might again split them, and
1433 * it would make skb timestamp tracking considerably
1434 * harder problem.
1435 */
1436 goto fallback;
1437 }
1438
1439 len = end_seq - TCP_SKB_CB(skb)->seq;
1440 BUG_ON(len < 0);
1441 BUG_ON(len > skb->len);
1442
1443 /* MSS boundaries should be honoured or else pcount will
1444 * severely break even though it makes things bit trickier.
1445 * Optimize common case to avoid most of the divides
1446 */
1447 mss = tcp_skb_mss(skb);
1448
1449 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1450 * drop this restriction as unnecessary
1451 */
1452 if (mss != tcp_skb_seglen(prev))
1453 goto fallback;
1454
1455 if (len == mss) {
1456 pcount = 1;
1457 } else if (len < mss) {
1458 goto noop;
1459 } else {
1460 pcount = len / mss;
1461 len = pcount * mss;
1462 }
1463 }
1464
1465 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1466 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1467 goto fallback;
1468
1469 if (!skb_shift(prev, skb, len))
1470 goto fallback;
1471 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1472 goto out;
1473
1474 /* Hole filled allows collapsing with the next as well, this is very
1475 * useful when hole on every nth skb pattern happens
1476 */
1477 if (prev == tcp_write_queue_tail(sk))
1478 goto out;
1479 skb = tcp_write_queue_next(sk, prev);
1480
1481 if (!skb_can_shift(skb) ||
1482 (skb == tcp_send_head(sk)) ||
1483 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1484 (mss != tcp_skb_seglen(skb)))
1485 goto out;
1486
1487 len = skb->len;
1488 if (skb_shift(prev, skb, len)) {
1489 pcount += tcp_skb_pcount(skb);
1490 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1491 }
1492
1493out:
1494 state->fack_count += pcount;
1495 return prev;
1496
1497noop:
1498 return skb;
1499
1500fallback:
1501 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1502 return NULL;
1503}
1504
1505static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1506 struct tcp_sack_block *next_dup,
1507 struct tcp_sacktag_state *state,
1508 u32 start_seq, u32 end_seq,
1509 bool dup_sack_in)
1510{
1511 struct tcp_sock *tp = tcp_sk(sk);
1512 struct sk_buff *tmp;
1513
1514 tcp_for_write_queue_from(skb, sk) {
1515 int in_sack = 0;
1516 bool dup_sack = dup_sack_in;
1517
1518 if (skb == tcp_send_head(sk))
1519 break;
1520
1521 /* queue is in-order => we can short-circuit the walk early */
1522 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1523 break;
1524
1525 if ((next_dup != NULL) &&
1526 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1527 in_sack = tcp_match_skb_to_sack(sk, skb,
1528 next_dup->start_seq,
1529 next_dup->end_seq);
1530 if (in_sack > 0)
1531 dup_sack = true;
1532 }
1533
1534 /* skb reference here is a bit tricky to get right, since
1535 * shifting can eat and free both this skb and the next,
1536 * so not even _safe variant of the loop is enough.
1537 */
1538 if (in_sack <= 0) {
1539 tmp = tcp_shift_skb_data(sk, skb, state,
1540 start_seq, end_seq, dup_sack);
1541 if (tmp != NULL) {
1542 if (tmp != skb) {
1543 skb = tmp;
1544 continue;
1545 }
1546
1547 in_sack = 0;
1548 } else {
1549 in_sack = tcp_match_skb_to_sack(sk, skb,
1550 start_seq,
1551 end_seq);
1552 }
1553 }
1554
1555 if (unlikely(in_sack < 0))
1556 break;
1557
1558 if (in_sack) {
1559 TCP_SKB_CB(skb)->sacked =
1560 tcp_sacktag_one(sk,
1561 state,
1562 TCP_SKB_CB(skb)->sacked,
1563 TCP_SKB_CB(skb)->seq,
1564 TCP_SKB_CB(skb)->end_seq,
1565 dup_sack,
1566 tcp_skb_pcount(skb),
1567 &skb->skb_mstamp);
1568
1569 if (!before(TCP_SKB_CB(skb)->seq,
1570 tcp_highest_sack_seq(tp)))
1571 tcp_advance_highest_sack(sk, skb);
1572 }
1573
1574 state->fack_count += tcp_skb_pcount(skb);
1575 }
1576 return skb;
1577}
1578
1579/* Avoid all extra work that is being done by sacktag while walking in
1580 * a normal way
1581 */
1582static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1583 struct tcp_sacktag_state *state,
1584 u32 skip_to_seq)
1585{
1586 tcp_for_write_queue_from(skb, sk) {
1587 if (skb == tcp_send_head(sk))
1588 break;
1589
1590 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1591 break;
1592
1593 state->fack_count += tcp_skb_pcount(skb);
1594 }
1595 return skb;
1596}
1597
1598static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1599 struct sock *sk,
1600 struct tcp_sack_block *next_dup,
1601 struct tcp_sacktag_state *state,
1602 u32 skip_to_seq)
1603{
1604 if (next_dup == NULL)
1605 return skb;
1606
1607 if (before(next_dup->start_seq, skip_to_seq)) {
1608 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1609 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1610 next_dup->start_seq, next_dup->end_seq,
1611 1);
1612 }
1613
1614 return skb;
1615}
1616
1617static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1618{
1619 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1620}
1621
1622static int
1623tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1624 u32 prior_snd_una, long *sack_rtt_us)
1625{
1626 struct tcp_sock *tp = tcp_sk(sk);
1627 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1628 TCP_SKB_CB(ack_skb)->sacked);
1629 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1630 struct tcp_sack_block sp[TCP_NUM_SACKS];
1631 struct tcp_sack_block *cache;
1632 struct tcp_sacktag_state state;
1633 struct sk_buff *skb;
1634 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1635 int used_sacks;
1636 bool found_dup_sack = false;
1637 int i, j;
1638 int first_sack_index;
1639
1640 state.flag = 0;
1641 state.reord = tp->packets_out;
1642 state.rtt_us = -1L;
1643
1644 if (!tp->sacked_out) {
1645 if (WARN_ON(tp->fackets_out))
1646 tp->fackets_out = 0;
1647 tcp_highest_sack_reset(sk);
1648 }
1649
1650 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1651 num_sacks, prior_snd_una);
1652 if (found_dup_sack)
1653 state.flag |= FLAG_DSACKING_ACK;
1654
1655 /* Eliminate too old ACKs, but take into
1656 * account more or less fresh ones, they can
1657 * contain valid SACK info.
1658 */
1659 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1660 return 0;
1661
1662 if (!tp->packets_out)
1663 goto out;
1664
1665 used_sacks = 0;
1666 first_sack_index = 0;
1667 for (i = 0; i < num_sacks; i++) {
1668 bool dup_sack = !i && found_dup_sack;
1669
1670 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1671 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1672
1673 if (!tcp_is_sackblock_valid(tp, dup_sack,
1674 sp[used_sacks].start_seq,
1675 sp[used_sacks].end_seq)) {
1676 int mib_idx;
1677
1678 if (dup_sack) {
1679 if (!tp->undo_marker)
1680 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1681 else
1682 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1683 } else {
1684 /* Don't count olds caused by ACK reordering */
1685 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1686 !after(sp[used_sacks].end_seq, tp->snd_una))
1687 continue;
1688 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1689 }
1690
1691 NET_INC_STATS_BH(sock_net(sk), mib_idx);
1692 if (i == 0)
1693 first_sack_index = -1;
1694 continue;
1695 }
1696
1697 /* Ignore very old stuff early */
1698 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1699 continue;
1700
1701 used_sacks++;
1702 }
1703
1704 /* order SACK blocks to allow in order walk of the retrans queue */
1705 for (i = used_sacks - 1; i > 0; i--) {
1706 for (j = 0; j < i; j++) {
1707 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1708 swap(sp[j], sp[j + 1]);
1709
1710 /* Track where the first SACK block goes to */
1711 if (j == first_sack_index)
1712 first_sack_index = j + 1;
1713 }
1714 }
1715 }
1716
1717 skb = tcp_write_queue_head(sk);
1718 state.fack_count = 0;
1719 i = 0;
1720
1721 if (!tp->sacked_out) {
1722 /* It's already past, so skip checking against it */
1723 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1724 } else {
1725 cache = tp->recv_sack_cache;
1726 /* Skip empty blocks in at head of the cache */
1727 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1728 !cache->end_seq)
1729 cache++;
1730 }
1731
1732 while (i < used_sacks) {
1733 u32 start_seq = sp[i].start_seq;
1734 u32 end_seq = sp[i].end_seq;
1735 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1736 struct tcp_sack_block *next_dup = NULL;
1737
1738 if (found_dup_sack && ((i + 1) == first_sack_index))
1739 next_dup = &sp[i + 1];
1740
1741 /* Skip too early cached blocks */
1742 while (tcp_sack_cache_ok(tp, cache) &&
1743 !before(start_seq, cache->end_seq))
1744 cache++;
1745
1746 /* Can skip some work by looking recv_sack_cache? */
1747 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1748 after(end_seq, cache->start_seq)) {
1749
1750 /* Head todo? */
1751 if (before(start_seq, cache->start_seq)) {
1752 skb = tcp_sacktag_skip(skb, sk, &state,
1753 start_seq);
1754 skb = tcp_sacktag_walk(skb, sk, next_dup,
1755 &state,
1756 start_seq,
1757 cache->start_seq,
1758 dup_sack);
1759 }
1760
1761 /* Rest of the block already fully processed? */
1762 if (!after(end_seq, cache->end_seq))
1763 goto advance_sp;
1764
1765 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1766 &state,
1767 cache->end_seq);
1768
1769 /* ...tail remains todo... */
1770 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1771 /* ...but better entrypoint exists! */
1772 skb = tcp_highest_sack(sk);
1773 if (skb == NULL)
1774 break;
1775 state.fack_count = tp->fackets_out;
1776 cache++;
1777 goto walk;
1778 }
1779
1780 skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
1781 /* Check overlap against next cached too (past this one already) */
1782 cache++;
1783 continue;
1784 }
1785
1786 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1787 skb = tcp_highest_sack(sk);
1788 if (skb == NULL)
1789 break;
1790 state.fack_count = tp->fackets_out;
1791 }
1792 skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
1793
1794walk:
1795 skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1796 start_seq, end_seq, dup_sack);
1797
1798advance_sp:
1799 i++;
1800 }
1801
1802 /* Clear the head of the cache sack blocks so we can skip it next time */
1803 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1804 tp->recv_sack_cache[i].start_seq = 0;
1805 tp->recv_sack_cache[i].end_seq = 0;
1806 }
1807 for (j = 0; j < used_sacks; j++)
1808 tp->recv_sack_cache[i++] = sp[j];
1809
1810 tcp_mark_lost_retrans(sk);
1811
1812 tcp_verify_left_out(tp);
1813
1814 if ((state.reord < tp->fackets_out) &&
1815 ((inet_csk(sk)->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker))
1816 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
1817
1818out:
1819
1820#if FASTRETRANS_DEBUG > 0
1821 WARN_ON((int)tp->sacked_out < 0);
1822 WARN_ON((int)tp->lost_out < 0);
1823 WARN_ON((int)tp->retrans_out < 0);
1824 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1825#endif
1826 *sack_rtt_us = state.rtt_us;
1827 return state.flag;
1828}
1829
1830/* Limits sacked_out so that sum with lost_out isn't ever larger than
1831 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1832 */
1833static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1834{
1835 u32 holes;
1836
1837 holes = max(tp->lost_out, 1U);
1838 holes = min(holes, tp->packets_out);
1839
1840 if ((tp->sacked_out + holes) > tp->packets_out) {
1841 tp->sacked_out = tp->packets_out - holes;
1842 return true;
1843 }
1844 return false;
1845}
1846
1847/* If we receive more dupacks than we expected counting segments
1848 * in assumption of absent reordering, interpret this as reordering.
1849 * The only another reason could be bug in receiver TCP.
1850 */
1851static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1852{
1853 struct tcp_sock *tp = tcp_sk(sk);
1854 if (tcp_limit_reno_sacked(tp))
1855 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1856}
1857
1858/* Emulate SACKs for SACKless connection: account for a new dupack. */
1859
1860static void tcp_add_reno_sack(struct sock *sk)
1861{
1862 struct tcp_sock *tp = tcp_sk(sk);
1863 tp->sacked_out++;
1864 tcp_check_reno_reordering(sk, 0);
1865 tcp_verify_left_out(tp);
1866}
1867
1868/* Account for ACK, ACKing some data in Reno Recovery phase. */
1869
1870static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1871{
1872 struct tcp_sock *tp = tcp_sk(sk);
1873
1874 if (acked > 0) {
1875 /* One ACK acked hole. The rest eat duplicate ACKs. */
1876 if (acked - 1 >= tp->sacked_out)
1877 tp->sacked_out = 0;
1878 else
1879 tp->sacked_out -= acked - 1;
1880 }
1881 tcp_check_reno_reordering(sk, acked);
1882 tcp_verify_left_out(tp);
1883}
1884
1885static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1886{
1887 tp->sacked_out = 0;
1888}
1889
1890static void tcp_clear_retrans_partial(struct tcp_sock *tp)
1891{
1892 tp->retrans_out = 0;
1893 tp->lost_out = 0;
1894
1895 tp->undo_marker = 0;
1896 tp->undo_retrans = 0;
1897}
1898
1899void tcp_clear_retrans(struct tcp_sock *tp)
1900{
1901 tcp_clear_retrans_partial(tp);
1902
1903 tp->fackets_out = 0;
1904 tp->sacked_out = 0;
1905}
1906
1907/* Enter Loss state. If "how" is not zero, forget all SACK information
1908 * and reset tags completely, otherwise preserve SACKs. If receiver
1909 * dropped its ofo queue, we will know this due to reneging detection.
1910 */
1911void tcp_enter_loss(struct sock *sk, int how)
1912{
1913 const struct inet_connection_sock *icsk = inet_csk(sk);
1914 struct tcp_sock *tp = tcp_sk(sk);
1915 struct sk_buff *skb;
1916 bool new_recovery = false;
1917
1918 /* Reduce ssthresh if it has not yet been made inside this window. */
1919 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1920 !after(tp->high_seq, tp->snd_una) ||
1921 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1922 new_recovery = true;
1923 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1924 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1925 tcp_ca_event(sk, CA_EVENT_LOSS);
1926 }
1927 tp->snd_cwnd = 1;
1928 tp->snd_cwnd_cnt = 0;
1929 tp->snd_cwnd_stamp = tcp_time_stamp;
1930
1931 tcp_clear_retrans_partial(tp);
1932
1933 if (tcp_is_reno(tp))
1934 tcp_reset_reno_sack(tp);
1935
1936 tp->undo_marker = tp->snd_una;
1937 if (how) {
1938 tp->sacked_out = 0;
1939 tp->fackets_out = 0;
1940 }
1941 tcp_clear_all_retrans_hints(tp);
1942
1943 tcp_for_write_queue(skb, sk) {
1944 if (skb == tcp_send_head(sk))
1945 break;
1946
1947 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1948 tp->undo_marker = 0;
1949
1950 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1951 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1952 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1953 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1954 tp->lost_out += tcp_skb_pcount(skb);
1955 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
1956 }
1957 }
1958 tcp_verify_left_out(tp);
1959
1960 /* Timeout in disordered state after receiving substantial DUPACKs
1961 * suggests that the degree of reordering is over-estimated.
1962 */
1963 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
1964 tp->sacked_out >= sysctl_tcp_reordering)
1965 tp->reordering = min_t(unsigned int, tp->reordering,
1966 sysctl_tcp_reordering);
1967 tcp_set_ca_state(sk, TCP_CA_Loss);
1968 tp->high_seq = tp->snd_nxt;
1969 TCP_ECN_queue_cwr(tp);
1970
1971 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1972 * loss recovery is underway except recurring timeout(s) on
1973 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1974 */
1975 tp->frto = sysctl_tcp_frto &&
1976 (new_recovery || icsk->icsk_retransmits) &&
1977 !inet_csk(sk)->icsk_mtup.probe_size;
1978}
1979
1980/* If ACK arrived pointing to a remembered SACK, it means that our
1981 * remembered SACKs do not reflect real state of receiver i.e.
1982 * receiver _host_ is heavily congested (or buggy).
1983 *
1984 * Do processing similar to RTO timeout.
1985 */
1986static bool tcp_check_sack_reneging(struct sock *sk, int flag)
1987{
1988 if (flag & FLAG_SACK_RENEGING) {
1989 struct inet_connection_sock *icsk = inet_csk(sk);
1990 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1991
1992 tcp_enter_loss(sk, 1);
1993 icsk->icsk_retransmits++;
1994 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
1995 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1996 icsk->icsk_rto, TCP_RTO_MAX);
1997 return true;
1998 }
1999 return false;
2000}
2001
2002static inline int tcp_fackets_out(const struct tcp_sock *tp)
2003{
2004 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2005}
2006
2007/* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2008 * counter when SACK is enabled (without SACK, sacked_out is used for
2009 * that purpose).
2010 *
2011 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2012 * segments up to the highest received SACK block so far and holes in
2013 * between them.
2014 *
2015 * With reordering, holes may still be in flight, so RFC3517 recovery
2016 * uses pure sacked_out (total number of SACKed segments) even though
2017 * it violates the RFC that uses duplicate ACKs, often these are equal
2018 * but when e.g. out-of-window ACKs or packet duplication occurs,
2019 * they differ. Since neither occurs due to loss, TCP should really
2020 * ignore them.
2021 */
2022static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2023{
2024 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2025}
2026
2027static bool tcp_pause_early_retransmit(struct sock *sk, int flag)
2028{
2029 struct tcp_sock *tp = tcp_sk(sk);
2030 unsigned long delay;
2031
2032 /* Delay early retransmit and entering fast recovery for
2033 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
2034 * available, or RTO is scheduled to fire first.
2035 */
2036 if (sysctl_tcp_early_retrans < 2 || sysctl_tcp_early_retrans > 3 ||
2037 (flag & FLAG_ECE) || !tp->srtt_us)
2038 return false;
2039
2040 delay = max(usecs_to_jiffies(tp->srtt_us >> 5),
2041 msecs_to_jiffies(2));
2042
2043 if (!time_after(inet_csk(sk)->icsk_timeout, (jiffies + delay)))
2044 return false;
2045
2046 inet_csk_reset_xmit_timer(sk, ICSK_TIME_EARLY_RETRANS, delay,
2047 TCP_RTO_MAX);
2048 return true;
2049}
2050
2051/* Linux NewReno/SACK/FACK/ECN state machine.
2052 * --------------------------------------
2053 *
2054 * "Open" Normal state, no dubious events, fast path.
2055 * "Disorder" In all the respects it is "Open",
2056 * but requires a bit more attention. It is entered when
2057 * we see some SACKs or dupacks. It is split of "Open"
2058 * mainly to move some processing from fast path to slow one.
2059 * "CWR" CWND was reduced due to some Congestion Notification event.
2060 * It can be ECN, ICMP source quench, local device congestion.
2061 * "Recovery" CWND was reduced, we are fast-retransmitting.
2062 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2063 *
2064 * tcp_fastretrans_alert() is entered:
2065 * - each incoming ACK, if state is not "Open"
2066 * - when arrived ACK is unusual, namely:
2067 * * SACK
2068 * * Duplicate ACK.
2069 * * ECN ECE.
2070 *
2071 * Counting packets in flight is pretty simple.
2072 *
2073 * in_flight = packets_out - left_out + retrans_out
2074 *
2075 * packets_out is SND.NXT-SND.UNA counted in packets.
2076 *
2077 * retrans_out is number of retransmitted segments.
2078 *
2079 * left_out is number of segments left network, but not ACKed yet.
2080 *
2081 * left_out = sacked_out + lost_out
2082 *
2083 * sacked_out: Packets, which arrived to receiver out of order
2084 * and hence not ACKed. With SACKs this number is simply
2085 * amount of SACKed data. Even without SACKs
2086 * it is easy to give pretty reliable estimate of this number,
2087 * counting duplicate ACKs.
2088 *
2089 * lost_out: Packets lost by network. TCP has no explicit
2090 * "loss notification" feedback from network (for now).
2091 * It means that this number can be only _guessed_.
2092 * Actually, it is the heuristics to predict lossage that
2093 * distinguishes different algorithms.
2094 *
2095 * F.e. after RTO, when all the queue is considered as lost,
2096 * lost_out = packets_out and in_flight = retrans_out.
2097 *
2098 * Essentially, we have now two algorithms counting
2099 * lost packets.
2100 *
2101 * FACK: It is the simplest heuristics. As soon as we decided
2102 * that something is lost, we decide that _all_ not SACKed
2103 * packets until the most forward SACK are lost. I.e.
2104 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2105 * It is absolutely correct estimate, if network does not reorder
2106 * packets. And it loses any connection to reality when reordering
2107 * takes place. We use FACK by default until reordering
2108 * is suspected on the path to this destination.
2109 *
2110 * NewReno: when Recovery is entered, we assume that one segment
2111 * is lost (classic Reno). While we are in Recovery and
2112 * a partial ACK arrives, we assume that one more packet
2113 * is lost (NewReno). This heuristics are the same in NewReno
2114 * and SACK.
2115 *
2116 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2117 * deflation etc. CWND is real congestion window, never inflated, changes
2118 * only according to classic VJ rules.
2119 *
2120 * Really tricky (and requiring careful tuning) part of algorithm
2121 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2122 * The first determines the moment _when_ we should reduce CWND and,
2123 * hence, slow down forward transmission. In fact, it determines the moment
2124 * when we decide that hole is caused by loss, rather than by a reorder.
2125 *
2126 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2127 * holes, caused by lost packets.
2128 *
2129 * And the most logically complicated part of algorithm is undo
2130 * heuristics. We detect false retransmits due to both too early
2131 * fast retransmit (reordering) and underestimated RTO, analyzing
2132 * timestamps and D-SACKs. When we detect that some segments were
2133 * retransmitted by mistake and CWND reduction was wrong, we undo
2134 * window reduction and abort recovery phase. This logic is hidden
2135 * inside several functions named tcp_try_undo_<something>.
2136 */
2137
2138/* This function decides, when we should leave Disordered state
2139 * and enter Recovery phase, reducing congestion window.
2140 *
2141 * Main question: may we further continue forward transmission
2142 * with the same cwnd?
2143 */
2144static bool tcp_time_to_recover(struct sock *sk, int flag)
2145{
2146 struct tcp_sock *tp = tcp_sk(sk);
2147 __u32 packets_out;
2148
2149 /* Trick#1: The loss is proven. */
2150 if (tp->lost_out)
2151 return true;
2152
2153 /* Not-A-Trick#2 : Classic rule... */
2154 if (tcp_dupack_heuristics(tp) > tp->reordering)
2155 return true;
2156
2157 /* Trick#4: It is still not OK... But will it be useful to delay
2158 * recovery more?
2159 */
2160 packets_out = tp->packets_out;
2161 if (packets_out <= tp->reordering &&
2162 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2163 !tcp_may_send_now(sk)) {
2164 /* We have nothing to send. This connection is limited
2165 * either by receiver window or by application.
2166 */
2167 return true;
2168 }
2169
2170 /* If a thin stream is detected, retransmit after first
2171 * received dupack. Employ only if SACK is supported in order
2172 * to avoid possible corner-case series of spurious retransmissions
2173 * Use only if there are no unsent data.
2174 */
2175 if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2176 tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2177 tcp_is_sack(tp) && !tcp_send_head(sk))
2178 return true;
2179
2180 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious
2181 * retransmissions due to small network reorderings, we implement
2182 * Mitigation A.3 in the RFC and delay the retransmission for a short
2183 * interval if appropriate.
2184 */
2185 if (tp->do_early_retrans && !tp->retrans_out && tp->sacked_out &&
2186 (tp->packets_out >= (tp->sacked_out + 1) && tp->packets_out < 4) &&
2187 !tcp_may_send_now(sk))
2188 return !tcp_pause_early_retransmit(sk, flag);
2189
2190 return false;
2191}
2192
2193/* Detect loss in event "A" above by marking head of queue up as lost.
2194 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2195 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2196 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2197 * the maximum SACKed segments to pass before reaching this limit.
2198 */
2199static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2200{
2201 struct tcp_sock *tp = tcp_sk(sk);
2202 struct sk_buff *skb;
2203 int cnt, oldcnt;
2204 int err;
2205 unsigned int mss;
2206 /* Use SACK to deduce losses of new sequences sent during recovery */
2207 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
2208
2209 WARN_ON(packets > tp->packets_out);
2210 if (tp->lost_skb_hint) {
2211 skb = tp->lost_skb_hint;
2212 cnt = tp->lost_cnt_hint;
2213 /* Head already handled? */
2214 if (mark_head && skb != tcp_write_queue_head(sk))
2215 return;
2216 } else {
2217 skb = tcp_write_queue_head(sk);
2218 cnt = 0;
2219 }
2220
2221 tcp_for_write_queue_from(skb, sk) {
2222 if (skb == tcp_send_head(sk))
2223 break;
2224 /* TODO: do this better */
2225 /* this is not the most efficient way to do this... */
2226 tp->lost_skb_hint = skb;
2227 tp->lost_cnt_hint = cnt;
2228
2229 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2230 break;
2231
2232 oldcnt = cnt;
2233 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2234 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2235 cnt += tcp_skb_pcount(skb);
2236
2237 if (cnt > packets) {
2238 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2239 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2240 (oldcnt >= packets))
2241 break;
2242
2243 mss = skb_shinfo(skb)->gso_size;
2244 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2245 if (err < 0)
2246 break;
2247 cnt = packets;
2248 }
2249
2250 tcp_skb_mark_lost(tp, skb);
2251
2252 if (mark_head)
2253 break;
2254 }
2255 tcp_verify_left_out(tp);
2256}
2257
2258/* Account newly detected lost packet(s) */
2259
2260static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2261{
2262 struct tcp_sock *tp = tcp_sk(sk);
2263
2264 if (tcp_is_reno(tp)) {
2265 tcp_mark_head_lost(sk, 1, 1);
2266 } else if (tcp_is_fack(tp)) {
2267 int lost = tp->fackets_out - tp->reordering;
2268 if (lost <= 0)
2269 lost = 1;
2270 tcp_mark_head_lost(sk, lost, 0);
2271 } else {
2272 int sacked_upto = tp->sacked_out - tp->reordering;
2273 if (sacked_upto >= 0)
2274 tcp_mark_head_lost(sk, sacked_upto, 0);
2275 else if (fast_rexmit)
2276 tcp_mark_head_lost(sk, 1, 1);
2277 }
2278}
2279
2280/* CWND moderation, preventing bursts due to too big ACKs
2281 * in dubious situations.
2282 */
2283static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2284{
2285 tp->snd_cwnd = min(tp->snd_cwnd,
2286 tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2287 tp->snd_cwnd_stamp = tcp_time_stamp;
2288}
2289
2290/* Nothing was retransmitted or returned timestamp is less
2291 * than timestamp of the first retransmission.
2292 */
2293static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2294{
2295 return !tp->retrans_stamp ||
2296 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2297 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2298}
2299
2300/* Undo procedures. */
2301
2302#if FASTRETRANS_DEBUG > 1
2303static void DBGUNDO(struct sock *sk, const char *msg)
2304{
2305 struct tcp_sock *tp = tcp_sk(sk);
2306 struct inet_sock *inet = inet_sk(sk);
2307
2308 if (sk->sk_family == AF_INET) {
2309 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2310 msg,
2311 &inet->inet_daddr, ntohs(inet->inet_dport),
2312 tp->snd_cwnd, tcp_left_out(tp),
2313 tp->snd_ssthresh, tp->prior_ssthresh,
2314 tp->packets_out);
2315 }
2316#if IS_ENABLED(CONFIG_IPV6)
2317 else if (sk->sk_family == AF_INET6) {
2318 struct ipv6_pinfo *np = inet6_sk(sk);
2319 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2320 msg,
2321 &np->daddr, ntohs(inet->inet_dport),
2322 tp->snd_cwnd, tcp_left_out(tp),
2323 tp->snd_ssthresh, tp->prior_ssthresh,
2324 tp->packets_out);
2325 }
2326#endif
2327}
2328#else
2329#define DBGUNDO(x...) do { } while (0)
2330#endif
2331
2332static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2333{
2334 struct tcp_sock *tp = tcp_sk(sk);
2335
2336 if (unmark_loss) {
2337 struct sk_buff *skb;
2338
2339 tcp_for_write_queue(skb, sk) {
2340 if (skb == tcp_send_head(sk))
2341 break;
2342 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2343 }
2344 tp->lost_out = 0;
2345 tcp_clear_all_retrans_hints(tp);
2346 }
2347
2348 if (tp->prior_ssthresh) {
2349 const struct inet_connection_sock *icsk = inet_csk(sk);
2350
2351 if (icsk->icsk_ca_ops->undo_cwnd)
2352 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2353 else
2354 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2355
2356 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2357 tp->snd_ssthresh = tp->prior_ssthresh;
2358 TCP_ECN_withdraw_cwr(tp);
2359 }
2360 } else {
2361 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2362 }
2363 tp->snd_cwnd_stamp = tcp_time_stamp;
2364 tp->undo_marker = 0;
2365}
2366
2367static inline bool tcp_may_undo(const struct tcp_sock *tp)
2368{
2369 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2370}
2371
2372/* People celebrate: "We love our President!" */
2373static bool tcp_try_undo_recovery(struct sock *sk)
2374{
2375 struct tcp_sock *tp = tcp_sk(sk);
2376
2377 if (tcp_may_undo(tp)) {
2378 int mib_idx;
2379
2380 /* Happy end! We did not retransmit anything
2381 * or our original transmission succeeded.
2382 */
2383 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2384 tcp_undo_cwnd_reduction(sk, false);
2385 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2386 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2387 else
2388 mib_idx = LINUX_MIB_TCPFULLUNDO;
2389
2390 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2391 }
2392 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2393 /* Hold old state until something *above* high_seq
2394 * is ACKed. For Reno it is MUST to prevent false
2395 * fast retransmits (RFC2582). SACK TCP is safe. */
2396 tcp_moderate_cwnd(tp);
2397 return true;
2398 }
2399 tcp_set_ca_state(sk, TCP_CA_Open);
2400 return false;
2401}
2402
2403/* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2404static bool tcp_try_undo_dsack(struct sock *sk)
2405{
2406 struct tcp_sock *tp = tcp_sk(sk);
2407
2408 if (tp->undo_marker && !tp->undo_retrans) {
2409 DBGUNDO(sk, "D-SACK");
2410 tcp_undo_cwnd_reduction(sk, false);
2411 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2412 return true;
2413 }
2414 return false;
2415}
2416
2417/* We can clear retrans_stamp when there are no retransmissions in the
2418 * window. It would seem that it is trivially available for us in
2419 * tp->retrans_out, however, that kind of assumptions doesn't consider
2420 * what will happen if errors occur when sending retransmission for the
2421 * second time. ...It could the that such segment has only
2422 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2423 * the head skb is enough except for some reneging corner cases that
2424 * are not worth the effort.
2425 *
2426 * Main reason for all this complexity is the fact that connection dying
2427 * time now depends on the validity of the retrans_stamp, in particular,
2428 * that successive retransmissions of a segment must not advance
2429 * retrans_stamp under any conditions.
2430 */
2431static bool tcp_any_retrans_done(const struct sock *sk)
2432{
2433 const struct tcp_sock *tp = tcp_sk(sk);
2434 struct sk_buff *skb;
2435
2436 if (tp->retrans_out)
2437 return true;
2438
2439 skb = tcp_write_queue_head(sk);
2440 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2441 return true;
2442
2443 return false;
2444}
2445
2446/* Undo during loss recovery after partial ACK or using F-RTO. */
2447static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2448{
2449 struct tcp_sock *tp = tcp_sk(sk);
2450
2451 if (frto_undo || tcp_may_undo(tp)) {
2452 tcp_undo_cwnd_reduction(sk, true);
2453
2454 DBGUNDO(sk, "partial loss");
2455 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2456 if (frto_undo)
2457 NET_INC_STATS_BH(sock_net(sk),
2458 LINUX_MIB_TCPSPURIOUSRTOS);
2459 inet_csk(sk)->icsk_retransmits = 0;
2460 if (frto_undo || tcp_is_sack(tp))
2461 tcp_set_ca_state(sk, TCP_CA_Open);
2462 return true;
2463 }
2464 return false;
2465}
2466
2467/* The cwnd reduction in CWR and Recovery use the PRR algorithm
2468 * https://datatracker.ietf.org/doc/draft-ietf-tcpm-proportional-rate-reduction/
2469 * It computes the number of packets to send (sndcnt) based on packets newly
2470 * delivered:
2471 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2472 * cwnd reductions across a full RTT.
2473 * 2) If packets in flight is lower than ssthresh (such as due to excess
2474 * losses and/or application stalls), do not perform any further cwnd
2475 * reductions, but instead slow start up to ssthresh.
2476 */
2477static void tcp_init_cwnd_reduction(struct sock *sk, const bool set_ssthresh)
2478{
2479 struct tcp_sock *tp = tcp_sk(sk);
2480
2481 tp->high_seq = tp->snd_nxt;
2482 tp->tlp_high_seq = 0;
2483 tp->snd_cwnd_cnt = 0;
2484 tp->prior_cwnd = tp->snd_cwnd;
2485 tp->prr_delivered = 0;
2486 tp->prr_out = 0;
2487 if (set_ssthresh)
2488 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2489 TCP_ECN_queue_cwr(tp);
2490}
2491
2492static void tcp_cwnd_reduction(struct sock *sk, const int prior_unsacked,
2493 int fast_rexmit)
2494{
2495 struct tcp_sock *tp = tcp_sk(sk);
2496 int sndcnt = 0;
2497 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2498 int newly_acked_sacked = prior_unsacked -
2499 (tp->packets_out - tp->sacked_out);
2500
2501 tp->prr_delivered += newly_acked_sacked;
2502 if (tcp_packets_in_flight(tp) > tp->snd_ssthresh) {
2503 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2504 tp->prior_cwnd - 1;
2505 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2506 } else {
2507 sndcnt = min_t(int, delta,
2508 max_t(int, tp->prr_delivered - tp->prr_out,
2509 newly_acked_sacked) + 1);
2510 }
2511
2512 sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0));
2513 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2514}
2515
2516static inline void tcp_end_cwnd_reduction(struct sock *sk)
2517{
2518 struct tcp_sock *tp = tcp_sk(sk);
2519
2520 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2521 if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR ||
2522 (tp->undo_marker && tp->snd_ssthresh < TCP_INFINITE_SSTHRESH)) {
2523 tp->snd_cwnd = tp->snd_ssthresh;
2524 tp->snd_cwnd_stamp = tcp_time_stamp;
2525 }
2526 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2527}
2528
2529/* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2530void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
2531{
2532 struct tcp_sock *tp = tcp_sk(sk);
2533
2534 tp->prior_ssthresh = 0;
2535 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2536 tp->undo_marker = 0;
2537 tcp_init_cwnd_reduction(sk, set_ssthresh);
2538 tcp_set_ca_state(sk, TCP_CA_CWR);
2539 }
2540}
2541
2542static void tcp_try_keep_open(struct sock *sk)
2543{
2544 struct tcp_sock *tp = tcp_sk(sk);
2545 int state = TCP_CA_Open;
2546
2547 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2548 state = TCP_CA_Disorder;
2549
2550 if (inet_csk(sk)->icsk_ca_state != state) {
2551 tcp_set_ca_state(sk, state);
2552 tp->high_seq = tp->snd_nxt;
2553 }
2554}
2555
2556static void tcp_try_to_open(struct sock *sk, int flag, const int prior_unsacked)
2557{
2558 struct tcp_sock *tp = tcp_sk(sk);
2559
2560 tcp_verify_left_out(tp);
2561
2562 if (!tcp_any_retrans_done(sk))
2563 tp->retrans_stamp = 0;
2564
2565 if (flag & FLAG_ECE)
2566 tcp_enter_cwr(sk, 1);
2567
2568 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2569 tcp_try_keep_open(sk);
2570 } else {
2571 tcp_cwnd_reduction(sk, prior_unsacked, 0);
2572 }
2573}
2574
2575static void tcp_mtup_probe_failed(struct sock *sk)
2576{
2577 struct inet_connection_sock *icsk = inet_csk(sk);
2578
2579 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2580 icsk->icsk_mtup.probe_size = 0;
2581}
2582
2583static void tcp_mtup_probe_success(struct sock *sk)
2584{
2585 struct tcp_sock *tp = tcp_sk(sk);
2586 struct inet_connection_sock *icsk = inet_csk(sk);
2587
2588 /* FIXME: breaks with very large cwnd */
2589 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2590 tp->snd_cwnd = tp->snd_cwnd *
2591 tcp_mss_to_mtu(sk, tp->mss_cache) /
2592 icsk->icsk_mtup.probe_size;
2593 tp->snd_cwnd_cnt = 0;
2594 tp->snd_cwnd_stamp = tcp_time_stamp;
2595 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2596
2597 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2598 icsk->icsk_mtup.probe_size = 0;
2599 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2600}
2601
2602/* Do a simple retransmit without using the backoff mechanisms in
2603 * tcp_timer. This is used for path mtu discovery.
2604 * The socket is already locked here.
2605 */
2606void tcp_simple_retransmit(struct sock *sk)
2607{
2608 const struct inet_connection_sock *icsk = inet_csk(sk);
2609 struct tcp_sock *tp = tcp_sk(sk);
2610 struct sk_buff *skb;
2611 unsigned int mss = tcp_current_mss(sk);
2612 u32 prior_lost = tp->lost_out;
2613
2614 tcp_for_write_queue(skb, sk) {
2615 if (skb == tcp_send_head(sk))
2616 break;
2617 if (tcp_skb_seglen(skb) > mss &&
2618 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2619 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2620 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2621 tp->retrans_out -= tcp_skb_pcount(skb);
2622 }
2623 tcp_skb_mark_lost_uncond_verify(tp, skb);
2624 }
2625 }
2626
2627 tcp_clear_retrans_hints_partial(tp);
2628
2629 if (prior_lost == tp->lost_out)
2630 return;
2631
2632 if (tcp_is_reno(tp))
2633 tcp_limit_reno_sacked(tp);
2634
2635 tcp_verify_left_out(tp);
2636
2637 /* Don't muck with the congestion window here.
2638 * Reason is that we do not increase amount of _data_
2639 * in network, but units changed and effective
2640 * cwnd/ssthresh really reduced now.
2641 */
2642 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2643 tp->high_seq = tp->snd_nxt;
2644 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2645 tp->prior_ssthresh = 0;
2646 tp->undo_marker = 0;
2647 tcp_set_ca_state(sk, TCP_CA_Loss);
2648 }
2649 tcp_xmit_retransmit_queue(sk);
2650}
2651EXPORT_SYMBOL(tcp_simple_retransmit);
2652
2653static void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2654{
2655 struct tcp_sock *tp = tcp_sk(sk);
2656 int mib_idx;
2657
2658 if (tcp_is_reno(tp))
2659 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2660 else
2661 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2662
2663 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2664
2665 tp->prior_ssthresh = 0;
2666 tp->undo_marker = tp->snd_una;
2667 tp->undo_retrans = tp->retrans_out;
2668
2669 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2670 if (!ece_ack)
2671 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2672 tcp_init_cwnd_reduction(sk, true);
2673 }
2674 tcp_set_ca_state(sk, TCP_CA_Recovery);
2675}
2676
2677/* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2678 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2679 */
2680static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack)
2681{
2682 struct inet_connection_sock *icsk = inet_csk(sk);
2683 struct tcp_sock *tp = tcp_sk(sk);
2684 bool recovered = !before(tp->snd_una, tp->high_seq);
2685
2686 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2687 /* Step 3.b. A timeout is spurious if not all data are
2688 * lost, i.e., never-retransmitted data are (s)acked.
2689 */
2690 if (tcp_try_undo_loss(sk, flag & FLAG_ORIG_SACK_ACKED))
2691 return;
2692
2693 if (after(tp->snd_nxt, tp->high_seq) &&
2694 (flag & FLAG_DATA_SACKED || is_dupack)) {
2695 tp->frto = 0; /* Loss was real: 2nd part of step 3.a */
2696 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2697 tp->high_seq = tp->snd_nxt;
2698 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
2699 TCP_NAGLE_OFF);
2700 if (after(tp->snd_nxt, tp->high_seq))
2701 return; /* Step 2.b */
2702 tp->frto = 0;
2703 }
2704 }
2705
2706 if (recovered) {
2707 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2708 icsk->icsk_retransmits = 0;
2709 tcp_try_undo_recovery(sk);
2710 return;
2711 }
2712 if (flag & FLAG_DATA_ACKED)
2713 icsk->icsk_retransmits = 0;
2714 if (tcp_is_reno(tp)) {
2715 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2716 * delivered. Lower inflight to clock out (re)tranmissions.
2717 */
2718 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2719 tcp_add_reno_sack(sk);
2720 else if (flag & FLAG_SND_UNA_ADVANCED)
2721 tcp_reset_reno_sack(tp);
2722 }
2723 if (tcp_try_undo_loss(sk, false))
2724 return;
2725 tcp_xmit_retransmit_queue(sk);
2726}
2727
2728/* Undo during fast recovery after partial ACK. */
2729static bool tcp_try_undo_partial(struct sock *sk, const int acked,
2730 const int prior_unsacked)
2731{
2732 struct tcp_sock *tp = tcp_sk(sk);
2733
2734 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2735 /* Plain luck! Hole if filled with delayed
2736 * packet, rather than with a retransmit.
2737 */
2738 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2739
2740 /* We are getting evidence that the reordering degree is higher
2741 * than we realized. If there are no retransmits out then we
2742 * can undo. Otherwise we clock out new packets but do not
2743 * mark more packets lost or retransmit more.
2744 */
2745 if (tp->retrans_out) {
2746 tcp_cwnd_reduction(sk, prior_unsacked, 0);
2747 return true;
2748 }
2749
2750 if (!tcp_any_retrans_done(sk))
2751 tp->retrans_stamp = 0;
2752
2753 DBGUNDO(sk, "partial recovery");
2754 tcp_undo_cwnd_reduction(sk, true);
2755 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2756 tcp_try_keep_open(sk);
2757 return true;
2758 }
2759 return false;
2760}
2761
2762/* Process an event, which can update packets-in-flight not trivially.
2763 * Main goal of this function is to calculate new estimate for left_out,
2764 * taking into account both packets sitting in receiver's buffer and
2765 * packets lost by network.
2766 *
2767 * Besides that it does CWND reduction, when packet loss is detected
2768 * and changes state of machine.
2769 *
2770 * It does _not_ decide what to send, it is made in function
2771 * tcp_xmit_retransmit_queue().
2772 */
2773static void tcp_fastretrans_alert(struct sock *sk, const int acked,
2774 const int prior_unsacked,
2775 bool is_dupack, int flag)
2776{
2777 struct inet_connection_sock *icsk = inet_csk(sk);
2778 struct tcp_sock *tp = tcp_sk(sk);
2779 bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2780 (tcp_fackets_out(tp) > tp->reordering));
2781 int fast_rexmit = 0;
2782
2783 if (WARN_ON(!tp->packets_out && tp->sacked_out))
2784 tp->sacked_out = 0;
2785 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2786 tp->fackets_out = 0;
2787
2788 /* Now state machine starts.
2789 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2790 if (flag & FLAG_ECE)
2791 tp->prior_ssthresh = 0;
2792
2793 /* B. In all the states check for reneging SACKs. */
2794 if (tcp_check_sack_reneging(sk, flag))
2795 return;
2796
2797 /* C. Check consistency of the current state. */
2798 tcp_verify_left_out(tp);
2799
2800 /* D. Check state exit conditions. State can be terminated
2801 * when high_seq is ACKed. */
2802 if (icsk->icsk_ca_state == TCP_CA_Open) {
2803 WARN_ON(tp->retrans_out != 0);
2804 tp->retrans_stamp = 0;
2805 } else if (!before(tp->snd_una, tp->high_seq)) {
2806 switch (icsk->icsk_ca_state) {
2807 case TCP_CA_CWR:
2808 /* CWR is to be held something *above* high_seq
2809 * is ACKed for CWR bit to reach receiver. */
2810 if (tp->snd_una != tp->high_seq) {
2811 tcp_end_cwnd_reduction(sk);
2812 tcp_set_ca_state(sk, TCP_CA_Open);
2813 }
2814 break;
2815
2816 case TCP_CA_Recovery:
2817 if (tcp_is_reno(tp))
2818 tcp_reset_reno_sack(tp);
2819 if (tcp_try_undo_recovery(sk))
2820 return;
2821 tcp_end_cwnd_reduction(sk);
2822 break;
2823 }
2824 }
2825
2826 /* E. Process state. */
2827 switch (icsk->icsk_ca_state) {
2828 case TCP_CA_Recovery:
2829 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2830 if (tcp_is_reno(tp) && is_dupack)
2831 tcp_add_reno_sack(sk);
2832 } else {
2833 if (tcp_try_undo_partial(sk, acked, prior_unsacked))
2834 return;
2835 /* Partial ACK arrived. Force fast retransmit. */
2836 do_lost = tcp_is_reno(tp) ||
2837 tcp_fackets_out(tp) > tp->reordering;
2838 }
2839 if (tcp_try_undo_dsack(sk)) {
2840 tcp_try_keep_open(sk);
2841 return;
2842 }
2843 break;
2844 case TCP_CA_Loss:
2845 tcp_process_loss(sk, flag, is_dupack);
2846 if (icsk->icsk_ca_state != TCP_CA_Open)
2847 return;
2848 /* Fall through to processing in Open state. */
2849 default:
2850 if (tcp_is_reno(tp)) {
2851 if (flag & FLAG_SND_UNA_ADVANCED)
2852 tcp_reset_reno_sack(tp);
2853 if (is_dupack)
2854 tcp_add_reno_sack(sk);
2855 }
2856
2857 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2858 tcp_try_undo_dsack(sk);
2859
2860 if (!tcp_time_to_recover(sk, flag)) {
2861 tcp_try_to_open(sk, flag, prior_unsacked);
2862 return;
2863 }
2864
2865 /* MTU probe failure: don't reduce cwnd */
2866 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2867 icsk->icsk_mtup.probe_size &&
2868 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2869 tcp_mtup_probe_failed(sk);
2870 /* Restores the reduction we did in tcp_mtup_probe() */
2871 tp->snd_cwnd++;
2872 tcp_simple_retransmit(sk);
2873 return;
2874 }
2875
2876 /* Otherwise enter Recovery state */
2877 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2878 fast_rexmit = 1;
2879 }
2880
2881 if (do_lost)
2882 tcp_update_scoreboard(sk, fast_rexmit);
2883 tcp_cwnd_reduction(sk, prior_unsacked, fast_rexmit);
2884 tcp_xmit_retransmit_queue(sk);
2885}
2886
2887static inline bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2888 long seq_rtt_us, long sack_rtt_us)
2889{
2890 const struct tcp_sock *tp = tcp_sk(sk);
2891
2892 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2893 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2894 * Karn's algorithm forbids taking RTT if some retransmitted data
2895 * is acked (RFC6298).
2896 */
2897 if (flag & FLAG_RETRANS_DATA_ACKED)
2898 seq_rtt_us = -1L;
2899
2900 if (seq_rtt_us < 0)
2901 seq_rtt_us = sack_rtt_us;
2902
2903 /* RTTM Rule: A TSecr value received in a segment is used to
2904 * update the averaged RTT measurement only if the segment
2905 * acknowledges some new data, i.e., only if it advances the
2906 * left edge of the send window.
2907 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2908 */
2909 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2910 flag & FLAG_ACKED)
2911 seq_rtt_us = jiffies_to_usecs(tcp_time_stamp - tp->rx_opt.rcv_tsecr);
2912
2913 if (seq_rtt_us < 0)
2914 return false;
2915
2916 tcp_rtt_estimator(sk, seq_rtt_us);
2917 tcp_set_rto(sk);
2918
2919 /* RFC6298: only reset backoff on valid RTT measurement. */
2920 inet_csk(sk)->icsk_backoff = 0;
2921 return true;
2922}
2923
2924/* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2925static void tcp_synack_rtt_meas(struct sock *sk, const u32 synack_stamp)
2926{
2927 struct tcp_sock *tp = tcp_sk(sk);
2928 long seq_rtt_us = -1L;
2929
2930 if (synack_stamp && !tp->total_retrans)
2931 seq_rtt_us = jiffies_to_usecs(tcp_time_stamp - synack_stamp);
2932
2933 /* If the ACK acks both the SYNACK and the (Fast Open'd) data packets
2934 * sent in SYN_RECV, SYNACK RTT is the smooth RTT computed in tcp_ack()
2935 */
2936 if (!tp->srtt_us)
2937 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, seq_rtt_us, -1L);
2938}
2939
2940static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked, u32 in_flight)
2941{
2942 const struct inet_connection_sock *icsk = inet_csk(sk);
2943 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked, in_flight);
2944 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2945}
2946
2947/* Restart timer after forward progress on connection.
2948 * RFC2988 recommends to restart timer to now+rto.
2949 */
2950void tcp_rearm_rto(struct sock *sk)
2951{
2952 const struct inet_connection_sock *icsk = inet_csk(sk);
2953 struct tcp_sock *tp = tcp_sk(sk);
2954
2955 /* If the retrans timer is currently being used by Fast Open
2956 * for SYN-ACK retrans purpose, stay put.
2957 */
2958 if (tp->fastopen_rsk)
2959 return;
2960
2961 if (!tp->packets_out) {
2962 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2963 } else {
2964 u32 rto = inet_csk(sk)->icsk_rto;
2965 /* Offset the time elapsed after installing regular RTO */
2966 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
2967 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
2968 struct sk_buff *skb = tcp_write_queue_head(sk);
2969 const u32 rto_time_stamp = TCP_SKB_CB(skb)->when + rto;
2970 s32 delta = (s32)(rto_time_stamp - tcp_time_stamp);
2971 /* delta may not be positive if the socket is locked
2972 * when the retrans timer fires and is rescheduled.
2973 */
2974 if (delta > 0)
2975 rto = delta;
2976 }
2977 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
2978 TCP_RTO_MAX);
2979 }
2980}
2981
2982/* This function is called when the delayed ER timer fires. TCP enters
2983 * fast recovery and performs fast-retransmit.
2984 */
2985void tcp_resume_early_retransmit(struct sock *sk)
2986{
2987 struct tcp_sock *tp = tcp_sk(sk);
2988
2989 tcp_rearm_rto(sk);
2990
2991 /* Stop if ER is disabled after the delayed ER timer is scheduled */
2992 if (!tp->do_early_retrans)
2993 return;
2994
2995 tcp_enter_recovery(sk, false);
2996 tcp_update_scoreboard(sk, 1);
2997 tcp_xmit_retransmit_queue(sk);
2998}
2999
3000/* If we get here, the whole TSO packet has not been acked. */
3001static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3002{
3003 struct tcp_sock *tp = tcp_sk(sk);
3004 u32 packets_acked;
3005
3006 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3007
3008 packets_acked = tcp_skb_pcount(skb);
3009 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3010 return 0;
3011 packets_acked -= tcp_skb_pcount(skb);
3012
3013 if (packets_acked) {
3014 BUG_ON(tcp_skb_pcount(skb) == 0);
3015 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3016 }
3017
3018 return packets_acked;
3019}
3020
3021/* Remove acknowledged frames from the retransmission queue. If our packet
3022 * is before the ack sequence we can discard it as it's confirmed to have
3023 * arrived at the other end.
3024 */
3025static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3026 u32 prior_snd_una, long sack_rtt_us)
3027{
3028 const struct inet_connection_sock *icsk = inet_csk(sk);
3029 struct skb_mstamp first_ackt, last_ackt, now;
3030 struct tcp_sock *tp = tcp_sk(sk);
3031 u32 prior_sacked = tp->sacked_out;
3032 u32 reord = tp->packets_out;
3033 bool fully_acked = true;
3034 long ca_seq_rtt_us = -1L;
3035 long seq_rtt_us = -1L;
3036 struct sk_buff *skb;
3037 u32 pkts_acked = 0;
3038 bool rtt_update;
3039 int flag = 0;
3040
3041 first_ackt.v64 = 0;
3042
3043 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3044 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3045 u8 sacked = scb->sacked;
3046 u32 acked_pcount;
3047
3048 /* Determine how many packets and what bytes were acked, tso and else */
3049 if (after(scb->end_seq, tp->snd_una)) {
3050 if (tcp_skb_pcount(skb) == 1 ||
3051 !after(tp->snd_una, scb->seq))
3052 break;
3053
3054 acked_pcount = tcp_tso_acked(sk, skb);
3055 if (!acked_pcount)
3056 break;
3057
3058 fully_acked = false;
3059 } else {
3060 acked_pcount = tcp_skb_pcount(skb);
3061 }
3062
3063 if (sacked & TCPCB_RETRANS) {
3064 if (sacked & TCPCB_SACKED_RETRANS)
3065 tp->retrans_out -= acked_pcount;
3066 flag |= FLAG_RETRANS_DATA_ACKED;
3067 } else {
3068 last_ackt = skb->skb_mstamp;
3069 WARN_ON_ONCE(last_ackt.v64 == 0);
3070 if (!first_ackt.v64)
3071 first_ackt = last_ackt;
3072
3073 if (!(sacked & TCPCB_SACKED_ACKED))
3074 reord = min(pkts_acked, reord);
3075 if (!after(scb->end_seq, tp->high_seq))
3076 flag |= FLAG_ORIG_SACK_ACKED;
3077 }
3078
3079 if (sacked & TCPCB_SACKED_ACKED)
3080 tp->sacked_out -= acked_pcount;
3081 if (sacked & TCPCB_LOST)
3082 tp->lost_out -= acked_pcount;
3083
3084 tp->packets_out -= acked_pcount;
3085 pkts_acked += acked_pcount;
3086
3087 /* Initial outgoing SYN's get put onto the write_queue
3088 * just like anything else we transmit. It is not
3089 * true data, and if we misinform our callers that
3090 * this ACK acks real data, we will erroneously exit
3091 * connection startup slow start one packet too
3092 * quickly. This is severely frowned upon behavior.
3093 */
3094 if (!(scb->tcp_flags & TCPHDR_SYN)) {
3095 flag |= FLAG_DATA_ACKED;
3096 } else {
3097 flag |= FLAG_SYN_ACKED;
3098 tp->retrans_stamp = 0;
3099 }
3100
3101 if (!fully_acked)
3102 break;
3103
3104 tcp_unlink_write_queue(skb, sk);
3105 sk_wmem_free_skb(sk, skb);
3106 if (skb == tp->retransmit_skb_hint)
3107 tp->retransmit_skb_hint = NULL;
3108 if (skb == tp->lost_skb_hint)
3109 tp->lost_skb_hint = NULL;
3110 }
3111
3112 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3113 tp->snd_up = tp->snd_una;
3114
3115 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3116 flag |= FLAG_SACK_RENEGING;
3117
3118 skb_mstamp_get(&now);
3119 if (first_ackt.v64) {
3120 seq_rtt_us = skb_mstamp_us_delta(&now, &first_ackt);
3121 ca_seq_rtt_us = skb_mstamp_us_delta(&now, &last_ackt);
3122 }
3123
3124 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us);
3125
3126 if (flag & FLAG_ACKED) {
3127 const struct tcp_congestion_ops *ca_ops
3128 = inet_csk(sk)->icsk_ca_ops;
3129
3130 tcp_rearm_rto(sk);
3131 if (unlikely(icsk->icsk_mtup.probe_size &&
3132 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3133 tcp_mtup_probe_success(sk);
3134 }
3135
3136 if (tcp_is_reno(tp)) {
3137 tcp_remove_reno_sacks(sk, pkts_acked);
3138 } else {
3139 int delta;
3140
3141 /* Non-retransmitted hole got filled? That's reordering */
3142 if (reord < prior_fackets)
3143 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3144
3145 delta = tcp_is_fack(tp) ? pkts_acked :
3146 prior_sacked - tp->sacked_out;
3147 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3148 }
3149
3150 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3151
3152 if (ca_ops->pkts_acked)
3153 ca_ops->pkts_acked(sk, pkts_acked, ca_seq_rtt_us);
3154
3155 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3156 sack_rtt_us > skb_mstamp_us_delta(&now, &skb->skb_mstamp)) {
3157 /* Do not re-arm RTO if the sack RTT is measured from data sent
3158 * after when the head was last (re)transmitted. Otherwise the
3159 * timeout may continue to extend in loss recovery.
3160 */
3161 tcp_rearm_rto(sk);
3162 }
3163
3164#if FASTRETRANS_DEBUG > 0
3165 WARN_ON((int)tp->sacked_out < 0);
3166 WARN_ON((int)tp->lost_out < 0);
3167 WARN_ON((int)tp->retrans_out < 0);
3168 if (!tp->packets_out && tcp_is_sack(tp)) {
3169 icsk = inet_csk(sk);
3170 if (tp->lost_out) {
3171 pr_debug("Leak l=%u %d\n",
3172 tp->lost_out, icsk->icsk_ca_state);
3173 tp->lost_out = 0;
3174 }
3175 if (tp->sacked_out) {
3176 pr_debug("Leak s=%u %d\n",
3177 tp->sacked_out, icsk->icsk_ca_state);
3178 tp->sacked_out = 0;
3179 }
3180 if (tp->retrans_out) {
3181 pr_debug("Leak r=%u %d\n",
3182 tp->retrans_out, icsk->icsk_ca_state);
3183 tp->retrans_out = 0;
3184 }
3185 }
3186#endif
3187 return flag;
3188}
3189
3190static void tcp_ack_probe(struct sock *sk)
3191{
3192 const struct tcp_sock *tp = tcp_sk(sk);
3193 struct inet_connection_sock *icsk = inet_csk(sk);
3194
3195 /* Was it a usable window open? */
3196
3197 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3198 icsk->icsk_backoff = 0;
3199 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3200 /* Socket must be waked up by subsequent tcp_data_snd_check().
3201 * This function is not for random using!
3202 */
3203 } else {
3204 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3205 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3206 TCP_RTO_MAX);
3207 }
3208}
3209
3210static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3211{
3212 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3213 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3214}
3215
3216/* Decide wheather to run the increase function of congestion control. */
3217static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3218{
3219 if (tcp_in_cwnd_reduction(sk))
3220 return false;
3221
3222 /* If reordering is high then always grow cwnd whenever data is
3223 * delivered regardless of its ordering. Otherwise stay conservative
3224 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3225 * new SACK or ECE mark may first advance cwnd here and later reduce
3226 * cwnd in tcp_fastretrans_alert() based on more states.
3227 */
3228 if (tcp_sk(sk)->reordering > sysctl_tcp_reordering)
3229 return flag & FLAG_FORWARD_PROGRESS;
3230
3231 return flag & FLAG_DATA_ACKED;
3232}
3233
3234/* Check that window update is acceptable.
3235 * The function assumes that snd_una<=ack<=snd_next.
3236 */
3237static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3238 const u32 ack, const u32 ack_seq,
3239 const u32 nwin)
3240{
3241 return after(ack, tp->snd_una) ||
3242 after(ack_seq, tp->snd_wl1) ||
3243 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3244}
3245
3246/* Update our send window.
3247 *
3248 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3249 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3250 */
3251static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3252 u32 ack_seq)
3253{
3254 struct tcp_sock *tp = tcp_sk(sk);
3255 int flag = 0;
3256 u32 nwin = ntohs(tcp_hdr(skb)->window);
3257
3258 if (likely(!tcp_hdr(skb)->syn))
3259 nwin <<= tp->rx_opt.snd_wscale;
3260
3261 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3262 flag |= FLAG_WIN_UPDATE;
3263 tcp_update_wl(tp, ack_seq);
3264
3265 if (tp->snd_wnd != nwin) {
3266 tp->snd_wnd = nwin;
3267
3268 /* Note, it is the only place, where
3269 * fast path is recovered for sending TCP.
3270 */
3271 tp->pred_flags = 0;
3272 tcp_fast_path_check(sk);
3273
3274 if (nwin > tp->max_window) {
3275 tp->max_window = nwin;
3276 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3277 }
3278 }
3279 }
3280
3281 tp->snd_una = ack;
3282
3283 return flag;
3284}
3285
3286/* RFC 5961 7 [ACK Throttling] */
3287static void tcp_send_challenge_ack(struct sock *sk)
3288{
3289 /* unprotected vars, we dont care of overwrites */
3290 static u32 challenge_timestamp;
3291 static unsigned int challenge_count;
3292 u32 now = jiffies / HZ;
3293
3294 if (now != challenge_timestamp) {
3295 challenge_timestamp = now;
3296 challenge_count = 0;
3297 }
3298 if (++challenge_count <= sysctl_tcp_challenge_ack_limit) {
3299 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK);
3300 tcp_send_ack(sk);
3301 }
3302}
3303
3304static void tcp_store_ts_recent(struct tcp_sock *tp)
3305{
3306 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3307 tp->rx_opt.ts_recent_stamp = get_seconds();
3308}
3309
3310static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3311{
3312 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3313 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3314 * extra check below makes sure this can only happen
3315 * for pure ACK frames. -DaveM
3316 *
3317 * Not only, also it occurs for expired timestamps.
3318 */
3319
3320 if (tcp_paws_check(&tp->rx_opt, 0))
3321 tcp_store_ts_recent(tp);
3322 }
3323}
3324
3325/* This routine deals with acks during a TLP episode.
3326 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3327 */
3328static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3329{
3330 struct tcp_sock *tp = tcp_sk(sk);
3331 bool is_tlp_dupack = (ack == tp->tlp_high_seq) &&
3332 !(flag & (FLAG_SND_UNA_ADVANCED |
3333 FLAG_NOT_DUP | FLAG_DATA_SACKED));
3334
3335 /* Mark the end of TLP episode on receiving TLP dupack or when
3336 * ack is after tlp_high_seq.
3337 */
3338 if (is_tlp_dupack) {
3339 tp->tlp_high_seq = 0;
3340 return;
3341 }
3342
3343 if (after(ack, tp->tlp_high_seq)) {
3344 tp->tlp_high_seq = 0;
3345 /* Don't reduce cwnd if DSACK arrives for TLP retrans. */
3346 if (!(flag & FLAG_DSACKING_ACK)) {
3347 tcp_init_cwnd_reduction(sk, true);
3348 tcp_set_ca_state(sk, TCP_CA_CWR);
3349 tcp_end_cwnd_reduction(sk);
3350 tcp_try_keep_open(sk);
3351 NET_INC_STATS_BH(sock_net(sk),
3352 LINUX_MIB_TCPLOSSPROBERECOVERY);
3353 }
3354 }
3355}
3356
3357/* This routine deals with incoming acks, but not outgoing ones. */
3358static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3359{
3360 struct inet_connection_sock *icsk = inet_csk(sk);
3361 struct tcp_sock *tp = tcp_sk(sk);
3362 u32 prior_snd_una = tp->snd_una;
3363 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3364 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3365 bool is_dupack = false;
3366 u32 prior_in_flight;
3367 u32 prior_fackets;
3368 int prior_packets = tp->packets_out;
3369 const int prior_unsacked = tp->packets_out - tp->sacked_out;
3370 int acked = 0; /* Number of packets newly acked */
3371 long sack_rtt_us = -1L;
3372
3373 /* If the ack is older than previous acks
3374 * then we can probably ignore it.
3375 */
3376 if (before(ack, prior_snd_una)) {
3377 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3378 if (before(ack, prior_snd_una - tp->max_window)) {
3379 tcp_send_challenge_ack(sk);
3380 return -1;
3381 }
3382 goto old_ack;
3383 }
3384
3385 /* If the ack includes data we haven't sent yet, discard
3386 * this segment (RFC793 Section 3.9).
3387 */
3388 if (after(ack, tp->snd_nxt))
3389 goto invalid_ack;
3390
3391 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
3392 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
3393 tcp_rearm_rto(sk);
3394
3395 if (after(ack, prior_snd_una))
3396 flag |= FLAG_SND_UNA_ADVANCED;
3397
3398 prior_fackets = tp->fackets_out;
3399 prior_in_flight = tcp_packets_in_flight(tp);
3400
3401 /* ts_recent update must be made after we are sure that the packet
3402 * is in window.
3403 */
3404 if (flag & FLAG_UPDATE_TS_RECENT)
3405 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3406
3407 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3408 /* Window is constant, pure forward advance.
3409 * No more checks are required.
3410 * Note, we use the fact that SND.UNA>=SND.WL2.
3411 */
3412 tcp_update_wl(tp, ack_seq);
3413 tp->snd_una = ack;
3414 flag |= FLAG_WIN_UPDATE;
3415
3416 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3417
3418 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3419 } else {
3420 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3421 flag |= FLAG_DATA;
3422 else
3423 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3424
3425 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3426
3427 if (TCP_SKB_CB(skb)->sacked)
3428 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3429 &sack_rtt_us);
3430
3431 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3432 flag |= FLAG_ECE;
3433
3434 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3435 }
3436
3437 /* We passed data and got it acked, remove any soft error
3438 * log. Something worked...
3439 */
3440 sk->sk_err_soft = 0;
3441 icsk->icsk_probes_out = 0;
3442 tp->rcv_tstamp = tcp_time_stamp;
3443 if (!prior_packets)
3444 goto no_queue;
3445
3446 /* See if we can take anything off of the retransmit queue. */
3447 acked = tp->packets_out;
3448 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una,
3449 sack_rtt_us);
3450 acked -= tp->packets_out;
3451
3452 /* Advance cwnd if state allows */
3453 if (tcp_may_raise_cwnd(sk, flag))
3454 tcp_cong_avoid(sk, ack, acked, prior_in_flight);
3455
3456 if (tcp_ack_is_dubious(sk, flag)) {
3457 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3458 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3459 is_dupack, flag);
3460 }
3461 if (tp->tlp_high_seq)
3462 tcp_process_tlp_ack(sk, ack, flag);
3463
3464 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) {
3465 struct dst_entry *dst = __sk_dst_get(sk);
3466 if (dst)
3467 dst_confirm(dst);
3468 }
3469
3470 if (icsk->icsk_pending == ICSK_TIME_RETRANS)
3471 tcp_schedule_loss_probe(sk);
3472 tcp_update_pacing_rate(sk);
3473 return 1;
3474
3475no_queue:
3476 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3477 if (flag & FLAG_DSACKING_ACK)
3478 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3479 is_dupack, flag);
3480 /* If this ack opens up a zero window, clear backoff. It was
3481 * being used to time the probes, and is probably far higher than
3482 * it needs to be for normal retransmission.
3483 */
3484 if (tcp_send_head(sk))
3485 tcp_ack_probe(sk);
3486
3487 if (tp->tlp_high_seq)
3488 tcp_process_tlp_ack(sk, ack, flag);
3489 return 1;
3490
3491invalid_ack:
3492 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3493 return -1;
3494
3495old_ack:
3496 /* If data was SACKed, tag it and see if we should send more data.
3497 * If data was DSACKed, see if we can undo a cwnd reduction.
3498 */
3499 if (TCP_SKB_CB(skb)->sacked) {
3500 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3501 &sack_rtt_us);
3502 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3503 is_dupack, flag);
3504 }
3505
3506 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3507 return 0;
3508}
3509
3510/* Look for tcp options. Normally only called on SYN and SYNACK packets.
3511 * But, this can also be called on packets in the established flow when
3512 * the fast version below fails.
3513 */
3514void tcp_parse_options(const struct sk_buff *skb,
3515 struct tcp_options_received *opt_rx, int estab,
3516 struct tcp_fastopen_cookie *foc)
3517{
3518 const unsigned char *ptr;
3519 const struct tcphdr *th = tcp_hdr(skb);
3520 int length = (th->doff * 4) - sizeof(struct tcphdr);
3521
3522 ptr = (const unsigned char *)(th + 1);
3523 opt_rx->saw_tstamp = 0;
3524
3525 while (length > 0) {
3526 int opcode = *ptr++;
3527 int opsize;
3528
3529 switch (opcode) {
3530 case TCPOPT_EOL:
3531 return;
3532 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3533 length--;
3534 continue;
3535 default:
3536 opsize = *ptr++;
3537 if (opsize < 2) /* "silly options" */
3538 return;
3539 if (opsize > length)
3540 return; /* don't parse partial options */
3541 switch (opcode) {
3542 case TCPOPT_MSS:
3543 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3544 u16 in_mss = get_unaligned_be16(ptr);
3545 if (in_mss) {
3546 if (opt_rx->user_mss &&
3547 opt_rx->user_mss < in_mss)
3548 in_mss = opt_rx->user_mss;
3549 opt_rx->mss_clamp = in_mss;
3550 }
3551 }
3552 break;
3553 case TCPOPT_WINDOW:
3554 if (opsize == TCPOLEN_WINDOW && th->syn &&
3555 !estab && sysctl_tcp_window_scaling) {
3556 __u8 snd_wscale = *(__u8 *)ptr;
3557 opt_rx->wscale_ok = 1;
3558 if (snd_wscale > 14) {
3559 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3560 __func__,
3561 snd_wscale);
3562 snd_wscale = 14;
3563 }
3564 opt_rx->snd_wscale = snd_wscale;
3565 }
3566 break;
3567 case TCPOPT_TIMESTAMP:
3568 if ((opsize == TCPOLEN_TIMESTAMP) &&
3569 ((estab && opt_rx->tstamp_ok) ||
3570 (!estab && sysctl_tcp_timestamps))) {
3571 opt_rx->saw_tstamp = 1;
3572 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3573 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3574 }
3575 break;
3576 case TCPOPT_SACK_PERM:
3577 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3578 !estab && sysctl_tcp_sack) {
3579 opt_rx->sack_ok = TCP_SACK_SEEN;
3580 tcp_sack_reset(opt_rx);
3581 }
3582 break;
3583
3584 case TCPOPT_SACK:
3585 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3586 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3587 opt_rx->sack_ok) {
3588 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3589 }
3590 break;
3591#ifdef CONFIG_TCP_MD5SIG
3592 case TCPOPT_MD5SIG:
3593 /*
3594 * The MD5 Hash has already been
3595 * checked (see tcp_v{4,6}_do_rcv()).
3596 */
3597 break;
3598#endif
3599 case TCPOPT_EXP:
3600 /* Fast Open option shares code 254 using a
3601 * 16 bits magic number. It's valid only in
3602 * SYN or SYN-ACK with an even size.
3603 */
3604 if (opsize < TCPOLEN_EXP_FASTOPEN_BASE ||
3605 get_unaligned_be16(ptr) != TCPOPT_FASTOPEN_MAGIC ||
3606 foc == NULL || !th->syn || (opsize & 1))
3607 break;
3608 foc->len = opsize - TCPOLEN_EXP_FASTOPEN_BASE;
3609 if (foc->len >= TCP_FASTOPEN_COOKIE_MIN &&
3610 foc->len <= TCP_FASTOPEN_COOKIE_MAX)
3611 memcpy(foc->val, ptr + 2, foc->len);
3612 else if (foc->len != 0)
3613 foc->len = -1;
3614 break;
3615
3616 }
3617 ptr += opsize-2;
3618 length -= opsize;
3619 }
3620 }
3621}
3622EXPORT_SYMBOL(tcp_parse_options);
3623
3624static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3625{
3626 const __be32 *ptr = (const __be32 *)(th + 1);
3627
3628 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3629 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3630 tp->rx_opt.saw_tstamp = 1;
3631 ++ptr;
3632 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3633 ++ptr;
3634 if (*ptr)
3635 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3636 else
3637 tp->rx_opt.rcv_tsecr = 0;
3638 return true;
3639 }
3640 return false;
3641}
3642
3643/* Fast parse options. This hopes to only see timestamps.
3644 * If it is wrong it falls back on tcp_parse_options().
3645 */
3646static bool tcp_fast_parse_options(const struct sk_buff *skb,
3647 const struct tcphdr *th, struct tcp_sock *tp)
3648{
3649 /* In the spirit of fast parsing, compare doff directly to constant
3650 * values. Because equality is used, short doff can be ignored here.
3651 */
3652 if (th->doff == (sizeof(*th) / 4)) {
3653 tp->rx_opt.saw_tstamp = 0;
3654 return false;
3655 } else if (tp->rx_opt.tstamp_ok &&
3656 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3657 if (tcp_parse_aligned_timestamp(tp, th))
3658 return true;
3659 }
3660
3661 tcp_parse_options(skb, &tp->rx_opt, 1, NULL);
3662 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3663 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3664
3665 return true;
3666}
3667
3668#ifdef CONFIG_TCP_MD5SIG
3669/*
3670 * Parse MD5 Signature option
3671 */
3672const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3673{
3674 int length = (th->doff << 2) - sizeof(*th);
3675 const u8 *ptr = (const u8 *)(th + 1);
3676
3677 /* If the TCP option is too short, we can short cut */
3678 if (length < TCPOLEN_MD5SIG)
3679 return NULL;
3680
3681 while (length > 0) {
3682 int opcode = *ptr++;
3683 int opsize;
3684
3685 switch (opcode) {
3686 case TCPOPT_EOL:
3687 return NULL;
3688 case TCPOPT_NOP:
3689 length--;
3690 continue;
3691 default:
3692 opsize = *ptr++;
3693 if (opsize < 2 || opsize > length)
3694 return NULL;
3695 if (opcode == TCPOPT_MD5SIG)
3696 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3697 }
3698 ptr += opsize - 2;
3699 length -= opsize;
3700 }
3701 return NULL;
3702}
3703EXPORT_SYMBOL(tcp_parse_md5sig_option);
3704#endif
3705
3706/* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3707 *
3708 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3709 * it can pass through stack. So, the following predicate verifies that
3710 * this segment is not used for anything but congestion avoidance or
3711 * fast retransmit. Moreover, we even are able to eliminate most of such
3712 * second order effects, if we apply some small "replay" window (~RTO)
3713 * to timestamp space.
3714 *
3715 * All these measures still do not guarantee that we reject wrapped ACKs
3716 * on networks with high bandwidth, when sequence space is recycled fastly,
3717 * but it guarantees that such events will be very rare and do not affect
3718 * connection seriously. This doesn't look nice, but alas, PAWS is really
3719 * buggy extension.
3720 *
3721 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3722 * states that events when retransmit arrives after original data are rare.
3723 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3724 * the biggest problem on large power networks even with minor reordering.
3725 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3726 * up to bandwidth of 18Gigabit/sec. 8) ]
3727 */
3728
3729static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3730{
3731 const struct tcp_sock *tp = tcp_sk(sk);
3732 const struct tcphdr *th = tcp_hdr(skb);
3733 u32 seq = TCP_SKB_CB(skb)->seq;
3734 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3735
3736 return (/* 1. Pure ACK with correct sequence number. */
3737 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3738
3739 /* 2. ... and duplicate ACK. */
3740 ack == tp->snd_una &&
3741
3742 /* 3. ... and does not update window. */
3743 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3744
3745 /* 4. ... and sits in replay window. */
3746 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3747}
3748
3749static inline bool tcp_paws_discard(const struct sock *sk,
3750 const struct sk_buff *skb)
3751{
3752 const struct tcp_sock *tp = tcp_sk(sk);
3753
3754 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
3755 !tcp_disordered_ack(sk, skb);
3756}
3757
3758/* Check segment sequence number for validity.
3759 *
3760 * Segment controls are considered valid, if the segment
3761 * fits to the window after truncation to the window. Acceptability
3762 * of data (and SYN, FIN, of course) is checked separately.
3763 * See tcp_data_queue(), for example.
3764 *
3765 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3766 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3767 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3768 * (borrowed from freebsd)
3769 */
3770
3771static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
3772{
3773 return !before(end_seq, tp->rcv_wup) &&
3774 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3775}
3776
3777/* When we get a reset we do this. */
3778void tcp_reset(struct sock *sk)
3779{
3780 /* We want the right error as BSD sees it (and indeed as we do). */
3781 switch (sk->sk_state) {
3782 case TCP_SYN_SENT:
3783 sk->sk_err = ECONNREFUSED;
3784 break;
3785 case TCP_CLOSE_WAIT:
3786 sk->sk_err = EPIPE;
3787 break;
3788 case TCP_CLOSE:
3789 return;
3790 default:
3791 sk->sk_err = ECONNRESET;
3792 }
3793 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3794 smp_wmb();
3795
3796 if (!sock_flag(sk, SOCK_DEAD))
3797 sk->sk_error_report(sk);
3798
3799 tcp_done(sk);
3800}
3801
3802/*
3803 * Process the FIN bit. This now behaves as it is supposed to work
3804 * and the FIN takes effect when it is validly part of sequence
3805 * space. Not before when we get holes.
3806 *
3807 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3808 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3809 * TIME-WAIT)
3810 *
3811 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3812 * close and we go into CLOSING (and later onto TIME-WAIT)
3813 *
3814 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3815 */
3816static void tcp_fin(struct sock *sk)
3817{
3818 struct tcp_sock *tp = tcp_sk(sk);
3819 const struct dst_entry *dst;
3820
3821 inet_csk_schedule_ack(sk);
3822
3823 sk->sk_shutdown |= RCV_SHUTDOWN;
3824 sock_set_flag(sk, SOCK_DONE);
3825
3826 switch (sk->sk_state) {
3827 case TCP_SYN_RECV:
3828 case TCP_ESTABLISHED:
3829 /* Move to CLOSE_WAIT */
3830 tcp_set_state(sk, TCP_CLOSE_WAIT);
3831 dst = __sk_dst_get(sk);
3832 if (!dst || !dst_metric(dst, RTAX_QUICKACK))
3833 inet_csk(sk)->icsk_ack.pingpong = 1;
3834 break;
3835
3836 case TCP_CLOSE_WAIT:
3837 case TCP_CLOSING:
3838 /* Received a retransmission of the FIN, do
3839 * nothing.
3840 */
3841 break;
3842 case TCP_LAST_ACK:
3843 /* RFC793: Remain in the LAST-ACK state. */
3844 break;
3845
3846 case TCP_FIN_WAIT1:
3847 /* This case occurs when a simultaneous close
3848 * happens, we must ack the received FIN and
3849 * enter the CLOSING state.
3850 */
3851 tcp_send_ack(sk);
3852 tcp_set_state(sk, TCP_CLOSING);
3853 break;
3854 case TCP_FIN_WAIT2:
3855 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3856 tcp_send_ack(sk);
3857 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3858 break;
3859 default:
3860 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3861 * cases we should never reach this piece of code.
3862 */
3863 pr_err("%s: Impossible, sk->sk_state=%d\n",
3864 __func__, sk->sk_state);
3865 break;
3866 }
3867
3868 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3869 * Probably, we should reset in this case. For now drop them.
3870 */
3871 __skb_queue_purge(&tp->out_of_order_queue);
3872 if (tcp_is_sack(tp))
3873 tcp_sack_reset(&tp->rx_opt);
3874 sk_mem_reclaim(sk);
3875
3876 if (!sock_flag(sk, SOCK_DEAD)) {
3877 sk->sk_state_change(sk);
3878
3879 /* Do not send POLL_HUP for half duplex close. */
3880 if (sk->sk_shutdown == SHUTDOWN_MASK ||
3881 sk->sk_state == TCP_CLOSE)
3882 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
3883 else
3884 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
3885 }
3886}
3887
3888static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
3889 u32 end_seq)
3890{
3891 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3892 if (before(seq, sp->start_seq))
3893 sp->start_seq = seq;
3894 if (after(end_seq, sp->end_seq))
3895 sp->end_seq = end_seq;
3896 return true;
3897 }
3898 return false;
3899}
3900
3901static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
3902{
3903 struct tcp_sock *tp = tcp_sk(sk);
3904
3905 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3906 int mib_idx;
3907
3908 if (before(seq, tp->rcv_nxt))
3909 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
3910 else
3911 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
3912
3913 NET_INC_STATS_BH(sock_net(sk), mib_idx);
3914
3915 tp->rx_opt.dsack = 1;
3916 tp->duplicate_sack[0].start_seq = seq;
3917 tp->duplicate_sack[0].end_seq = end_seq;
3918 }
3919}
3920
3921static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
3922{
3923 struct tcp_sock *tp = tcp_sk(sk);
3924
3925 if (!tp->rx_opt.dsack)
3926 tcp_dsack_set(sk, seq, end_seq);
3927 else
3928 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3929}
3930
3931static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
3932{
3933 struct tcp_sock *tp = tcp_sk(sk);
3934
3935 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3936 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3937 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
3938 tcp_enter_quickack_mode(sk);
3939
3940 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3941 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3942
3943 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3944 end_seq = tp->rcv_nxt;
3945 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
3946 }
3947 }
3948
3949 tcp_send_ack(sk);
3950}
3951
3952/* These routines update the SACK block as out-of-order packets arrive or
3953 * in-order packets close up the sequence space.
3954 */
3955static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
3956{
3957 int this_sack;
3958 struct tcp_sack_block *sp = &tp->selective_acks[0];
3959 struct tcp_sack_block *swalk = sp + 1;
3960
3961 /* See if the recent change to the first SACK eats into
3962 * or hits the sequence space of other SACK blocks, if so coalesce.
3963 */
3964 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
3965 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
3966 int i;
3967
3968 /* Zap SWALK, by moving every further SACK up by one slot.
3969 * Decrease num_sacks.
3970 */
3971 tp->rx_opt.num_sacks--;
3972 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
3973 sp[i] = sp[i + 1];
3974 continue;
3975 }
3976 this_sack++, swalk++;
3977 }
3978}
3979
3980static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
3981{
3982 struct tcp_sock *tp = tcp_sk(sk);
3983 struct tcp_sack_block *sp = &tp->selective_acks[0];
3984 int cur_sacks = tp->rx_opt.num_sacks;
3985 int this_sack;
3986
3987 if (!cur_sacks)
3988 goto new_sack;
3989
3990 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
3991 if (tcp_sack_extend(sp, seq, end_seq)) {
3992 /* Rotate this_sack to the first one. */
3993 for (; this_sack > 0; this_sack--, sp--)
3994 swap(*sp, *(sp - 1));
3995 if (cur_sacks > 1)
3996 tcp_sack_maybe_coalesce(tp);
3997 return;
3998 }
3999 }
4000
4001 /* Could not find an adjacent existing SACK, build a new one,
4002 * put it at the front, and shift everyone else down. We
4003 * always know there is at least one SACK present already here.
4004 *
4005 * If the sack array is full, forget about the last one.
4006 */
4007 if (this_sack >= TCP_NUM_SACKS) {
4008 this_sack--;
4009 tp->rx_opt.num_sacks--;
4010 sp--;
4011 }
4012 for (; this_sack > 0; this_sack--, sp--)
4013 *sp = *(sp - 1);
4014
4015new_sack:
4016 /* Build the new head SACK, and we're done. */
4017 sp->start_seq = seq;
4018 sp->end_seq = end_seq;
4019 tp->rx_opt.num_sacks++;
4020}
4021
4022/* RCV.NXT advances, some SACKs should be eaten. */
4023
4024static void tcp_sack_remove(struct tcp_sock *tp)
4025{
4026 struct tcp_sack_block *sp = &tp->selective_acks[0];
4027 int num_sacks = tp->rx_opt.num_sacks;
4028 int this_sack;
4029
4030 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4031 if (skb_queue_empty(&tp->out_of_order_queue)) {
4032 tp->rx_opt.num_sacks = 0;
4033 return;
4034 }
4035
4036 for (this_sack = 0; this_sack < num_sacks;) {
4037 /* Check if the start of the sack is covered by RCV.NXT. */
4038 if (!before(tp->rcv_nxt, sp->start_seq)) {
4039 int i;
4040
4041 /* RCV.NXT must cover all the block! */
4042 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4043
4044 /* Zap this SACK, by moving forward any other SACKS. */
4045 for (i = this_sack+1; i < num_sacks; i++)
4046 tp->selective_acks[i-1] = tp->selective_acks[i];
4047 num_sacks--;
4048 continue;
4049 }
4050 this_sack++;
4051 sp++;
4052 }
4053 tp->rx_opt.num_sacks = num_sacks;
4054}
4055
4056/* This one checks to see if we can put data from the
4057 * out_of_order queue into the receive_queue.
4058 */
4059static void tcp_ofo_queue(struct sock *sk)
4060{
4061 struct tcp_sock *tp = tcp_sk(sk);
4062 __u32 dsack_high = tp->rcv_nxt;
4063 struct sk_buff *skb;
4064
4065 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4066 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4067 break;
4068
4069 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4070 __u32 dsack = dsack_high;
4071 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4072 dsack_high = TCP_SKB_CB(skb)->end_seq;
4073 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4074 }
4075
4076 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4077 SOCK_DEBUG(sk, "ofo packet was already received\n");
4078 __skb_unlink(skb, &tp->out_of_order_queue);
4079 __kfree_skb(skb);
4080 continue;
4081 }
4082 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4083 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4084 TCP_SKB_CB(skb)->end_seq);
4085
4086 __skb_unlink(skb, &tp->out_of_order_queue);
4087 __skb_queue_tail(&sk->sk_receive_queue, skb);
4088 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4089 if (tcp_hdr(skb)->fin)
4090 tcp_fin(sk);
4091 }
4092}
4093
4094static bool tcp_prune_ofo_queue(struct sock *sk);
4095static int tcp_prune_queue(struct sock *sk);
4096
4097static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4098 unsigned int size)
4099{
4100 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4101 !sk_rmem_schedule(sk, skb, size)) {
4102
4103 if (tcp_prune_queue(sk) < 0)
4104 return -1;
4105
4106 if (!sk_rmem_schedule(sk, skb, size)) {
4107 if (!tcp_prune_ofo_queue(sk))
4108 return -1;
4109
4110 if (!sk_rmem_schedule(sk, skb, size))
4111 return -1;
4112 }
4113 }
4114 return 0;
4115}
4116
4117/**
4118 * tcp_try_coalesce - try to merge skb to prior one
4119 * @sk: socket
4120 * @to: prior buffer
4121 * @from: buffer to add in queue
4122 * @fragstolen: pointer to boolean
4123 *
4124 * Before queueing skb @from after @to, try to merge them
4125 * to reduce overall memory use and queue lengths, if cost is small.
4126 * Packets in ofo or receive queues can stay a long time.
4127 * Better try to coalesce them right now to avoid future collapses.
4128 * Returns true if caller should free @from instead of queueing it
4129 */
4130static bool tcp_try_coalesce(struct sock *sk,
4131 struct sk_buff *to,
4132 struct sk_buff *from,
4133 bool *fragstolen)
4134{
4135 int delta;
4136
4137 *fragstolen = false;
4138
4139 if (tcp_hdr(from)->fin)
4140 return false;
4141
4142 /* Its possible this segment overlaps with prior segment in queue */
4143 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4144 return false;
4145
4146 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4147 return false;
4148
4149 atomic_add(delta, &sk->sk_rmem_alloc);
4150 sk_mem_charge(sk, delta);
4151 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4152 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4153 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4154 return true;
4155}
4156
4157static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4158{
4159 struct tcp_sock *tp = tcp_sk(sk);
4160 struct sk_buff *skb1;
4161 u32 seq, end_seq;
4162
4163 TCP_ECN_check_ce(tp, skb);
4164
4165 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4166 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFODROP);
4167 __kfree_skb(skb);
4168 return;
4169 }
4170
4171 /* Disable header prediction. */
4172 tp->pred_flags = 0;
4173 inet_csk_schedule_ack(sk);
4174
4175 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4176 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4177 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4178
4179 skb1 = skb_peek_tail(&tp->out_of_order_queue);
4180 if (!skb1) {
4181 /* Initial out of order segment, build 1 SACK. */
4182 if (tcp_is_sack(tp)) {
4183 tp->rx_opt.num_sacks = 1;
4184 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4185 tp->selective_acks[0].end_seq =
4186 TCP_SKB_CB(skb)->end_seq;
4187 }
4188 __skb_queue_head(&tp->out_of_order_queue, skb);
4189 goto end;
4190 }
4191
4192 seq = TCP_SKB_CB(skb)->seq;
4193 end_seq = TCP_SKB_CB(skb)->end_seq;
4194
4195 if (seq == TCP_SKB_CB(skb1)->end_seq) {
4196 bool fragstolen;
4197
4198 if (!tcp_try_coalesce(sk, skb1, skb, &fragstolen)) {
4199 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4200 } else {
4201 tcp_grow_window(sk, skb);
4202 kfree_skb_partial(skb, fragstolen);
4203 skb = NULL;
4204 }
4205
4206 if (!tp->rx_opt.num_sacks ||
4207 tp->selective_acks[0].end_seq != seq)
4208 goto add_sack;
4209
4210 /* Common case: data arrive in order after hole. */
4211 tp->selective_acks[0].end_seq = end_seq;
4212 goto end;
4213 }
4214
4215 /* Find place to insert this segment. */
4216 while (1) {
4217 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4218 break;
4219 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4220 skb1 = NULL;
4221 break;
4222 }
4223 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4224 }
4225
4226 /* Do skb overlap to previous one? */
4227 if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4228 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4229 /* All the bits are present. Drop. */
4230 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4231 __kfree_skb(skb);
4232 skb = NULL;
4233 tcp_dsack_set(sk, seq, end_seq);
4234 goto add_sack;
4235 }
4236 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4237 /* Partial overlap. */
4238 tcp_dsack_set(sk, seq,
4239 TCP_SKB_CB(skb1)->end_seq);
4240 } else {
4241 if (skb_queue_is_first(&tp->out_of_order_queue,
4242 skb1))
4243 skb1 = NULL;
4244 else
4245 skb1 = skb_queue_prev(
4246 &tp->out_of_order_queue,
4247 skb1);
4248 }
4249 }
4250 if (!skb1)
4251 __skb_queue_head(&tp->out_of_order_queue, skb);
4252 else
4253 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4254
4255 /* And clean segments covered by new one as whole. */
4256 while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4257 skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4258
4259 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4260 break;
4261 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4262 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4263 end_seq);
4264 break;
4265 }
4266 __skb_unlink(skb1, &tp->out_of_order_queue);
4267 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4268 TCP_SKB_CB(skb1)->end_seq);
4269 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4270 __kfree_skb(skb1);
4271 }
4272
4273add_sack:
4274 if (tcp_is_sack(tp))
4275 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4276end:
4277 if (skb) {
4278 tcp_grow_window(sk, skb);
4279 skb_set_owner_r(skb, sk);
4280 }
4281}
4282
4283static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4284 bool *fragstolen)
4285{
4286 int eaten;
4287 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4288
4289 __skb_pull(skb, hdrlen);
4290 eaten = (tail &&
4291 tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0;
4292 tcp_sk(sk)->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4293 if (!eaten) {
4294 __skb_queue_tail(&sk->sk_receive_queue, skb);
4295 skb_set_owner_r(skb, sk);
4296 }
4297 return eaten;
4298}
4299
4300int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4301{
4302 struct sk_buff *skb = NULL;
4303 struct tcphdr *th;
4304 bool fragstolen;
4305
4306 if (size == 0)
4307 return 0;
4308
4309 skb = alloc_skb(size + sizeof(*th), sk->sk_allocation);
4310 if (!skb)
4311 goto err;
4312
4313 if (tcp_try_rmem_schedule(sk, skb, size + sizeof(*th)))
4314 goto err_free;
4315
4316 th = (struct tcphdr *)skb_put(skb, sizeof(*th));
4317 skb_reset_transport_header(skb);
4318 memset(th, 0, sizeof(*th));
4319
4320 if (memcpy_fromiovec(skb_put(skb, size), msg->msg_iov, size))
4321 goto err_free;
4322
4323 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4324 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4325 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4326
4327 if (tcp_queue_rcv(sk, skb, sizeof(*th), &fragstolen)) {
4328 WARN_ON_ONCE(fragstolen); /* should not happen */
4329 __kfree_skb(skb);
4330 }
4331 return size;
4332
4333err_free:
4334 kfree_skb(skb);
4335err:
4336 return -ENOMEM;
4337}
4338
4339static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4340{
4341 const struct tcphdr *th = tcp_hdr(skb);
4342 struct tcp_sock *tp = tcp_sk(sk);
4343 int eaten = -1;
4344 bool fragstolen = false;
4345
4346 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4347 goto drop;
4348
4349 skb_dst_drop(skb);
4350 __skb_pull(skb, th->doff * 4);
4351
4352 TCP_ECN_accept_cwr(tp, skb);
4353
4354 tp->rx_opt.dsack = 0;
4355
4356 /* Queue data for delivery to the user.
4357 * Packets in sequence go to the receive queue.
4358 * Out of sequence packets to the out_of_order_queue.
4359 */
4360 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4361 if (tcp_receive_window(tp) == 0)
4362 goto out_of_window;
4363
4364 /* Ok. In sequence. In window. */
4365 if (tp->ucopy.task == current &&
4366 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4367 sock_owned_by_user(sk) && !tp->urg_data) {
4368 int chunk = min_t(unsigned int, skb->len,
4369 tp->ucopy.len);
4370
4371 __set_current_state(TASK_RUNNING);
4372
4373 local_bh_enable();
4374 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
4375 tp->ucopy.len -= chunk;
4376 tp->copied_seq += chunk;
4377 eaten = (chunk == skb->len);
4378 tcp_rcv_space_adjust(sk);
4379 }
4380 local_bh_disable();
4381 }
4382
4383 if (eaten <= 0) {
4384queue_and_out:
4385 if (eaten < 0 &&
4386 tcp_try_rmem_schedule(sk, skb, skb->truesize))
4387 goto drop;
4388
4389 eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4390 }
4391 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4392 if (skb->len)
4393 tcp_event_data_recv(sk, skb);
4394 if (th->fin)
4395 tcp_fin(sk);
4396
4397 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4398 tcp_ofo_queue(sk);
4399
4400 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4401 * gap in queue is filled.
4402 */
4403 if (skb_queue_empty(&tp->out_of_order_queue))
4404 inet_csk(sk)->icsk_ack.pingpong = 0;
4405 }
4406
4407 if (tp->rx_opt.num_sacks)
4408 tcp_sack_remove(tp);
4409
4410 tcp_fast_path_check(sk);
4411
4412 if (eaten > 0)
4413 kfree_skb_partial(skb, fragstolen);
4414 if (!sock_flag(sk, SOCK_DEAD))
4415 sk->sk_data_ready(sk);
4416 return;
4417 }
4418
4419 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4420 /* A retransmit, 2nd most common case. Force an immediate ack. */
4421 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4422 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4423
4424out_of_window:
4425 tcp_enter_quickack_mode(sk);
4426 inet_csk_schedule_ack(sk);
4427drop:
4428 __kfree_skb(skb);
4429 return;
4430 }
4431
4432 /* Out of window. F.e. zero window probe. */
4433 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4434 goto out_of_window;
4435
4436 tcp_enter_quickack_mode(sk);
4437
4438 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4439 /* Partial packet, seq < rcv_next < end_seq */
4440 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4441 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4442 TCP_SKB_CB(skb)->end_seq);
4443
4444 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4445
4446 /* If window is closed, drop tail of packet. But after
4447 * remembering D-SACK for its head made in previous line.
4448 */
4449 if (!tcp_receive_window(tp))
4450 goto out_of_window;
4451 goto queue_and_out;
4452 }
4453
4454 tcp_data_queue_ofo(sk, skb);
4455}
4456
4457static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4458 struct sk_buff_head *list)
4459{
4460 struct sk_buff *next = NULL;
4461
4462 if (!skb_queue_is_last(list, skb))
4463 next = skb_queue_next(list, skb);
4464
4465 __skb_unlink(skb, list);
4466 __kfree_skb(skb);
4467 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4468
4469 return next;
4470}
4471
4472/* Collapse contiguous sequence of skbs head..tail with
4473 * sequence numbers start..end.
4474 *
4475 * If tail is NULL, this means until the end of the list.
4476 *
4477 * Segments with FIN/SYN are not collapsed (only because this
4478 * simplifies code)
4479 */
4480static void
4481tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4482 struct sk_buff *head, struct sk_buff *tail,
4483 u32 start, u32 end)
4484{
4485 struct sk_buff *skb, *n;
4486 bool end_of_skbs;
4487
4488 /* First, check that queue is collapsible and find
4489 * the point where collapsing can be useful. */
4490 skb = head;
4491restart:
4492 end_of_skbs = true;
4493 skb_queue_walk_from_safe(list, skb, n) {
4494 if (skb == tail)
4495 break;
4496 /* No new bits? It is possible on ofo queue. */
4497 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4498 skb = tcp_collapse_one(sk, skb, list);
4499 if (!skb)
4500 break;
4501 goto restart;
4502 }
4503
4504 /* The first skb to collapse is:
4505 * - not SYN/FIN and
4506 * - bloated or contains data before "start" or
4507 * overlaps to the next one.
4508 */
4509 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
4510 (tcp_win_from_space(skb->truesize) > skb->len ||
4511 before(TCP_SKB_CB(skb)->seq, start))) {
4512 end_of_skbs = false;
4513 break;
4514 }
4515
4516 if (!skb_queue_is_last(list, skb)) {
4517 struct sk_buff *next = skb_queue_next(list, skb);
4518 if (next != tail &&
4519 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4520 end_of_skbs = false;
4521 break;
4522 }
4523 }
4524
4525 /* Decided to skip this, advance start seq. */
4526 start = TCP_SKB_CB(skb)->end_seq;
4527 }
4528 if (end_of_skbs || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
4529 return;
4530
4531 while (before(start, end)) {
4532 struct sk_buff *nskb;
4533 unsigned int header = skb_headroom(skb);
4534 int copy = SKB_MAX_ORDER(header, 0);
4535
4536 /* Too big header? This can happen with IPv6. */
4537 if (copy < 0)
4538 return;
4539 if (end - start < copy)
4540 copy = end - start;
4541 nskb = alloc_skb(copy + header, GFP_ATOMIC);
4542 if (!nskb)
4543 return;
4544
4545 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
4546 skb_set_network_header(nskb, (skb_network_header(skb) -
4547 skb->head));
4548 skb_set_transport_header(nskb, (skb_transport_header(skb) -
4549 skb->head));
4550 skb_reserve(nskb, header);
4551 memcpy(nskb->head, skb->head, header);
4552 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4553 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4554 __skb_queue_before(list, skb, nskb);
4555 skb_set_owner_r(nskb, sk);
4556
4557 /* Copy data, releasing collapsed skbs. */
4558 while (copy > 0) {
4559 int offset = start - TCP_SKB_CB(skb)->seq;
4560 int size = TCP_SKB_CB(skb)->end_seq - start;
4561
4562 BUG_ON(offset < 0);
4563 if (size > 0) {
4564 size = min(copy, size);
4565 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4566 BUG();
4567 TCP_SKB_CB(nskb)->end_seq += size;
4568 copy -= size;
4569 start += size;
4570 }
4571 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4572 skb = tcp_collapse_one(sk, skb, list);
4573 if (!skb ||
4574 skb == tail ||
4575 tcp_hdr(skb)->syn ||
4576 tcp_hdr(skb)->fin)
4577 return;
4578 }
4579 }
4580 }
4581}
4582
4583/* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4584 * and tcp_collapse() them until all the queue is collapsed.
4585 */
4586static void tcp_collapse_ofo_queue(struct sock *sk)
4587{
4588 struct tcp_sock *tp = tcp_sk(sk);
4589 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4590 struct sk_buff *head;
4591 u32 start, end;
4592
4593 if (skb == NULL)
4594 return;
4595
4596 start = TCP_SKB_CB(skb)->seq;
4597 end = TCP_SKB_CB(skb)->end_seq;
4598 head = skb;
4599
4600 for (;;) {
4601 struct sk_buff *next = NULL;
4602
4603 if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4604 next = skb_queue_next(&tp->out_of_order_queue, skb);
4605 skb = next;
4606
4607 /* Segment is terminated when we see gap or when
4608 * we are at the end of all the queue. */
4609 if (!skb ||
4610 after(TCP_SKB_CB(skb)->seq, end) ||
4611 before(TCP_SKB_CB(skb)->end_seq, start)) {
4612 tcp_collapse(sk, &tp->out_of_order_queue,
4613 head, skb, start, end);
4614 head = skb;
4615 if (!skb)
4616 break;
4617 /* Start new segment */
4618 start = TCP_SKB_CB(skb)->seq;
4619 end = TCP_SKB_CB(skb)->end_seq;
4620 } else {
4621 if (before(TCP_SKB_CB(skb)->seq, start))
4622 start = TCP_SKB_CB(skb)->seq;
4623 if (after(TCP_SKB_CB(skb)->end_seq, end))
4624 end = TCP_SKB_CB(skb)->end_seq;
4625 }
4626 }
4627}
4628
4629/*
4630 * Purge the out-of-order queue.
4631 * Return true if queue was pruned.
4632 */
4633static bool tcp_prune_ofo_queue(struct sock *sk)
4634{
4635 struct tcp_sock *tp = tcp_sk(sk);
4636 bool res = false;
4637
4638 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4639 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4640 __skb_queue_purge(&tp->out_of_order_queue);
4641
4642 /* Reset SACK state. A conforming SACK implementation will
4643 * do the same at a timeout based retransmit. When a connection
4644 * is in a sad state like this, we care only about integrity
4645 * of the connection not performance.
4646 */
4647 if (tp->rx_opt.sack_ok)
4648 tcp_sack_reset(&tp->rx_opt);
4649 sk_mem_reclaim(sk);
4650 res = true;
4651 }
4652 return res;
4653}
4654
4655/* Reduce allocated memory if we can, trying to get
4656 * the socket within its memory limits again.
4657 *
4658 * Return less than zero if we should start dropping frames
4659 * until the socket owning process reads some of the data
4660 * to stabilize the situation.
4661 */
4662static int tcp_prune_queue(struct sock *sk)
4663{
4664 struct tcp_sock *tp = tcp_sk(sk);
4665
4666 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4667
4668 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4669
4670 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4671 tcp_clamp_window(sk);
4672 else if (sk_under_memory_pressure(sk))
4673 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4674
4675 tcp_collapse_ofo_queue(sk);
4676 if (!skb_queue_empty(&sk->sk_receive_queue))
4677 tcp_collapse(sk, &sk->sk_receive_queue,
4678 skb_peek(&sk->sk_receive_queue),
4679 NULL,
4680 tp->copied_seq, tp->rcv_nxt);
4681 sk_mem_reclaim(sk);
4682
4683 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4684 return 0;
4685
4686 /* Collapsing did not help, destructive actions follow.
4687 * This must not ever occur. */
4688
4689 tcp_prune_ofo_queue(sk);
4690
4691 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4692 return 0;
4693
4694 /* If we are really being abused, tell the caller to silently
4695 * drop receive data on the floor. It will get retransmitted
4696 * and hopefully then we'll have sufficient space.
4697 */
4698 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4699
4700 /* Massive buffer overcommit. */
4701 tp->pred_flags = 0;
4702 return -1;
4703}
4704
4705/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4706 * As additional protections, we do not touch cwnd in retransmission phases,
4707 * and if application hit its sndbuf limit recently.
4708 */
4709void tcp_cwnd_application_limited(struct sock *sk)
4710{
4711 struct tcp_sock *tp = tcp_sk(sk);
4712
4713 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4714 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4715 /* Limited by application or receiver window. */
4716 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4717 u32 win_used = max(tp->snd_cwnd_used, init_win);
4718 if (win_used < tp->snd_cwnd) {
4719 tp->snd_ssthresh = tcp_current_ssthresh(sk);
4720 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4721 }
4722 tp->snd_cwnd_used = 0;
4723 }
4724 tp->snd_cwnd_stamp = tcp_time_stamp;
4725}
4726
4727static bool tcp_should_expand_sndbuf(const struct sock *sk)
4728{
4729 const struct tcp_sock *tp = tcp_sk(sk);
4730
4731 /* If the user specified a specific send buffer setting, do
4732 * not modify it.
4733 */
4734 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4735 return false;
4736
4737 /* If we are under global TCP memory pressure, do not expand. */
4738 if (sk_under_memory_pressure(sk))
4739 return false;
4740
4741 /* If we are under soft global TCP memory pressure, do not expand. */
4742 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
4743 return false;
4744
4745 /* If we filled the congestion window, do not expand. */
4746 if (tp->packets_out >= tp->snd_cwnd)
4747 return false;
4748
4749 return true;
4750}
4751
4752/* When incoming ACK allowed to free some skb from write_queue,
4753 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4754 * on the exit from tcp input handler.
4755 *
4756 * PROBLEM: sndbuf expansion does not work well with largesend.
4757 */
4758static void tcp_new_space(struct sock *sk)
4759{
4760 struct tcp_sock *tp = tcp_sk(sk);
4761
4762 if (tcp_should_expand_sndbuf(sk)) {
4763 tcp_sndbuf_expand(sk);
4764 tp->snd_cwnd_stamp = tcp_time_stamp;
4765 }
4766
4767 sk->sk_write_space(sk);
4768}
4769
4770static void tcp_check_space(struct sock *sk)
4771{
4772 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4773 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4774 if (sk->sk_socket &&
4775 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4776 tcp_new_space(sk);
4777 }
4778}
4779
4780static inline void tcp_data_snd_check(struct sock *sk)
4781{
4782 tcp_push_pending_frames(sk);
4783 tcp_check_space(sk);
4784}
4785
4786/*
4787 * Check if sending an ack is needed.
4788 */
4789static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4790{
4791 struct tcp_sock *tp = tcp_sk(sk);
4792
4793 /* More than one full frame received... */
4794 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
4795 /* ... and right edge of window advances far enough.
4796 * (tcp_recvmsg() will send ACK otherwise). Or...
4797 */
4798 __tcp_select_window(sk) >= tp->rcv_wnd) ||
4799 /* We ACK each frame or... */
4800 tcp_in_quickack_mode(sk) ||
4801 /* We have out of order data. */
4802 (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
4803 /* Then ack it now */
4804 tcp_send_ack(sk);
4805 } else {
4806 /* Else, send delayed ack. */
4807 tcp_send_delayed_ack(sk);
4808 }
4809}
4810
4811static inline void tcp_ack_snd_check(struct sock *sk)
4812{
4813 if (!inet_csk_ack_scheduled(sk)) {
4814 /* We sent a data segment already. */
4815 return;
4816 }
4817 __tcp_ack_snd_check(sk, 1);
4818}
4819
4820/*
4821 * This routine is only called when we have urgent data
4822 * signaled. Its the 'slow' part of tcp_urg. It could be
4823 * moved inline now as tcp_urg is only called from one
4824 * place. We handle URGent data wrong. We have to - as
4825 * BSD still doesn't use the correction from RFC961.
4826 * For 1003.1g we should support a new option TCP_STDURG to permit
4827 * either form (or just set the sysctl tcp_stdurg).
4828 */
4829
4830static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
4831{
4832 struct tcp_sock *tp = tcp_sk(sk);
4833 u32 ptr = ntohs(th->urg_ptr);
4834
4835 if (ptr && !sysctl_tcp_stdurg)
4836 ptr--;
4837 ptr += ntohl(th->seq);
4838
4839 /* Ignore urgent data that we've already seen and read. */
4840 if (after(tp->copied_seq, ptr))
4841 return;
4842
4843 /* Do not replay urg ptr.
4844 *
4845 * NOTE: interesting situation not covered by specs.
4846 * Misbehaving sender may send urg ptr, pointing to segment,
4847 * which we already have in ofo queue. We are not able to fetch
4848 * such data and will stay in TCP_URG_NOTYET until will be eaten
4849 * by recvmsg(). Seems, we are not obliged to handle such wicked
4850 * situations. But it is worth to think about possibility of some
4851 * DoSes using some hypothetical application level deadlock.
4852 */
4853 if (before(ptr, tp->rcv_nxt))
4854 return;
4855
4856 /* Do we already have a newer (or duplicate) urgent pointer? */
4857 if (tp->urg_data && !after(ptr, tp->urg_seq))
4858 return;
4859
4860 /* Tell the world about our new urgent pointer. */
4861 sk_send_sigurg(sk);
4862
4863 /* We may be adding urgent data when the last byte read was
4864 * urgent. To do this requires some care. We cannot just ignore
4865 * tp->copied_seq since we would read the last urgent byte again
4866 * as data, nor can we alter copied_seq until this data arrives
4867 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4868 *
4869 * NOTE. Double Dutch. Rendering to plain English: author of comment
4870 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4871 * and expect that both A and B disappear from stream. This is _wrong_.
4872 * Though this happens in BSD with high probability, this is occasional.
4873 * Any application relying on this is buggy. Note also, that fix "works"
4874 * only in this artificial test. Insert some normal data between A and B and we will
4875 * decline of BSD again. Verdict: it is better to remove to trap
4876 * buggy users.
4877 */
4878 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4879 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
4880 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4881 tp->copied_seq++;
4882 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4883 __skb_unlink(skb, &sk->sk_receive_queue);
4884 __kfree_skb(skb);
4885 }
4886 }
4887
4888 tp->urg_data = TCP_URG_NOTYET;
4889 tp->urg_seq = ptr;
4890
4891 /* Disable header prediction. */
4892 tp->pred_flags = 0;
4893}
4894
4895/* This is the 'fast' part of urgent handling. */
4896static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
4897{
4898 struct tcp_sock *tp = tcp_sk(sk);
4899
4900 /* Check if we get a new urgent pointer - normally not. */
4901 if (th->urg)
4902 tcp_check_urg(sk, th);
4903
4904 /* Do we wait for any urgent data? - normally not... */
4905 if (tp->urg_data == TCP_URG_NOTYET) {
4906 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4907 th->syn;
4908
4909 /* Is the urgent pointer pointing into this packet? */
4910 if (ptr < skb->len) {
4911 u8 tmp;
4912 if (skb_copy_bits(skb, ptr, &tmp, 1))
4913 BUG();
4914 tp->urg_data = TCP_URG_VALID | tmp;
4915 if (!sock_flag(sk, SOCK_DEAD))
4916 sk->sk_data_ready(sk);
4917 }
4918 }
4919}
4920
4921static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4922{
4923 struct tcp_sock *tp = tcp_sk(sk);
4924 int chunk = skb->len - hlen;
4925 int err;
4926
4927 local_bh_enable();
4928 if (skb_csum_unnecessary(skb))
4929 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4930 else
4931 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4932 tp->ucopy.iov);
4933
4934 if (!err) {
4935 tp->ucopy.len -= chunk;
4936 tp->copied_seq += chunk;
4937 tcp_rcv_space_adjust(sk);
4938 }
4939
4940 local_bh_disable();
4941 return err;
4942}
4943
4944static __sum16 __tcp_checksum_complete_user(struct sock *sk,
4945 struct sk_buff *skb)
4946{
4947 __sum16 result;
4948
4949 if (sock_owned_by_user(sk)) {
4950 local_bh_enable();
4951 result = __tcp_checksum_complete(skb);
4952 local_bh_disable();
4953 } else {
4954 result = __tcp_checksum_complete(skb);
4955 }
4956 return result;
4957}
4958
4959static inline bool tcp_checksum_complete_user(struct sock *sk,
4960 struct sk_buff *skb)
4961{
4962 return !skb_csum_unnecessary(skb) &&
4963 __tcp_checksum_complete_user(sk, skb);
4964}
4965
4966#ifdef CONFIG_NET_DMA
4967static bool tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb,
4968 int hlen)
4969{
4970 struct tcp_sock *tp = tcp_sk(sk);
4971 int chunk = skb->len - hlen;
4972 int dma_cookie;
4973 bool copied_early = false;
4974
4975 if (tp->ucopy.wakeup)
4976 return false;
4977
4978 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
4979 tp->ucopy.dma_chan = net_dma_find_channel();
4980
4981 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
4982
4983 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
4984 skb, hlen,
4985 tp->ucopy.iov, chunk,
4986 tp->ucopy.pinned_list);
4987
4988 if (dma_cookie < 0)
4989 goto out;
4990
4991 tp->ucopy.dma_cookie = dma_cookie;
4992 copied_early = true;
4993
4994 tp->ucopy.len -= chunk;
4995 tp->copied_seq += chunk;
4996 tcp_rcv_space_adjust(sk);
4997
4998 if ((tp->ucopy.len == 0) ||
4999 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
5000 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
5001 tp->ucopy.wakeup = 1;
5002 sk->sk_data_ready(sk);
5003 }
5004 } else if (chunk > 0) {
5005 tp->ucopy.wakeup = 1;
5006 sk->sk_data_ready(sk);
5007 }
5008out:
5009 return copied_early;
5010}
5011#endif /* CONFIG_NET_DMA */
5012
5013/* Does PAWS and seqno based validation of an incoming segment, flags will
5014 * play significant role here.
5015 */
5016static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5017 const struct tcphdr *th, int syn_inerr)
5018{
5019 struct tcp_sock *tp = tcp_sk(sk);
5020
5021 /* RFC1323: H1. Apply PAWS check first. */
5022 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
5023 tcp_paws_discard(sk, skb)) {
5024 if (!th->rst) {
5025 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5026 tcp_send_dupack(sk, skb);
5027 goto discard;
5028 }
5029 /* Reset is accepted even if it did not pass PAWS. */
5030 }
5031
5032 /* Step 1: check sequence number */
5033 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5034 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5035 * (RST) segments are validated by checking their SEQ-fields."
5036 * And page 69: "If an incoming segment is not acceptable,
5037 * an acknowledgment should be sent in reply (unless the RST
5038 * bit is set, if so drop the segment and return)".
5039 */
5040 if (!th->rst) {
5041 if (th->syn)
5042 goto syn_challenge;
5043 tcp_send_dupack(sk, skb);
5044 }
5045 goto discard;
5046 }
5047
5048 /* Step 2: check RST bit */
5049 if (th->rst) {
5050 /* RFC 5961 3.2 :
5051 * If sequence number exactly matches RCV.NXT, then
5052 * RESET the connection
5053 * else
5054 * Send a challenge ACK
5055 */
5056 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt)
5057 tcp_reset(sk);
5058 else
5059 tcp_send_challenge_ack(sk);
5060 goto discard;
5061 }
5062
5063 /* step 3: check security and precedence [ignored] */
5064
5065 /* step 4: Check for a SYN
5066 * RFC 5691 4.2 : Send a challenge ack
5067 */
5068 if (th->syn) {
5069syn_challenge:
5070 if (syn_inerr)
5071 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5072 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5073 tcp_send_challenge_ack(sk);
5074 goto discard;
5075 }
5076
5077 return true;
5078
5079discard:
5080 __kfree_skb(skb);
5081 return false;
5082}
5083
5084/*
5085 * TCP receive function for the ESTABLISHED state.
5086 *
5087 * It is split into a fast path and a slow path. The fast path is
5088 * disabled when:
5089 * - A zero window was announced from us - zero window probing
5090 * is only handled properly in the slow path.
5091 * - Out of order segments arrived.
5092 * - Urgent data is expected.
5093 * - There is no buffer space left
5094 * - Unexpected TCP flags/window values/header lengths are received
5095 * (detected by checking the TCP header against pred_flags)
5096 * - Data is sent in both directions. Fast path only supports pure senders
5097 * or pure receivers (this means either the sequence number or the ack
5098 * value must stay constant)
5099 * - Unexpected TCP option.
5100 *
5101 * When these conditions are not satisfied it drops into a standard
5102 * receive procedure patterned after RFC793 to handle all cases.
5103 * The first three cases are guaranteed by proper pred_flags setting,
5104 * the rest is checked inline. Fast processing is turned on in
5105 * tcp_data_queue when everything is OK.
5106 */
5107void tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5108 const struct tcphdr *th, unsigned int len)
5109{
5110 struct tcp_sock *tp = tcp_sk(sk);
5111
5112 if (unlikely(sk->sk_rx_dst == NULL))
5113 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5114 /*
5115 * Header prediction.
5116 * The code loosely follows the one in the famous
5117 * "30 instruction TCP receive" Van Jacobson mail.
5118 *
5119 * Van's trick is to deposit buffers into socket queue
5120 * on a device interrupt, to call tcp_recv function
5121 * on the receive process context and checksum and copy
5122 * the buffer to user space. smart...
5123 *
5124 * Our current scheme is not silly either but we take the
5125 * extra cost of the net_bh soft interrupt processing...
5126 * We do checksum and copy also but from device to kernel.
5127 */
5128
5129 tp->rx_opt.saw_tstamp = 0;
5130
5131 /* pred_flags is 0xS?10 << 16 + snd_wnd
5132 * if header_prediction is to be made
5133 * 'S' will always be tp->tcp_header_len >> 2
5134 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5135 * turn it off (when there are holes in the receive
5136 * space for instance)
5137 * PSH flag is ignored.
5138 */
5139
5140 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5141 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5142 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5143 int tcp_header_len = tp->tcp_header_len;
5144
5145 /* Timestamp header prediction: tcp_header_len
5146 * is automatically equal to th->doff*4 due to pred_flags
5147 * match.
5148 */
5149
5150 /* Check timestamp */
5151 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5152 /* No? Slow path! */
5153 if (!tcp_parse_aligned_timestamp(tp, th))
5154 goto slow_path;
5155
5156 /* If PAWS failed, check it more carefully in slow path */
5157 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5158 goto slow_path;
5159
5160 /* DO NOT update ts_recent here, if checksum fails
5161 * and timestamp was corrupted part, it will result
5162 * in a hung connection since we will drop all
5163 * future packets due to the PAWS test.
5164 */
5165 }
5166
5167 if (len <= tcp_header_len) {
5168 /* Bulk data transfer: sender */
5169 if (len == tcp_header_len) {
5170 /* Predicted packet is in window by definition.
5171 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5172 * Hence, check seq<=rcv_wup reduces to:
5173 */
5174 if (tcp_header_len ==
5175 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5176 tp->rcv_nxt == tp->rcv_wup)
5177 tcp_store_ts_recent(tp);
5178
5179 /* We know that such packets are checksummed
5180 * on entry.
5181 */
5182 tcp_ack(sk, skb, 0);
5183 __kfree_skb(skb);
5184 tcp_data_snd_check(sk);
5185 return;
5186 } else { /* Header too small */
5187 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5188 goto discard;
5189 }
5190 } else {
5191 int eaten = 0;
5192 int copied_early = 0;
5193 bool fragstolen = false;
5194
5195 if (tp->copied_seq == tp->rcv_nxt &&
5196 len - tcp_header_len <= tp->ucopy.len) {
5197#ifdef CONFIG_NET_DMA
5198 if (tp->ucopy.task == current &&
5199 sock_owned_by_user(sk) &&
5200 tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
5201 copied_early = 1;
5202 eaten = 1;
5203 }
5204#endif
5205 if (tp->ucopy.task == current &&
5206 sock_owned_by_user(sk) && !copied_early) {
5207 __set_current_state(TASK_RUNNING);
5208
5209 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
5210 eaten = 1;
5211 }
5212 if (eaten) {
5213 /* Predicted packet is in window by definition.
5214 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5215 * Hence, check seq<=rcv_wup reduces to:
5216 */
5217 if (tcp_header_len ==
5218 (sizeof(struct tcphdr) +
5219 TCPOLEN_TSTAMP_ALIGNED) &&
5220 tp->rcv_nxt == tp->rcv_wup)
5221 tcp_store_ts_recent(tp);
5222
5223 tcp_rcv_rtt_measure_ts(sk, skb);
5224
5225 __skb_pull(skb, tcp_header_len);
5226 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5227 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5228 }
5229 if (copied_early)
5230 tcp_cleanup_rbuf(sk, skb->len);
5231 }
5232 if (!eaten) {
5233 if (tcp_checksum_complete_user(sk, skb))
5234 goto csum_error;
5235
5236 if ((int)skb->truesize > sk->sk_forward_alloc)
5237 goto step5;
5238
5239 /* Predicted packet is in window by definition.
5240 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5241 * Hence, check seq<=rcv_wup reduces to:
5242 */
5243 if (tcp_header_len ==
5244 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5245 tp->rcv_nxt == tp->rcv_wup)
5246 tcp_store_ts_recent(tp);
5247
5248 tcp_rcv_rtt_measure_ts(sk, skb);
5249
5250 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5251
5252 /* Bulk data transfer: receiver */
5253 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5254 &fragstolen);
5255 }
5256
5257 tcp_event_data_recv(sk, skb);
5258
5259 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5260 /* Well, only one small jumplet in fast path... */
5261 tcp_ack(sk, skb, FLAG_DATA);
5262 tcp_data_snd_check(sk);
5263 if (!inet_csk_ack_scheduled(sk))
5264 goto no_ack;
5265 }
5266
5267 if (!copied_early || tp->rcv_nxt != tp->rcv_wup)
5268 __tcp_ack_snd_check(sk, 0);
5269no_ack:
5270#ifdef CONFIG_NET_DMA
5271 if (copied_early)
5272 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
5273 else
5274#endif
5275 if (eaten)
5276 kfree_skb_partial(skb, fragstolen);
5277 sk->sk_data_ready(sk);
5278 return;
5279 }
5280 }
5281
5282slow_path:
5283 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5284 goto csum_error;
5285
5286 if (!th->ack && !th->rst)
5287 goto discard;
5288
5289 /*
5290 * Standard slow path.
5291 */
5292
5293 if (!tcp_validate_incoming(sk, skb, th, 1))
5294 return;
5295
5296step5:
5297 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5298 goto discard;
5299
5300 tcp_rcv_rtt_measure_ts(sk, skb);
5301
5302 /* Process urgent data. */
5303 tcp_urg(sk, skb, th);
5304
5305 /* step 7: process the segment text */
5306 tcp_data_queue(sk, skb);
5307
5308 tcp_data_snd_check(sk);
5309 tcp_ack_snd_check(sk);
5310 return;
5311
5312csum_error:
5313 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_CSUMERRORS);
5314 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5315
5316discard:
5317 __kfree_skb(skb);
5318}
5319EXPORT_SYMBOL(tcp_rcv_established);
5320
5321void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5322{
5323 struct tcp_sock *tp = tcp_sk(sk);
5324 struct inet_connection_sock *icsk = inet_csk(sk);
5325
5326 tcp_set_state(sk, TCP_ESTABLISHED);
5327
5328 if (skb != NULL) {
5329 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5330 security_inet_conn_established(sk, skb);
5331 }
5332
5333 /* Make sure socket is routed, for correct metrics. */
5334 icsk->icsk_af_ops->rebuild_header(sk);
5335
5336 tcp_init_metrics(sk);
5337
5338 tcp_init_congestion_control(sk);
5339
5340 /* Prevent spurious tcp_cwnd_restart() on first data
5341 * packet.
5342 */
5343 tp->lsndtime = tcp_time_stamp;
5344
5345 tcp_init_buffer_space(sk);
5346
5347 if (sock_flag(sk, SOCK_KEEPOPEN))
5348 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5349
5350 if (!tp->rx_opt.snd_wscale)
5351 __tcp_fast_path_on(tp, tp->snd_wnd);
5352 else
5353 tp->pred_flags = 0;
5354
5355 if (!sock_flag(sk, SOCK_DEAD)) {
5356 sk->sk_state_change(sk);
5357 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5358 }
5359}
5360
5361static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5362 struct tcp_fastopen_cookie *cookie)
5363{
5364 struct tcp_sock *tp = tcp_sk(sk);
5365 struct sk_buff *data = tp->syn_data ? tcp_write_queue_head(sk) : NULL;
5366 u16 mss = tp->rx_opt.mss_clamp;
5367 bool syn_drop;
5368
5369 if (mss == tp->rx_opt.user_mss) {
5370 struct tcp_options_received opt;
5371
5372 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5373 tcp_clear_options(&opt);
5374 opt.user_mss = opt.mss_clamp = 0;
5375 tcp_parse_options(synack, &opt, 0, NULL);
5376 mss = opt.mss_clamp;
5377 }
5378
5379 if (!tp->syn_fastopen) /* Ignore an unsolicited cookie */
5380 cookie->len = -1;
5381
5382 /* The SYN-ACK neither has cookie nor acknowledges the data. Presumably
5383 * the remote receives only the retransmitted (regular) SYNs: either
5384 * the original SYN-data or the corresponding SYN-ACK is lost.
5385 */
5386 syn_drop = (cookie->len <= 0 && data && tp->total_retrans);
5387
5388 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop);
5389
5390 if (data) { /* Retransmit unacked data in SYN */
5391 tcp_for_write_queue_from(data, sk) {
5392 if (data == tcp_send_head(sk) ||
5393 __tcp_retransmit_skb(sk, data))
5394 break;
5395 }
5396 tcp_rearm_rto(sk);
5397 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVEFAIL);
5398 return true;
5399 }
5400 tp->syn_data_acked = tp->syn_data;
5401 if (tp->syn_data_acked)
5402 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
5403 return false;
5404}
5405
5406static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5407 const struct tcphdr *th, unsigned int len)
5408{
5409 struct inet_connection_sock *icsk = inet_csk(sk);
5410 struct tcp_sock *tp = tcp_sk(sk);
5411 struct tcp_fastopen_cookie foc = { .len = -1 };
5412 int saved_clamp = tp->rx_opt.mss_clamp;
5413
5414 tcp_parse_options(skb, &tp->rx_opt, 0, &foc);
5415 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5416 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5417
5418 if (th->ack) {
5419 /* rfc793:
5420 * "If the state is SYN-SENT then
5421 * first check the ACK bit
5422 * If the ACK bit is set
5423 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5424 * a reset (unless the RST bit is set, if so drop
5425 * the segment and return)"
5426 */
5427 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5428 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5429 goto reset_and_undo;
5430
5431 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5432 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5433 tcp_time_stamp)) {
5434 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5435 goto reset_and_undo;
5436 }
5437
5438 /* Now ACK is acceptable.
5439 *
5440 * "If the RST bit is set
5441 * If the ACK was acceptable then signal the user "error:
5442 * connection reset", drop the segment, enter CLOSED state,
5443 * delete TCB, and return."
5444 */
5445
5446 if (th->rst) {
5447 tcp_reset(sk);
5448 goto discard;
5449 }
5450
5451 /* rfc793:
5452 * "fifth, if neither of the SYN or RST bits is set then
5453 * drop the segment and return."
5454 *
5455 * See note below!
5456 * --ANK(990513)
5457 */
5458 if (!th->syn)
5459 goto discard_and_undo;
5460
5461 /* rfc793:
5462 * "If the SYN bit is on ...
5463 * are acceptable then ...
5464 * (our SYN has been ACKed), change the connection
5465 * state to ESTABLISHED..."
5466 */
5467
5468 TCP_ECN_rcv_synack(tp, th);
5469
5470 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5471 tcp_ack(sk, skb, FLAG_SLOWPATH);
5472
5473 /* Ok.. it's good. Set up sequence numbers and
5474 * move to established.
5475 */
5476 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5477 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5478
5479 /* RFC1323: The window in SYN & SYN/ACK segments is
5480 * never scaled.
5481 */
5482 tp->snd_wnd = ntohs(th->window);
5483
5484 if (!tp->rx_opt.wscale_ok) {
5485 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5486 tp->window_clamp = min(tp->window_clamp, 65535U);
5487 }
5488
5489 if (tp->rx_opt.saw_tstamp) {
5490 tp->rx_opt.tstamp_ok = 1;
5491 tp->tcp_header_len =
5492 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5493 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5494 tcp_store_ts_recent(tp);
5495 } else {
5496 tp->tcp_header_len = sizeof(struct tcphdr);
5497 }
5498
5499 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5500 tcp_enable_fack(tp);
5501
5502 tcp_mtup_init(sk);
5503 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5504 tcp_initialize_rcv_mss(sk);
5505
5506 /* Remember, tcp_poll() does not lock socket!
5507 * Change state from SYN-SENT only after copied_seq
5508 * is initialized. */
5509 tp->copied_seq = tp->rcv_nxt;
5510
5511 smp_mb();
5512
5513 tcp_finish_connect(sk, skb);
5514
5515 if ((tp->syn_fastopen || tp->syn_data) &&
5516 tcp_rcv_fastopen_synack(sk, skb, &foc))
5517 return -1;
5518
5519 if (sk->sk_write_pending ||
5520 icsk->icsk_accept_queue.rskq_defer_accept ||
5521 icsk->icsk_ack.pingpong) {
5522 /* Save one ACK. Data will be ready after
5523 * several ticks, if write_pending is set.
5524 *
5525 * It may be deleted, but with this feature tcpdumps
5526 * look so _wonderfully_ clever, that I was not able
5527 * to stand against the temptation 8) --ANK
5528 */
5529 inet_csk_schedule_ack(sk);
5530 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5531 tcp_enter_quickack_mode(sk);
5532 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5533 TCP_DELACK_MAX, TCP_RTO_MAX);
5534
5535discard:
5536 __kfree_skb(skb);
5537 return 0;
5538 } else {
5539 tcp_send_ack(sk);
5540 }
5541 return -1;
5542 }
5543
5544 /* No ACK in the segment */
5545
5546 if (th->rst) {
5547 /* rfc793:
5548 * "If the RST bit is set
5549 *
5550 * Otherwise (no ACK) drop the segment and return."
5551 */
5552
5553 goto discard_and_undo;
5554 }
5555
5556 /* PAWS check. */
5557 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5558 tcp_paws_reject(&tp->rx_opt, 0))
5559 goto discard_and_undo;
5560
5561 if (th->syn) {
5562 /* We see SYN without ACK. It is attempt of
5563 * simultaneous connect with crossed SYNs.
5564 * Particularly, it can be connect to self.
5565 */
5566 tcp_set_state(sk, TCP_SYN_RECV);
5567
5568 if (tp->rx_opt.saw_tstamp) {
5569 tp->rx_opt.tstamp_ok = 1;
5570 tcp_store_ts_recent(tp);
5571 tp->tcp_header_len =
5572 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5573 } else {
5574 tp->tcp_header_len = sizeof(struct tcphdr);
5575 }
5576
5577 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5578 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5579
5580 /* RFC1323: The window in SYN & SYN/ACK segments is
5581 * never scaled.
5582 */
5583 tp->snd_wnd = ntohs(th->window);
5584 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5585 tp->max_window = tp->snd_wnd;
5586
5587 TCP_ECN_rcv_syn(tp, th);
5588
5589 tcp_mtup_init(sk);
5590 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5591 tcp_initialize_rcv_mss(sk);
5592
5593 tcp_send_synack(sk);
5594#if 0
5595 /* Note, we could accept data and URG from this segment.
5596 * There are no obstacles to make this (except that we must
5597 * either change tcp_recvmsg() to prevent it from returning data
5598 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5599 *
5600 * However, if we ignore data in ACKless segments sometimes,
5601 * we have no reasons to accept it sometimes.
5602 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5603 * is not flawless. So, discard packet for sanity.
5604 * Uncomment this return to process the data.
5605 */
5606 return -1;
5607#else
5608 goto discard;
5609#endif
5610 }
5611 /* "fifth, if neither of the SYN or RST bits is set then
5612 * drop the segment and return."
5613 */
5614
5615discard_and_undo:
5616 tcp_clear_options(&tp->rx_opt);
5617 tp->rx_opt.mss_clamp = saved_clamp;
5618 goto discard;
5619
5620reset_and_undo:
5621 tcp_clear_options(&tp->rx_opt);
5622 tp->rx_opt.mss_clamp = saved_clamp;
5623 return 1;
5624}
5625
5626/*
5627 * This function implements the receiving procedure of RFC 793 for
5628 * all states except ESTABLISHED and TIME_WAIT.
5629 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5630 * address independent.
5631 */
5632
5633int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5634 const struct tcphdr *th, unsigned int len)
5635{
5636 struct tcp_sock *tp = tcp_sk(sk);
5637 struct inet_connection_sock *icsk = inet_csk(sk);
5638 struct request_sock *req;
5639 int queued = 0;
5640 bool acceptable;
5641 u32 synack_stamp;
5642
5643 tp->rx_opt.saw_tstamp = 0;
5644
5645 switch (sk->sk_state) {
5646 case TCP_CLOSE:
5647 goto discard;
5648
5649 case TCP_LISTEN:
5650 if (th->ack)
5651 return 1;
5652
5653 if (th->rst)
5654 goto discard;
5655
5656 if (th->syn) {
5657 if (th->fin)
5658 goto discard;
5659 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5660 return 1;
5661
5662 /* Now we have several options: In theory there is
5663 * nothing else in the frame. KA9Q has an option to
5664 * send data with the syn, BSD accepts data with the
5665 * syn up to the [to be] advertised window and
5666 * Solaris 2.1 gives you a protocol error. For now
5667 * we just ignore it, that fits the spec precisely
5668 * and avoids incompatibilities. It would be nice in
5669 * future to drop through and process the data.
5670 *
5671 * Now that TTCP is starting to be used we ought to
5672 * queue this data.
5673 * But, this leaves one open to an easy denial of
5674 * service attack, and SYN cookies can't defend
5675 * against this problem. So, we drop the data
5676 * in the interest of security over speed unless
5677 * it's still in use.
5678 */
5679 kfree_skb(skb);
5680 return 0;
5681 }
5682 goto discard;
5683
5684 case TCP_SYN_SENT:
5685 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5686 if (queued >= 0)
5687 return queued;
5688
5689 /* Do step6 onward by hand. */
5690 tcp_urg(sk, skb, th);
5691 __kfree_skb(skb);
5692 tcp_data_snd_check(sk);
5693 return 0;
5694 }
5695
5696 req = tp->fastopen_rsk;
5697 if (req != NULL) {
5698 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
5699 sk->sk_state != TCP_FIN_WAIT1);
5700
5701 if (tcp_check_req(sk, skb, req, NULL, true) == NULL)
5702 goto discard;
5703 }
5704
5705 if (!th->ack && !th->rst)
5706 goto discard;
5707
5708 if (!tcp_validate_incoming(sk, skb, th, 0))
5709 return 0;
5710
5711 /* step 5: check the ACK field */
5712 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
5713 FLAG_UPDATE_TS_RECENT) > 0;
5714
5715 switch (sk->sk_state) {
5716 case TCP_SYN_RECV:
5717 if (!acceptable)
5718 return 1;
5719
5720 /* Once we leave TCP_SYN_RECV, we no longer need req
5721 * so release it.
5722 */
5723 if (req) {
5724 synack_stamp = tcp_rsk(req)->snt_synack;
5725 tp->total_retrans = req->num_retrans;
5726 reqsk_fastopen_remove(sk, req, false);
5727 } else {
5728 synack_stamp = tp->lsndtime;
5729 /* Make sure socket is routed, for correct metrics. */
5730 icsk->icsk_af_ops->rebuild_header(sk);
5731 tcp_init_congestion_control(sk);
5732
5733 tcp_mtup_init(sk);
5734 tp->copied_seq = tp->rcv_nxt;
5735 tcp_init_buffer_space(sk);
5736 }
5737 smp_mb();
5738 tcp_set_state(sk, TCP_ESTABLISHED);
5739 sk->sk_state_change(sk);
5740
5741 /* Note, that this wakeup is only for marginal crossed SYN case.
5742 * Passively open sockets are not waked up, because
5743 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5744 */
5745 if (sk->sk_socket)
5746 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5747
5748 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5749 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
5750 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5751 tcp_synack_rtt_meas(sk, synack_stamp);
5752
5753 if (tp->rx_opt.tstamp_ok)
5754 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5755
5756 if (req) {
5757 /* Re-arm the timer because data may have been sent out.
5758 * This is similar to the regular data transmission case
5759 * when new data has just been ack'ed.
5760 *
5761 * (TFO) - we could try to be more aggressive and
5762 * retransmitting any data sooner based on when they
5763 * are sent out.
5764 */
5765 tcp_rearm_rto(sk);
5766 } else
5767 tcp_init_metrics(sk);
5768
5769 tcp_update_pacing_rate(sk);
5770
5771 /* Prevent spurious tcp_cwnd_restart() on first data packet */
5772 tp->lsndtime = tcp_time_stamp;
5773
5774 tcp_initialize_rcv_mss(sk);
5775 tcp_fast_path_on(tp);
5776 break;
5777
5778 case TCP_FIN_WAIT1: {
5779 struct dst_entry *dst;
5780 int tmo;
5781
5782 /* If we enter the TCP_FIN_WAIT1 state and we are a
5783 * Fast Open socket and this is the first acceptable
5784 * ACK we have received, this would have acknowledged
5785 * our SYNACK so stop the SYNACK timer.
5786 */
5787 if (req != NULL) {
5788 /* Return RST if ack_seq is invalid.
5789 * Note that RFC793 only says to generate a
5790 * DUPACK for it but for TCP Fast Open it seems
5791 * better to treat this case like TCP_SYN_RECV
5792 * above.
5793 */
5794 if (!acceptable)
5795 return 1;
5796 /* We no longer need the request sock. */
5797 reqsk_fastopen_remove(sk, req, false);
5798 tcp_rearm_rto(sk);
5799 }
5800 if (tp->snd_una != tp->write_seq)
5801 break;
5802
5803 tcp_set_state(sk, TCP_FIN_WAIT2);
5804 sk->sk_shutdown |= SEND_SHUTDOWN;
5805
5806 dst = __sk_dst_get(sk);
5807 if (dst)
5808 dst_confirm(dst);
5809
5810 if (!sock_flag(sk, SOCK_DEAD)) {
5811 /* Wake up lingering close() */
5812 sk->sk_state_change(sk);
5813 break;
5814 }
5815
5816 if (tp->linger2 < 0 ||
5817 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5818 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5819 tcp_done(sk);
5820 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5821 return 1;
5822 }
5823
5824 tmo = tcp_fin_time(sk);
5825 if (tmo > TCP_TIMEWAIT_LEN) {
5826 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5827 } else if (th->fin || sock_owned_by_user(sk)) {
5828 /* Bad case. We could lose such FIN otherwise.
5829 * It is not a big problem, but it looks confusing
5830 * and not so rare event. We still can lose it now,
5831 * if it spins in bh_lock_sock(), but it is really
5832 * marginal case.
5833 */
5834 inet_csk_reset_keepalive_timer(sk, tmo);
5835 } else {
5836 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5837 goto discard;
5838 }
5839 break;
5840 }
5841
5842 case TCP_CLOSING:
5843 if (tp->snd_una == tp->write_seq) {
5844 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5845 goto discard;
5846 }
5847 break;
5848
5849 case TCP_LAST_ACK:
5850 if (tp->snd_una == tp->write_seq) {
5851 tcp_update_metrics(sk);
5852 tcp_done(sk);
5853 goto discard;
5854 }
5855 break;
5856 }
5857
5858 /* step 6: check the URG bit */
5859 tcp_urg(sk, skb, th);
5860
5861 /* step 7: process the segment text */
5862 switch (sk->sk_state) {
5863 case TCP_CLOSE_WAIT:
5864 case TCP_CLOSING:
5865 case TCP_LAST_ACK:
5866 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5867 break;
5868 case TCP_FIN_WAIT1:
5869 case TCP_FIN_WAIT2:
5870 /* RFC 793 says to queue data in these states,
5871 * RFC 1122 says we MUST send a reset.
5872 * BSD 4.4 also does reset.
5873 */
5874 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5875 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5876 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5877 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5878 tcp_reset(sk);
5879 return 1;
5880 }
5881 }
5882 /* Fall through */
5883 case TCP_ESTABLISHED:
5884 tcp_data_queue(sk, skb);
5885 queued = 1;
5886 break;
5887 }
5888
5889 /* tcp_data could move socket to TIME-WAIT */
5890 if (sk->sk_state != TCP_CLOSE) {
5891 tcp_data_snd_check(sk);
5892 tcp_ack_snd_check(sk);
5893 }
5894
5895 if (!queued) {
5896discard:
5897 __kfree_skb(skb);
5898 }
5899 return 0;
5900}
5901EXPORT_SYMBOL(tcp_rcv_state_process);
1// SPDX-License-Identifier: GPL-2.0
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:
24 * Pedro Roque : Fast Retransmit/Recovery.
25 * Two receive queues.
26 * Retransmit queue handled by TCP.
27 * Better retransmit timer handling.
28 * New congestion avoidance.
29 * Header prediction.
30 * Variable renaming.
31 *
32 * Eric : Fast Retransmit.
33 * Randy Scott : MSS option defines.
34 * Eric Schenk : Fixes to slow start algorithm.
35 * Eric Schenk : Yet another double ACK bug.
36 * Eric Schenk : Delayed ACK bug fixes.
37 * Eric Schenk : Floyd style fast retrans war avoidance.
38 * David S. Miller : Don't allow zero congestion window.
39 * Eric Schenk : Fix retransmitter so that it sends
40 * next packet on ack of previous packet.
41 * Andi Kleen : Moved open_request checking here
42 * and process RSTs for open_requests.
43 * Andi Kleen : Better prune_queue, and other fixes.
44 * Andrey Savochkin: Fix RTT measurements in the presence of
45 * timestamps.
46 * Andrey Savochkin: Check sequence numbers correctly when
47 * removing SACKs due to in sequence incoming
48 * data segments.
49 * Andi Kleen: Make sure we never ack data there is not
50 * enough room for. Also make this condition
51 * a fatal error if it might still happen.
52 * Andi Kleen: Add tcp_measure_rcv_mss to make
53 * connections with MSS<min(MTU,ann. MSS)
54 * work without delayed acks.
55 * Andi Kleen: Process packets with PSH set in the
56 * fast path.
57 * J Hadi Salim: ECN support
58 * Andrei Gurtov,
59 * Pasi Sarolahti,
60 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
61 * engine. Lots of bugs are found.
62 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
63 */
64
65#define pr_fmt(fmt) "TCP: " fmt
66
67#include <linux/mm.h>
68#include <linux/slab.h>
69#include <linux/module.h>
70#include <linux/sysctl.h>
71#include <linux/kernel.h>
72#include <linux/prefetch.h>
73#include <net/dst.h>
74#include <net/tcp.h>
75#include <net/inet_common.h>
76#include <linux/ipsec.h>
77#include <asm/unaligned.h>
78#include <linux/errqueue.h>
79#include <trace/events/tcp.h>
80#include <linux/static_key.h>
81
82int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
83
84#define FLAG_DATA 0x01 /* Incoming frame contained data. */
85#define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
86#define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
87#define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
88#define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
89#define FLAG_DATA_SACKED 0x20 /* New SACK. */
90#define FLAG_ECE 0x40 /* ECE in this ACK */
91#define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
92#define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
93#define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
94#define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
95#define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
96#define FLAG_SET_XMIT_TIMER 0x1000 /* Set TLP or RTO timer */
97#define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
98#define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
99#define FLAG_NO_CHALLENGE_ACK 0x8000 /* do not call tcp_send_challenge_ack() */
100#define FLAG_ACK_MAYBE_DELAYED 0x10000 /* Likely a delayed ACK */
101
102#define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
103#define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
104#define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE|FLAG_DSACKING_ACK)
105#define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
106
107#define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
108#define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
109
110#define REXMIT_NONE 0 /* no loss recovery to do */
111#define REXMIT_LOST 1 /* retransmit packets marked lost */
112#define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */
113
114static void tcp_gro_dev_warn(struct sock *sk, const struct sk_buff *skb,
115 unsigned int len)
116{
117 static bool __once __read_mostly;
118
119 if (!__once) {
120 struct net_device *dev;
121
122 __once = true;
123
124 rcu_read_lock();
125 dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif);
126 if (!dev || len >= dev->mtu)
127 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n",
128 dev ? dev->name : "Unknown driver");
129 rcu_read_unlock();
130 }
131}
132
133/* Adapt the MSS value used to make delayed ack decision to the
134 * real world.
135 */
136static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
137{
138 struct inet_connection_sock *icsk = inet_csk(sk);
139 const unsigned int lss = icsk->icsk_ack.last_seg_size;
140 unsigned int len;
141
142 icsk->icsk_ack.last_seg_size = 0;
143
144 /* skb->len may jitter because of SACKs, even if peer
145 * sends good full-sized frames.
146 */
147 len = skb_shinfo(skb)->gso_size ? : skb->len;
148 if (len >= icsk->icsk_ack.rcv_mss) {
149 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len,
150 tcp_sk(sk)->advmss);
151 /* Account for possibly-removed options */
152 if (unlikely(len > icsk->icsk_ack.rcv_mss +
153 MAX_TCP_OPTION_SPACE))
154 tcp_gro_dev_warn(sk, skb, len);
155 } else {
156 /* Otherwise, we make more careful check taking into account,
157 * that SACKs block is variable.
158 *
159 * "len" is invariant segment length, including TCP header.
160 */
161 len += skb->data - skb_transport_header(skb);
162 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
163 /* If PSH is not set, packet should be
164 * full sized, provided peer TCP is not badly broken.
165 * This observation (if it is correct 8)) allows
166 * to handle super-low mtu links fairly.
167 */
168 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
169 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
170 /* Subtract also invariant (if peer is RFC compliant),
171 * tcp header plus fixed timestamp option length.
172 * Resulting "len" is MSS free of SACK jitter.
173 */
174 len -= tcp_sk(sk)->tcp_header_len;
175 icsk->icsk_ack.last_seg_size = len;
176 if (len == lss) {
177 icsk->icsk_ack.rcv_mss = len;
178 return;
179 }
180 }
181 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
182 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
183 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
184 }
185}
186
187static void tcp_incr_quickack(struct sock *sk)
188{
189 struct inet_connection_sock *icsk = inet_csk(sk);
190 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
191
192 if (quickacks == 0)
193 quickacks = 2;
194 if (quickacks > icsk->icsk_ack.quick)
195 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
196}
197
198static void tcp_enter_quickack_mode(struct sock *sk)
199{
200 struct inet_connection_sock *icsk = inet_csk(sk);
201 tcp_incr_quickack(sk);
202 icsk->icsk_ack.pingpong = 0;
203 icsk->icsk_ack.ato = TCP_ATO_MIN;
204}
205
206/* Send ACKs quickly, if "quick" count is not exhausted
207 * and the session is not interactive.
208 */
209
210static bool tcp_in_quickack_mode(struct sock *sk)
211{
212 const struct inet_connection_sock *icsk = inet_csk(sk);
213 const struct dst_entry *dst = __sk_dst_get(sk);
214
215 return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
216 (icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong);
217}
218
219static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
220{
221 if (tp->ecn_flags & TCP_ECN_OK)
222 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
223}
224
225static void tcp_ecn_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
226{
227 if (tcp_hdr(skb)->cwr)
228 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
229}
230
231static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
232{
233 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
234}
235
236static void __tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
237{
238 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
239 case INET_ECN_NOT_ECT:
240 /* Funny extension: if ECT is not set on a segment,
241 * and we already seen ECT on a previous segment,
242 * it is probably a retransmit.
243 */
244 if (tp->ecn_flags & TCP_ECN_SEEN)
245 tcp_enter_quickack_mode((struct sock *)tp);
246 break;
247 case INET_ECN_CE:
248 if (tcp_ca_needs_ecn((struct sock *)tp))
249 tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_IS_CE);
250
251 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
252 /* Better not delay acks, sender can have a very low cwnd */
253 tcp_enter_quickack_mode((struct sock *)tp);
254 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
255 }
256 tp->ecn_flags |= TCP_ECN_SEEN;
257 break;
258 default:
259 if (tcp_ca_needs_ecn((struct sock *)tp))
260 tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_NO_CE);
261 tp->ecn_flags |= TCP_ECN_SEEN;
262 break;
263 }
264}
265
266static void tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
267{
268 if (tp->ecn_flags & TCP_ECN_OK)
269 __tcp_ecn_check_ce(tp, skb);
270}
271
272static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
273{
274 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
275 tp->ecn_flags &= ~TCP_ECN_OK;
276}
277
278static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
279{
280 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
281 tp->ecn_flags &= ~TCP_ECN_OK;
282}
283
284static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
285{
286 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
287 return true;
288 return false;
289}
290
291/* Buffer size and advertised window tuning.
292 *
293 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
294 */
295
296static void tcp_sndbuf_expand(struct sock *sk)
297{
298 const struct tcp_sock *tp = tcp_sk(sk);
299 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
300 int sndmem, per_mss;
301 u32 nr_segs;
302
303 /* Worst case is non GSO/TSO : each frame consumes one skb
304 * and skb->head is kmalloced using power of two area of memory
305 */
306 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
307 MAX_TCP_HEADER +
308 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
309
310 per_mss = roundup_pow_of_two(per_mss) +
311 SKB_DATA_ALIGN(sizeof(struct sk_buff));
312
313 nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
314 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
315
316 /* Fast Recovery (RFC 5681 3.2) :
317 * Cubic needs 1.7 factor, rounded to 2 to include
318 * extra cushion (application might react slowly to EPOLLOUT)
319 */
320 sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2;
321 sndmem *= nr_segs * per_mss;
322
323 if (sk->sk_sndbuf < sndmem)
324 sk->sk_sndbuf = min(sndmem, sock_net(sk)->ipv4.sysctl_tcp_wmem[2]);
325}
326
327/* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
328 *
329 * All tcp_full_space() is split to two parts: "network" buffer, allocated
330 * forward and advertised in receiver window (tp->rcv_wnd) and
331 * "application buffer", required to isolate scheduling/application
332 * latencies from network.
333 * window_clamp is maximal advertised window. It can be less than
334 * tcp_full_space(), in this case tcp_full_space() - window_clamp
335 * is reserved for "application" buffer. The less window_clamp is
336 * the smoother our behaviour from viewpoint of network, but the lower
337 * throughput and the higher sensitivity of the connection to losses. 8)
338 *
339 * rcv_ssthresh is more strict window_clamp used at "slow start"
340 * phase to predict further behaviour of this connection.
341 * It is used for two goals:
342 * - to enforce header prediction at sender, even when application
343 * requires some significant "application buffer". It is check #1.
344 * - to prevent pruning of receive queue because of misprediction
345 * of receiver window. Check #2.
346 *
347 * The scheme does not work when sender sends good segments opening
348 * window and then starts to feed us spaghetti. But it should work
349 * in common situations. Otherwise, we have to rely on queue collapsing.
350 */
351
352/* Slow part of check#2. */
353static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
354{
355 struct tcp_sock *tp = tcp_sk(sk);
356 /* Optimize this! */
357 int truesize = tcp_win_from_space(sk, skb->truesize) >> 1;
358 int window = tcp_win_from_space(sk, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]) >> 1;
359
360 while (tp->rcv_ssthresh <= window) {
361 if (truesize <= skb->len)
362 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
363
364 truesize >>= 1;
365 window >>= 1;
366 }
367 return 0;
368}
369
370static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
371{
372 struct tcp_sock *tp = tcp_sk(sk);
373
374 /* Check #1 */
375 if (tp->rcv_ssthresh < tp->window_clamp &&
376 (int)tp->rcv_ssthresh < tcp_space(sk) &&
377 !tcp_under_memory_pressure(sk)) {
378 int incr;
379
380 /* Check #2. Increase window, if skb with such overhead
381 * will fit to rcvbuf in future.
382 */
383 if (tcp_win_from_space(sk, skb->truesize) <= skb->len)
384 incr = 2 * tp->advmss;
385 else
386 incr = __tcp_grow_window(sk, skb);
387
388 if (incr) {
389 incr = max_t(int, incr, 2 * skb->len);
390 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
391 tp->window_clamp);
392 inet_csk(sk)->icsk_ack.quick |= 1;
393 }
394 }
395}
396
397/* 3. Tuning rcvbuf, when connection enters established state. */
398static void tcp_fixup_rcvbuf(struct sock *sk)
399{
400 u32 mss = tcp_sk(sk)->advmss;
401 int rcvmem;
402
403 rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) *
404 tcp_default_init_rwnd(mss);
405
406 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
407 * Allow enough cushion so that sender is not limited by our window
408 */
409 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf)
410 rcvmem <<= 2;
411
412 if (sk->sk_rcvbuf < rcvmem)
413 sk->sk_rcvbuf = min(rcvmem, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
414}
415
416/* 4. Try to fixup all. It is made immediately after connection enters
417 * established state.
418 */
419void tcp_init_buffer_space(struct sock *sk)
420{
421 int tcp_app_win = sock_net(sk)->ipv4.sysctl_tcp_app_win;
422 struct tcp_sock *tp = tcp_sk(sk);
423 int maxwin;
424
425 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
426 tcp_fixup_rcvbuf(sk);
427 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
428 tcp_sndbuf_expand(sk);
429
430 tp->rcvq_space.space = tp->rcv_wnd;
431 tcp_mstamp_refresh(tp);
432 tp->rcvq_space.time = tp->tcp_mstamp;
433 tp->rcvq_space.seq = tp->copied_seq;
434
435 maxwin = tcp_full_space(sk);
436
437 if (tp->window_clamp >= maxwin) {
438 tp->window_clamp = maxwin;
439
440 if (tcp_app_win && maxwin > 4 * tp->advmss)
441 tp->window_clamp = max(maxwin -
442 (maxwin >> tcp_app_win),
443 4 * tp->advmss);
444 }
445
446 /* Force reservation of one segment. */
447 if (tcp_app_win &&
448 tp->window_clamp > 2 * tp->advmss &&
449 tp->window_clamp + tp->advmss > maxwin)
450 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
451
452 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
453 tp->snd_cwnd_stamp = tcp_jiffies32;
454}
455
456/* 5. Recalculate window clamp after socket hit its memory bounds. */
457static void tcp_clamp_window(struct sock *sk)
458{
459 struct tcp_sock *tp = tcp_sk(sk);
460 struct inet_connection_sock *icsk = inet_csk(sk);
461 struct net *net = sock_net(sk);
462
463 icsk->icsk_ack.quick = 0;
464
465 if (sk->sk_rcvbuf < net->ipv4.sysctl_tcp_rmem[2] &&
466 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
467 !tcp_under_memory_pressure(sk) &&
468 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
469 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
470 net->ipv4.sysctl_tcp_rmem[2]);
471 }
472 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
473 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
474}
475
476/* Initialize RCV_MSS value.
477 * RCV_MSS is an our guess about MSS used by the peer.
478 * We haven't any direct information about the MSS.
479 * It's better to underestimate the RCV_MSS rather than overestimate.
480 * Overestimations make us ACKing less frequently than needed.
481 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
482 */
483void tcp_initialize_rcv_mss(struct sock *sk)
484{
485 const struct tcp_sock *tp = tcp_sk(sk);
486 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
487
488 hint = min(hint, tp->rcv_wnd / 2);
489 hint = min(hint, TCP_MSS_DEFAULT);
490 hint = max(hint, TCP_MIN_MSS);
491
492 inet_csk(sk)->icsk_ack.rcv_mss = hint;
493}
494EXPORT_SYMBOL(tcp_initialize_rcv_mss);
495
496/* Receiver "autotuning" code.
497 *
498 * The algorithm for RTT estimation w/o timestamps is based on
499 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
500 * <http://public.lanl.gov/radiant/pubs.html#DRS>
501 *
502 * More detail on this code can be found at
503 * <http://staff.psc.edu/jheffner/>,
504 * though this reference is out of date. A new paper
505 * is pending.
506 */
507static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
508{
509 u32 new_sample = tp->rcv_rtt_est.rtt_us;
510 long m = sample;
511
512 if (new_sample != 0) {
513 /* If we sample in larger samples in the non-timestamp
514 * case, we could grossly overestimate the RTT especially
515 * with chatty applications or bulk transfer apps which
516 * are stalled on filesystem I/O.
517 *
518 * Also, since we are only going for a minimum in the
519 * non-timestamp case, we do not smooth things out
520 * else with timestamps disabled convergence takes too
521 * long.
522 */
523 if (!win_dep) {
524 m -= (new_sample >> 3);
525 new_sample += m;
526 } else {
527 m <<= 3;
528 if (m < new_sample)
529 new_sample = m;
530 }
531 } else {
532 /* No previous measure. */
533 new_sample = m << 3;
534 }
535
536 tp->rcv_rtt_est.rtt_us = new_sample;
537}
538
539static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
540{
541 u32 delta_us;
542
543 if (tp->rcv_rtt_est.time == 0)
544 goto new_measure;
545 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
546 return;
547 delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time);
548 if (!delta_us)
549 delta_us = 1;
550 tcp_rcv_rtt_update(tp, delta_us, 1);
551
552new_measure:
553 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
554 tp->rcv_rtt_est.time = tp->tcp_mstamp;
555}
556
557static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
558 const struct sk_buff *skb)
559{
560 struct tcp_sock *tp = tcp_sk(sk);
561
562 if (tp->rx_opt.rcv_tsecr &&
563 (TCP_SKB_CB(skb)->end_seq -
564 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss)) {
565 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
566 u32 delta_us;
567
568 if (!delta)
569 delta = 1;
570 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
571 tcp_rcv_rtt_update(tp, delta_us, 0);
572 }
573}
574
575/*
576 * This function should be called every time data is copied to user space.
577 * It calculates the appropriate TCP receive buffer space.
578 */
579void tcp_rcv_space_adjust(struct sock *sk)
580{
581 struct tcp_sock *tp = tcp_sk(sk);
582 u32 copied;
583 int time;
584
585 tcp_mstamp_refresh(tp);
586 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time);
587 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0)
588 return;
589
590 /* Number of bytes copied to user in last RTT */
591 copied = tp->copied_seq - tp->rcvq_space.seq;
592 if (copied <= tp->rcvq_space.space)
593 goto new_measure;
594
595 /* A bit of theory :
596 * copied = bytes received in previous RTT, our base window
597 * To cope with packet losses, we need a 2x factor
598 * To cope with slow start, and sender growing its cwin by 100 %
599 * every RTT, we need a 4x factor, because the ACK we are sending
600 * now is for the next RTT, not the current one :
601 * <prev RTT . ><current RTT .. ><next RTT .... >
602 */
603
604 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf &&
605 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
606 int rcvmem, rcvbuf;
607 u64 rcvwin, grow;
608
609 /* minimal window to cope with packet losses, assuming
610 * steady state. Add some cushion because of small variations.
611 */
612 rcvwin = ((u64)copied << 1) + 16 * tp->advmss;
613
614 /* Accommodate for sender rate increase (eg. slow start) */
615 grow = rcvwin * (copied - tp->rcvq_space.space);
616 do_div(grow, tp->rcvq_space.space);
617 rcvwin += (grow << 1);
618
619 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
620 while (tcp_win_from_space(sk, rcvmem) < tp->advmss)
621 rcvmem += 128;
622
623 do_div(rcvwin, tp->advmss);
624 rcvbuf = min_t(u64, rcvwin * rcvmem,
625 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
626 if (rcvbuf > sk->sk_rcvbuf) {
627 sk->sk_rcvbuf = rcvbuf;
628
629 /* Make the window clamp follow along. */
630 tp->window_clamp = tcp_win_from_space(sk, rcvbuf);
631 }
632 }
633 tp->rcvq_space.space = copied;
634
635new_measure:
636 tp->rcvq_space.seq = tp->copied_seq;
637 tp->rcvq_space.time = tp->tcp_mstamp;
638}
639
640/* There is something which you must keep in mind when you analyze the
641 * behavior of the tp->ato delayed ack timeout interval. When a
642 * connection starts up, we want to ack as quickly as possible. The
643 * problem is that "good" TCP's do slow start at the beginning of data
644 * transmission. The means that until we send the first few ACK's the
645 * sender will sit on his end and only queue most of his data, because
646 * he can only send snd_cwnd unacked packets at any given time. For
647 * each ACK we send, he increments snd_cwnd and transmits more of his
648 * queue. -DaveM
649 */
650static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
651{
652 struct tcp_sock *tp = tcp_sk(sk);
653 struct inet_connection_sock *icsk = inet_csk(sk);
654 u32 now;
655
656 inet_csk_schedule_ack(sk);
657
658 tcp_measure_rcv_mss(sk, skb);
659
660 tcp_rcv_rtt_measure(tp);
661
662 now = tcp_jiffies32;
663
664 if (!icsk->icsk_ack.ato) {
665 /* The _first_ data packet received, initialize
666 * delayed ACK engine.
667 */
668 tcp_incr_quickack(sk);
669 icsk->icsk_ack.ato = TCP_ATO_MIN;
670 } else {
671 int m = now - icsk->icsk_ack.lrcvtime;
672
673 if (m <= TCP_ATO_MIN / 2) {
674 /* The fastest case is the first. */
675 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
676 } else if (m < icsk->icsk_ack.ato) {
677 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
678 if (icsk->icsk_ack.ato > icsk->icsk_rto)
679 icsk->icsk_ack.ato = icsk->icsk_rto;
680 } else if (m > icsk->icsk_rto) {
681 /* Too long gap. Apparently sender failed to
682 * restart window, so that we send ACKs quickly.
683 */
684 tcp_incr_quickack(sk);
685 sk_mem_reclaim(sk);
686 }
687 }
688 icsk->icsk_ack.lrcvtime = now;
689
690 tcp_ecn_check_ce(tp, skb);
691
692 if (skb->len >= 128)
693 tcp_grow_window(sk, skb);
694}
695
696/* Called to compute a smoothed rtt estimate. The data fed to this
697 * routine either comes from timestamps, or from segments that were
698 * known _not_ to have been retransmitted [see Karn/Partridge
699 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
700 * piece by Van Jacobson.
701 * NOTE: the next three routines used to be one big routine.
702 * To save cycles in the RFC 1323 implementation it was better to break
703 * it up into three procedures. -- erics
704 */
705static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
706{
707 struct tcp_sock *tp = tcp_sk(sk);
708 long m = mrtt_us; /* RTT */
709 u32 srtt = tp->srtt_us;
710
711 /* The following amusing code comes from Jacobson's
712 * article in SIGCOMM '88. Note that rtt and mdev
713 * are scaled versions of rtt and mean deviation.
714 * This is designed to be as fast as possible
715 * m stands for "measurement".
716 *
717 * On a 1990 paper the rto value is changed to:
718 * RTO = rtt + 4 * mdev
719 *
720 * Funny. This algorithm seems to be very broken.
721 * These formulae increase RTO, when it should be decreased, increase
722 * too slowly, when it should be increased quickly, decrease too quickly
723 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
724 * does not matter how to _calculate_ it. Seems, it was trap
725 * that VJ failed to avoid. 8)
726 */
727 if (srtt != 0) {
728 m -= (srtt >> 3); /* m is now error in rtt est */
729 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
730 if (m < 0) {
731 m = -m; /* m is now abs(error) */
732 m -= (tp->mdev_us >> 2); /* similar update on mdev */
733 /* This is similar to one of Eifel findings.
734 * Eifel blocks mdev updates when rtt decreases.
735 * This solution is a bit different: we use finer gain
736 * for mdev in this case (alpha*beta).
737 * Like Eifel it also prevents growth of rto,
738 * but also it limits too fast rto decreases,
739 * happening in pure Eifel.
740 */
741 if (m > 0)
742 m >>= 3;
743 } else {
744 m -= (tp->mdev_us >> 2); /* similar update on mdev */
745 }
746 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
747 if (tp->mdev_us > tp->mdev_max_us) {
748 tp->mdev_max_us = tp->mdev_us;
749 if (tp->mdev_max_us > tp->rttvar_us)
750 tp->rttvar_us = tp->mdev_max_us;
751 }
752 if (after(tp->snd_una, tp->rtt_seq)) {
753 if (tp->mdev_max_us < tp->rttvar_us)
754 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
755 tp->rtt_seq = tp->snd_nxt;
756 tp->mdev_max_us = tcp_rto_min_us(sk);
757 }
758 } else {
759 /* no previous measure. */
760 srtt = m << 3; /* take the measured time to be rtt */
761 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
762 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
763 tp->mdev_max_us = tp->rttvar_us;
764 tp->rtt_seq = tp->snd_nxt;
765 }
766 tp->srtt_us = max(1U, srtt);
767}
768
769static void tcp_update_pacing_rate(struct sock *sk)
770{
771 const struct tcp_sock *tp = tcp_sk(sk);
772 u64 rate;
773
774 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
775 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
776
777 /* current rate is (cwnd * mss) / srtt
778 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
779 * In Congestion Avoidance phase, set it to 120 % the current rate.
780 *
781 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
782 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
783 * end of slow start and should slow down.
784 */
785 if (tp->snd_cwnd < tp->snd_ssthresh / 2)
786 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio;
787 else
788 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio;
789
790 rate *= max(tp->snd_cwnd, tp->packets_out);
791
792 if (likely(tp->srtt_us))
793 do_div(rate, tp->srtt_us);
794
795 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate
796 * without any lock. We want to make sure compiler wont store
797 * intermediate values in this location.
798 */
799 WRITE_ONCE(sk->sk_pacing_rate, min_t(u64, rate,
800 sk->sk_max_pacing_rate));
801}
802
803/* Calculate rto without backoff. This is the second half of Van Jacobson's
804 * routine referred to above.
805 */
806static void tcp_set_rto(struct sock *sk)
807{
808 const struct tcp_sock *tp = tcp_sk(sk);
809 /* Old crap is replaced with new one. 8)
810 *
811 * More seriously:
812 * 1. If rtt variance happened to be less 50msec, it is hallucination.
813 * It cannot be less due to utterly erratic ACK generation made
814 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
815 * to do with delayed acks, because at cwnd>2 true delack timeout
816 * is invisible. Actually, Linux-2.4 also generates erratic
817 * ACKs in some circumstances.
818 */
819 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
820
821 /* 2. Fixups made earlier cannot be right.
822 * If we do not estimate RTO correctly without them,
823 * all the algo is pure shit and should be replaced
824 * with correct one. It is exactly, which we pretend to do.
825 */
826
827 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
828 * guarantees that rto is higher.
829 */
830 tcp_bound_rto(sk);
831}
832
833__u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
834{
835 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
836
837 if (!cwnd)
838 cwnd = TCP_INIT_CWND;
839 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
840}
841
842/* Take a notice that peer is sending D-SACKs */
843static void tcp_dsack_seen(struct tcp_sock *tp)
844{
845 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
846 tp->rack.dsack_seen = 1;
847}
848
849/* It's reordering when higher sequence was delivered (i.e. sacked) before
850 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering
851 * distance is approximated in full-mss packet distance ("reordering").
852 */
853static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq,
854 const int ts)
855{
856 struct tcp_sock *tp = tcp_sk(sk);
857 const u32 mss = tp->mss_cache;
858 u32 fack, metric;
859
860 fack = tcp_highest_sack_seq(tp);
861 if (!before(low_seq, fack))
862 return;
863
864 metric = fack - low_seq;
865 if ((metric > tp->reordering * mss) && mss) {
866#if FASTRETRANS_DEBUG > 1
867 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
868 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
869 tp->reordering,
870 0,
871 tp->sacked_out,
872 tp->undo_marker ? tp->undo_retrans : 0);
873#endif
874 tp->reordering = min_t(u32, (metric + mss - 1) / mss,
875 sock_net(sk)->ipv4.sysctl_tcp_max_reordering);
876 }
877
878 tp->rack.reord = 1;
879 /* This exciting event is worth to be remembered. 8) */
880 NET_INC_STATS(sock_net(sk),
881 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER);
882}
883
884/* This must be called before lost_out is incremented */
885static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
886{
887 if (!tp->retransmit_skb_hint ||
888 before(TCP_SKB_CB(skb)->seq,
889 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
890 tp->retransmit_skb_hint = skb;
891}
892
893/* Sum the number of packets on the wire we have marked as lost.
894 * There are two cases we care about here:
895 * a) Packet hasn't been marked lost (nor retransmitted),
896 * and this is the first loss.
897 * b) Packet has been marked both lost and retransmitted,
898 * and this means we think it was lost again.
899 */
900static void tcp_sum_lost(struct tcp_sock *tp, struct sk_buff *skb)
901{
902 __u8 sacked = TCP_SKB_CB(skb)->sacked;
903
904 if (!(sacked & TCPCB_LOST) ||
905 ((sacked & TCPCB_LOST) && (sacked & TCPCB_SACKED_RETRANS)))
906 tp->lost += tcp_skb_pcount(skb);
907}
908
909static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
910{
911 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
912 tcp_verify_retransmit_hint(tp, skb);
913
914 tp->lost_out += tcp_skb_pcount(skb);
915 tcp_sum_lost(tp, skb);
916 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
917 }
918}
919
920void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb)
921{
922 tcp_verify_retransmit_hint(tp, skb);
923
924 tcp_sum_lost(tp, skb);
925 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
926 tp->lost_out += tcp_skb_pcount(skb);
927 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
928 }
929}
930
931/* This procedure tags the retransmission queue when SACKs arrive.
932 *
933 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
934 * Packets in queue with these bits set are counted in variables
935 * sacked_out, retrans_out and lost_out, correspondingly.
936 *
937 * Valid combinations are:
938 * Tag InFlight Description
939 * 0 1 - orig segment is in flight.
940 * S 0 - nothing flies, orig reached receiver.
941 * L 0 - nothing flies, orig lost by net.
942 * R 2 - both orig and retransmit are in flight.
943 * L|R 1 - orig is lost, retransmit is in flight.
944 * S|R 1 - orig reached receiver, retrans is still in flight.
945 * (L|S|R is logically valid, it could occur when L|R is sacked,
946 * but it is equivalent to plain S and code short-curcuits it to S.
947 * L|S is logically invalid, it would mean -1 packet in flight 8))
948 *
949 * These 6 states form finite state machine, controlled by the following events:
950 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
951 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
952 * 3. Loss detection event of two flavors:
953 * A. Scoreboard estimator decided the packet is lost.
954 * A'. Reno "three dupacks" marks head of queue lost.
955 * B. SACK arrives sacking SND.NXT at the moment, when the
956 * segment was retransmitted.
957 * 4. D-SACK added new rule: D-SACK changes any tag to S.
958 *
959 * It is pleasant to note, that state diagram turns out to be commutative,
960 * so that we are allowed not to be bothered by order of our actions,
961 * when multiple events arrive simultaneously. (see the function below).
962 *
963 * Reordering detection.
964 * --------------------
965 * Reordering metric is maximal distance, which a packet can be displaced
966 * in packet stream. With SACKs we can estimate it:
967 *
968 * 1. SACK fills old hole and the corresponding segment was not
969 * ever retransmitted -> reordering. Alas, we cannot use it
970 * when segment was retransmitted.
971 * 2. The last flaw is solved with D-SACK. D-SACK arrives
972 * for retransmitted and already SACKed segment -> reordering..
973 * Both of these heuristics are not used in Loss state, when we cannot
974 * account for retransmits accurately.
975 *
976 * SACK block validation.
977 * ----------------------
978 *
979 * SACK block range validation checks that the received SACK block fits to
980 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
981 * Note that SND.UNA is not included to the range though being valid because
982 * it means that the receiver is rather inconsistent with itself reporting
983 * SACK reneging when it should advance SND.UNA. Such SACK block this is
984 * perfectly valid, however, in light of RFC2018 which explicitly states
985 * that "SACK block MUST reflect the newest segment. Even if the newest
986 * segment is going to be discarded ...", not that it looks very clever
987 * in case of head skb. Due to potentional receiver driven attacks, we
988 * choose to avoid immediate execution of a walk in write queue due to
989 * reneging and defer head skb's loss recovery to standard loss recovery
990 * procedure that will eventually trigger (nothing forbids us doing this).
991 *
992 * Implements also blockage to start_seq wrap-around. Problem lies in the
993 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
994 * there's no guarantee that it will be before snd_nxt (n). The problem
995 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
996 * wrap (s_w):
997 *
998 * <- outs wnd -> <- wrapzone ->
999 * u e n u_w e_w s n_w
1000 * | | | | | | |
1001 * |<------------+------+----- TCP seqno space --------------+---------->|
1002 * ...-- <2^31 ->| |<--------...
1003 * ...---- >2^31 ------>| |<--------...
1004 *
1005 * Current code wouldn't be vulnerable but it's better still to discard such
1006 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1007 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1008 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1009 * equal to the ideal case (infinite seqno space without wrap caused issues).
1010 *
1011 * With D-SACK the lower bound is extended to cover sequence space below
1012 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1013 * again, D-SACK block must not to go across snd_una (for the same reason as
1014 * for the normal SACK blocks, explained above). But there all simplicity
1015 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1016 * fully below undo_marker they do not affect behavior in anyway and can
1017 * therefore be safely ignored. In rare cases (which are more or less
1018 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1019 * fragmentation and packet reordering past skb's retransmission. To consider
1020 * them correctly, the acceptable range must be extended even more though
1021 * the exact amount is rather hard to quantify. However, tp->max_window can
1022 * be used as an exaggerated estimate.
1023 */
1024static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1025 u32 start_seq, u32 end_seq)
1026{
1027 /* Too far in future, or reversed (interpretation is ambiguous) */
1028 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1029 return false;
1030
1031 /* Nasty start_seq wrap-around check (see comments above) */
1032 if (!before(start_seq, tp->snd_nxt))
1033 return false;
1034
1035 /* In outstanding window? ...This is valid exit for D-SACKs too.
1036 * start_seq == snd_una is non-sensical (see comments above)
1037 */
1038 if (after(start_seq, tp->snd_una))
1039 return true;
1040
1041 if (!is_dsack || !tp->undo_marker)
1042 return false;
1043
1044 /* ...Then it's D-SACK, and must reside below snd_una completely */
1045 if (after(end_seq, tp->snd_una))
1046 return false;
1047
1048 if (!before(start_seq, tp->undo_marker))
1049 return true;
1050
1051 /* Too old */
1052 if (!after(end_seq, tp->undo_marker))
1053 return false;
1054
1055 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1056 * start_seq < undo_marker and end_seq >= undo_marker.
1057 */
1058 return !before(start_seq, end_seq - tp->max_window);
1059}
1060
1061static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1062 struct tcp_sack_block_wire *sp, int num_sacks,
1063 u32 prior_snd_una)
1064{
1065 struct tcp_sock *tp = tcp_sk(sk);
1066 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1067 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1068 bool dup_sack = false;
1069
1070 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1071 dup_sack = true;
1072 tcp_dsack_seen(tp);
1073 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1074 } else if (num_sacks > 1) {
1075 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1076 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1077
1078 if (!after(end_seq_0, end_seq_1) &&
1079 !before(start_seq_0, start_seq_1)) {
1080 dup_sack = true;
1081 tcp_dsack_seen(tp);
1082 NET_INC_STATS(sock_net(sk),
1083 LINUX_MIB_TCPDSACKOFORECV);
1084 }
1085 }
1086
1087 /* D-SACK for already forgotten data... Do dumb counting. */
1088 if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
1089 !after(end_seq_0, prior_snd_una) &&
1090 after(end_seq_0, tp->undo_marker))
1091 tp->undo_retrans--;
1092
1093 return dup_sack;
1094}
1095
1096struct tcp_sacktag_state {
1097 u32 reord;
1098 /* Timestamps for earliest and latest never-retransmitted segment
1099 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1100 * but congestion control should still get an accurate delay signal.
1101 */
1102 u64 first_sackt;
1103 u64 last_sackt;
1104 struct rate_sample *rate;
1105 int flag;
1106 unsigned int mss_now;
1107};
1108
1109/* Check if skb is fully within the SACK block. In presence of GSO skbs,
1110 * the incoming SACK may not exactly match but we can find smaller MSS
1111 * aligned portion of it that matches. Therefore we might need to fragment
1112 * which may fail and creates some hassle (caller must handle error case
1113 * returns).
1114 *
1115 * FIXME: this could be merged to shift decision code
1116 */
1117static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1118 u32 start_seq, u32 end_seq)
1119{
1120 int err;
1121 bool in_sack;
1122 unsigned int pkt_len;
1123 unsigned int mss;
1124
1125 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1126 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1127
1128 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1129 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1130 mss = tcp_skb_mss(skb);
1131 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1132
1133 if (!in_sack) {
1134 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1135 if (pkt_len < mss)
1136 pkt_len = mss;
1137 } else {
1138 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1139 if (pkt_len < mss)
1140 return -EINVAL;
1141 }
1142
1143 /* Round if necessary so that SACKs cover only full MSSes
1144 * and/or the remaining small portion (if present)
1145 */
1146 if (pkt_len > mss) {
1147 unsigned int new_len = (pkt_len / mss) * mss;
1148 if (!in_sack && new_len < pkt_len)
1149 new_len += mss;
1150 pkt_len = new_len;
1151 }
1152
1153 if (pkt_len >= skb->len && !in_sack)
1154 return 0;
1155
1156 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
1157 pkt_len, mss, GFP_ATOMIC);
1158 if (err < 0)
1159 return err;
1160 }
1161
1162 return in_sack;
1163}
1164
1165/* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1166static u8 tcp_sacktag_one(struct sock *sk,
1167 struct tcp_sacktag_state *state, u8 sacked,
1168 u32 start_seq, u32 end_seq,
1169 int dup_sack, int pcount,
1170 u64 xmit_time)
1171{
1172 struct tcp_sock *tp = tcp_sk(sk);
1173
1174 /* Account D-SACK for retransmitted packet. */
1175 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1176 if (tp->undo_marker && tp->undo_retrans > 0 &&
1177 after(end_seq, tp->undo_marker))
1178 tp->undo_retrans--;
1179 if ((sacked & TCPCB_SACKED_ACKED) &&
1180 before(start_seq, state->reord))
1181 state->reord = start_seq;
1182 }
1183
1184 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1185 if (!after(end_seq, tp->snd_una))
1186 return sacked;
1187
1188 if (!(sacked & TCPCB_SACKED_ACKED)) {
1189 tcp_rack_advance(tp, sacked, end_seq, xmit_time);
1190
1191 if (sacked & TCPCB_SACKED_RETRANS) {
1192 /* If the segment is not tagged as lost,
1193 * we do not clear RETRANS, believing
1194 * that retransmission is still in flight.
1195 */
1196 if (sacked & TCPCB_LOST) {
1197 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1198 tp->lost_out -= pcount;
1199 tp->retrans_out -= pcount;
1200 }
1201 } else {
1202 if (!(sacked & TCPCB_RETRANS)) {
1203 /* New sack for not retransmitted frame,
1204 * which was in hole. It is reordering.
1205 */
1206 if (before(start_seq,
1207 tcp_highest_sack_seq(tp)) &&
1208 before(start_seq, state->reord))
1209 state->reord = start_seq;
1210
1211 if (!after(end_seq, tp->high_seq))
1212 state->flag |= FLAG_ORIG_SACK_ACKED;
1213 if (state->first_sackt == 0)
1214 state->first_sackt = xmit_time;
1215 state->last_sackt = xmit_time;
1216 }
1217
1218 if (sacked & TCPCB_LOST) {
1219 sacked &= ~TCPCB_LOST;
1220 tp->lost_out -= pcount;
1221 }
1222 }
1223
1224 sacked |= TCPCB_SACKED_ACKED;
1225 state->flag |= FLAG_DATA_SACKED;
1226 tp->sacked_out += pcount;
1227 tp->delivered += pcount; /* Out-of-order packets delivered */
1228
1229 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1230 if (tp->lost_skb_hint &&
1231 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1232 tp->lost_cnt_hint += pcount;
1233 }
1234
1235 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1236 * frames and clear it. undo_retrans is decreased above, L|R frames
1237 * are accounted above as well.
1238 */
1239 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1240 sacked &= ~TCPCB_SACKED_RETRANS;
1241 tp->retrans_out -= pcount;
1242 }
1243
1244 return sacked;
1245}
1246
1247/* Shift newly-SACKed bytes from this skb to the immediately previous
1248 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1249 */
1250static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev,
1251 struct sk_buff *skb,
1252 struct tcp_sacktag_state *state,
1253 unsigned int pcount, int shifted, int mss,
1254 bool dup_sack)
1255{
1256 struct tcp_sock *tp = tcp_sk(sk);
1257 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1258 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1259
1260 BUG_ON(!pcount);
1261
1262 /* Adjust counters and hints for the newly sacked sequence
1263 * range but discard the return value since prev is already
1264 * marked. We must tag the range first because the seq
1265 * advancement below implicitly advances
1266 * tcp_highest_sack_seq() when skb is highest_sack.
1267 */
1268 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1269 start_seq, end_seq, dup_sack, pcount,
1270 skb->skb_mstamp);
1271 tcp_rate_skb_delivered(sk, skb, state->rate);
1272
1273 if (skb == tp->lost_skb_hint)
1274 tp->lost_cnt_hint += pcount;
1275
1276 TCP_SKB_CB(prev)->end_seq += shifted;
1277 TCP_SKB_CB(skb)->seq += shifted;
1278
1279 tcp_skb_pcount_add(prev, pcount);
1280 BUG_ON(tcp_skb_pcount(skb) < pcount);
1281 tcp_skb_pcount_add(skb, -pcount);
1282
1283 /* When we're adding to gso_segs == 1, gso_size will be zero,
1284 * in theory this shouldn't be necessary but as long as DSACK
1285 * code can come after this skb later on it's better to keep
1286 * setting gso_size to something.
1287 */
1288 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1289 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1290
1291 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1292 if (tcp_skb_pcount(skb) <= 1)
1293 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1294
1295 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1296 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1297
1298 if (skb->len > 0) {
1299 BUG_ON(!tcp_skb_pcount(skb));
1300 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1301 return false;
1302 }
1303
1304 /* Whole SKB was eaten :-) */
1305
1306 if (skb == tp->retransmit_skb_hint)
1307 tp->retransmit_skb_hint = prev;
1308 if (skb == tp->lost_skb_hint) {
1309 tp->lost_skb_hint = prev;
1310 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1311 }
1312
1313 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1314 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
1315 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1316 TCP_SKB_CB(prev)->end_seq++;
1317
1318 if (skb == tcp_highest_sack(sk))
1319 tcp_advance_highest_sack(sk, skb);
1320
1321 tcp_skb_collapse_tstamp(prev, skb);
1322 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp))
1323 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0;
1324
1325 tcp_rtx_queue_unlink_and_free(skb, sk);
1326
1327 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
1328
1329 return true;
1330}
1331
1332/* I wish gso_size would have a bit more sane initialization than
1333 * something-or-zero which complicates things
1334 */
1335static int tcp_skb_seglen(const struct sk_buff *skb)
1336{
1337 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1338}
1339
1340/* Shifting pages past head area doesn't work */
1341static int skb_can_shift(const struct sk_buff *skb)
1342{
1343 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1344}
1345
1346/* Try collapsing SACK blocks spanning across multiple skbs to a single
1347 * skb.
1348 */
1349static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1350 struct tcp_sacktag_state *state,
1351 u32 start_seq, u32 end_seq,
1352 bool dup_sack)
1353{
1354 struct tcp_sock *tp = tcp_sk(sk);
1355 struct sk_buff *prev;
1356 int mss;
1357 int pcount = 0;
1358 int len;
1359 int in_sack;
1360
1361 /* Normally R but no L won't result in plain S */
1362 if (!dup_sack &&
1363 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1364 goto fallback;
1365 if (!skb_can_shift(skb))
1366 goto fallback;
1367 /* This frame is about to be dropped (was ACKed). */
1368 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1369 goto fallback;
1370
1371 /* Can only happen with delayed DSACK + discard craziness */
1372 prev = skb_rb_prev(skb);
1373 if (!prev)
1374 goto fallback;
1375
1376 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1377 goto fallback;
1378
1379 if (!tcp_skb_can_collapse_to(prev))
1380 goto fallback;
1381
1382 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1383 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1384
1385 if (in_sack) {
1386 len = skb->len;
1387 pcount = tcp_skb_pcount(skb);
1388 mss = tcp_skb_seglen(skb);
1389
1390 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1391 * drop this restriction as unnecessary
1392 */
1393 if (mss != tcp_skb_seglen(prev))
1394 goto fallback;
1395 } else {
1396 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1397 goto noop;
1398 /* CHECKME: This is non-MSS split case only?, this will
1399 * cause skipped skbs due to advancing loop btw, original
1400 * has that feature too
1401 */
1402 if (tcp_skb_pcount(skb) <= 1)
1403 goto noop;
1404
1405 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1406 if (!in_sack) {
1407 /* TODO: head merge to next could be attempted here
1408 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1409 * though it might not be worth of the additional hassle
1410 *
1411 * ...we can probably just fallback to what was done
1412 * previously. We could try merging non-SACKed ones
1413 * as well but it probably isn't going to buy off
1414 * because later SACKs might again split them, and
1415 * it would make skb timestamp tracking considerably
1416 * harder problem.
1417 */
1418 goto fallback;
1419 }
1420
1421 len = end_seq - TCP_SKB_CB(skb)->seq;
1422 BUG_ON(len < 0);
1423 BUG_ON(len > skb->len);
1424
1425 /* MSS boundaries should be honoured or else pcount will
1426 * severely break even though it makes things bit trickier.
1427 * Optimize common case to avoid most of the divides
1428 */
1429 mss = tcp_skb_mss(skb);
1430
1431 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1432 * drop this restriction as unnecessary
1433 */
1434 if (mss != tcp_skb_seglen(prev))
1435 goto fallback;
1436
1437 if (len == mss) {
1438 pcount = 1;
1439 } else if (len < mss) {
1440 goto noop;
1441 } else {
1442 pcount = len / mss;
1443 len = pcount * mss;
1444 }
1445 }
1446
1447 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1448 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1449 goto fallback;
1450
1451 if (!skb_shift(prev, skb, len))
1452 goto fallback;
1453 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack))
1454 goto out;
1455
1456 /* Hole filled allows collapsing with the next as well, this is very
1457 * useful when hole on every nth skb pattern happens
1458 */
1459 skb = skb_rb_next(prev);
1460 if (!skb)
1461 goto out;
1462
1463 if (!skb_can_shift(skb) ||
1464 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1465 (mss != tcp_skb_seglen(skb)))
1466 goto out;
1467
1468 len = skb->len;
1469 if (skb_shift(prev, skb, len)) {
1470 pcount += tcp_skb_pcount(skb);
1471 tcp_shifted_skb(sk, prev, skb, state, tcp_skb_pcount(skb),
1472 len, mss, 0);
1473 }
1474
1475out:
1476 return prev;
1477
1478noop:
1479 return skb;
1480
1481fallback:
1482 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1483 return NULL;
1484}
1485
1486static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1487 struct tcp_sack_block *next_dup,
1488 struct tcp_sacktag_state *state,
1489 u32 start_seq, u32 end_seq,
1490 bool dup_sack_in)
1491{
1492 struct tcp_sock *tp = tcp_sk(sk);
1493 struct sk_buff *tmp;
1494
1495 skb_rbtree_walk_from(skb) {
1496 int in_sack = 0;
1497 bool dup_sack = dup_sack_in;
1498
1499 /* queue is in-order => we can short-circuit the walk early */
1500 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1501 break;
1502
1503 if (next_dup &&
1504 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1505 in_sack = tcp_match_skb_to_sack(sk, skb,
1506 next_dup->start_seq,
1507 next_dup->end_seq);
1508 if (in_sack > 0)
1509 dup_sack = true;
1510 }
1511
1512 /* skb reference here is a bit tricky to get right, since
1513 * shifting can eat and free both this skb and the next,
1514 * so not even _safe variant of the loop is enough.
1515 */
1516 if (in_sack <= 0) {
1517 tmp = tcp_shift_skb_data(sk, skb, state,
1518 start_seq, end_seq, dup_sack);
1519 if (tmp) {
1520 if (tmp != skb) {
1521 skb = tmp;
1522 continue;
1523 }
1524
1525 in_sack = 0;
1526 } else {
1527 in_sack = tcp_match_skb_to_sack(sk, skb,
1528 start_seq,
1529 end_seq);
1530 }
1531 }
1532
1533 if (unlikely(in_sack < 0))
1534 break;
1535
1536 if (in_sack) {
1537 TCP_SKB_CB(skb)->sacked =
1538 tcp_sacktag_one(sk,
1539 state,
1540 TCP_SKB_CB(skb)->sacked,
1541 TCP_SKB_CB(skb)->seq,
1542 TCP_SKB_CB(skb)->end_seq,
1543 dup_sack,
1544 tcp_skb_pcount(skb),
1545 skb->skb_mstamp);
1546 tcp_rate_skb_delivered(sk, skb, state->rate);
1547 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1548 list_del_init(&skb->tcp_tsorted_anchor);
1549
1550 if (!before(TCP_SKB_CB(skb)->seq,
1551 tcp_highest_sack_seq(tp)))
1552 tcp_advance_highest_sack(sk, skb);
1553 }
1554 }
1555 return skb;
1556}
1557
1558static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk,
1559 struct tcp_sacktag_state *state,
1560 u32 seq)
1561{
1562 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node;
1563 struct sk_buff *skb;
1564
1565 while (*p) {
1566 parent = *p;
1567 skb = rb_to_skb(parent);
1568 if (before(seq, TCP_SKB_CB(skb)->seq)) {
1569 p = &parent->rb_left;
1570 continue;
1571 }
1572 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) {
1573 p = &parent->rb_right;
1574 continue;
1575 }
1576 return skb;
1577 }
1578 return NULL;
1579}
1580
1581static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1582 struct tcp_sacktag_state *state,
1583 u32 skip_to_seq)
1584{
1585 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq))
1586 return skb;
1587
1588 return tcp_sacktag_bsearch(sk, state, skip_to_seq);
1589}
1590
1591static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1592 struct sock *sk,
1593 struct tcp_sack_block *next_dup,
1594 struct tcp_sacktag_state *state,
1595 u32 skip_to_seq)
1596{
1597 if (!next_dup)
1598 return skb;
1599
1600 if (before(next_dup->start_seq, skip_to_seq)) {
1601 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1602 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1603 next_dup->start_seq, next_dup->end_seq,
1604 1);
1605 }
1606
1607 return skb;
1608}
1609
1610static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1611{
1612 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1613}
1614
1615static int
1616tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1617 u32 prior_snd_una, struct tcp_sacktag_state *state)
1618{
1619 struct tcp_sock *tp = tcp_sk(sk);
1620 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1621 TCP_SKB_CB(ack_skb)->sacked);
1622 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1623 struct tcp_sack_block sp[TCP_NUM_SACKS];
1624 struct tcp_sack_block *cache;
1625 struct sk_buff *skb;
1626 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1627 int used_sacks;
1628 bool found_dup_sack = false;
1629 int i, j;
1630 int first_sack_index;
1631
1632 state->flag = 0;
1633 state->reord = tp->snd_nxt;
1634
1635 if (!tp->sacked_out)
1636 tcp_highest_sack_reset(sk);
1637
1638 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1639 num_sacks, prior_snd_una);
1640 if (found_dup_sack) {
1641 state->flag |= FLAG_DSACKING_ACK;
1642 tp->delivered++; /* A spurious retransmission is delivered */
1643 }
1644
1645 /* Eliminate too old ACKs, but take into
1646 * account more or less fresh ones, they can
1647 * contain valid SACK info.
1648 */
1649 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1650 return 0;
1651
1652 if (!tp->packets_out)
1653 goto out;
1654
1655 used_sacks = 0;
1656 first_sack_index = 0;
1657 for (i = 0; i < num_sacks; i++) {
1658 bool dup_sack = !i && found_dup_sack;
1659
1660 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1661 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1662
1663 if (!tcp_is_sackblock_valid(tp, dup_sack,
1664 sp[used_sacks].start_seq,
1665 sp[used_sacks].end_seq)) {
1666 int mib_idx;
1667
1668 if (dup_sack) {
1669 if (!tp->undo_marker)
1670 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1671 else
1672 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1673 } else {
1674 /* Don't count olds caused by ACK reordering */
1675 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1676 !after(sp[used_sacks].end_seq, tp->snd_una))
1677 continue;
1678 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1679 }
1680
1681 NET_INC_STATS(sock_net(sk), mib_idx);
1682 if (i == 0)
1683 first_sack_index = -1;
1684 continue;
1685 }
1686
1687 /* Ignore very old stuff early */
1688 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1689 continue;
1690
1691 used_sacks++;
1692 }
1693
1694 /* order SACK blocks to allow in order walk of the retrans queue */
1695 for (i = used_sacks - 1; i > 0; i--) {
1696 for (j = 0; j < i; j++) {
1697 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1698 swap(sp[j], sp[j + 1]);
1699
1700 /* Track where the first SACK block goes to */
1701 if (j == first_sack_index)
1702 first_sack_index = j + 1;
1703 }
1704 }
1705 }
1706
1707 state->mss_now = tcp_current_mss(sk);
1708 skb = NULL;
1709 i = 0;
1710
1711 if (!tp->sacked_out) {
1712 /* It's already past, so skip checking against it */
1713 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1714 } else {
1715 cache = tp->recv_sack_cache;
1716 /* Skip empty blocks in at head of the cache */
1717 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1718 !cache->end_seq)
1719 cache++;
1720 }
1721
1722 while (i < used_sacks) {
1723 u32 start_seq = sp[i].start_seq;
1724 u32 end_seq = sp[i].end_seq;
1725 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1726 struct tcp_sack_block *next_dup = NULL;
1727
1728 if (found_dup_sack && ((i + 1) == first_sack_index))
1729 next_dup = &sp[i + 1];
1730
1731 /* Skip too early cached blocks */
1732 while (tcp_sack_cache_ok(tp, cache) &&
1733 !before(start_seq, cache->end_seq))
1734 cache++;
1735
1736 /* Can skip some work by looking recv_sack_cache? */
1737 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1738 after(end_seq, cache->start_seq)) {
1739
1740 /* Head todo? */
1741 if (before(start_seq, cache->start_seq)) {
1742 skb = tcp_sacktag_skip(skb, sk, state,
1743 start_seq);
1744 skb = tcp_sacktag_walk(skb, sk, next_dup,
1745 state,
1746 start_seq,
1747 cache->start_seq,
1748 dup_sack);
1749 }
1750
1751 /* Rest of the block already fully processed? */
1752 if (!after(end_seq, cache->end_seq))
1753 goto advance_sp;
1754
1755 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1756 state,
1757 cache->end_seq);
1758
1759 /* ...tail remains todo... */
1760 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1761 /* ...but better entrypoint exists! */
1762 skb = tcp_highest_sack(sk);
1763 if (!skb)
1764 break;
1765 cache++;
1766 goto walk;
1767 }
1768
1769 skb = tcp_sacktag_skip(skb, sk, state, cache->end_seq);
1770 /* Check overlap against next cached too (past this one already) */
1771 cache++;
1772 continue;
1773 }
1774
1775 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1776 skb = tcp_highest_sack(sk);
1777 if (!skb)
1778 break;
1779 }
1780 skb = tcp_sacktag_skip(skb, sk, state, start_seq);
1781
1782walk:
1783 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
1784 start_seq, end_seq, dup_sack);
1785
1786advance_sp:
1787 i++;
1788 }
1789
1790 /* Clear the head of the cache sack blocks so we can skip it next time */
1791 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1792 tp->recv_sack_cache[i].start_seq = 0;
1793 tp->recv_sack_cache[i].end_seq = 0;
1794 }
1795 for (j = 0; j < used_sacks; j++)
1796 tp->recv_sack_cache[i++] = sp[j];
1797
1798 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker)
1799 tcp_check_sack_reordering(sk, state->reord, 0);
1800
1801 tcp_verify_left_out(tp);
1802out:
1803
1804#if FASTRETRANS_DEBUG > 0
1805 WARN_ON((int)tp->sacked_out < 0);
1806 WARN_ON((int)tp->lost_out < 0);
1807 WARN_ON((int)tp->retrans_out < 0);
1808 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1809#endif
1810 return state->flag;
1811}
1812
1813/* Limits sacked_out so that sum with lost_out isn't ever larger than
1814 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1815 */
1816static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1817{
1818 u32 holes;
1819
1820 holes = max(tp->lost_out, 1U);
1821 holes = min(holes, tp->packets_out);
1822
1823 if ((tp->sacked_out + holes) > tp->packets_out) {
1824 tp->sacked_out = tp->packets_out - holes;
1825 return true;
1826 }
1827 return false;
1828}
1829
1830/* If we receive more dupacks than we expected counting segments
1831 * in assumption of absent reordering, interpret this as reordering.
1832 * The only another reason could be bug in receiver TCP.
1833 */
1834static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1835{
1836 struct tcp_sock *tp = tcp_sk(sk);
1837
1838 if (!tcp_limit_reno_sacked(tp))
1839 return;
1840
1841 tp->reordering = min_t(u32, tp->packets_out + addend,
1842 sock_net(sk)->ipv4.sysctl_tcp_max_reordering);
1843 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER);
1844}
1845
1846/* Emulate SACKs for SACKless connection: account for a new dupack. */
1847
1848static void tcp_add_reno_sack(struct sock *sk)
1849{
1850 struct tcp_sock *tp = tcp_sk(sk);
1851 u32 prior_sacked = tp->sacked_out;
1852
1853 tp->sacked_out++;
1854 tcp_check_reno_reordering(sk, 0);
1855 if (tp->sacked_out > prior_sacked)
1856 tp->delivered++; /* Some out-of-order packet is delivered */
1857 tcp_verify_left_out(tp);
1858}
1859
1860/* Account for ACK, ACKing some data in Reno Recovery phase. */
1861
1862static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1863{
1864 struct tcp_sock *tp = tcp_sk(sk);
1865
1866 if (acked > 0) {
1867 /* One ACK acked hole. The rest eat duplicate ACKs. */
1868 tp->delivered += max_t(int, acked - tp->sacked_out, 1);
1869 if (acked - 1 >= tp->sacked_out)
1870 tp->sacked_out = 0;
1871 else
1872 tp->sacked_out -= acked - 1;
1873 }
1874 tcp_check_reno_reordering(sk, acked);
1875 tcp_verify_left_out(tp);
1876}
1877
1878static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1879{
1880 tp->sacked_out = 0;
1881}
1882
1883void tcp_clear_retrans(struct tcp_sock *tp)
1884{
1885 tp->retrans_out = 0;
1886 tp->lost_out = 0;
1887 tp->undo_marker = 0;
1888 tp->undo_retrans = -1;
1889 tp->sacked_out = 0;
1890}
1891
1892static inline void tcp_init_undo(struct tcp_sock *tp)
1893{
1894 tp->undo_marker = tp->snd_una;
1895 /* Retransmission still in flight may cause DSACKs later. */
1896 tp->undo_retrans = tp->retrans_out ? : -1;
1897}
1898
1899/* Enter Loss state. If we detect SACK reneging, forget all SACK information
1900 * and reset tags completely, otherwise preserve SACKs. If receiver
1901 * dropped its ofo queue, we will know this due to reneging detection.
1902 */
1903void tcp_enter_loss(struct sock *sk)
1904{
1905 const struct inet_connection_sock *icsk = inet_csk(sk);
1906 struct tcp_sock *tp = tcp_sk(sk);
1907 struct net *net = sock_net(sk);
1908 struct sk_buff *skb;
1909 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
1910 bool is_reneg; /* is receiver reneging on SACKs? */
1911 bool mark_lost;
1912
1913 /* Reduce ssthresh if it has not yet been made inside this window. */
1914 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1915 !after(tp->high_seq, tp->snd_una) ||
1916 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1917 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1918 tp->prior_cwnd = tp->snd_cwnd;
1919 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1920 tcp_ca_event(sk, CA_EVENT_LOSS);
1921 tcp_init_undo(tp);
1922 }
1923 tp->snd_cwnd = 1;
1924 tp->snd_cwnd_cnt = 0;
1925 tp->snd_cwnd_stamp = tcp_jiffies32;
1926
1927 tp->retrans_out = 0;
1928 tp->lost_out = 0;
1929
1930 if (tcp_is_reno(tp))
1931 tcp_reset_reno_sack(tp);
1932
1933 skb = tcp_rtx_queue_head(sk);
1934 is_reneg = skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED);
1935 if (is_reneg) {
1936 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1937 tp->sacked_out = 0;
1938 /* Mark SACK reneging until we recover from this loss event. */
1939 tp->is_sack_reneg = 1;
1940 }
1941 tcp_clear_all_retrans_hints(tp);
1942
1943 skb_rbtree_walk_from(skb) {
1944 mark_lost = (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
1945 is_reneg);
1946 if (mark_lost)
1947 tcp_sum_lost(tp, skb);
1948 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1949 if (mark_lost) {
1950 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1951 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1952 tp->lost_out += tcp_skb_pcount(skb);
1953 }
1954 }
1955 tcp_verify_left_out(tp);
1956
1957 /* Timeout in disordered state after receiving substantial DUPACKs
1958 * suggests that the degree of reordering is over-estimated.
1959 */
1960 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
1961 tp->sacked_out >= net->ipv4.sysctl_tcp_reordering)
1962 tp->reordering = min_t(unsigned int, tp->reordering,
1963 net->ipv4.sysctl_tcp_reordering);
1964 tcp_set_ca_state(sk, TCP_CA_Loss);
1965 tp->high_seq = tp->snd_nxt;
1966 tcp_ecn_queue_cwr(tp);
1967
1968 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1969 * loss recovery is underway except recurring timeout(s) on
1970 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1971 */
1972 tp->frto = net->ipv4.sysctl_tcp_frto &&
1973 (new_recovery || icsk->icsk_retransmits) &&
1974 !inet_csk(sk)->icsk_mtup.probe_size;
1975}
1976
1977/* If ACK arrived pointing to a remembered SACK, it means that our
1978 * remembered SACKs do not reflect real state of receiver i.e.
1979 * receiver _host_ is heavily congested (or buggy).
1980 *
1981 * To avoid big spurious retransmission bursts due to transient SACK
1982 * scoreboard oddities that look like reneging, we give the receiver a
1983 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
1984 * restore sanity to the SACK scoreboard. If the apparent reneging
1985 * persists until this RTO then we'll clear the SACK scoreboard.
1986 */
1987static bool tcp_check_sack_reneging(struct sock *sk, int flag)
1988{
1989 if (flag & FLAG_SACK_RENEGING) {
1990 struct tcp_sock *tp = tcp_sk(sk);
1991 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
1992 msecs_to_jiffies(10));
1993
1994 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1995 delay, TCP_RTO_MAX);
1996 return true;
1997 }
1998 return false;
1999}
2000
2001/* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2002 * counter when SACK is enabled (without SACK, sacked_out is used for
2003 * that purpose).
2004 *
2005 * With reordering, holes may still be in flight, so RFC3517 recovery
2006 * uses pure sacked_out (total number of SACKed segments) even though
2007 * it violates the RFC that uses duplicate ACKs, often these are equal
2008 * but when e.g. out-of-window ACKs or packet duplication occurs,
2009 * they differ. Since neither occurs due to loss, TCP should really
2010 * ignore them.
2011 */
2012static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2013{
2014 return tp->sacked_out + 1;
2015}
2016
2017/* Linux NewReno/SACK/ECN state machine.
2018 * --------------------------------------
2019 *
2020 * "Open" Normal state, no dubious events, fast path.
2021 * "Disorder" In all the respects it is "Open",
2022 * but requires a bit more attention. It is entered when
2023 * we see some SACKs or dupacks. It is split of "Open"
2024 * mainly to move some processing from fast path to slow one.
2025 * "CWR" CWND was reduced due to some Congestion Notification event.
2026 * It can be ECN, ICMP source quench, local device congestion.
2027 * "Recovery" CWND was reduced, we are fast-retransmitting.
2028 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2029 *
2030 * tcp_fastretrans_alert() is entered:
2031 * - each incoming ACK, if state is not "Open"
2032 * - when arrived ACK is unusual, namely:
2033 * * SACK
2034 * * Duplicate ACK.
2035 * * ECN ECE.
2036 *
2037 * Counting packets in flight is pretty simple.
2038 *
2039 * in_flight = packets_out - left_out + retrans_out
2040 *
2041 * packets_out is SND.NXT-SND.UNA counted in packets.
2042 *
2043 * retrans_out is number of retransmitted segments.
2044 *
2045 * left_out is number of segments left network, but not ACKed yet.
2046 *
2047 * left_out = sacked_out + lost_out
2048 *
2049 * sacked_out: Packets, which arrived to receiver out of order
2050 * and hence not ACKed. With SACKs this number is simply
2051 * amount of SACKed data. Even without SACKs
2052 * it is easy to give pretty reliable estimate of this number,
2053 * counting duplicate ACKs.
2054 *
2055 * lost_out: Packets lost by network. TCP has no explicit
2056 * "loss notification" feedback from network (for now).
2057 * It means that this number can be only _guessed_.
2058 * Actually, it is the heuristics to predict lossage that
2059 * distinguishes different algorithms.
2060 *
2061 * F.e. after RTO, when all the queue is considered as lost,
2062 * lost_out = packets_out and in_flight = retrans_out.
2063 *
2064 * Essentially, we have now a few algorithms detecting
2065 * lost packets.
2066 *
2067 * If the receiver supports SACK:
2068 *
2069 * RFC6675/3517: It is the conventional algorithm. A packet is
2070 * considered lost if the number of higher sequence packets
2071 * SACKed is greater than or equal the DUPACK thoreshold
2072 * (reordering). This is implemented in tcp_mark_head_lost and
2073 * tcp_update_scoreboard.
2074 *
2075 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2076 * (2017-) that checks timing instead of counting DUPACKs.
2077 * Essentially a packet is considered lost if it's not S/ACKed
2078 * after RTT + reordering_window, where both metrics are
2079 * dynamically measured and adjusted. This is implemented in
2080 * tcp_rack_mark_lost.
2081 *
2082 * If the receiver does not support SACK:
2083 *
2084 * NewReno (RFC6582): in Recovery we assume that one segment
2085 * is lost (classic Reno). While we are in Recovery and
2086 * a partial ACK arrives, we assume that one more packet
2087 * is lost (NewReno). This heuristics are the same in NewReno
2088 * and SACK.
2089 *
2090 * Really tricky (and requiring careful tuning) part of algorithm
2091 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2092 * The first determines the moment _when_ we should reduce CWND and,
2093 * hence, slow down forward transmission. In fact, it determines the moment
2094 * when we decide that hole is caused by loss, rather than by a reorder.
2095 *
2096 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2097 * holes, caused by lost packets.
2098 *
2099 * And the most logically complicated part of algorithm is undo
2100 * heuristics. We detect false retransmits due to both too early
2101 * fast retransmit (reordering) and underestimated RTO, analyzing
2102 * timestamps and D-SACKs. When we detect that some segments were
2103 * retransmitted by mistake and CWND reduction was wrong, we undo
2104 * window reduction and abort recovery phase. This logic is hidden
2105 * inside several functions named tcp_try_undo_<something>.
2106 */
2107
2108/* This function decides, when we should leave Disordered state
2109 * and enter Recovery phase, reducing congestion window.
2110 *
2111 * Main question: may we further continue forward transmission
2112 * with the same cwnd?
2113 */
2114static bool tcp_time_to_recover(struct sock *sk, int flag)
2115{
2116 struct tcp_sock *tp = tcp_sk(sk);
2117
2118 /* Trick#1: The loss is proven. */
2119 if (tp->lost_out)
2120 return true;
2121
2122 /* Not-A-Trick#2 : Classic rule... */
2123 if (tcp_dupack_heuristics(tp) > tp->reordering)
2124 return true;
2125
2126 return false;
2127}
2128
2129/* Detect loss in event "A" above by marking head of queue up as lost.
2130 * For non-SACK(Reno) senders, the first "packets" number of segments
2131 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2132 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2133 * the maximum SACKed segments to pass before reaching this limit.
2134 */
2135static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2136{
2137 struct tcp_sock *tp = tcp_sk(sk);
2138 struct sk_buff *skb;
2139 int cnt, oldcnt, lost;
2140 unsigned int mss;
2141 /* Use SACK to deduce losses of new sequences sent during recovery */
2142 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
2143
2144 WARN_ON(packets > tp->packets_out);
2145 skb = tp->lost_skb_hint;
2146 if (skb) {
2147 /* Head already handled? */
2148 if (mark_head && after(TCP_SKB_CB(skb)->seq, tp->snd_una))
2149 return;
2150 cnt = tp->lost_cnt_hint;
2151 } else {
2152 skb = tcp_rtx_queue_head(sk);
2153 cnt = 0;
2154 }
2155
2156 skb_rbtree_walk_from(skb) {
2157 /* TODO: do this better */
2158 /* this is not the most efficient way to do this... */
2159 tp->lost_skb_hint = skb;
2160 tp->lost_cnt_hint = cnt;
2161
2162 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2163 break;
2164
2165 oldcnt = cnt;
2166 if (tcp_is_reno(tp) ||
2167 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2168 cnt += tcp_skb_pcount(skb);
2169
2170 if (cnt > packets) {
2171 if (tcp_is_sack(tp) ||
2172 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2173 (oldcnt >= packets))
2174 break;
2175
2176 mss = tcp_skb_mss(skb);
2177 /* If needed, chop off the prefix to mark as lost. */
2178 lost = (packets - oldcnt) * mss;
2179 if (lost < skb->len &&
2180 tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2181 lost, mss, GFP_ATOMIC) < 0)
2182 break;
2183 cnt = packets;
2184 }
2185
2186 tcp_skb_mark_lost(tp, skb);
2187
2188 if (mark_head)
2189 break;
2190 }
2191 tcp_verify_left_out(tp);
2192}
2193
2194/* Account newly detected lost packet(s) */
2195
2196static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2197{
2198 struct tcp_sock *tp = tcp_sk(sk);
2199
2200 if (tcp_is_reno(tp)) {
2201 tcp_mark_head_lost(sk, 1, 1);
2202 } else {
2203 int sacked_upto = tp->sacked_out - tp->reordering;
2204 if (sacked_upto >= 0)
2205 tcp_mark_head_lost(sk, sacked_upto, 0);
2206 else if (fast_rexmit)
2207 tcp_mark_head_lost(sk, 1, 1);
2208 }
2209}
2210
2211static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2212{
2213 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2214 before(tp->rx_opt.rcv_tsecr, when);
2215}
2216
2217/* skb is spurious retransmitted if the returned timestamp echo
2218 * reply is prior to the skb transmission time
2219 */
2220static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2221 const struct sk_buff *skb)
2222{
2223 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2224 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb));
2225}
2226
2227/* Nothing was retransmitted or returned timestamp is less
2228 * than timestamp of the first retransmission.
2229 */
2230static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2231{
2232 return !tp->retrans_stamp ||
2233 tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
2234}
2235
2236/* Undo procedures. */
2237
2238/* We can clear retrans_stamp when there are no retransmissions in the
2239 * window. It would seem that it is trivially available for us in
2240 * tp->retrans_out, however, that kind of assumptions doesn't consider
2241 * what will happen if errors occur when sending retransmission for the
2242 * second time. ...It could the that such segment has only
2243 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2244 * the head skb is enough except for some reneging corner cases that
2245 * are not worth the effort.
2246 *
2247 * Main reason for all this complexity is the fact that connection dying
2248 * time now depends on the validity of the retrans_stamp, in particular,
2249 * that successive retransmissions of a segment must not advance
2250 * retrans_stamp under any conditions.
2251 */
2252static bool tcp_any_retrans_done(const struct sock *sk)
2253{
2254 const struct tcp_sock *tp = tcp_sk(sk);
2255 struct sk_buff *skb;
2256
2257 if (tp->retrans_out)
2258 return true;
2259
2260 skb = tcp_rtx_queue_head(sk);
2261 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2262 return true;
2263
2264 return false;
2265}
2266
2267static void DBGUNDO(struct sock *sk, const char *msg)
2268{
2269#if FASTRETRANS_DEBUG > 1
2270 struct tcp_sock *tp = tcp_sk(sk);
2271 struct inet_sock *inet = inet_sk(sk);
2272
2273 if (sk->sk_family == AF_INET) {
2274 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2275 msg,
2276 &inet->inet_daddr, ntohs(inet->inet_dport),
2277 tp->snd_cwnd, tcp_left_out(tp),
2278 tp->snd_ssthresh, tp->prior_ssthresh,
2279 tp->packets_out);
2280 }
2281#if IS_ENABLED(CONFIG_IPV6)
2282 else if (sk->sk_family == AF_INET6) {
2283 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2284 msg,
2285 &sk->sk_v6_daddr, ntohs(inet->inet_dport),
2286 tp->snd_cwnd, tcp_left_out(tp),
2287 tp->snd_ssthresh, tp->prior_ssthresh,
2288 tp->packets_out);
2289 }
2290#endif
2291#endif
2292}
2293
2294static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2295{
2296 struct tcp_sock *tp = tcp_sk(sk);
2297
2298 if (unmark_loss) {
2299 struct sk_buff *skb;
2300
2301 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2302 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2303 }
2304 tp->lost_out = 0;
2305 tcp_clear_all_retrans_hints(tp);
2306 }
2307
2308 if (tp->prior_ssthresh) {
2309 const struct inet_connection_sock *icsk = inet_csk(sk);
2310
2311 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2312
2313 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2314 tp->snd_ssthresh = tp->prior_ssthresh;
2315 tcp_ecn_withdraw_cwr(tp);
2316 }
2317 }
2318 tp->snd_cwnd_stamp = tcp_jiffies32;
2319 tp->undo_marker = 0;
2320 tp->rack.advanced = 1; /* Force RACK to re-exam losses */
2321}
2322
2323static inline bool tcp_may_undo(const struct tcp_sock *tp)
2324{
2325 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2326}
2327
2328/* People celebrate: "We love our President!" */
2329static bool tcp_try_undo_recovery(struct sock *sk)
2330{
2331 struct tcp_sock *tp = tcp_sk(sk);
2332
2333 if (tcp_may_undo(tp)) {
2334 int mib_idx;
2335
2336 /* Happy end! We did not retransmit anything
2337 * or our original transmission succeeded.
2338 */
2339 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2340 tcp_undo_cwnd_reduction(sk, false);
2341 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2342 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2343 else
2344 mib_idx = LINUX_MIB_TCPFULLUNDO;
2345
2346 NET_INC_STATS(sock_net(sk), mib_idx);
2347 } else if (tp->rack.reo_wnd_persist) {
2348 tp->rack.reo_wnd_persist--;
2349 }
2350 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2351 /* Hold old state until something *above* high_seq
2352 * is ACKed. For Reno it is MUST to prevent false
2353 * fast retransmits (RFC2582). SACK TCP is safe. */
2354 if (!tcp_any_retrans_done(sk))
2355 tp->retrans_stamp = 0;
2356 return true;
2357 }
2358 tcp_set_ca_state(sk, TCP_CA_Open);
2359 tp->is_sack_reneg = 0;
2360 return false;
2361}
2362
2363/* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2364static bool tcp_try_undo_dsack(struct sock *sk)
2365{
2366 struct tcp_sock *tp = tcp_sk(sk);
2367
2368 if (tp->undo_marker && !tp->undo_retrans) {
2369 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH,
2370 tp->rack.reo_wnd_persist + 1);
2371 DBGUNDO(sk, "D-SACK");
2372 tcp_undo_cwnd_reduction(sk, false);
2373 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2374 return true;
2375 }
2376 return false;
2377}
2378
2379/* Undo during loss recovery after partial ACK or using F-RTO. */
2380static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2381{
2382 struct tcp_sock *tp = tcp_sk(sk);
2383
2384 if (frto_undo || tcp_may_undo(tp)) {
2385 tcp_undo_cwnd_reduction(sk, true);
2386
2387 DBGUNDO(sk, "partial loss");
2388 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2389 if (frto_undo)
2390 NET_INC_STATS(sock_net(sk),
2391 LINUX_MIB_TCPSPURIOUSRTOS);
2392 inet_csk(sk)->icsk_retransmits = 0;
2393 if (frto_undo || tcp_is_sack(tp)) {
2394 tcp_set_ca_state(sk, TCP_CA_Open);
2395 tp->is_sack_reneg = 0;
2396 }
2397 return true;
2398 }
2399 return false;
2400}
2401
2402/* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2403 * It computes the number of packets to send (sndcnt) based on packets newly
2404 * delivered:
2405 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2406 * cwnd reductions across a full RTT.
2407 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2408 * But when the retransmits are acked without further losses, PRR
2409 * slow starts cwnd up to ssthresh to speed up the recovery.
2410 */
2411static void tcp_init_cwnd_reduction(struct sock *sk)
2412{
2413 struct tcp_sock *tp = tcp_sk(sk);
2414
2415 tp->high_seq = tp->snd_nxt;
2416 tp->tlp_high_seq = 0;
2417 tp->snd_cwnd_cnt = 0;
2418 tp->prior_cwnd = tp->snd_cwnd;
2419 tp->prr_delivered = 0;
2420 tp->prr_out = 0;
2421 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2422 tcp_ecn_queue_cwr(tp);
2423}
2424
2425void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int flag)
2426{
2427 struct tcp_sock *tp = tcp_sk(sk);
2428 int sndcnt = 0;
2429 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2430
2431 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2432 return;
2433
2434 tp->prr_delivered += newly_acked_sacked;
2435 if (delta < 0) {
2436 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2437 tp->prior_cwnd - 1;
2438 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2439 } else if ((flag & FLAG_RETRANS_DATA_ACKED) &&
2440 !(flag & FLAG_LOST_RETRANS)) {
2441 sndcnt = min_t(int, delta,
2442 max_t(int, tp->prr_delivered - tp->prr_out,
2443 newly_acked_sacked) + 1);
2444 } else {
2445 sndcnt = min(delta, newly_acked_sacked);
2446 }
2447 /* Force a fast retransmit upon entering fast recovery */
2448 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2449 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2450}
2451
2452static inline void tcp_end_cwnd_reduction(struct sock *sk)
2453{
2454 struct tcp_sock *tp = tcp_sk(sk);
2455
2456 if (inet_csk(sk)->icsk_ca_ops->cong_control)
2457 return;
2458
2459 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2460 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH &&
2461 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) {
2462 tp->snd_cwnd = tp->snd_ssthresh;
2463 tp->snd_cwnd_stamp = tcp_jiffies32;
2464 }
2465 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2466}
2467
2468/* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2469void tcp_enter_cwr(struct sock *sk)
2470{
2471 struct tcp_sock *tp = tcp_sk(sk);
2472
2473 tp->prior_ssthresh = 0;
2474 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2475 tp->undo_marker = 0;
2476 tcp_init_cwnd_reduction(sk);
2477 tcp_set_ca_state(sk, TCP_CA_CWR);
2478 }
2479}
2480EXPORT_SYMBOL(tcp_enter_cwr);
2481
2482static void tcp_try_keep_open(struct sock *sk)
2483{
2484 struct tcp_sock *tp = tcp_sk(sk);
2485 int state = TCP_CA_Open;
2486
2487 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2488 state = TCP_CA_Disorder;
2489
2490 if (inet_csk(sk)->icsk_ca_state != state) {
2491 tcp_set_ca_state(sk, state);
2492 tp->high_seq = tp->snd_nxt;
2493 }
2494}
2495
2496static void tcp_try_to_open(struct sock *sk, int flag)
2497{
2498 struct tcp_sock *tp = tcp_sk(sk);
2499
2500 tcp_verify_left_out(tp);
2501
2502 if (!tcp_any_retrans_done(sk))
2503 tp->retrans_stamp = 0;
2504
2505 if (flag & FLAG_ECE)
2506 tcp_enter_cwr(sk);
2507
2508 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2509 tcp_try_keep_open(sk);
2510 }
2511}
2512
2513static void tcp_mtup_probe_failed(struct sock *sk)
2514{
2515 struct inet_connection_sock *icsk = inet_csk(sk);
2516
2517 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2518 icsk->icsk_mtup.probe_size = 0;
2519 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2520}
2521
2522static void tcp_mtup_probe_success(struct sock *sk)
2523{
2524 struct tcp_sock *tp = tcp_sk(sk);
2525 struct inet_connection_sock *icsk = inet_csk(sk);
2526
2527 /* FIXME: breaks with very large cwnd */
2528 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2529 tp->snd_cwnd = tp->snd_cwnd *
2530 tcp_mss_to_mtu(sk, tp->mss_cache) /
2531 icsk->icsk_mtup.probe_size;
2532 tp->snd_cwnd_cnt = 0;
2533 tp->snd_cwnd_stamp = tcp_jiffies32;
2534 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2535
2536 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2537 icsk->icsk_mtup.probe_size = 0;
2538 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2539 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2540}
2541
2542/* Do a simple retransmit without using the backoff mechanisms in
2543 * tcp_timer. This is used for path mtu discovery.
2544 * The socket is already locked here.
2545 */
2546void tcp_simple_retransmit(struct sock *sk)
2547{
2548 const struct inet_connection_sock *icsk = inet_csk(sk);
2549 struct tcp_sock *tp = tcp_sk(sk);
2550 struct sk_buff *skb;
2551 unsigned int mss = tcp_current_mss(sk);
2552
2553 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2554 if (tcp_skb_seglen(skb) > mss &&
2555 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2556 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2557 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2558 tp->retrans_out -= tcp_skb_pcount(skb);
2559 }
2560 tcp_skb_mark_lost_uncond_verify(tp, skb);
2561 }
2562 }
2563
2564 tcp_clear_retrans_hints_partial(tp);
2565
2566 if (!tp->lost_out)
2567 return;
2568
2569 if (tcp_is_reno(tp))
2570 tcp_limit_reno_sacked(tp);
2571
2572 tcp_verify_left_out(tp);
2573
2574 /* Don't muck with the congestion window here.
2575 * Reason is that we do not increase amount of _data_
2576 * in network, but units changed and effective
2577 * cwnd/ssthresh really reduced now.
2578 */
2579 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2580 tp->high_seq = tp->snd_nxt;
2581 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2582 tp->prior_ssthresh = 0;
2583 tp->undo_marker = 0;
2584 tcp_set_ca_state(sk, TCP_CA_Loss);
2585 }
2586 tcp_xmit_retransmit_queue(sk);
2587}
2588EXPORT_SYMBOL(tcp_simple_retransmit);
2589
2590void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2591{
2592 struct tcp_sock *tp = tcp_sk(sk);
2593 int mib_idx;
2594
2595 if (tcp_is_reno(tp))
2596 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2597 else
2598 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2599
2600 NET_INC_STATS(sock_net(sk), mib_idx);
2601
2602 tp->prior_ssthresh = 0;
2603 tcp_init_undo(tp);
2604
2605 if (!tcp_in_cwnd_reduction(sk)) {
2606 if (!ece_ack)
2607 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2608 tcp_init_cwnd_reduction(sk);
2609 }
2610 tcp_set_ca_state(sk, TCP_CA_Recovery);
2611}
2612
2613/* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2614 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2615 */
2616static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack,
2617 int *rexmit)
2618{
2619 struct tcp_sock *tp = tcp_sk(sk);
2620 bool recovered = !before(tp->snd_una, tp->high_seq);
2621
2622 if ((flag & FLAG_SND_UNA_ADVANCED) &&
2623 tcp_try_undo_loss(sk, false))
2624 return;
2625
2626 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2627 /* Step 3.b. A timeout is spurious if not all data are
2628 * lost, i.e., never-retransmitted data are (s)acked.
2629 */
2630 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2631 tcp_try_undo_loss(sk, true))
2632 return;
2633
2634 if (after(tp->snd_nxt, tp->high_seq)) {
2635 if (flag & FLAG_DATA_SACKED || is_dupack)
2636 tp->frto = 0; /* Step 3.a. loss was real */
2637 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2638 tp->high_seq = tp->snd_nxt;
2639 /* Step 2.b. Try send new data (but deferred until cwnd
2640 * is updated in tcp_ack()). Otherwise fall back to
2641 * the conventional recovery.
2642 */
2643 if (!tcp_write_queue_empty(sk) &&
2644 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2645 *rexmit = REXMIT_NEW;
2646 return;
2647 }
2648 tp->frto = 0;
2649 }
2650 }
2651
2652 if (recovered) {
2653 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2654 tcp_try_undo_recovery(sk);
2655 return;
2656 }
2657 if (tcp_is_reno(tp)) {
2658 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2659 * delivered. Lower inflight to clock out (re)tranmissions.
2660 */
2661 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2662 tcp_add_reno_sack(sk);
2663 else if (flag & FLAG_SND_UNA_ADVANCED)
2664 tcp_reset_reno_sack(tp);
2665 }
2666 *rexmit = REXMIT_LOST;
2667}
2668
2669/* Undo during fast recovery after partial ACK. */
2670static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una)
2671{
2672 struct tcp_sock *tp = tcp_sk(sk);
2673
2674 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2675 /* Plain luck! Hole if filled with delayed
2676 * packet, rather than with a retransmit. Check reordering.
2677 */
2678 tcp_check_sack_reordering(sk, prior_snd_una, 1);
2679
2680 /* We are getting evidence that the reordering degree is higher
2681 * than we realized. If there are no retransmits out then we
2682 * can undo. Otherwise we clock out new packets but do not
2683 * mark more packets lost or retransmit more.
2684 */
2685 if (tp->retrans_out)
2686 return true;
2687
2688 if (!tcp_any_retrans_done(sk))
2689 tp->retrans_stamp = 0;
2690
2691 DBGUNDO(sk, "partial recovery");
2692 tcp_undo_cwnd_reduction(sk, true);
2693 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2694 tcp_try_keep_open(sk);
2695 return true;
2696 }
2697 return false;
2698}
2699
2700static void tcp_rack_identify_loss(struct sock *sk, int *ack_flag)
2701{
2702 struct tcp_sock *tp = tcp_sk(sk);
2703
2704 /* Use RACK to detect loss */
2705 if (sock_net(sk)->ipv4.sysctl_tcp_recovery & TCP_RACK_LOSS_DETECTION) {
2706 u32 prior_retrans = tp->retrans_out;
2707
2708 tcp_rack_mark_lost(sk);
2709 if (prior_retrans > tp->retrans_out)
2710 *ack_flag |= FLAG_LOST_RETRANS;
2711 }
2712}
2713
2714static bool tcp_force_fast_retransmit(struct sock *sk)
2715{
2716 struct tcp_sock *tp = tcp_sk(sk);
2717
2718 return after(tcp_highest_sack_seq(tp),
2719 tp->snd_una + tp->reordering * tp->mss_cache);
2720}
2721
2722/* Process an event, which can update packets-in-flight not trivially.
2723 * Main goal of this function is to calculate new estimate for left_out,
2724 * taking into account both packets sitting in receiver's buffer and
2725 * packets lost by network.
2726 *
2727 * Besides that it updates the congestion state when packet loss or ECN
2728 * is detected. But it does not reduce the cwnd, it is done by the
2729 * congestion control later.
2730 *
2731 * It does _not_ decide what to send, it is made in function
2732 * tcp_xmit_retransmit_queue().
2733 */
2734static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una,
2735 bool is_dupack, int *ack_flag, int *rexmit)
2736{
2737 struct inet_connection_sock *icsk = inet_csk(sk);
2738 struct tcp_sock *tp = tcp_sk(sk);
2739 int fast_rexmit = 0, flag = *ack_flag;
2740 bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2741 tcp_force_fast_retransmit(sk));
2742
2743 if (!tp->packets_out && tp->sacked_out)
2744 tp->sacked_out = 0;
2745
2746 /* Now state machine starts.
2747 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2748 if (flag & FLAG_ECE)
2749 tp->prior_ssthresh = 0;
2750
2751 /* B. In all the states check for reneging SACKs. */
2752 if (tcp_check_sack_reneging(sk, flag))
2753 return;
2754
2755 /* C. Check consistency of the current state. */
2756 tcp_verify_left_out(tp);
2757
2758 /* D. Check state exit conditions. State can be terminated
2759 * when high_seq is ACKed. */
2760 if (icsk->icsk_ca_state == TCP_CA_Open) {
2761 WARN_ON(tp->retrans_out != 0);
2762 tp->retrans_stamp = 0;
2763 } else if (!before(tp->snd_una, tp->high_seq)) {
2764 switch (icsk->icsk_ca_state) {
2765 case TCP_CA_CWR:
2766 /* CWR is to be held something *above* high_seq
2767 * is ACKed for CWR bit to reach receiver. */
2768 if (tp->snd_una != tp->high_seq) {
2769 tcp_end_cwnd_reduction(sk);
2770 tcp_set_ca_state(sk, TCP_CA_Open);
2771 }
2772 break;
2773
2774 case TCP_CA_Recovery:
2775 if (tcp_is_reno(tp))
2776 tcp_reset_reno_sack(tp);
2777 if (tcp_try_undo_recovery(sk))
2778 return;
2779 tcp_end_cwnd_reduction(sk);
2780 break;
2781 }
2782 }
2783
2784 /* E. Process state. */
2785 switch (icsk->icsk_ca_state) {
2786 case TCP_CA_Recovery:
2787 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2788 if (tcp_is_reno(tp) && is_dupack)
2789 tcp_add_reno_sack(sk);
2790 } else {
2791 if (tcp_try_undo_partial(sk, prior_snd_una))
2792 return;
2793 /* Partial ACK arrived. Force fast retransmit. */
2794 do_lost = tcp_is_reno(tp) ||
2795 tcp_force_fast_retransmit(sk);
2796 }
2797 if (tcp_try_undo_dsack(sk)) {
2798 tcp_try_keep_open(sk);
2799 return;
2800 }
2801 tcp_rack_identify_loss(sk, ack_flag);
2802 break;
2803 case TCP_CA_Loss:
2804 tcp_process_loss(sk, flag, is_dupack, rexmit);
2805 tcp_rack_identify_loss(sk, ack_flag);
2806 if (!(icsk->icsk_ca_state == TCP_CA_Open ||
2807 (*ack_flag & FLAG_LOST_RETRANS)))
2808 return;
2809 /* Change state if cwnd is undone or retransmits are lost */
2810 /* fall through */
2811 default:
2812 if (tcp_is_reno(tp)) {
2813 if (flag & FLAG_SND_UNA_ADVANCED)
2814 tcp_reset_reno_sack(tp);
2815 if (is_dupack)
2816 tcp_add_reno_sack(sk);
2817 }
2818
2819 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2820 tcp_try_undo_dsack(sk);
2821
2822 tcp_rack_identify_loss(sk, ack_flag);
2823 if (!tcp_time_to_recover(sk, flag)) {
2824 tcp_try_to_open(sk, flag);
2825 return;
2826 }
2827
2828 /* MTU probe failure: don't reduce cwnd */
2829 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2830 icsk->icsk_mtup.probe_size &&
2831 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2832 tcp_mtup_probe_failed(sk);
2833 /* Restores the reduction we did in tcp_mtup_probe() */
2834 tp->snd_cwnd++;
2835 tcp_simple_retransmit(sk);
2836 return;
2837 }
2838
2839 /* Otherwise enter Recovery state */
2840 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2841 fast_rexmit = 1;
2842 }
2843
2844 if (do_lost)
2845 tcp_update_scoreboard(sk, fast_rexmit);
2846 *rexmit = REXMIT_LOST;
2847}
2848
2849static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag)
2850{
2851 u32 wlen = sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen * HZ;
2852 struct tcp_sock *tp = tcp_sk(sk);
2853
2854 if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) {
2855 /* If the remote keeps returning delayed ACKs, eventually
2856 * the min filter would pick it up and overestimate the
2857 * prop. delay when it expires. Skip suspected delayed ACKs.
2858 */
2859 return;
2860 }
2861 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32,
2862 rtt_us ? : jiffies_to_usecs(1));
2863}
2864
2865static bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2866 long seq_rtt_us, long sack_rtt_us,
2867 long ca_rtt_us, struct rate_sample *rs)
2868{
2869 const struct tcp_sock *tp = tcp_sk(sk);
2870
2871 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2872 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2873 * Karn's algorithm forbids taking RTT if some retransmitted data
2874 * is acked (RFC6298).
2875 */
2876 if (seq_rtt_us < 0)
2877 seq_rtt_us = sack_rtt_us;
2878
2879 /* RTTM Rule: A TSecr value received in a segment is used to
2880 * update the averaged RTT measurement only if the segment
2881 * acknowledges some new data, i.e., only if it advances the
2882 * left edge of the send window.
2883 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2884 */
2885 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2886 flag & FLAG_ACKED) {
2887 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
2888 u32 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
2889
2890 seq_rtt_us = ca_rtt_us = delta_us;
2891 }
2892 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */
2893 if (seq_rtt_us < 0)
2894 return false;
2895
2896 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2897 * always taken together with ACK, SACK, or TS-opts. Any negative
2898 * values will be skipped with the seq_rtt_us < 0 check above.
2899 */
2900 tcp_update_rtt_min(sk, ca_rtt_us, flag);
2901 tcp_rtt_estimator(sk, seq_rtt_us);
2902 tcp_set_rto(sk);
2903
2904 /* RFC6298: only reset backoff on valid RTT measurement. */
2905 inet_csk(sk)->icsk_backoff = 0;
2906 return true;
2907}
2908
2909/* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2910void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
2911{
2912 struct rate_sample rs;
2913 long rtt_us = -1L;
2914
2915 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack)
2916 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack);
2917
2918 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs);
2919}
2920
2921
2922static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
2923{
2924 const struct inet_connection_sock *icsk = inet_csk(sk);
2925
2926 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
2927 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32;
2928}
2929
2930/* Restart timer after forward progress on connection.
2931 * RFC2988 recommends to restart timer to now+rto.
2932 */
2933void tcp_rearm_rto(struct sock *sk)
2934{
2935 const struct inet_connection_sock *icsk = inet_csk(sk);
2936 struct tcp_sock *tp = tcp_sk(sk);
2937
2938 /* If the retrans timer is currently being used by Fast Open
2939 * for SYN-ACK retrans purpose, stay put.
2940 */
2941 if (tp->fastopen_rsk)
2942 return;
2943
2944 if (!tp->packets_out) {
2945 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2946 } else {
2947 u32 rto = inet_csk(sk)->icsk_rto;
2948 /* Offset the time elapsed after installing regular RTO */
2949 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
2950 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
2951 s64 delta_us = tcp_rto_delta_us(sk);
2952 /* delta_us may not be positive if the socket is locked
2953 * when the retrans timer fires and is rescheduled.
2954 */
2955 rto = usecs_to_jiffies(max_t(int, delta_us, 1));
2956 }
2957 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
2958 TCP_RTO_MAX);
2959 }
2960}
2961
2962/* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
2963static void tcp_set_xmit_timer(struct sock *sk)
2964{
2965 if (!tcp_schedule_loss_probe(sk, true))
2966 tcp_rearm_rto(sk);
2967}
2968
2969/* If we get here, the whole TSO packet has not been acked. */
2970static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
2971{
2972 struct tcp_sock *tp = tcp_sk(sk);
2973 u32 packets_acked;
2974
2975 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
2976
2977 packets_acked = tcp_skb_pcount(skb);
2978 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2979 return 0;
2980 packets_acked -= tcp_skb_pcount(skb);
2981
2982 if (packets_acked) {
2983 BUG_ON(tcp_skb_pcount(skb) == 0);
2984 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
2985 }
2986
2987 return packets_acked;
2988}
2989
2990static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
2991 u32 prior_snd_una)
2992{
2993 const struct skb_shared_info *shinfo;
2994
2995 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
2996 if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
2997 return;
2998
2999 shinfo = skb_shinfo(skb);
3000 if (!before(shinfo->tskey, prior_snd_una) &&
3001 before(shinfo->tskey, tcp_sk(sk)->snd_una)) {
3002 tcp_skb_tsorted_save(skb) {
3003 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3004 } tcp_skb_tsorted_restore(skb);
3005 }
3006}
3007
3008/* Remove acknowledged frames from the retransmission queue. If our packet
3009 * is before the ack sequence we can discard it as it's confirmed to have
3010 * arrived at the other end.
3011 */
3012static int tcp_clean_rtx_queue(struct sock *sk, u32 prior_fack,
3013 u32 prior_snd_una,
3014 struct tcp_sacktag_state *sack)
3015{
3016 const struct inet_connection_sock *icsk = inet_csk(sk);
3017 u64 first_ackt, last_ackt;
3018 struct tcp_sock *tp = tcp_sk(sk);
3019 u32 prior_sacked = tp->sacked_out;
3020 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */
3021 struct sk_buff *skb, *next;
3022 bool fully_acked = true;
3023 long sack_rtt_us = -1L;
3024 long seq_rtt_us = -1L;
3025 long ca_rtt_us = -1L;
3026 u32 pkts_acked = 0;
3027 u32 last_in_flight = 0;
3028 bool rtt_update;
3029 int flag = 0;
3030
3031 first_ackt = 0;
3032
3033 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) {
3034 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3035 const u32 start_seq = scb->seq;
3036 u8 sacked = scb->sacked;
3037 u32 acked_pcount;
3038
3039 tcp_ack_tstamp(sk, skb, prior_snd_una);
3040
3041 /* Determine how many packets and what bytes were acked, tso and else */
3042 if (after(scb->end_seq, tp->snd_una)) {
3043 if (tcp_skb_pcount(skb) == 1 ||
3044 !after(tp->snd_una, scb->seq))
3045 break;
3046
3047 acked_pcount = tcp_tso_acked(sk, skb);
3048 if (!acked_pcount)
3049 break;
3050 fully_acked = false;
3051 } else {
3052 acked_pcount = tcp_skb_pcount(skb);
3053 }
3054
3055 if (unlikely(sacked & TCPCB_RETRANS)) {
3056 if (sacked & TCPCB_SACKED_RETRANS)
3057 tp->retrans_out -= acked_pcount;
3058 flag |= FLAG_RETRANS_DATA_ACKED;
3059 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3060 last_ackt = skb->skb_mstamp;
3061 WARN_ON_ONCE(last_ackt == 0);
3062 if (!first_ackt)
3063 first_ackt = last_ackt;
3064
3065 last_in_flight = TCP_SKB_CB(skb)->tx.in_flight;
3066 if (before(start_seq, reord))
3067 reord = start_seq;
3068 if (!after(scb->end_seq, tp->high_seq))
3069 flag |= FLAG_ORIG_SACK_ACKED;
3070 }
3071
3072 if (sacked & TCPCB_SACKED_ACKED) {
3073 tp->sacked_out -= acked_pcount;
3074 } else if (tcp_is_sack(tp)) {
3075 tp->delivered += acked_pcount;
3076 if (!tcp_skb_spurious_retrans(tp, skb))
3077 tcp_rack_advance(tp, sacked, scb->end_seq,
3078 skb->skb_mstamp);
3079 }
3080 if (sacked & TCPCB_LOST)
3081 tp->lost_out -= acked_pcount;
3082
3083 tp->packets_out -= acked_pcount;
3084 pkts_acked += acked_pcount;
3085 tcp_rate_skb_delivered(sk, skb, sack->rate);
3086
3087 /* Initial outgoing SYN's get put onto the write_queue
3088 * just like anything else we transmit. It is not
3089 * true data, and if we misinform our callers that
3090 * this ACK acks real data, we will erroneously exit
3091 * connection startup slow start one packet too
3092 * quickly. This is severely frowned upon behavior.
3093 */
3094 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3095 flag |= FLAG_DATA_ACKED;
3096 } else {
3097 flag |= FLAG_SYN_ACKED;
3098 tp->retrans_stamp = 0;
3099 }
3100
3101 if (!fully_acked)
3102 break;
3103
3104 next = skb_rb_next(skb);
3105 if (unlikely(skb == tp->retransmit_skb_hint))
3106 tp->retransmit_skb_hint = NULL;
3107 if (unlikely(skb == tp->lost_skb_hint))
3108 tp->lost_skb_hint = NULL;
3109 tcp_rtx_queue_unlink_and_free(skb, sk);
3110 }
3111
3112 if (!skb)
3113 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3114
3115 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3116 tp->snd_up = tp->snd_una;
3117
3118 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3119 flag |= FLAG_SACK_RENEGING;
3120
3121 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3122 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt);
3123 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt);
3124
3125 if (pkts_acked == 1 && last_in_flight < tp->mss_cache &&
3126 last_in_flight && !prior_sacked && fully_acked &&
3127 sack->rate->prior_delivered + 1 == tp->delivered &&
3128 !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) {
3129 /* Conservatively mark a delayed ACK. It's typically
3130 * from a lone runt packet over the round trip to
3131 * a receiver w/o out-of-order or CE events.
3132 */
3133 flag |= FLAG_ACK_MAYBE_DELAYED;
3134 }
3135 }
3136 if (sack->first_sackt) {
3137 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt);
3138 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt);
3139 }
3140 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3141 ca_rtt_us, sack->rate);
3142
3143 if (flag & FLAG_ACKED) {
3144 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3145 if (unlikely(icsk->icsk_mtup.probe_size &&
3146 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3147 tcp_mtup_probe_success(sk);
3148 }
3149
3150 if (tcp_is_reno(tp)) {
3151 tcp_remove_reno_sacks(sk, pkts_acked);
3152 } else {
3153 int delta;
3154
3155 /* Non-retransmitted hole got filled? That's reordering */
3156 if (before(reord, prior_fack))
3157 tcp_check_sack_reordering(sk, reord, 0);
3158
3159 delta = prior_sacked - tp->sacked_out;
3160 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3161 }
3162 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3163 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp, skb->skb_mstamp)) {
3164 /* Do not re-arm RTO if the sack RTT is measured from data sent
3165 * after when the head was last (re)transmitted. Otherwise the
3166 * timeout may continue to extend in loss recovery.
3167 */
3168 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3169 }
3170
3171 if (icsk->icsk_ca_ops->pkts_acked) {
3172 struct ack_sample sample = { .pkts_acked = pkts_acked,
3173 .rtt_us = sack->rate->rtt_us,
3174 .in_flight = last_in_flight };
3175
3176 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3177 }
3178
3179#if FASTRETRANS_DEBUG > 0
3180 WARN_ON((int)tp->sacked_out < 0);
3181 WARN_ON((int)tp->lost_out < 0);
3182 WARN_ON((int)tp->retrans_out < 0);
3183 if (!tp->packets_out && tcp_is_sack(tp)) {
3184 icsk = inet_csk(sk);
3185 if (tp->lost_out) {
3186 pr_debug("Leak l=%u %d\n",
3187 tp->lost_out, icsk->icsk_ca_state);
3188 tp->lost_out = 0;
3189 }
3190 if (tp->sacked_out) {
3191 pr_debug("Leak s=%u %d\n",
3192 tp->sacked_out, icsk->icsk_ca_state);
3193 tp->sacked_out = 0;
3194 }
3195 if (tp->retrans_out) {
3196 pr_debug("Leak r=%u %d\n",
3197 tp->retrans_out, icsk->icsk_ca_state);
3198 tp->retrans_out = 0;
3199 }
3200 }
3201#endif
3202 return flag;
3203}
3204
3205static void tcp_ack_probe(struct sock *sk)
3206{
3207 struct inet_connection_sock *icsk = inet_csk(sk);
3208 struct sk_buff *head = tcp_send_head(sk);
3209 const struct tcp_sock *tp = tcp_sk(sk);
3210
3211 /* Was it a usable window open? */
3212 if (!head)
3213 return;
3214 if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) {
3215 icsk->icsk_backoff = 0;
3216 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3217 /* Socket must be waked up by subsequent tcp_data_snd_check().
3218 * This function is not for random using!
3219 */
3220 } else {
3221 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3222
3223 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3224 when, TCP_RTO_MAX);
3225 }
3226}
3227
3228static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3229{
3230 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3231 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3232}
3233
3234/* Decide wheather to run the increase function of congestion control. */
3235static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3236{
3237 /* If reordering is high then always grow cwnd whenever data is
3238 * delivered regardless of its ordering. Otherwise stay conservative
3239 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3240 * new SACK or ECE mark may first advance cwnd here and later reduce
3241 * cwnd in tcp_fastretrans_alert() based on more states.
3242 */
3243 if (tcp_sk(sk)->reordering > sock_net(sk)->ipv4.sysctl_tcp_reordering)
3244 return flag & FLAG_FORWARD_PROGRESS;
3245
3246 return flag & FLAG_DATA_ACKED;
3247}
3248
3249/* The "ultimate" congestion control function that aims to replace the rigid
3250 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3251 * It's called toward the end of processing an ACK with precise rate
3252 * information. All transmission or retransmission are delayed afterwards.
3253 */
3254static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3255 int flag, const struct rate_sample *rs)
3256{
3257 const struct inet_connection_sock *icsk = inet_csk(sk);
3258
3259 if (icsk->icsk_ca_ops->cong_control) {
3260 icsk->icsk_ca_ops->cong_control(sk, rs);
3261 return;
3262 }
3263
3264 if (tcp_in_cwnd_reduction(sk)) {
3265 /* Reduce cwnd if state mandates */
3266 tcp_cwnd_reduction(sk, acked_sacked, flag);
3267 } else if (tcp_may_raise_cwnd(sk, flag)) {
3268 /* Advance cwnd if state allows */
3269 tcp_cong_avoid(sk, ack, acked_sacked);
3270 }
3271 tcp_update_pacing_rate(sk);
3272}
3273
3274/* Check that window update is acceptable.
3275 * The function assumes that snd_una<=ack<=snd_next.
3276 */
3277static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3278 const u32 ack, const u32 ack_seq,
3279 const u32 nwin)
3280{
3281 return after(ack, tp->snd_una) ||
3282 after(ack_seq, tp->snd_wl1) ||
3283 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3284}
3285
3286/* If we update tp->snd_una, also update tp->bytes_acked */
3287static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3288{
3289 u32 delta = ack - tp->snd_una;
3290
3291 sock_owned_by_me((struct sock *)tp);
3292 tp->bytes_acked += delta;
3293 tp->snd_una = ack;
3294}
3295
3296/* If we update tp->rcv_nxt, also update tp->bytes_received */
3297static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3298{
3299 u32 delta = seq - tp->rcv_nxt;
3300
3301 sock_owned_by_me((struct sock *)tp);
3302 tp->bytes_received += delta;
3303 tp->rcv_nxt = seq;
3304}
3305
3306/* Update our send window.
3307 *
3308 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3309 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3310 */
3311static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3312 u32 ack_seq)
3313{
3314 struct tcp_sock *tp = tcp_sk(sk);
3315 int flag = 0;
3316 u32 nwin = ntohs(tcp_hdr(skb)->window);
3317
3318 if (likely(!tcp_hdr(skb)->syn))
3319 nwin <<= tp->rx_opt.snd_wscale;
3320
3321 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3322 flag |= FLAG_WIN_UPDATE;
3323 tcp_update_wl(tp, ack_seq);
3324
3325 if (tp->snd_wnd != nwin) {
3326 tp->snd_wnd = nwin;
3327
3328 /* Note, it is the only place, where
3329 * fast path is recovered for sending TCP.
3330 */
3331 tp->pred_flags = 0;
3332 tcp_fast_path_check(sk);
3333
3334 if (!tcp_write_queue_empty(sk))
3335 tcp_slow_start_after_idle_check(sk);
3336
3337 if (nwin > tp->max_window) {
3338 tp->max_window = nwin;
3339 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3340 }
3341 }
3342 }
3343
3344 tcp_snd_una_update(tp, ack);
3345
3346 return flag;
3347}
3348
3349static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3350 u32 *last_oow_ack_time)
3351{
3352 if (*last_oow_ack_time) {
3353 s32 elapsed = (s32)(tcp_jiffies32 - *last_oow_ack_time);
3354
3355 if (0 <= elapsed && elapsed < net->ipv4.sysctl_tcp_invalid_ratelimit) {
3356 NET_INC_STATS(net, mib_idx);
3357 return true; /* rate-limited: don't send yet! */
3358 }
3359 }
3360
3361 *last_oow_ack_time = tcp_jiffies32;
3362
3363 return false; /* not rate-limited: go ahead, send dupack now! */
3364}
3365
3366/* Return true if we're currently rate-limiting out-of-window ACKs and
3367 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3368 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3369 * attacks that send repeated SYNs or ACKs for the same connection. To
3370 * do this, we do not send a duplicate SYNACK or ACK if the remote
3371 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3372 */
3373bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3374 int mib_idx, u32 *last_oow_ack_time)
3375{
3376 /* Data packets without SYNs are not likely part of an ACK loop. */
3377 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3378 !tcp_hdr(skb)->syn)
3379 return false;
3380
3381 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3382}
3383
3384/* RFC 5961 7 [ACK Throttling] */
3385static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
3386{
3387 /* unprotected vars, we dont care of overwrites */
3388 static u32 challenge_timestamp;
3389 static unsigned int challenge_count;
3390 struct tcp_sock *tp = tcp_sk(sk);
3391 struct net *net = sock_net(sk);
3392 u32 count, now;
3393
3394 /* First check our per-socket dupack rate limit. */
3395 if (__tcp_oow_rate_limited(net,
3396 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3397 &tp->last_oow_ack_time))
3398 return;
3399
3400 /* Then check host-wide RFC 5961 rate limit. */
3401 now = jiffies / HZ;
3402 if (now != challenge_timestamp) {
3403 u32 ack_limit = net->ipv4.sysctl_tcp_challenge_ack_limit;
3404 u32 half = (ack_limit + 1) >> 1;
3405
3406 challenge_timestamp = now;
3407 WRITE_ONCE(challenge_count, half + prandom_u32_max(ack_limit));
3408 }
3409 count = READ_ONCE(challenge_count);
3410 if (count > 0) {
3411 WRITE_ONCE(challenge_count, count - 1);
3412 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK);
3413 tcp_send_ack(sk);
3414 }
3415}
3416
3417static void tcp_store_ts_recent(struct tcp_sock *tp)
3418{
3419 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3420 tp->rx_opt.ts_recent_stamp = get_seconds();
3421}
3422
3423static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3424{
3425 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3426 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3427 * extra check below makes sure this can only happen
3428 * for pure ACK frames. -DaveM
3429 *
3430 * Not only, also it occurs for expired timestamps.
3431 */
3432
3433 if (tcp_paws_check(&tp->rx_opt, 0))
3434 tcp_store_ts_recent(tp);
3435 }
3436}
3437
3438/* This routine deals with acks during a TLP episode.
3439 * We mark the end of a TLP episode on receiving TLP dupack or when
3440 * ack is after tlp_high_seq.
3441 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3442 */
3443static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3444{
3445 struct tcp_sock *tp = tcp_sk(sk);
3446
3447 if (before(ack, tp->tlp_high_seq))
3448 return;
3449
3450 if (flag & FLAG_DSACKING_ACK) {
3451 /* This DSACK means original and TLP probe arrived; no loss */
3452 tp->tlp_high_seq = 0;
3453 } else if (after(ack, tp->tlp_high_seq)) {
3454 /* ACK advances: there was a loss, so reduce cwnd. Reset
3455 * tlp_high_seq in tcp_init_cwnd_reduction()
3456 */
3457 tcp_init_cwnd_reduction(sk);
3458 tcp_set_ca_state(sk, TCP_CA_CWR);
3459 tcp_end_cwnd_reduction(sk);
3460 tcp_try_keep_open(sk);
3461 NET_INC_STATS(sock_net(sk),
3462 LINUX_MIB_TCPLOSSPROBERECOVERY);
3463 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3464 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3465 /* Pure dupack: original and TLP probe arrived; no loss */
3466 tp->tlp_high_seq = 0;
3467 }
3468}
3469
3470static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3471{
3472 const struct inet_connection_sock *icsk = inet_csk(sk);
3473
3474 if (icsk->icsk_ca_ops->in_ack_event)
3475 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3476}
3477
3478/* Congestion control has updated the cwnd already. So if we're in
3479 * loss recovery then now we do any new sends (for FRTO) or
3480 * retransmits (for CA_Loss or CA_recovery) that make sense.
3481 */
3482static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3483{
3484 struct tcp_sock *tp = tcp_sk(sk);
3485
3486 if (rexmit == REXMIT_NONE)
3487 return;
3488
3489 if (unlikely(rexmit == 2)) {
3490 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3491 TCP_NAGLE_OFF);
3492 if (after(tp->snd_nxt, tp->high_seq))
3493 return;
3494 tp->frto = 0;
3495 }
3496 tcp_xmit_retransmit_queue(sk);
3497}
3498
3499/* This routine deals with incoming acks, but not outgoing ones. */
3500static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3501{
3502 struct inet_connection_sock *icsk = inet_csk(sk);
3503 struct tcp_sock *tp = tcp_sk(sk);
3504 struct tcp_sacktag_state sack_state;
3505 struct rate_sample rs = { .prior_delivered = 0 };
3506 u32 prior_snd_una = tp->snd_una;
3507 bool is_sack_reneg = tp->is_sack_reneg;
3508 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3509 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3510 bool is_dupack = false;
3511 int prior_packets = tp->packets_out;
3512 u32 delivered = tp->delivered;
3513 u32 lost = tp->lost;
3514 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3515 u32 prior_fack;
3516
3517 sack_state.first_sackt = 0;
3518 sack_state.rate = &rs;
3519
3520 /* We very likely will need to access rtx queue. */
3521 prefetch(sk->tcp_rtx_queue.rb_node);
3522
3523 /* If the ack is older than previous acks
3524 * then we can probably ignore it.
3525 */
3526 if (before(ack, prior_snd_una)) {
3527 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3528 if (before(ack, prior_snd_una - tp->max_window)) {
3529 if (!(flag & FLAG_NO_CHALLENGE_ACK))
3530 tcp_send_challenge_ack(sk, skb);
3531 return -1;
3532 }
3533 goto old_ack;
3534 }
3535
3536 /* If the ack includes data we haven't sent yet, discard
3537 * this segment (RFC793 Section 3.9).
3538 */
3539 if (after(ack, tp->snd_nxt))
3540 goto invalid_ack;
3541
3542 if (after(ack, prior_snd_una)) {
3543 flag |= FLAG_SND_UNA_ADVANCED;
3544 icsk->icsk_retransmits = 0;
3545 }
3546
3547 prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una;
3548 rs.prior_in_flight = tcp_packets_in_flight(tp);
3549
3550 /* ts_recent update must be made after we are sure that the packet
3551 * is in window.
3552 */
3553 if (flag & FLAG_UPDATE_TS_RECENT)
3554 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3555
3556 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3557 /* Window is constant, pure forward advance.
3558 * No more checks are required.
3559 * Note, we use the fact that SND.UNA>=SND.WL2.
3560 */
3561 tcp_update_wl(tp, ack_seq);
3562 tcp_snd_una_update(tp, ack);
3563 flag |= FLAG_WIN_UPDATE;
3564
3565 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3566
3567 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3568 } else {
3569 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3570
3571 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3572 flag |= FLAG_DATA;
3573 else
3574 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3575
3576 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3577
3578 if (TCP_SKB_CB(skb)->sacked)
3579 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3580 &sack_state);
3581
3582 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3583 flag |= FLAG_ECE;
3584 ack_ev_flags |= CA_ACK_ECE;
3585 }
3586
3587 if (flag & FLAG_WIN_UPDATE)
3588 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3589
3590 tcp_in_ack_event(sk, ack_ev_flags);
3591 }
3592
3593 /* We passed data and got it acked, remove any soft error
3594 * log. Something worked...
3595 */
3596 sk->sk_err_soft = 0;
3597 icsk->icsk_probes_out = 0;
3598 tp->rcv_tstamp = tcp_jiffies32;
3599 if (!prior_packets)
3600 goto no_queue;
3601
3602 /* See if we can take anything off of the retransmit queue. */
3603 flag |= tcp_clean_rtx_queue(sk, prior_fack, prior_snd_una, &sack_state);
3604
3605 tcp_rack_update_reo_wnd(sk, &rs);
3606
3607 if (tp->tlp_high_seq)
3608 tcp_process_tlp_ack(sk, ack, flag);
3609 /* If needed, reset TLP/RTO timer; RACK may later override this. */
3610 if (flag & FLAG_SET_XMIT_TIMER)
3611 tcp_set_xmit_timer(sk);
3612
3613 if (tcp_ack_is_dubious(sk, flag)) {
3614 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3615 tcp_fastretrans_alert(sk, prior_snd_una, is_dupack, &flag,
3616 &rexmit);
3617 }
3618
3619 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3620 sk_dst_confirm(sk);
3621
3622 delivered = tp->delivered - delivered; /* freshly ACKed or SACKed */
3623 lost = tp->lost - lost; /* freshly marked lost */
3624 rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED);
3625 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate);
3626 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate);
3627 tcp_xmit_recovery(sk, rexmit);
3628 return 1;
3629
3630no_queue:
3631 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3632 if (flag & FLAG_DSACKING_ACK)
3633 tcp_fastretrans_alert(sk, prior_snd_una, is_dupack, &flag,
3634 &rexmit);
3635 /* If this ack opens up a zero window, clear backoff. It was
3636 * being used to time the probes, and is probably far higher than
3637 * it needs to be for normal retransmission.
3638 */
3639 tcp_ack_probe(sk);
3640
3641 if (tp->tlp_high_seq)
3642 tcp_process_tlp_ack(sk, ack, flag);
3643 return 1;
3644
3645invalid_ack:
3646 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3647 return -1;
3648
3649old_ack:
3650 /* If data was SACKed, tag it and see if we should send more data.
3651 * If data was DSACKed, see if we can undo a cwnd reduction.
3652 */
3653 if (TCP_SKB_CB(skb)->sacked) {
3654 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3655 &sack_state);
3656 tcp_fastretrans_alert(sk, prior_snd_una, is_dupack, &flag,
3657 &rexmit);
3658 tcp_xmit_recovery(sk, rexmit);
3659 }
3660
3661 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3662 return 0;
3663}
3664
3665static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3666 bool syn, struct tcp_fastopen_cookie *foc,
3667 bool exp_opt)
3668{
3669 /* Valid only in SYN or SYN-ACK with an even length. */
3670 if (!foc || !syn || len < 0 || (len & 1))
3671 return;
3672
3673 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3674 len <= TCP_FASTOPEN_COOKIE_MAX)
3675 memcpy(foc->val, cookie, len);
3676 else if (len != 0)
3677 len = -1;
3678 foc->len = len;
3679 foc->exp = exp_opt;
3680}
3681
3682static void smc_parse_options(const struct tcphdr *th,
3683 struct tcp_options_received *opt_rx,
3684 const unsigned char *ptr,
3685 int opsize)
3686{
3687#if IS_ENABLED(CONFIG_SMC)
3688 if (static_branch_unlikely(&tcp_have_smc)) {
3689 if (th->syn && !(opsize & 1) &&
3690 opsize >= TCPOLEN_EXP_SMC_BASE &&
3691 get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC)
3692 opt_rx->smc_ok = 1;
3693 }
3694#endif
3695}
3696
3697/* Look for tcp options. Normally only called on SYN and SYNACK packets.
3698 * But, this can also be called on packets in the established flow when
3699 * the fast version below fails.
3700 */
3701void tcp_parse_options(const struct net *net,
3702 const struct sk_buff *skb,
3703 struct tcp_options_received *opt_rx, int estab,
3704 struct tcp_fastopen_cookie *foc)
3705{
3706 const unsigned char *ptr;
3707 const struct tcphdr *th = tcp_hdr(skb);
3708 int length = (th->doff * 4) - sizeof(struct tcphdr);
3709
3710 ptr = (const unsigned char *)(th + 1);
3711 opt_rx->saw_tstamp = 0;
3712
3713 while (length > 0) {
3714 int opcode = *ptr++;
3715 int opsize;
3716
3717 switch (opcode) {
3718 case TCPOPT_EOL:
3719 return;
3720 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3721 length--;
3722 continue;
3723 default:
3724 opsize = *ptr++;
3725 if (opsize < 2) /* "silly options" */
3726 return;
3727 if (opsize > length)
3728 return; /* don't parse partial options */
3729 switch (opcode) {
3730 case TCPOPT_MSS:
3731 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3732 u16 in_mss = get_unaligned_be16(ptr);
3733 if (in_mss) {
3734 if (opt_rx->user_mss &&
3735 opt_rx->user_mss < in_mss)
3736 in_mss = opt_rx->user_mss;
3737 opt_rx->mss_clamp = in_mss;
3738 }
3739 }
3740 break;
3741 case TCPOPT_WINDOW:
3742 if (opsize == TCPOLEN_WINDOW && th->syn &&
3743 !estab && net->ipv4.sysctl_tcp_window_scaling) {
3744 __u8 snd_wscale = *(__u8 *)ptr;
3745 opt_rx->wscale_ok = 1;
3746 if (snd_wscale > TCP_MAX_WSCALE) {
3747 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
3748 __func__,
3749 snd_wscale,
3750 TCP_MAX_WSCALE);
3751 snd_wscale = TCP_MAX_WSCALE;
3752 }
3753 opt_rx->snd_wscale = snd_wscale;
3754 }
3755 break;
3756 case TCPOPT_TIMESTAMP:
3757 if ((opsize == TCPOLEN_TIMESTAMP) &&
3758 ((estab && opt_rx->tstamp_ok) ||
3759 (!estab && net->ipv4.sysctl_tcp_timestamps))) {
3760 opt_rx->saw_tstamp = 1;
3761 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3762 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3763 }
3764 break;
3765 case TCPOPT_SACK_PERM:
3766 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3767 !estab && net->ipv4.sysctl_tcp_sack) {
3768 opt_rx->sack_ok = TCP_SACK_SEEN;
3769 tcp_sack_reset(opt_rx);
3770 }
3771 break;
3772
3773 case TCPOPT_SACK:
3774 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3775 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3776 opt_rx->sack_ok) {
3777 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3778 }
3779 break;
3780#ifdef CONFIG_TCP_MD5SIG
3781 case TCPOPT_MD5SIG:
3782 /*
3783 * The MD5 Hash has already been
3784 * checked (see tcp_v{4,6}_do_rcv()).
3785 */
3786 break;
3787#endif
3788 case TCPOPT_FASTOPEN:
3789 tcp_parse_fastopen_option(
3790 opsize - TCPOLEN_FASTOPEN_BASE,
3791 ptr, th->syn, foc, false);
3792 break;
3793
3794 case TCPOPT_EXP:
3795 /* Fast Open option shares code 254 using a
3796 * 16 bits magic number.
3797 */
3798 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
3799 get_unaligned_be16(ptr) ==
3800 TCPOPT_FASTOPEN_MAGIC)
3801 tcp_parse_fastopen_option(opsize -
3802 TCPOLEN_EXP_FASTOPEN_BASE,
3803 ptr + 2, th->syn, foc, true);
3804 else
3805 smc_parse_options(th, opt_rx, ptr,
3806 opsize);
3807 break;
3808
3809 }
3810 ptr += opsize-2;
3811 length -= opsize;
3812 }
3813 }
3814}
3815EXPORT_SYMBOL(tcp_parse_options);
3816
3817static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3818{
3819 const __be32 *ptr = (const __be32 *)(th + 1);
3820
3821 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3822 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3823 tp->rx_opt.saw_tstamp = 1;
3824 ++ptr;
3825 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3826 ++ptr;
3827 if (*ptr)
3828 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3829 else
3830 tp->rx_opt.rcv_tsecr = 0;
3831 return true;
3832 }
3833 return false;
3834}
3835
3836/* Fast parse options. This hopes to only see timestamps.
3837 * If it is wrong it falls back on tcp_parse_options().
3838 */
3839static bool tcp_fast_parse_options(const struct net *net,
3840 const struct sk_buff *skb,
3841 const struct tcphdr *th, struct tcp_sock *tp)
3842{
3843 /* In the spirit of fast parsing, compare doff directly to constant
3844 * values. Because equality is used, short doff can be ignored here.
3845 */
3846 if (th->doff == (sizeof(*th) / 4)) {
3847 tp->rx_opt.saw_tstamp = 0;
3848 return false;
3849 } else if (tp->rx_opt.tstamp_ok &&
3850 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3851 if (tcp_parse_aligned_timestamp(tp, th))
3852 return true;
3853 }
3854
3855 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL);
3856 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3857 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3858
3859 return true;
3860}
3861
3862#ifdef CONFIG_TCP_MD5SIG
3863/*
3864 * Parse MD5 Signature option
3865 */
3866const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3867{
3868 int length = (th->doff << 2) - sizeof(*th);
3869 const u8 *ptr = (const u8 *)(th + 1);
3870
3871 /* If not enough data remaining, we can short cut */
3872 while (length >= TCPOLEN_MD5SIG) {
3873 int opcode = *ptr++;
3874 int opsize;
3875
3876 switch (opcode) {
3877 case TCPOPT_EOL:
3878 return NULL;
3879 case TCPOPT_NOP:
3880 length--;
3881 continue;
3882 default:
3883 opsize = *ptr++;
3884 if (opsize < 2 || opsize > length)
3885 return NULL;
3886 if (opcode == TCPOPT_MD5SIG)
3887 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3888 }
3889 ptr += opsize - 2;
3890 length -= opsize;
3891 }
3892 return NULL;
3893}
3894EXPORT_SYMBOL(tcp_parse_md5sig_option);
3895#endif
3896
3897/* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3898 *
3899 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3900 * it can pass through stack. So, the following predicate verifies that
3901 * this segment is not used for anything but congestion avoidance or
3902 * fast retransmit. Moreover, we even are able to eliminate most of such
3903 * second order effects, if we apply some small "replay" window (~RTO)
3904 * to timestamp space.
3905 *
3906 * All these measures still do not guarantee that we reject wrapped ACKs
3907 * on networks with high bandwidth, when sequence space is recycled fastly,
3908 * but it guarantees that such events will be very rare and do not affect
3909 * connection seriously. This doesn't look nice, but alas, PAWS is really
3910 * buggy extension.
3911 *
3912 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3913 * states that events when retransmit arrives after original data are rare.
3914 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3915 * the biggest problem on large power networks even with minor reordering.
3916 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3917 * up to bandwidth of 18Gigabit/sec. 8) ]
3918 */
3919
3920static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3921{
3922 const struct tcp_sock *tp = tcp_sk(sk);
3923 const struct tcphdr *th = tcp_hdr(skb);
3924 u32 seq = TCP_SKB_CB(skb)->seq;
3925 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3926
3927 return (/* 1. Pure ACK with correct sequence number. */
3928 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3929
3930 /* 2. ... and duplicate ACK. */
3931 ack == tp->snd_una &&
3932
3933 /* 3. ... and does not update window. */
3934 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3935
3936 /* 4. ... and sits in replay window. */
3937 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3938}
3939
3940static inline bool tcp_paws_discard(const struct sock *sk,
3941 const struct sk_buff *skb)
3942{
3943 const struct tcp_sock *tp = tcp_sk(sk);
3944
3945 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
3946 !tcp_disordered_ack(sk, skb);
3947}
3948
3949/* Check segment sequence number for validity.
3950 *
3951 * Segment controls are considered valid, if the segment
3952 * fits to the window after truncation to the window. Acceptability
3953 * of data (and SYN, FIN, of course) is checked separately.
3954 * See tcp_data_queue(), for example.
3955 *
3956 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3957 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3958 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3959 * (borrowed from freebsd)
3960 */
3961
3962static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
3963{
3964 return !before(end_seq, tp->rcv_wup) &&
3965 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3966}
3967
3968/* When we get a reset we do this. */
3969void tcp_reset(struct sock *sk)
3970{
3971 trace_tcp_receive_reset(sk);
3972
3973 /* We want the right error as BSD sees it (and indeed as we do). */
3974 switch (sk->sk_state) {
3975 case TCP_SYN_SENT:
3976 sk->sk_err = ECONNREFUSED;
3977 break;
3978 case TCP_CLOSE_WAIT:
3979 sk->sk_err = EPIPE;
3980 break;
3981 case TCP_CLOSE:
3982 return;
3983 default:
3984 sk->sk_err = ECONNRESET;
3985 }
3986 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3987 smp_wmb();
3988
3989 tcp_write_queue_purge(sk);
3990 tcp_done(sk);
3991
3992 if (!sock_flag(sk, SOCK_DEAD))
3993 sk->sk_error_report(sk);
3994}
3995
3996/*
3997 * Process the FIN bit. This now behaves as it is supposed to work
3998 * and the FIN takes effect when it is validly part of sequence
3999 * space. Not before when we get holes.
4000 *
4001 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4002 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4003 * TIME-WAIT)
4004 *
4005 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4006 * close and we go into CLOSING (and later onto TIME-WAIT)
4007 *
4008 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4009 */
4010void tcp_fin(struct sock *sk)
4011{
4012 struct tcp_sock *tp = tcp_sk(sk);
4013
4014 inet_csk_schedule_ack(sk);
4015
4016 sk->sk_shutdown |= RCV_SHUTDOWN;
4017 sock_set_flag(sk, SOCK_DONE);
4018
4019 switch (sk->sk_state) {
4020 case TCP_SYN_RECV:
4021 case TCP_ESTABLISHED:
4022 /* Move to CLOSE_WAIT */
4023 tcp_set_state(sk, TCP_CLOSE_WAIT);
4024 inet_csk(sk)->icsk_ack.pingpong = 1;
4025 break;
4026
4027 case TCP_CLOSE_WAIT:
4028 case TCP_CLOSING:
4029 /* Received a retransmission of the FIN, do
4030 * nothing.
4031 */
4032 break;
4033 case TCP_LAST_ACK:
4034 /* RFC793: Remain in the LAST-ACK state. */
4035 break;
4036
4037 case TCP_FIN_WAIT1:
4038 /* This case occurs when a simultaneous close
4039 * happens, we must ack the received FIN and
4040 * enter the CLOSING state.
4041 */
4042 tcp_send_ack(sk);
4043 tcp_set_state(sk, TCP_CLOSING);
4044 break;
4045 case TCP_FIN_WAIT2:
4046 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4047 tcp_send_ack(sk);
4048 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4049 break;
4050 default:
4051 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4052 * cases we should never reach this piece of code.
4053 */
4054 pr_err("%s: Impossible, sk->sk_state=%d\n",
4055 __func__, sk->sk_state);
4056 break;
4057 }
4058
4059 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4060 * Probably, we should reset in this case. For now drop them.
4061 */
4062 skb_rbtree_purge(&tp->out_of_order_queue);
4063 if (tcp_is_sack(tp))
4064 tcp_sack_reset(&tp->rx_opt);
4065 sk_mem_reclaim(sk);
4066
4067 if (!sock_flag(sk, SOCK_DEAD)) {
4068 sk->sk_state_change(sk);
4069
4070 /* Do not send POLL_HUP for half duplex close. */
4071 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4072 sk->sk_state == TCP_CLOSE)
4073 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4074 else
4075 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4076 }
4077}
4078
4079static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4080 u32 end_seq)
4081{
4082 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4083 if (before(seq, sp->start_seq))
4084 sp->start_seq = seq;
4085 if (after(end_seq, sp->end_seq))
4086 sp->end_seq = end_seq;
4087 return true;
4088 }
4089 return false;
4090}
4091
4092static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4093{
4094 struct tcp_sock *tp = tcp_sk(sk);
4095
4096 if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) {
4097 int mib_idx;
4098
4099 if (before(seq, tp->rcv_nxt))
4100 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4101 else
4102 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4103
4104 NET_INC_STATS(sock_net(sk), mib_idx);
4105
4106 tp->rx_opt.dsack = 1;
4107 tp->duplicate_sack[0].start_seq = seq;
4108 tp->duplicate_sack[0].end_seq = end_seq;
4109 }
4110}
4111
4112static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4113{
4114 struct tcp_sock *tp = tcp_sk(sk);
4115
4116 if (!tp->rx_opt.dsack)
4117 tcp_dsack_set(sk, seq, end_seq);
4118 else
4119 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4120}
4121
4122static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4123{
4124 struct tcp_sock *tp = tcp_sk(sk);
4125
4126 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4127 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4128 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4129 tcp_enter_quickack_mode(sk);
4130
4131 if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) {
4132 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4133
4134 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4135 end_seq = tp->rcv_nxt;
4136 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4137 }
4138 }
4139
4140 tcp_send_ack(sk);
4141}
4142
4143/* These routines update the SACK block as out-of-order packets arrive or
4144 * in-order packets close up the sequence space.
4145 */
4146static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4147{
4148 int this_sack;
4149 struct tcp_sack_block *sp = &tp->selective_acks[0];
4150 struct tcp_sack_block *swalk = sp + 1;
4151
4152 /* See if the recent change to the first SACK eats into
4153 * or hits the sequence space of other SACK blocks, if so coalesce.
4154 */
4155 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4156 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4157 int i;
4158
4159 /* Zap SWALK, by moving every further SACK up by one slot.
4160 * Decrease num_sacks.
4161 */
4162 tp->rx_opt.num_sacks--;
4163 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4164 sp[i] = sp[i + 1];
4165 continue;
4166 }
4167 this_sack++, swalk++;
4168 }
4169}
4170
4171static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4172{
4173 struct tcp_sock *tp = tcp_sk(sk);
4174 struct tcp_sack_block *sp = &tp->selective_acks[0];
4175 int cur_sacks = tp->rx_opt.num_sacks;
4176 int this_sack;
4177
4178 if (!cur_sacks)
4179 goto new_sack;
4180
4181 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4182 if (tcp_sack_extend(sp, seq, end_seq)) {
4183 /* Rotate this_sack to the first one. */
4184 for (; this_sack > 0; this_sack--, sp--)
4185 swap(*sp, *(sp - 1));
4186 if (cur_sacks > 1)
4187 tcp_sack_maybe_coalesce(tp);
4188 return;
4189 }
4190 }
4191
4192 /* Could not find an adjacent existing SACK, build a new one,
4193 * put it at the front, and shift everyone else down. We
4194 * always know there is at least one SACK present already here.
4195 *
4196 * If the sack array is full, forget about the last one.
4197 */
4198 if (this_sack >= TCP_NUM_SACKS) {
4199 this_sack--;
4200 tp->rx_opt.num_sacks--;
4201 sp--;
4202 }
4203 for (; this_sack > 0; this_sack--, sp--)
4204 *sp = *(sp - 1);
4205
4206new_sack:
4207 /* Build the new head SACK, and we're done. */
4208 sp->start_seq = seq;
4209 sp->end_seq = end_seq;
4210 tp->rx_opt.num_sacks++;
4211}
4212
4213/* RCV.NXT advances, some SACKs should be eaten. */
4214
4215static void tcp_sack_remove(struct tcp_sock *tp)
4216{
4217 struct tcp_sack_block *sp = &tp->selective_acks[0];
4218 int num_sacks = tp->rx_opt.num_sacks;
4219 int this_sack;
4220
4221 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4222 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4223 tp->rx_opt.num_sacks = 0;
4224 return;
4225 }
4226
4227 for (this_sack = 0; this_sack < num_sacks;) {
4228 /* Check if the start of the sack is covered by RCV.NXT. */
4229 if (!before(tp->rcv_nxt, sp->start_seq)) {
4230 int i;
4231
4232 /* RCV.NXT must cover all the block! */
4233 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4234
4235 /* Zap this SACK, by moving forward any other SACKS. */
4236 for (i = this_sack+1; i < num_sacks; i++)
4237 tp->selective_acks[i-1] = tp->selective_acks[i];
4238 num_sacks--;
4239 continue;
4240 }
4241 this_sack++;
4242 sp++;
4243 }
4244 tp->rx_opt.num_sacks = num_sacks;
4245}
4246
4247/**
4248 * tcp_try_coalesce - try to merge skb to prior one
4249 * @sk: socket
4250 * @dest: destination queue
4251 * @to: prior buffer
4252 * @from: buffer to add in queue
4253 * @fragstolen: pointer to boolean
4254 *
4255 * Before queueing skb @from after @to, try to merge them
4256 * to reduce overall memory use and queue lengths, if cost is small.
4257 * Packets in ofo or receive queues can stay a long time.
4258 * Better try to coalesce them right now to avoid future collapses.
4259 * Returns true if caller should free @from instead of queueing it
4260 */
4261static bool tcp_try_coalesce(struct sock *sk,
4262 struct sk_buff *to,
4263 struct sk_buff *from,
4264 bool *fragstolen)
4265{
4266 int delta;
4267
4268 *fragstolen = false;
4269
4270 /* Its possible this segment overlaps with prior segment in queue */
4271 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4272 return false;
4273
4274 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4275 return false;
4276
4277 atomic_add(delta, &sk->sk_rmem_alloc);
4278 sk_mem_charge(sk, delta);
4279 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4280 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4281 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4282 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4283
4284 if (TCP_SKB_CB(from)->has_rxtstamp) {
4285 TCP_SKB_CB(to)->has_rxtstamp = true;
4286 to->tstamp = from->tstamp;
4287 }
4288
4289 return true;
4290}
4291
4292static void tcp_drop(struct sock *sk, struct sk_buff *skb)
4293{
4294 sk_drops_add(sk, skb);
4295 __kfree_skb(skb);
4296}
4297
4298/* This one checks to see if we can put data from the
4299 * out_of_order queue into the receive_queue.
4300 */
4301static void tcp_ofo_queue(struct sock *sk)
4302{
4303 struct tcp_sock *tp = tcp_sk(sk);
4304 __u32 dsack_high = tp->rcv_nxt;
4305 bool fin, fragstolen, eaten;
4306 struct sk_buff *skb, *tail;
4307 struct rb_node *p;
4308
4309 p = rb_first(&tp->out_of_order_queue);
4310 while (p) {
4311 skb = rb_to_skb(p);
4312 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4313 break;
4314
4315 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4316 __u32 dsack = dsack_high;
4317 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4318 dsack_high = TCP_SKB_CB(skb)->end_seq;
4319 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4320 }
4321 p = rb_next(p);
4322 rb_erase(&skb->rbnode, &tp->out_of_order_queue);
4323
4324 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
4325 SOCK_DEBUG(sk, "ofo packet was already received\n");
4326 tcp_drop(sk, skb);
4327 continue;
4328 }
4329 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4330 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4331 TCP_SKB_CB(skb)->end_seq);
4332
4333 tail = skb_peek_tail(&sk->sk_receive_queue);
4334 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4335 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4336 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
4337 if (!eaten)
4338 __skb_queue_tail(&sk->sk_receive_queue, skb);
4339 else
4340 kfree_skb_partial(skb, fragstolen);
4341
4342 if (unlikely(fin)) {
4343 tcp_fin(sk);
4344 /* tcp_fin() purges tp->out_of_order_queue,
4345 * so we must end this loop right now.
4346 */
4347 break;
4348 }
4349 }
4350}
4351
4352static bool tcp_prune_ofo_queue(struct sock *sk);
4353static int tcp_prune_queue(struct sock *sk);
4354
4355static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4356 unsigned int size)
4357{
4358 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4359 !sk_rmem_schedule(sk, skb, size)) {
4360
4361 if (tcp_prune_queue(sk) < 0)
4362 return -1;
4363
4364 while (!sk_rmem_schedule(sk, skb, size)) {
4365 if (!tcp_prune_ofo_queue(sk))
4366 return -1;
4367 }
4368 }
4369 return 0;
4370}
4371
4372static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4373{
4374 struct tcp_sock *tp = tcp_sk(sk);
4375 struct rb_node **p, *parent;
4376 struct sk_buff *skb1;
4377 u32 seq, end_seq;
4378 bool fragstolen;
4379
4380 tcp_ecn_check_ce(tp, skb);
4381
4382 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4383 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
4384 tcp_drop(sk, skb);
4385 return;
4386 }
4387
4388 /* Disable header prediction. */
4389 tp->pred_flags = 0;
4390 inet_csk_schedule_ack(sk);
4391
4392 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4393 seq = TCP_SKB_CB(skb)->seq;
4394 end_seq = TCP_SKB_CB(skb)->end_seq;
4395 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4396 tp->rcv_nxt, seq, end_seq);
4397
4398 p = &tp->out_of_order_queue.rb_node;
4399 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4400 /* Initial out of order segment, build 1 SACK. */
4401 if (tcp_is_sack(tp)) {
4402 tp->rx_opt.num_sacks = 1;
4403 tp->selective_acks[0].start_seq = seq;
4404 tp->selective_acks[0].end_seq = end_seq;
4405 }
4406 rb_link_node(&skb->rbnode, NULL, p);
4407 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4408 tp->ooo_last_skb = skb;
4409 goto end;
4410 }
4411
4412 /* In the typical case, we are adding an skb to the end of the list.
4413 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4414 */
4415 if (tcp_try_coalesce(sk, tp->ooo_last_skb,
4416 skb, &fragstolen)) {
4417coalesce_done:
4418 tcp_grow_window(sk, skb);
4419 kfree_skb_partial(skb, fragstolen);
4420 skb = NULL;
4421 goto add_sack;
4422 }
4423 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4424 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
4425 parent = &tp->ooo_last_skb->rbnode;
4426 p = &parent->rb_right;
4427 goto insert;
4428 }
4429
4430 /* Find place to insert this segment. Handle overlaps on the way. */
4431 parent = NULL;
4432 while (*p) {
4433 parent = *p;
4434 skb1 = rb_to_skb(parent);
4435 if (before(seq, TCP_SKB_CB(skb1)->seq)) {
4436 p = &parent->rb_left;
4437 continue;
4438 }
4439 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4440 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4441 /* All the bits are present. Drop. */
4442 NET_INC_STATS(sock_net(sk),
4443 LINUX_MIB_TCPOFOMERGE);
4444 __kfree_skb(skb);
4445 skb = NULL;
4446 tcp_dsack_set(sk, seq, end_seq);
4447 goto add_sack;
4448 }
4449 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4450 /* Partial overlap. */
4451 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
4452 } else {
4453 /* skb's seq == skb1's seq and skb covers skb1.
4454 * Replace skb1 with skb.
4455 */
4456 rb_replace_node(&skb1->rbnode, &skb->rbnode,
4457 &tp->out_of_order_queue);
4458 tcp_dsack_extend(sk,
4459 TCP_SKB_CB(skb1)->seq,
4460 TCP_SKB_CB(skb1)->end_seq);
4461 NET_INC_STATS(sock_net(sk),
4462 LINUX_MIB_TCPOFOMERGE);
4463 __kfree_skb(skb1);
4464 goto merge_right;
4465 }
4466 } else if (tcp_try_coalesce(sk, skb1,
4467 skb, &fragstolen)) {
4468 goto coalesce_done;
4469 }
4470 p = &parent->rb_right;
4471 }
4472insert:
4473 /* Insert segment into RB tree. */
4474 rb_link_node(&skb->rbnode, parent, p);
4475 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4476
4477merge_right:
4478 /* Remove other segments covered by skb. */
4479 while ((skb1 = skb_rb_next(skb)) != NULL) {
4480 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4481 break;
4482 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4483 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4484 end_seq);
4485 break;
4486 }
4487 rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
4488 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4489 TCP_SKB_CB(skb1)->end_seq);
4490 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4491 tcp_drop(sk, skb1);
4492 }
4493 /* If there is no skb after us, we are the last_skb ! */
4494 if (!skb1)
4495 tp->ooo_last_skb = skb;
4496
4497add_sack:
4498 if (tcp_is_sack(tp))
4499 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4500end:
4501 if (skb) {
4502 tcp_grow_window(sk, skb);
4503 skb_condense(skb);
4504 skb_set_owner_r(skb, sk);
4505 }
4506}
4507
4508static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4509 bool *fragstolen)
4510{
4511 int eaten;
4512 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4513
4514 __skb_pull(skb, hdrlen);
4515 eaten = (tail &&
4516 tcp_try_coalesce(sk, tail,
4517 skb, fragstolen)) ? 1 : 0;
4518 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4519 if (!eaten) {
4520 __skb_queue_tail(&sk->sk_receive_queue, skb);
4521 skb_set_owner_r(skb, sk);
4522 }
4523 return eaten;
4524}
4525
4526int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4527{
4528 struct sk_buff *skb;
4529 int err = -ENOMEM;
4530 int data_len = 0;
4531 bool fragstolen;
4532
4533 if (size == 0)
4534 return 0;
4535
4536 if (size > PAGE_SIZE) {
4537 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4538
4539 data_len = npages << PAGE_SHIFT;
4540 size = data_len + (size & ~PAGE_MASK);
4541 }
4542 skb = alloc_skb_with_frags(size - data_len, data_len,
4543 PAGE_ALLOC_COSTLY_ORDER,
4544 &err, sk->sk_allocation);
4545 if (!skb)
4546 goto err;
4547
4548 skb_put(skb, size - data_len);
4549 skb->data_len = data_len;
4550 skb->len = size;
4551
4552 if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4553 goto err_free;
4554
4555 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
4556 if (err)
4557 goto err_free;
4558
4559 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4560 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4561 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4562
4563 if (tcp_queue_rcv(sk, skb, 0, &fragstolen)) {
4564 WARN_ON_ONCE(fragstolen); /* should not happen */
4565 __kfree_skb(skb);
4566 }
4567 return size;
4568
4569err_free:
4570 kfree_skb(skb);
4571err:
4572 return err;
4573
4574}
4575
4576static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4577{
4578 struct tcp_sock *tp = tcp_sk(sk);
4579 bool fragstolen;
4580 int eaten;
4581
4582 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
4583 __kfree_skb(skb);
4584 return;
4585 }
4586 skb_dst_drop(skb);
4587 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
4588
4589 tcp_ecn_accept_cwr(tp, skb);
4590
4591 tp->rx_opt.dsack = 0;
4592
4593 /* Queue data for delivery to the user.
4594 * Packets in sequence go to the receive queue.
4595 * Out of sequence packets to the out_of_order_queue.
4596 */
4597 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4598 if (tcp_receive_window(tp) == 0)
4599 goto out_of_window;
4600
4601 /* Ok. In sequence. In window. */
4602queue_and_out:
4603 if (skb_queue_len(&sk->sk_receive_queue) == 0)
4604 sk_forced_mem_schedule(sk, skb->truesize);
4605 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4606 goto drop;
4607
4608 eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4609 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4610 if (skb->len)
4611 tcp_event_data_recv(sk, skb);
4612 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4613 tcp_fin(sk);
4614
4615 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4616 tcp_ofo_queue(sk);
4617
4618 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4619 * gap in queue is filled.
4620 */
4621 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
4622 inet_csk(sk)->icsk_ack.pingpong = 0;
4623 }
4624
4625 if (tp->rx_opt.num_sacks)
4626 tcp_sack_remove(tp);
4627
4628 tcp_fast_path_check(sk);
4629
4630 if (eaten > 0)
4631 kfree_skb_partial(skb, fragstolen);
4632 if (!sock_flag(sk, SOCK_DEAD))
4633 sk->sk_data_ready(sk);
4634 return;
4635 }
4636
4637 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4638 /* A retransmit, 2nd most common case. Force an immediate ack. */
4639 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4640 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4641
4642out_of_window:
4643 tcp_enter_quickack_mode(sk);
4644 inet_csk_schedule_ack(sk);
4645drop:
4646 tcp_drop(sk, skb);
4647 return;
4648 }
4649
4650 /* Out of window. F.e. zero window probe. */
4651 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4652 goto out_of_window;
4653
4654 tcp_enter_quickack_mode(sk);
4655
4656 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4657 /* Partial packet, seq < rcv_next < end_seq */
4658 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4659 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4660 TCP_SKB_CB(skb)->end_seq);
4661
4662 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4663
4664 /* If window is closed, drop tail of packet. But after
4665 * remembering D-SACK for its head made in previous line.
4666 */
4667 if (!tcp_receive_window(tp))
4668 goto out_of_window;
4669 goto queue_and_out;
4670 }
4671
4672 tcp_data_queue_ofo(sk, skb);
4673}
4674
4675static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
4676{
4677 if (list)
4678 return !skb_queue_is_last(list, skb) ? skb->next : NULL;
4679
4680 return skb_rb_next(skb);
4681}
4682
4683static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4684 struct sk_buff_head *list,
4685 struct rb_root *root)
4686{
4687 struct sk_buff *next = tcp_skb_next(skb, list);
4688
4689 if (list)
4690 __skb_unlink(skb, list);
4691 else
4692 rb_erase(&skb->rbnode, root);
4693
4694 __kfree_skb(skb);
4695 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4696
4697 return next;
4698}
4699
4700/* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
4701void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
4702{
4703 struct rb_node **p = &root->rb_node;
4704 struct rb_node *parent = NULL;
4705 struct sk_buff *skb1;
4706
4707 while (*p) {
4708 parent = *p;
4709 skb1 = rb_to_skb(parent);
4710 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
4711 p = &parent->rb_left;
4712 else
4713 p = &parent->rb_right;
4714 }
4715 rb_link_node(&skb->rbnode, parent, p);
4716 rb_insert_color(&skb->rbnode, root);
4717}
4718
4719/* Collapse contiguous sequence of skbs head..tail with
4720 * sequence numbers start..end.
4721 *
4722 * If tail is NULL, this means until the end of the queue.
4723 *
4724 * Segments with FIN/SYN are not collapsed (only because this
4725 * simplifies code)
4726 */
4727static void
4728tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
4729 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
4730{
4731 struct sk_buff *skb = head, *n;
4732 struct sk_buff_head tmp;
4733 bool end_of_skbs;
4734
4735 /* First, check that queue is collapsible and find
4736 * the point where collapsing can be useful.
4737 */
4738restart:
4739 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
4740 n = tcp_skb_next(skb, list);
4741
4742 /* No new bits? It is possible on ofo queue. */
4743 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4744 skb = tcp_collapse_one(sk, skb, list, root);
4745 if (!skb)
4746 break;
4747 goto restart;
4748 }
4749
4750 /* The first skb to collapse is:
4751 * - not SYN/FIN and
4752 * - bloated or contains data before "start" or
4753 * overlaps to the next one.
4754 */
4755 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
4756 (tcp_win_from_space(sk, skb->truesize) > skb->len ||
4757 before(TCP_SKB_CB(skb)->seq, start))) {
4758 end_of_skbs = false;
4759 break;
4760 }
4761
4762 if (n && n != tail &&
4763 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
4764 end_of_skbs = false;
4765 break;
4766 }
4767
4768 /* Decided to skip this, advance start seq. */
4769 start = TCP_SKB_CB(skb)->end_seq;
4770 }
4771 if (end_of_skbs ||
4772 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4773 return;
4774
4775 __skb_queue_head_init(&tmp);
4776
4777 while (before(start, end)) {
4778 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
4779 struct sk_buff *nskb;
4780
4781 nskb = alloc_skb(copy, GFP_ATOMIC);
4782 if (!nskb)
4783 break;
4784
4785 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4786 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4787 if (list)
4788 __skb_queue_before(list, skb, nskb);
4789 else
4790 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
4791 skb_set_owner_r(nskb, sk);
4792
4793 /* Copy data, releasing collapsed skbs. */
4794 while (copy > 0) {
4795 int offset = start - TCP_SKB_CB(skb)->seq;
4796 int size = TCP_SKB_CB(skb)->end_seq - start;
4797
4798 BUG_ON(offset < 0);
4799 if (size > 0) {
4800 size = min(copy, size);
4801 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4802 BUG();
4803 TCP_SKB_CB(nskb)->end_seq += size;
4804 copy -= size;
4805 start += size;
4806 }
4807 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4808 skb = tcp_collapse_one(sk, skb, list, root);
4809 if (!skb ||
4810 skb == tail ||
4811 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4812 goto end;
4813 }
4814 }
4815 }
4816end:
4817 skb_queue_walk_safe(&tmp, skb, n)
4818 tcp_rbtree_insert(root, skb);
4819}
4820
4821/* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4822 * and tcp_collapse() them until all the queue is collapsed.
4823 */
4824static void tcp_collapse_ofo_queue(struct sock *sk)
4825{
4826 struct tcp_sock *tp = tcp_sk(sk);
4827 struct sk_buff *skb, *head;
4828 u32 start, end;
4829
4830 skb = skb_rb_first(&tp->out_of_order_queue);
4831new_range:
4832 if (!skb) {
4833 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue);
4834 return;
4835 }
4836 start = TCP_SKB_CB(skb)->seq;
4837 end = TCP_SKB_CB(skb)->end_seq;
4838
4839 for (head = skb;;) {
4840 skb = skb_rb_next(skb);
4841
4842 /* Range is terminated when we see a gap or when
4843 * we are at the queue end.
4844 */
4845 if (!skb ||
4846 after(TCP_SKB_CB(skb)->seq, end) ||
4847 before(TCP_SKB_CB(skb)->end_seq, start)) {
4848 tcp_collapse(sk, NULL, &tp->out_of_order_queue,
4849 head, skb, start, end);
4850 goto new_range;
4851 }
4852
4853 if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
4854 start = TCP_SKB_CB(skb)->seq;
4855 if (after(TCP_SKB_CB(skb)->end_seq, end))
4856 end = TCP_SKB_CB(skb)->end_seq;
4857 }
4858}
4859
4860/*
4861 * Clean the out-of-order queue to make room.
4862 * We drop high sequences packets to :
4863 * 1) Let a chance for holes to be filled.
4864 * 2) not add too big latencies if thousands of packets sit there.
4865 * (But if application shrinks SO_RCVBUF, we could still end up
4866 * freeing whole queue here)
4867 *
4868 * Return true if queue has shrunk.
4869 */
4870static bool tcp_prune_ofo_queue(struct sock *sk)
4871{
4872 struct tcp_sock *tp = tcp_sk(sk);
4873 struct rb_node *node, *prev;
4874
4875 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
4876 return false;
4877
4878 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
4879 node = &tp->ooo_last_skb->rbnode;
4880 do {
4881 prev = rb_prev(node);
4882 rb_erase(node, &tp->out_of_order_queue);
4883 tcp_drop(sk, rb_to_skb(node));
4884 sk_mem_reclaim(sk);
4885 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
4886 !tcp_under_memory_pressure(sk))
4887 break;
4888 node = prev;
4889 } while (node);
4890 tp->ooo_last_skb = rb_to_skb(prev);
4891
4892 /* Reset SACK state. A conforming SACK implementation will
4893 * do the same at a timeout based retransmit. When a connection
4894 * is in a sad state like this, we care only about integrity
4895 * of the connection not performance.
4896 */
4897 if (tp->rx_opt.sack_ok)
4898 tcp_sack_reset(&tp->rx_opt);
4899 return true;
4900}
4901
4902/* Reduce allocated memory if we can, trying to get
4903 * the socket within its memory limits again.
4904 *
4905 * Return less than zero if we should start dropping frames
4906 * until the socket owning process reads some of the data
4907 * to stabilize the situation.
4908 */
4909static int tcp_prune_queue(struct sock *sk)
4910{
4911 struct tcp_sock *tp = tcp_sk(sk);
4912
4913 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4914
4915 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
4916
4917 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4918 tcp_clamp_window(sk);
4919 else if (tcp_under_memory_pressure(sk))
4920 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4921
4922 tcp_collapse_ofo_queue(sk);
4923 if (!skb_queue_empty(&sk->sk_receive_queue))
4924 tcp_collapse(sk, &sk->sk_receive_queue, NULL,
4925 skb_peek(&sk->sk_receive_queue),
4926 NULL,
4927 tp->copied_seq, tp->rcv_nxt);
4928 sk_mem_reclaim(sk);
4929
4930 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4931 return 0;
4932
4933 /* Collapsing did not help, destructive actions follow.
4934 * This must not ever occur. */
4935
4936 tcp_prune_ofo_queue(sk);
4937
4938 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4939 return 0;
4940
4941 /* If we are really being abused, tell the caller to silently
4942 * drop receive data on the floor. It will get retransmitted
4943 * and hopefully then we'll have sufficient space.
4944 */
4945 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
4946
4947 /* Massive buffer overcommit. */
4948 tp->pred_flags = 0;
4949 return -1;
4950}
4951
4952static bool tcp_should_expand_sndbuf(const struct sock *sk)
4953{
4954 const struct tcp_sock *tp = tcp_sk(sk);
4955
4956 /* If the user specified a specific send buffer setting, do
4957 * not modify it.
4958 */
4959 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4960 return false;
4961
4962 /* If we are under global TCP memory pressure, do not expand. */
4963 if (tcp_under_memory_pressure(sk))
4964 return false;
4965
4966 /* If we are under soft global TCP memory pressure, do not expand. */
4967 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
4968 return false;
4969
4970 /* If we filled the congestion window, do not expand. */
4971 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
4972 return false;
4973
4974 return true;
4975}
4976
4977/* When incoming ACK allowed to free some skb from write_queue,
4978 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4979 * on the exit from tcp input handler.
4980 *
4981 * PROBLEM: sndbuf expansion does not work well with largesend.
4982 */
4983static void tcp_new_space(struct sock *sk)
4984{
4985 struct tcp_sock *tp = tcp_sk(sk);
4986
4987 if (tcp_should_expand_sndbuf(sk)) {
4988 tcp_sndbuf_expand(sk);
4989 tp->snd_cwnd_stamp = tcp_jiffies32;
4990 }
4991
4992 sk->sk_write_space(sk);
4993}
4994
4995static void tcp_check_space(struct sock *sk)
4996{
4997 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4998 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4999 /* pairs with tcp_poll() */
5000 smp_mb();
5001 if (sk->sk_socket &&
5002 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
5003 tcp_new_space(sk);
5004 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5005 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
5006 }
5007 }
5008}
5009
5010static inline void tcp_data_snd_check(struct sock *sk)
5011{
5012 tcp_push_pending_frames(sk);
5013 tcp_check_space(sk);
5014}
5015
5016/*
5017 * Check if sending an ack is needed.
5018 */
5019static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5020{
5021 struct tcp_sock *tp = tcp_sk(sk);
5022
5023 /* More than one full frame received... */
5024 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5025 /* ... and right edge of window advances far enough.
5026 * (tcp_recvmsg() will send ACK otherwise). Or...
5027 */
5028 __tcp_select_window(sk) >= tp->rcv_wnd) ||
5029 /* We ACK each frame or... */
5030 tcp_in_quickack_mode(sk) ||
5031 /* We have out of order data. */
5032 (ofo_possible && !RB_EMPTY_ROOT(&tp->out_of_order_queue))) {
5033 /* Then ack it now */
5034 tcp_send_ack(sk);
5035 } else {
5036 /* Else, send delayed ack. */
5037 tcp_send_delayed_ack(sk);
5038 }
5039}
5040
5041static inline void tcp_ack_snd_check(struct sock *sk)
5042{
5043 if (!inet_csk_ack_scheduled(sk)) {
5044 /* We sent a data segment already. */
5045 return;
5046 }
5047 __tcp_ack_snd_check(sk, 1);
5048}
5049
5050/*
5051 * This routine is only called when we have urgent data
5052 * signaled. Its the 'slow' part of tcp_urg. It could be
5053 * moved inline now as tcp_urg is only called from one
5054 * place. We handle URGent data wrong. We have to - as
5055 * BSD still doesn't use the correction from RFC961.
5056 * For 1003.1g we should support a new option TCP_STDURG to permit
5057 * either form (or just set the sysctl tcp_stdurg).
5058 */
5059
5060static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5061{
5062 struct tcp_sock *tp = tcp_sk(sk);
5063 u32 ptr = ntohs(th->urg_ptr);
5064
5065 if (ptr && !sock_net(sk)->ipv4.sysctl_tcp_stdurg)
5066 ptr--;
5067 ptr += ntohl(th->seq);
5068
5069 /* Ignore urgent data that we've already seen and read. */
5070 if (after(tp->copied_seq, ptr))
5071 return;
5072
5073 /* Do not replay urg ptr.
5074 *
5075 * NOTE: interesting situation not covered by specs.
5076 * Misbehaving sender may send urg ptr, pointing to segment,
5077 * which we already have in ofo queue. We are not able to fetch
5078 * such data and will stay in TCP_URG_NOTYET until will be eaten
5079 * by recvmsg(). Seems, we are not obliged to handle such wicked
5080 * situations. But it is worth to think about possibility of some
5081 * DoSes using some hypothetical application level deadlock.
5082 */
5083 if (before(ptr, tp->rcv_nxt))
5084 return;
5085
5086 /* Do we already have a newer (or duplicate) urgent pointer? */
5087 if (tp->urg_data && !after(ptr, tp->urg_seq))
5088 return;
5089
5090 /* Tell the world about our new urgent pointer. */
5091 sk_send_sigurg(sk);
5092
5093 /* We may be adding urgent data when the last byte read was
5094 * urgent. To do this requires some care. We cannot just ignore
5095 * tp->copied_seq since we would read the last urgent byte again
5096 * as data, nor can we alter copied_seq until this data arrives
5097 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5098 *
5099 * NOTE. Double Dutch. Rendering to plain English: author of comment
5100 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5101 * and expect that both A and B disappear from stream. This is _wrong_.
5102 * Though this happens in BSD with high probability, this is occasional.
5103 * Any application relying on this is buggy. Note also, that fix "works"
5104 * only in this artificial test. Insert some normal data between A and B and we will
5105 * decline of BSD again. Verdict: it is better to remove to trap
5106 * buggy users.
5107 */
5108 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5109 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5110 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5111 tp->copied_seq++;
5112 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5113 __skb_unlink(skb, &sk->sk_receive_queue);
5114 __kfree_skb(skb);
5115 }
5116 }
5117
5118 tp->urg_data = TCP_URG_NOTYET;
5119 tp->urg_seq = ptr;
5120
5121 /* Disable header prediction. */
5122 tp->pred_flags = 0;
5123}
5124
5125/* This is the 'fast' part of urgent handling. */
5126static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5127{
5128 struct tcp_sock *tp = tcp_sk(sk);
5129
5130 /* Check if we get a new urgent pointer - normally not. */
5131 if (th->urg)
5132 tcp_check_urg(sk, th);
5133
5134 /* Do we wait for any urgent data? - normally not... */
5135 if (tp->urg_data == TCP_URG_NOTYET) {
5136 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5137 th->syn;
5138
5139 /* Is the urgent pointer pointing into this packet? */
5140 if (ptr < skb->len) {
5141 u8 tmp;
5142 if (skb_copy_bits(skb, ptr, &tmp, 1))
5143 BUG();
5144 tp->urg_data = TCP_URG_VALID | tmp;
5145 if (!sock_flag(sk, SOCK_DEAD))
5146 sk->sk_data_ready(sk);
5147 }
5148 }
5149}
5150
5151/* Accept RST for rcv_nxt - 1 after a FIN.
5152 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5153 * FIN is sent followed by a RST packet. The RST is sent with the same
5154 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5155 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5156 * ACKs on the closed socket. In addition middleboxes can drop either the
5157 * challenge ACK or a subsequent RST.
5158 */
5159static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb)
5160{
5161 struct tcp_sock *tp = tcp_sk(sk);
5162
5163 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) &&
5164 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK |
5165 TCPF_CLOSING));
5166}
5167
5168/* Does PAWS and seqno based validation of an incoming segment, flags will
5169 * play significant role here.
5170 */
5171static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5172 const struct tcphdr *th, int syn_inerr)
5173{
5174 struct tcp_sock *tp = tcp_sk(sk);
5175 bool rst_seq_match = false;
5176
5177 /* RFC1323: H1. Apply PAWS check first. */
5178 if (tcp_fast_parse_options(sock_net(sk), skb, th, tp) &&
5179 tp->rx_opt.saw_tstamp &&
5180 tcp_paws_discard(sk, skb)) {
5181 if (!th->rst) {
5182 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5183 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5184 LINUX_MIB_TCPACKSKIPPEDPAWS,
5185 &tp->last_oow_ack_time))
5186 tcp_send_dupack(sk, skb);
5187 goto discard;
5188 }
5189 /* Reset is accepted even if it did not pass PAWS. */
5190 }
5191
5192 /* Step 1: check sequence number */
5193 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5194 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5195 * (RST) segments are validated by checking their SEQ-fields."
5196 * And page 69: "If an incoming segment is not acceptable,
5197 * an acknowledgment should be sent in reply (unless the RST
5198 * bit is set, if so drop the segment and return)".
5199 */
5200 if (!th->rst) {
5201 if (th->syn)
5202 goto syn_challenge;
5203 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5204 LINUX_MIB_TCPACKSKIPPEDSEQ,
5205 &tp->last_oow_ack_time))
5206 tcp_send_dupack(sk, skb);
5207 } else if (tcp_reset_check(sk, skb)) {
5208 tcp_reset(sk);
5209 }
5210 goto discard;
5211 }
5212
5213 /* Step 2: check RST bit */
5214 if (th->rst) {
5215 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5216 * FIN and SACK too if available):
5217 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5218 * the right-most SACK block,
5219 * then
5220 * RESET the connection
5221 * else
5222 * Send a challenge ACK
5223 */
5224 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt ||
5225 tcp_reset_check(sk, skb)) {
5226 rst_seq_match = true;
5227 } else if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
5228 struct tcp_sack_block *sp = &tp->selective_acks[0];
5229 int max_sack = sp[0].end_seq;
5230 int this_sack;
5231
5232 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
5233 ++this_sack) {
5234 max_sack = after(sp[this_sack].end_seq,
5235 max_sack) ?
5236 sp[this_sack].end_seq : max_sack;
5237 }
5238
5239 if (TCP_SKB_CB(skb)->seq == max_sack)
5240 rst_seq_match = true;
5241 }
5242
5243 if (rst_seq_match)
5244 tcp_reset(sk);
5245 else {
5246 /* Disable TFO if RST is out-of-order
5247 * and no data has been received
5248 * for current active TFO socket
5249 */
5250 if (tp->syn_fastopen && !tp->data_segs_in &&
5251 sk->sk_state == TCP_ESTABLISHED)
5252 tcp_fastopen_active_disable(sk);
5253 tcp_send_challenge_ack(sk, skb);
5254 }
5255 goto discard;
5256 }
5257
5258 /* step 3: check security and precedence [ignored] */
5259
5260 /* step 4: Check for a SYN
5261 * RFC 5961 4.2 : Send a challenge ack
5262 */
5263 if (th->syn) {
5264syn_challenge:
5265 if (syn_inerr)
5266 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5267 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5268 tcp_send_challenge_ack(sk, skb);
5269 goto discard;
5270 }
5271
5272 return true;
5273
5274discard:
5275 tcp_drop(sk, skb);
5276 return false;
5277}
5278
5279/*
5280 * TCP receive function for the ESTABLISHED state.
5281 *
5282 * It is split into a fast path and a slow path. The fast path is
5283 * disabled when:
5284 * - A zero window was announced from us - zero window probing
5285 * is only handled properly in the slow path.
5286 * - Out of order segments arrived.
5287 * - Urgent data is expected.
5288 * - There is no buffer space left
5289 * - Unexpected TCP flags/window values/header lengths are received
5290 * (detected by checking the TCP header against pred_flags)
5291 * - Data is sent in both directions. Fast path only supports pure senders
5292 * or pure receivers (this means either the sequence number or the ack
5293 * value must stay constant)
5294 * - Unexpected TCP option.
5295 *
5296 * When these conditions are not satisfied it drops into a standard
5297 * receive procedure patterned after RFC793 to handle all cases.
5298 * The first three cases are guaranteed by proper pred_flags setting,
5299 * the rest is checked inline. Fast processing is turned on in
5300 * tcp_data_queue when everything is OK.
5301 */
5302void tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5303 const struct tcphdr *th)
5304{
5305 unsigned int len = skb->len;
5306 struct tcp_sock *tp = tcp_sk(sk);
5307
5308 /* TCP congestion window tracking */
5309 trace_tcp_probe(sk, skb);
5310
5311 tcp_mstamp_refresh(tp);
5312 if (unlikely(!sk->sk_rx_dst))
5313 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5314 /*
5315 * Header prediction.
5316 * The code loosely follows the one in the famous
5317 * "30 instruction TCP receive" Van Jacobson mail.
5318 *
5319 * Van's trick is to deposit buffers into socket queue
5320 * on a device interrupt, to call tcp_recv function
5321 * on the receive process context and checksum and copy
5322 * the buffer to user space. smart...
5323 *
5324 * Our current scheme is not silly either but we take the
5325 * extra cost of the net_bh soft interrupt processing...
5326 * We do checksum and copy also but from device to kernel.
5327 */
5328
5329 tp->rx_opt.saw_tstamp = 0;
5330
5331 /* pred_flags is 0xS?10 << 16 + snd_wnd
5332 * if header_prediction is to be made
5333 * 'S' will always be tp->tcp_header_len >> 2
5334 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5335 * turn it off (when there are holes in the receive
5336 * space for instance)
5337 * PSH flag is ignored.
5338 */
5339
5340 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5341 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5342 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5343 int tcp_header_len = tp->tcp_header_len;
5344
5345 /* Timestamp header prediction: tcp_header_len
5346 * is automatically equal to th->doff*4 due to pred_flags
5347 * match.
5348 */
5349
5350 /* Check timestamp */
5351 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5352 /* No? Slow path! */
5353 if (!tcp_parse_aligned_timestamp(tp, th))
5354 goto slow_path;
5355
5356 /* If PAWS failed, check it more carefully in slow path */
5357 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5358 goto slow_path;
5359
5360 /* DO NOT update ts_recent here, if checksum fails
5361 * and timestamp was corrupted part, it will result
5362 * in a hung connection since we will drop all
5363 * future packets due to the PAWS test.
5364 */
5365 }
5366
5367 if (len <= tcp_header_len) {
5368 /* Bulk data transfer: sender */
5369 if (len == tcp_header_len) {
5370 /* Predicted packet is in window by definition.
5371 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5372 * Hence, check seq<=rcv_wup reduces to:
5373 */
5374 if (tcp_header_len ==
5375 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5376 tp->rcv_nxt == tp->rcv_wup)
5377 tcp_store_ts_recent(tp);
5378
5379 /* We know that such packets are checksummed
5380 * on entry.
5381 */
5382 tcp_ack(sk, skb, 0);
5383 __kfree_skb(skb);
5384 tcp_data_snd_check(sk);
5385 return;
5386 } else { /* Header too small */
5387 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5388 goto discard;
5389 }
5390 } else {
5391 int eaten = 0;
5392 bool fragstolen = false;
5393
5394 if (tcp_checksum_complete(skb))
5395 goto csum_error;
5396
5397 if ((int)skb->truesize > sk->sk_forward_alloc)
5398 goto step5;
5399
5400 /* Predicted packet is in window by definition.
5401 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5402 * Hence, check seq<=rcv_wup reduces to:
5403 */
5404 if (tcp_header_len ==
5405 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5406 tp->rcv_nxt == tp->rcv_wup)
5407 tcp_store_ts_recent(tp);
5408
5409 tcp_rcv_rtt_measure_ts(sk, skb);
5410
5411 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
5412
5413 /* Bulk data transfer: receiver */
5414 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5415 &fragstolen);
5416
5417 tcp_event_data_recv(sk, skb);
5418
5419 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5420 /* Well, only one small jumplet in fast path... */
5421 tcp_ack(sk, skb, FLAG_DATA);
5422 tcp_data_snd_check(sk);
5423 if (!inet_csk_ack_scheduled(sk))
5424 goto no_ack;
5425 }
5426
5427 __tcp_ack_snd_check(sk, 0);
5428no_ack:
5429 if (eaten)
5430 kfree_skb_partial(skb, fragstolen);
5431 sk->sk_data_ready(sk);
5432 return;
5433 }
5434 }
5435
5436slow_path:
5437 if (len < (th->doff << 2) || tcp_checksum_complete(skb))
5438 goto csum_error;
5439
5440 if (!th->ack && !th->rst && !th->syn)
5441 goto discard;
5442
5443 /*
5444 * Standard slow path.
5445 */
5446
5447 if (!tcp_validate_incoming(sk, skb, th, 1))
5448 return;
5449
5450step5:
5451 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5452 goto discard;
5453
5454 tcp_rcv_rtt_measure_ts(sk, skb);
5455
5456 /* Process urgent data. */
5457 tcp_urg(sk, skb, th);
5458
5459 /* step 7: process the segment text */
5460 tcp_data_queue(sk, skb);
5461
5462 tcp_data_snd_check(sk);
5463 tcp_ack_snd_check(sk);
5464 return;
5465
5466csum_error:
5467 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
5468 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5469
5470discard:
5471 tcp_drop(sk, skb);
5472}
5473EXPORT_SYMBOL(tcp_rcv_established);
5474
5475void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5476{
5477 struct tcp_sock *tp = tcp_sk(sk);
5478 struct inet_connection_sock *icsk = inet_csk(sk);
5479
5480 tcp_set_state(sk, TCP_ESTABLISHED);
5481 icsk->icsk_ack.lrcvtime = tcp_jiffies32;
5482
5483 if (skb) {
5484 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5485 security_inet_conn_established(sk, skb);
5486 }
5487
5488 tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB);
5489
5490 /* Prevent spurious tcp_cwnd_restart() on first data
5491 * packet.
5492 */
5493 tp->lsndtime = tcp_jiffies32;
5494
5495 if (sock_flag(sk, SOCK_KEEPOPEN))
5496 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5497
5498 if (!tp->rx_opt.snd_wscale)
5499 __tcp_fast_path_on(tp, tp->snd_wnd);
5500 else
5501 tp->pred_flags = 0;
5502}
5503
5504static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5505 struct tcp_fastopen_cookie *cookie)
5506{
5507 struct tcp_sock *tp = tcp_sk(sk);
5508 struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL;
5509 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
5510 bool syn_drop = false;
5511
5512 if (mss == tp->rx_opt.user_mss) {
5513 struct tcp_options_received opt;
5514
5515 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5516 tcp_clear_options(&opt);
5517 opt.user_mss = opt.mss_clamp = 0;
5518 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL);
5519 mss = opt.mss_clamp;
5520 }
5521
5522 if (!tp->syn_fastopen) {
5523 /* Ignore an unsolicited cookie */
5524 cookie->len = -1;
5525 } else if (tp->total_retrans) {
5526 /* SYN timed out and the SYN-ACK neither has a cookie nor
5527 * acknowledges data. Presumably the remote received only
5528 * the retransmitted (regular) SYNs: either the original
5529 * SYN-data or the corresponding SYN-ACK was dropped.
5530 */
5531 syn_drop = (cookie->len < 0 && data);
5532 } else if (cookie->len < 0 && !tp->syn_data) {
5533 /* We requested a cookie but didn't get it. If we did not use
5534 * the (old) exp opt format then try so next time (try_exp=1).
5535 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5536 */
5537 try_exp = tp->syn_fastopen_exp ? 2 : 1;
5538 }
5539
5540 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
5541
5542 if (data) { /* Retransmit unacked data in SYN */
5543 skb_rbtree_walk_from(data) {
5544 if (__tcp_retransmit_skb(sk, data, 1))
5545 break;
5546 }
5547 tcp_rearm_rto(sk);
5548 NET_INC_STATS(sock_net(sk),
5549 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
5550 return true;
5551 }
5552 tp->syn_data_acked = tp->syn_data;
5553 if (tp->syn_data_acked)
5554 NET_INC_STATS(sock_net(sk),
5555 LINUX_MIB_TCPFASTOPENACTIVE);
5556
5557 tcp_fastopen_add_skb(sk, synack);
5558
5559 return false;
5560}
5561
5562static void smc_check_reset_syn(struct tcp_sock *tp)
5563{
5564#if IS_ENABLED(CONFIG_SMC)
5565 if (static_branch_unlikely(&tcp_have_smc)) {
5566 if (tp->syn_smc && !tp->rx_opt.smc_ok)
5567 tp->syn_smc = 0;
5568 }
5569#endif
5570}
5571
5572static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5573 const struct tcphdr *th)
5574{
5575 struct inet_connection_sock *icsk = inet_csk(sk);
5576 struct tcp_sock *tp = tcp_sk(sk);
5577 struct tcp_fastopen_cookie foc = { .len = -1 };
5578 int saved_clamp = tp->rx_opt.mss_clamp;
5579 bool fastopen_fail;
5580
5581 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc);
5582 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5583 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5584
5585 if (th->ack) {
5586 /* rfc793:
5587 * "If the state is SYN-SENT then
5588 * first check the ACK bit
5589 * If the ACK bit is set
5590 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5591 * a reset (unless the RST bit is set, if so drop
5592 * the segment and return)"
5593 */
5594 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5595 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5596 goto reset_and_undo;
5597
5598 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5599 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5600 tcp_time_stamp(tp))) {
5601 NET_INC_STATS(sock_net(sk),
5602 LINUX_MIB_PAWSACTIVEREJECTED);
5603 goto reset_and_undo;
5604 }
5605
5606 /* Now ACK is acceptable.
5607 *
5608 * "If the RST bit is set
5609 * If the ACK was acceptable then signal the user "error:
5610 * connection reset", drop the segment, enter CLOSED state,
5611 * delete TCB, and return."
5612 */
5613
5614 if (th->rst) {
5615 tcp_reset(sk);
5616 goto discard;
5617 }
5618
5619 /* rfc793:
5620 * "fifth, if neither of the SYN or RST bits is set then
5621 * drop the segment and return."
5622 *
5623 * See note below!
5624 * --ANK(990513)
5625 */
5626 if (!th->syn)
5627 goto discard_and_undo;
5628
5629 /* rfc793:
5630 * "If the SYN bit is on ...
5631 * are acceptable then ...
5632 * (our SYN has been ACKed), change the connection
5633 * state to ESTABLISHED..."
5634 */
5635
5636 tcp_ecn_rcv_synack(tp, th);
5637
5638 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5639 tcp_ack(sk, skb, FLAG_SLOWPATH);
5640
5641 /* Ok.. it's good. Set up sequence numbers and
5642 * move to established.
5643 */
5644 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5645 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5646
5647 /* RFC1323: The window in SYN & SYN/ACK segments is
5648 * never scaled.
5649 */
5650 tp->snd_wnd = ntohs(th->window);
5651
5652 if (!tp->rx_opt.wscale_ok) {
5653 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5654 tp->window_clamp = min(tp->window_clamp, 65535U);
5655 }
5656
5657 if (tp->rx_opt.saw_tstamp) {
5658 tp->rx_opt.tstamp_ok = 1;
5659 tp->tcp_header_len =
5660 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5661 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5662 tcp_store_ts_recent(tp);
5663 } else {
5664 tp->tcp_header_len = sizeof(struct tcphdr);
5665 }
5666
5667 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5668 tcp_initialize_rcv_mss(sk);
5669
5670 /* Remember, tcp_poll() does not lock socket!
5671 * Change state from SYN-SENT only after copied_seq
5672 * is initialized. */
5673 tp->copied_seq = tp->rcv_nxt;
5674
5675 smc_check_reset_syn(tp);
5676
5677 smp_mb();
5678
5679 tcp_finish_connect(sk, skb);
5680
5681 fastopen_fail = (tp->syn_fastopen || tp->syn_data) &&
5682 tcp_rcv_fastopen_synack(sk, skb, &foc);
5683
5684 if (!sock_flag(sk, SOCK_DEAD)) {
5685 sk->sk_state_change(sk);
5686 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5687 }
5688 if (fastopen_fail)
5689 return -1;
5690 if (sk->sk_write_pending ||
5691 icsk->icsk_accept_queue.rskq_defer_accept ||
5692 icsk->icsk_ack.pingpong) {
5693 /* Save one ACK. Data will be ready after
5694 * several ticks, if write_pending is set.
5695 *
5696 * It may be deleted, but with this feature tcpdumps
5697 * look so _wonderfully_ clever, that I was not able
5698 * to stand against the temptation 8) --ANK
5699 */
5700 inet_csk_schedule_ack(sk);
5701 tcp_enter_quickack_mode(sk);
5702 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5703 TCP_DELACK_MAX, TCP_RTO_MAX);
5704
5705discard:
5706 tcp_drop(sk, skb);
5707 return 0;
5708 } else {
5709 tcp_send_ack(sk);
5710 }
5711 return -1;
5712 }
5713
5714 /* No ACK in the segment */
5715
5716 if (th->rst) {
5717 /* rfc793:
5718 * "If the RST bit is set
5719 *
5720 * Otherwise (no ACK) drop the segment and return."
5721 */
5722
5723 goto discard_and_undo;
5724 }
5725
5726 /* PAWS check. */
5727 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5728 tcp_paws_reject(&tp->rx_opt, 0))
5729 goto discard_and_undo;
5730
5731 if (th->syn) {
5732 /* We see SYN without ACK. It is attempt of
5733 * simultaneous connect with crossed SYNs.
5734 * Particularly, it can be connect to self.
5735 */
5736 tcp_set_state(sk, TCP_SYN_RECV);
5737
5738 if (tp->rx_opt.saw_tstamp) {
5739 tp->rx_opt.tstamp_ok = 1;
5740 tcp_store_ts_recent(tp);
5741 tp->tcp_header_len =
5742 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5743 } else {
5744 tp->tcp_header_len = sizeof(struct tcphdr);
5745 }
5746
5747 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5748 tp->copied_seq = tp->rcv_nxt;
5749 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5750
5751 /* RFC1323: The window in SYN & SYN/ACK segments is
5752 * never scaled.
5753 */
5754 tp->snd_wnd = ntohs(th->window);
5755 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5756 tp->max_window = tp->snd_wnd;
5757
5758 tcp_ecn_rcv_syn(tp, th);
5759
5760 tcp_mtup_init(sk);
5761 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5762 tcp_initialize_rcv_mss(sk);
5763
5764 tcp_send_synack(sk);
5765#if 0
5766 /* Note, we could accept data and URG from this segment.
5767 * There are no obstacles to make this (except that we must
5768 * either change tcp_recvmsg() to prevent it from returning data
5769 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5770 *
5771 * However, if we ignore data in ACKless segments sometimes,
5772 * we have no reasons to accept it sometimes.
5773 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5774 * is not flawless. So, discard packet for sanity.
5775 * Uncomment this return to process the data.
5776 */
5777 return -1;
5778#else
5779 goto discard;
5780#endif
5781 }
5782 /* "fifth, if neither of the SYN or RST bits is set then
5783 * drop the segment and return."
5784 */
5785
5786discard_and_undo:
5787 tcp_clear_options(&tp->rx_opt);
5788 tp->rx_opt.mss_clamp = saved_clamp;
5789 goto discard;
5790
5791reset_and_undo:
5792 tcp_clear_options(&tp->rx_opt);
5793 tp->rx_opt.mss_clamp = saved_clamp;
5794 return 1;
5795}
5796
5797/*
5798 * This function implements the receiving procedure of RFC 793 for
5799 * all states except ESTABLISHED and TIME_WAIT.
5800 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5801 * address independent.
5802 */
5803
5804int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
5805{
5806 struct tcp_sock *tp = tcp_sk(sk);
5807 struct inet_connection_sock *icsk = inet_csk(sk);
5808 const struct tcphdr *th = tcp_hdr(skb);
5809 struct request_sock *req;
5810 int queued = 0;
5811 bool acceptable;
5812
5813 switch (sk->sk_state) {
5814 case TCP_CLOSE:
5815 goto discard;
5816
5817 case TCP_LISTEN:
5818 if (th->ack)
5819 return 1;
5820
5821 if (th->rst)
5822 goto discard;
5823
5824 if (th->syn) {
5825 if (th->fin)
5826 goto discard;
5827 /* It is possible that we process SYN packets from backlog,
5828 * so we need to make sure to disable BH right there.
5829 */
5830 local_bh_disable();
5831 acceptable = icsk->icsk_af_ops->conn_request(sk, skb) >= 0;
5832 local_bh_enable();
5833
5834 if (!acceptable)
5835 return 1;
5836 consume_skb(skb);
5837 return 0;
5838 }
5839 goto discard;
5840
5841 case TCP_SYN_SENT:
5842 tp->rx_opt.saw_tstamp = 0;
5843 tcp_mstamp_refresh(tp);
5844 queued = tcp_rcv_synsent_state_process(sk, skb, th);
5845 if (queued >= 0)
5846 return queued;
5847
5848 /* Do step6 onward by hand. */
5849 tcp_urg(sk, skb, th);
5850 __kfree_skb(skb);
5851 tcp_data_snd_check(sk);
5852 return 0;
5853 }
5854
5855 tcp_mstamp_refresh(tp);
5856 tp->rx_opt.saw_tstamp = 0;
5857 req = tp->fastopen_rsk;
5858 if (req) {
5859 bool req_stolen;
5860
5861 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
5862 sk->sk_state != TCP_FIN_WAIT1);
5863
5864 if (!tcp_check_req(sk, skb, req, true, &req_stolen))
5865 goto discard;
5866 }
5867
5868 if (!th->ack && !th->rst && !th->syn)
5869 goto discard;
5870
5871 if (!tcp_validate_incoming(sk, skb, th, 0))
5872 return 0;
5873
5874 /* step 5: check the ACK field */
5875 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
5876 FLAG_UPDATE_TS_RECENT |
5877 FLAG_NO_CHALLENGE_ACK) > 0;
5878
5879 if (!acceptable) {
5880 if (sk->sk_state == TCP_SYN_RECV)
5881 return 1; /* send one RST */
5882 tcp_send_challenge_ack(sk, skb);
5883 goto discard;
5884 }
5885 switch (sk->sk_state) {
5886 case TCP_SYN_RECV:
5887 if (!tp->srtt_us)
5888 tcp_synack_rtt_meas(sk, req);
5889
5890 /* Once we leave TCP_SYN_RECV, we no longer need req
5891 * so release it.
5892 */
5893 if (req) {
5894 inet_csk(sk)->icsk_retransmits = 0;
5895 reqsk_fastopen_remove(sk, req, false);
5896 /* Re-arm the timer because data may have been sent out.
5897 * This is similar to the regular data transmission case
5898 * when new data has just been ack'ed.
5899 *
5900 * (TFO) - we could try to be more aggressive and
5901 * retransmitting any data sooner based on when they
5902 * are sent out.
5903 */
5904 tcp_rearm_rto(sk);
5905 } else {
5906 tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB);
5907 tp->copied_seq = tp->rcv_nxt;
5908 }
5909 smp_mb();
5910 tcp_set_state(sk, TCP_ESTABLISHED);
5911 sk->sk_state_change(sk);
5912
5913 /* Note, that this wakeup is only for marginal crossed SYN case.
5914 * Passively open sockets are not waked up, because
5915 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5916 */
5917 if (sk->sk_socket)
5918 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5919
5920 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5921 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
5922 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5923
5924 if (tp->rx_opt.tstamp_ok)
5925 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5926
5927 if (!inet_csk(sk)->icsk_ca_ops->cong_control)
5928 tcp_update_pacing_rate(sk);
5929
5930 /* Prevent spurious tcp_cwnd_restart() on first data packet */
5931 tp->lsndtime = tcp_jiffies32;
5932
5933 tcp_initialize_rcv_mss(sk);
5934 tcp_fast_path_on(tp);
5935 break;
5936
5937 case TCP_FIN_WAIT1: {
5938 int tmo;
5939
5940 /* If we enter the TCP_FIN_WAIT1 state and we are a
5941 * Fast Open socket and this is the first acceptable
5942 * ACK we have received, this would have acknowledged
5943 * our SYNACK so stop the SYNACK timer.
5944 */
5945 if (req) {
5946 /* We no longer need the request sock. */
5947 reqsk_fastopen_remove(sk, req, false);
5948 tcp_rearm_rto(sk);
5949 }
5950 if (tp->snd_una != tp->write_seq)
5951 break;
5952
5953 tcp_set_state(sk, TCP_FIN_WAIT2);
5954 sk->sk_shutdown |= SEND_SHUTDOWN;
5955
5956 sk_dst_confirm(sk);
5957
5958 if (!sock_flag(sk, SOCK_DEAD)) {
5959 /* Wake up lingering close() */
5960 sk->sk_state_change(sk);
5961 break;
5962 }
5963
5964 if (tp->linger2 < 0) {
5965 tcp_done(sk);
5966 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5967 return 1;
5968 }
5969 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5970 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5971 /* Receive out of order FIN after close() */
5972 if (tp->syn_fastopen && th->fin)
5973 tcp_fastopen_active_disable(sk);
5974 tcp_done(sk);
5975 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5976 return 1;
5977 }
5978
5979 tmo = tcp_fin_time(sk);
5980 if (tmo > TCP_TIMEWAIT_LEN) {
5981 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5982 } else if (th->fin || sock_owned_by_user(sk)) {
5983 /* Bad case. We could lose such FIN otherwise.
5984 * It is not a big problem, but it looks confusing
5985 * and not so rare event. We still can lose it now,
5986 * if it spins in bh_lock_sock(), but it is really
5987 * marginal case.
5988 */
5989 inet_csk_reset_keepalive_timer(sk, tmo);
5990 } else {
5991 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5992 goto discard;
5993 }
5994 break;
5995 }
5996
5997 case TCP_CLOSING:
5998 if (tp->snd_una == tp->write_seq) {
5999 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6000 goto discard;
6001 }
6002 break;
6003
6004 case TCP_LAST_ACK:
6005 if (tp->snd_una == tp->write_seq) {
6006 tcp_update_metrics(sk);
6007 tcp_done(sk);
6008 goto discard;
6009 }
6010 break;
6011 }
6012
6013 /* step 6: check the URG bit */
6014 tcp_urg(sk, skb, th);
6015
6016 /* step 7: process the segment text */
6017 switch (sk->sk_state) {
6018 case TCP_CLOSE_WAIT:
6019 case TCP_CLOSING:
6020 case TCP_LAST_ACK:
6021 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
6022 break;
6023 /* fall through */
6024 case TCP_FIN_WAIT1:
6025 case TCP_FIN_WAIT2:
6026 /* RFC 793 says to queue data in these states,
6027 * RFC 1122 says we MUST send a reset.
6028 * BSD 4.4 also does reset.
6029 */
6030 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6031 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6032 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6033 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6034 tcp_reset(sk);
6035 return 1;
6036 }
6037 }
6038 /* Fall through */
6039 case TCP_ESTABLISHED:
6040 tcp_data_queue(sk, skb);
6041 queued = 1;
6042 break;
6043 }
6044
6045 /* tcp_data could move socket to TIME-WAIT */
6046 if (sk->sk_state != TCP_CLOSE) {
6047 tcp_data_snd_check(sk);
6048 tcp_ack_snd_check(sk);
6049 }
6050
6051 if (!queued) {
6052discard:
6053 tcp_drop(sk, skb);
6054 }
6055 return 0;
6056}
6057EXPORT_SYMBOL(tcp_rcv_state_process);
6058
6059static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6060{
6061 struct inet_request_sock *ireq = inet_rsk(req);
6062
6063 if (family == AF_INET)
6064 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6065 &ireq->ir_rmt_addr, port);
6066#if IS_ENABLED(CONFIG_IPV6)
6067 else if (family == AF_INET6)
6068 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6069 &ireq->ir_v6_rmt_addr, port);
6070#endif
6071}
6072
6073/* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6074 *
6075 * If we receive a SYN packet with these bits set, it means a
6076 * network is playing bad games with TOS bits. In order to
6077 * avoid possible false congestion notifications, we disable
6078 * TCP ECN negotiation.
6079 *
6080 * Exception: tcp_ca wants ECN. This is required for DCTCP
6081 * congestion control: Linux DCTCP asserts ECT on all packets,
6082 * including SYN, which is most optimal solution; however,
6083 * others, such as FreeBSD do not.
6084 */
6085static void tcp_ecn_create_request(struct request_sock *req,
6086 const struct sk_buff *skb,
6087 const struct sock *listen_sk,
6088 const struct dst_entry *dst)
6089{
6090 const struct tcphdr *th = tcp_hdr(skb);
6091 const struct net *net = sock_net(listen_sk);
6092 bool th_ecn = th->ece && th->cwr;
6093 bool ect, ecn_ok;
6094 u32 ecn_ok_dst;
6095
6096 if (!th_ecn)
6097 return;
6098
6099 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
6100 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6101 ecn_ok = net->ipv4.sysctl_tcp_ecn || ecn_ok_dst;
6102
6103 if ((!ect && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
6104 (ecn_ok_dst & DST_FEATURE_ECN_CA) ||
6105 tcp_bpf_ca_needs_ecn((struct sock *)req))
6106 inet_rsk(req)->ecn_ok = 1;
6107}
6108
6109static void tcp_openreq_init(struct request_sock *req,
6110 const struct tcp_options_received *rx_opt,
6111 struct sk_buff *skb, const struct sock *sk)
6112{
6113 struct inet_request_sock *ireq = inet_rsk(req);
6114
6115 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
6116 req->cookie_ts = 0;
6117 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6118 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6119 tcp_rsk(req)->snt_synack = tcp_clock_us();
6120 tcp_rsk(req)->last_oow_ack_time = 0;
6121 req->mss = rx_opt->mss_clamp;
6122 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6123 ireq->tstamp_ok = rx_opt->tstamp_ok;
6124 ireq->sack_ok = rx_opt->sack_ok;
6125 ireq->snd_wscale = rx_opt->snd_wscale;
6126 ireq->wscale_ok = rx_opt->wscale_ok;
6127 ireq->acked = 0;
6128 ireq->ecn_ok = 0;
6129 ireq->ir_rmt_port = tcp_hdr(skb)->source;
6130 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6131 ireq->ir_mark = inet_request_mark(sk, skb);
6132#if IS_ENABLED(CONFIG_SMC)
6133 ireq->smc_ok = rx_opt->smc_ok;
6134#endif
6135}
6136
6137struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6138 struct sock *sk_listener,
6139 bool attach_listener)
6140{
6141 struct request_sock *req = reqsk_alloc(ops, sk_listener,
6142 attach_listener);
6143
6144 if (req) {
6145 struct inet_request_sock *ireq = inet_rsk(req);
6146
6147 ireq->ireq_opt = NULL;
6148#if IS_ENABLED(CONFIG_IPV6)
6149 ireq->pktopts = NULL;
6150#endif
6151 atomic64_set(&ireq->ir_cookie, 0);
6152 ireq->ireq_state = TCP_NEW_SYN_RECV;
6153 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
6154 ireq->ireq_family = sk_listener->sk_family;
6155 }
6156
6157 return req;
6158}
6159EXPORT_SYMBOL(inet_reqsk_alloc);
6160
6161/*
6162 * Return true if a syncookie should be sent
6163 */
6164static bool tcp_syn_flood_action(const struct sock *sk,
6165 const struct sk_buff *skb,
6166 const char *proto)
6167{
6168 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
6169 const char *msg = "Dropping request";
6170 bool want_cookie = false;
6171 struct net *net = sock_net(sk);
6172
6173#ifdef CONFIG_SYN_COOKIES
6174 if (net->ipv4.sysctl_tcp_syncookies) {
6175 msg = "Sending cookies";
6176 want_cookie = true;
6177 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
6178 } else
6179#endif
6180 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
6181
6182 if (!queue->synflood_warned &&
6183 net->ipv4.sysctl_tcp_syncookies != 2 &&
6184 xchg(&queue->synflood_warned, 1) == 0)
6185 pr_info("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6186 proto, ntohs(tcp_hdr(skb)->dest), msg);
6187
6188 return want_cookie;
6189}
6190
6191static void tcp_reqsk_record_syn(const struct sock *sk,
6192 struct request_sock *req,
6193 const struct sk_buff *skb)
6194{
6195 if (tcp_sk(sk)->save_syn) {
6196 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
6197 u32 *copy;
6198
6199 copy = kmalloc(len + sizeof(u32), GFP_ATOMIC);
6200 if (copy) {
6201 copy[0] = len;
6202 memcpy(©[1], skb_network_header(skb), len);
6203 req->saved_syn = copy;
6204 }
6205 }
6206}
6207
6208int tcp_conn_request(struct request_sock_ops *rsk_ops,
6209 const struct tcp_request_sock_ops *af_ops,
6210 struct sock *sk, struct sk_buff *skb)
6211{
6212 struct tcp_fastopen_cookie foc = { .len = -1 };
6213 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6214 struct tcp_options_received tmp_opt;
6215 struct tcp_sock *tp = tcp_sk(sk);
6216 struct net *net = sock_net(sk);
6217 struct sock *fastopen_sk = NULL;
6218 struct request_sock *req;
6219 bool want_cookie = false;
6220 struct dst_entry *dst;
6221 struct flowi fl;
6222
6223 /* TW buckets are converted to open requests without
6224 * limitations, they conserve resources and peer is
6225 * evidently real one.
6226 */
6227 if ((net->ipv4.sysctl_tcp_syncookies == 2 ||
6228 inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6229 want_cookie = tcp_syn_flood_action(sk, skb, rsk_ops->slab_name);
6230 if (!want_cookie)
6231 goto drop;
6232 }
6233
6234 if (sk_acceptq_is_full(sk)) {
6235 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6236 goto drop;
6237 }
6238
6239 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
6240 if (!req)
6241 goto drop;
6242
6243 tcp_rsk(req)->af_specific = af_ops;
6244 tcp_rsk(req)->ts_off = 0;
6245
6246 tcp_clear_options(&tmp_opt);
6247 tmp_opt.mss_clamp = af_ops->mss_clamp;
6248 tmp_opt.user_mss = tp->rx_opt.user_mss;
6249 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0,
6250 want_cookie ? NULL : &foc);
6251
6252 if (want_cookie && !tmp_opt.saw_tstamp)
6253 tcp_clear_options(&tmp_opt);
6254
6255 if (IS_ENABLED(CONFIG_SMC) && want_cookie)
6256 tmp_opt.smc_ok = 0;
6257
6258 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6259 tcp_openreq_init(req, &tmp_opt, skb, sk);
6260 inet_rsk(req)->no_srccheck = inet_sk(sk)->transparent;
6261
6262 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6263 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
6264
6265 af_ops->init_req(req, sk, skb);
6266
6267 if (security_inet_conn_request(sk, skb, req))
6268 goto drop_and_free;
6269
6270 if (tmp_opt.tstamp_ok)
6271 tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb);
6272
6273 dst = af_ops->route_req(sk, &fl, req);
6274 if (!dst)
6275 goto drop_and_free;
6276
6277 if (!want_cookie && !isn) {
6278 /* Kill the following clause, if you dislike this way. */
6279 if (!net->ipv4.sysctl_tcp_syncookies &&
6280 (net->ipv4.sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6281 (net->ipv4.sysctl_max_syn_backlog >> 2)) &&
6282 !tcp_peer_is_proven(req, dst)) {
6283 /* Without syncookies last quarter of
6284 * backlog is filled with destinations,
6285 * proven to be alive.
6286 * It means that we continue to communicate
6287 * to destinations, already remembered
6288 * to the moment of synflood.
6289 */
6290 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6291 rsk_ops->family);
6292 goto drop_and_release;
6293 }
6294
6295 isn = af_ops->init_seq(skb);
6296 }
6297
6298 tcp_ecn_create_request(req, skb, sk, dst);
6299
6300 if (want_cookie) {
6301 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6302 req->cookie_ts = tmp_opt.tstamp_ok;
6303 if (!tmp_opt.tstamp_ok)
6304 inet_rsk(req)->ecn_ok = 0;
6305 }
6306
6307 tcp_rsk(req)->snt_isn = isn;
6308 tcp_rsk(req)->txhash = net_tx_rndhash();
6309 tcp_openreq_init_rwin(req, sk, dst);
6310 if (!want_cookie) {
6311 tcp_reqsk_record_syn(sk, req, skb);
6312 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
6313 }
6314 if (fastopen_sk) {
6315 af_ops->send_synack(fastopen_sk, dst, &fl, req,
6316 &foc, TCP_SYNACK_FASTOPEN);
6317 /* Add the child socket directly into the accept queue */
6318 inet_csk_reqsk_queue_add(sk, req, fastopen_sk);
6319 sk->sk_data_ready(sk);
6320 bh_unlock_sock(fastopen_sk);
6321 sock_put(fastopen_sk);
6322 } else {
6323 tcp_rsk(req)->tfo_listener = false;
6324 if (!want_cookie)
6325 inet_csk_reqsk_queue_hash_add(sk, req,
6326 tcp_timeout_init((struct sock *)req));
6327 af_ops->send_synack(sk, dst, &fl, req, &foc,
6328 !want_cookie ? TCP_SYNACK_NORMAL :
6329 TCP_SYNACK_COOKIE);
6330 if (want_cookie) {
6331 reqsk_free(req);
6332 return 0;
6333 }
6334 }
6335 reqsk_put(req);
6336 return 0;
6337
6338drop_and_release:
6339 dst_release(dst);
6340drop_and_free:
6341 reqsk_free(req);
6342drop:
6343 tcp_listendrop(sk);
6344 return 0;
6345}
6346EXPORT_SYMBOL(tcp_conn_request);