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_ecn __read_mostly = 2;
85EXPORT_SYMBOL(sysctl_tcp_ecn);
86int sysctl_tcp_dsack __read_mostly = 1;
87int sysctl_tcp_app_win __read_mostly = 31;
88int sysctl_tcp_adv_win_scale __read_mostly = 1;
89EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
90
91int sysctl_tcp_stdurg __read_mostly;
92int sysctl_tcp_rfc1337 __read_mostly;
93int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
94int sysctl_tcp_frto __read_mostly = 2;
95int sysctl_tcp_frto_response __read_mostly;
96int sysctl_tcp_nometrics_save __read_mostly;
97
98int sysctl_tcp_thin_dupack __read_mostly;
99
100int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
101int sysctl_tcp_abc __read_mostly;
102int sysctl_tcp_early_retrans __read_mostly = 2;
103
104#define FLAG_DATA 0x01 /* Incoming frame contained data. */
105#define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
106#define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
107#define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
108#define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
109#define FLAG_DATA_SACKED 0x20 /* New SACK. */
110#define FLAG_ECE 0x40 /* ECE in this ACK */
111#define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
112#define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */
113#define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
114#define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
115#define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */
116#define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
117
118#define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
119#define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
120#define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
121#define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
122#define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
123
124#define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
125#define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
126
127/* Adapt the MSS value used to make delayed ack decision to the
128 * real world.
129 */
130static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
131{
132 struct inet_connection_sock *icsk = inet_csk(sk);
133 const unsigned int lss = icsk->icsk_ack.last_seg_size;
134 unsigned int len;
135
136 icsk->icsk_ack.last_seg_size = 0;
137
138 /* skb->len may jitter because of SACKs, even if peer
139 * sends good full-sized frames.
140 */
141 len = skb_shinfo(skb)->gso_size ? : skb->len;
142 if (len >= icsk->icsk_ack.rcv_mss) {
143 icsk->icsk_ack.rcv_mss = len;
144 } else {
145 /* Otherwise, we make more careful check taking into account,
146 * that SACKs block is variable.
147 *
148 * "len" is invariant segment length, including TCP header.
149 */
150 len += skb->data - skb_transport_header(skb);
151 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
152 /* If PSH is not set, packet should be
153 * full sized, provided peer TCP is not badly broken.
154 * This observation (if it is correct 8)) allows
155 * to handle super-low mtu links fairly.
156 */
157 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
158 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
159 /* Subtract also invariant (if peer is RFC compliant),
160 * tcp header plus fixed timestamp option length.
161 * Resulting "len" is MSS free of SACK jitter.
162 */
163 len -= tcp_sk(sk)->tcp_header_len;
164 icsk->icsk_ack.last_seg_size = len;
165 if (len == lss) {
166 icsk->icsk_ack.rcv_mss = len;
167 return;
168 }
169 }
170 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
171 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
172 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
173 }
174}
175
176static void tcp_incr_quickack(struct sock *sk)
177{
178 struct inet_connection_sock *icsk = inet_csk(sk);
179 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
180
181 if (quickacks == 0)
182 quickacks = 2;
183 if (quickacks > icsk->icsk_ack.quick)
184 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
185}
186
187static void tcp_enter_quickack_mode(struct sock *sk)
188{
189 struct inet_connection_sock *icsk = inet_csk(sk);
190 tcp_incr_quickack(sk);
191 icsk->icsk_ack.pingpong = 0;
192 icsk->icsk_ack.ato = TCP_ATO_MIN;
193}
194
195/* Send ACKs quickly, if "quick" count is not exhausted
196 * and the session is not interactive.
197 */
198
199static inline bool tcp_in_quickack_mode(const struct sock *sk)
200{
201 const struct inet_connection_sock *icsk = inet_csk(sk);
202
203 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
204}
205
206static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
207{
208 if (tp->ecn_flags & TCP_ECN_OK)
209 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
210}
211
212static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
213{
214 if (tcp_hdr(skb)->cwr)
215 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
216}
217
218static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
219{
220 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
221}
222
223static inline void TCP_ECN_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
224{
225 if (!(tp->ecn_flags & TCP_ECN_OK))
226 return;
227
228 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
229 case INET_ECN_NOT_ECT:
230 /* Funny extension: if ECT is not set on a segment,
231 * and we already seen ECT on a previous segment,
232 * it is probably a retransmit.
233 */
234 if (tp->ecn_flags & TCP_ECN_SEEN)
235 tcp_enter_quickack_mode((struct sock *)tp);
236 break;
237 case INET_ECN_CE:
238 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
239 /* fallinto */
240 default:
241 tp->ecn_flags |= TCP_ECN_SEEN;
242 }
243}
244
245static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
246{
247 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
248 tp->ecn_flags &= ~TCP_ECN_OK;
249}
250
251static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
252{
253 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
254 tp->ecn_flags &= ~TCP_ECN_OK;
255}
256
257static bool TCP_ECN_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
258{
259 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
260 return true;
261 return false;
262}
263
264/* Buffer size and advertised window tuning.
265 *
266 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
267 */
268
269static void tcp_fixup_sndbuf(struct sock *sk)
270{
271 int sndmem = SKB_TRUESIZE(tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER);
272
273 sndmem *= TCP_INIT_CWND;
274 if (sk->sk_sndbuf < sndmem)
275 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
276}
277
278/* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
279 *
280 * All tcp_full_space() is split to two parts: "network" buffer, allocated
281 * forward and advertised in receiver window (tp->rcv_wnd) and
282 * "application buffer", required to isolate scheduling/application
283 * latencies from network.
284 * window_clamp is maximal advertised window. It can be less than
285 * tcp_full_space(), in this case tcp_full_space() - window_clamp
286 * is reserved for "application" buffer. The less window_clamp is
287 * the smoother our behaviour from viewpoint of network, but the lower
288 * throughput and the higher sensitivity of the connection to losses. 8)
289 *
290 * rcv_ssthresh is more strict window_clamp used at "slow start"
291 * phase to predict further behaviour of this connection.
292 * It is used for two goals:
293 * - to enforce header prediction at sender, even when application
294 * requires some significant "application buffer". It is check #1.
295 * - to prevent pruning of receive queue because of misprediction
296 * of receiver window. Check #2.
297 *
298 * The scheme does not work when sender sends good segments opening
299 * window and then starts to feed us spaghetti. But it should work
300 * in common situations. Otherwise, we have to rely on queue collapsing.
301 */
302
303/* Slow part of check#2. */
304static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
305{
306 struct tcp_sock *tp = tcp_sk(sk);
307 /* Optimize this! */
308 int truesize = tcp_win_from_space(skb->truesize) >> 1;
309 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
310
311 while (tp->rcv_ssthresh <= window) {
312 if (truesize <= skb->len)
313 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
314
315 truesize >>= 1;
316 window >>= 1;
317 }
318 return 0;
319}
320
321static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
322{
323 struct tcp_sock *tp = tcp_sk(sk);
324
325 /* Check #1 */
326 if (tp->rcv_ssthresh < tp->window_clamp &&
327 (int)tp->rcv_ssthresh < tcp_space(sk) &&
328 !sk_under_memory_pressure(sk)) {
329 int incr;
330
331 /* Check #2. Increase window, if skb with such overhead
332 * will fit to rcvbuf in future.
333 */
334 if (tcp_win_from_space(skb->truesize) <= skb->len)
335 incr = 2 * tp->advmss;
336 else
337 incr = __tcp_grow_window(sk, skb);
338
339 if (incr) {
340 incr = max_t(int, incr, 2 * skb->len);
341 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
342 tp->window_clamp);
343 inet_csk(sk)->icsk_ack.quick |= 1;
344 }
345 }
346}
347
348/* 3. Tuning rcvbuf, when connection enters established state. */
349
350static void tcp_fixup_rcvbuf(struct sock *sk)
351{
352 u32 mss = tcp_sk(sk)->advmss;
353 u32 icwnd = TCP_DEFAULT_INIT_RCVWND;
354 int rcvmem;
355
356 /* Limit to 10 segments if mss <= 1460,
357 * or 14600/mss segments, with a minimum of two segments.
358 */
359 if (mss > 1460)
360 icwnd = max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2);
361
362 rcvmem = SKB_TRUESIZE(mss + MAX_TCP_HEADER);
363 while (tcp_win_from_space(rcvmem) < mss)
364 rcvmem += 128;
365
366 rcvmem *= icwnd;
367
368 if (sk->sk_rcvbuf < rcvmem)
369 sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
370}
371
372/* 4. Try to fixup all. It is made immediately after connection enters
373 * established state.
374 */
375static void tcp_init_buffer_space(struct sock *sk)
376{
377 struct tcp_sock *tp = tcp_sk(sk);
378 int maxwin;
379
380 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
381 tcp_fixup_rcvbuf(sk);
382 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
383 tcp_fixup_sndbuf(sk);
384
385 tp->rcvq_space.space = tp->rcv_wnd;
386
387 maxwin = tcp_full_space(sk);
388
389 if (tp->window_clamp >= maxwin) {
390 tp->window_clamp = maxwin;
391
392 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
393 tp->window_clamp = max(maxwin -
394 (maxwin >> sysctl_tcp_app_win),
395 4 * tp->advmss);
396 }
397
398 /* Force reservation of one segment. */
399 if (sysctl_tcp_app_win &&
400 tp->window_clamp > 2 * tp->advmss &&
401 tp->window_clamp + tp->advmss > maxwin)
402 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
403
404 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
405 tp->snd_cwnd_stamp = tcp_time_stamp;
406}
407
408/* 5. Recalculate window clamp after socket hit its memory bounds. */
409static void tcp_clamp_window(struct sock *sk)
410{
411 struct tcp_sock *tp = tcp_sk(sk);
412 struct inet_connection_sock *icsk = inet_csk(sk);
413
414 icsk->icsk_ack.quick = 0;
415
416 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
417 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
418 !sk_under_memory_pressure(sk) &&
419 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
420 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
421 sysctl_tcp_rmem[2]);
422 }
423 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
424 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
425}
426
427/* Initialize RCV_MSS value.
428 * RCV_MSS is an our guess about MSS used by the peer.
429 * We haven't any direct information about the MSS.
430 * It's better to underestimate the RCV_MSS rather than overestimate.
431 * Overestimations make us ACKing less frequently than needed.
432 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
433 */
434void tcp_initialize_rcv_mss(struct sock *sk)
435{
436 const struct tcp_sock *tp = tcp_sk(sk);
437 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
438
439 hint = min(hint, tp->rcv_wnd / 2);
440 hint = min(hint, TCP_MSS_DEFAULT);
441 hint = max(hint, TCP_MIN_MSS);
442
443 inet_csk(sk)->icsk_ack.rcv_mss = hint;
444}
445EXPORT_SYMBOL(tcp_initialize_rcv_mss);
446
447/* Receiver "autotuning" code.
448 *
449 * The algorithm for RTT estimation w/o timestamps is based on
450 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
451 * <http://public.lanl.gov/radiant/pubs.html#DRS>
452 *
453 * More detail on this code can be found at
454 * <http://staff.psc.edu/jheffner/>,
455 * though this reference is out of date. A new paper
456 * is pending.
457 */
458static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
459{
460 u32 new_sample = tp->rcv_rtt_est.rtt;
461 long m = sample;
462
463 if (m == 0)
464 m = 1;
465
466 if (new_sample != 0) {
467 /* If we sample in larger samples in the non-timestamp
468 * case, we could grossly overestimate the RTT especially
469 * with chatty applications or bulk transfer apps which
470 * are stalled on filesystem I/O.
471 *
472 * Also, since we are only going for a minimum in the
473 * non-timestamp case, we do not smooth things out
474 * else with timestamps disabled convergence takes too
475 * long.
476 */
477 if (!win_dep) {
478 m -= (new_sample >> 3);
479 new_sample += m;
480 } else {
481 m <<= 3;
482 if (m < new_sample)
483 new_sample = m;
484 }
485 } else {
486 /* No previous measure. */
487 new_sample = m << 3;
488 }
489
490 if (tp->rcv_rtt_est.rtt != new_sample)
491 tp->rcv_rtt_est.rtt = new_sample;
492}
493
494static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
495{
496 if (tp->rcv_rtt_est.time == 0)
497 goto new_measure;
498 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
499 return;
500 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1);
501
502new_measure:
503 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
504 tp->rcv_rtt_est.time = tcp_time_stamp;
505}
506
507static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
508 const struct sk_buff *skb)
509{
510 struct tcp_sock *tp = tcp_sk(sk);
511 if (tp->rx_opt.rcv_tsecr &&
512 (TCP_SKB_CB(skb)->end_seq -
513 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
514 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
515}
516
517/*
518 * This function should be called every time data is copied to user space.
519 * It calculates the appropriate TCP receive buffer space.
520 */
521void tcp_rcv_space_adjust(struct sock *sk)
522{
523 struct tcp_sock *tp = tcp_sk(sk);
524 int time;
525 int space;
526
527 if (tp->rcvq_space.time == 0)
528 goto new_measure;
529
530 time = tcp_time_stamp - tp->rcvq_space.time;
531 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
532 return;
533
534 space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
535
536 space = max(tp->rcvq_space.space, space);
537
538 if (tp->rcvq_space.space != space) {
539 int rcvmem;
540
541 tp->rcvq_space.space = space;
542
543 if (sysctl_tcp_moderate_rcvbuf &&
544 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
545 int new_clamp = space;
546
547 /* Receive space grows, normalize in order to
548 * take into account packet headers and sk_buff
549 * structure overhead.
550 */
551 space /= tp->advmss;
552 if (!space)
553 space = 1;
554 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
555 while (tcp_win_from_space(rcvmem) < tp->advmss)
556 rcvmem += 128;
557 space *= rcvmem;
558 space = min(space, sysctl_tcp_rmem[2]);
559 if (space > sk->sk_rcvbuf) {
560 sk->sk_rcvbuf = space;
561
562 /* Make the window clamp follow along. */
563 tp->window_clamp = new_clamp;
564 }
565 }
566 }
567
568new_measure:
569 tp->rcvq_space.seq = tp->copied_seq;
570 tp->rcvq_space.time = tcp_time_stamp;
571}
572
573/* There is something which you must keep in mind when you analyze the
574 * behavior of the tp->ato delayed ack timeout interval. When a
575 * connection starts up, we want to ack as quickly as possible. The
576 * problem is that "good" TCP's do slow start at the beginning of data
577 * transmission. The means that until we send the first few ACK's the
578 * sender will sit on his end and only queue most of his data, because
579 * he can only send snd_cwnd unacked packets at any given time. For
580 * each ACK we send, he increments snd_cwnd and transmits more of his
581 * queue. -DaveM
582 */
583static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
584{
585 struct tcp_sock *tp = tcp_sk(sk);
586 struct inet_connection_sock *icsk = inet_csk(sk);
587 u32 now;
588
589 inet_csk_schedule_ack(sk);
590
591 tcp_measure_rcv_mss(sk, skb);
592
593 tcp_rcv_rtt_measure(tp);
594
595 now = tcp_time_stamp;
596
597 if (!icsk->icsk_ack.ato) {
598 /* The _first_ data packet received, initialize
599 * delayed ACK engine.
600 */
601 tcp_incr_quickack(sk);
602 icsk->icsk_ack.ato = TCP_ATO_MIN;
603 } else {
604 int m = now - icsk->icsk_ack.lrcvtime;
605
606 if (m <= TCP_ATO_MIN / 2) {
607 /* The fastest case is the first. */
608 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
609 } else if (m < icsk->icsk_ack.ato) {
610 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
611 if (icsk->icsk_ack.ato > icsk->icsk_rto)
612 icsk->icsk_ack.ato = icsk->icsk_rto;
613 } else if (m > icsk->icsk_rto) {
614 /* Too long gap. Apparently sender failed to
615 * restart window, so that we send ACKs quickly.
616 */
617 tcp_incr_quickack(sk);
618 sk_mem_reclaim(sk);
619 }
620 }
621 icsk->icsk_ack.lrcvtime = now;
622
623 TCP_ECN_check_ce(tp, skb);
624
625 if (skb->len >= 128)
626 tcp_grow_window(sk, skb);
627}
628
629/* Called to compute a smoothed rtt estimate. The data fed to this
630 * routine either comes from timestamps, or from segments that were
631 * known _not_ to have been retransmitted [see Karn/Partridge
632 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
633 * piece by Van Jacobson.
634 * NOTE: the next three routines used to be one big routine.
635 * To save cycles in the RFC 1323 implementation it was better to break
636 * it up into three procedures. -- erics
637 */
638static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
639{
640 struct tcp_sock *tp = tcp_sk(sk);
641 long m = mrtt; /* RTT */
642
643 /* The following amusing code comes from Jacobson's
644 * article in SIGCOMM '88. Note that rtt and mdev
645 * are scaled versions of rtt and mean deviation.
646 * This is designed to be as fast as possible
647 * m stands for "measurement".
648 *
649 * On a 1990 paper the rto value is changed to:
650 * RTO = rtt + 4 * mdev
651 *
652 * Funny. This algorithm seems to be very broken.
653 * These formulae increase RTO, when it should be decreased, increase
654 * too slowly, when it should be increased quickly, decrease too quickly
655 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
656 * does not matter how to _calculate_ it. Seems, it was trap
657 * that VJ failed to avoid. 8)
658 */
659 if (m == 0)
660 m = 1;
661 if (tp->srtt != 0) {
662 m -= (tp->srtt >> 3); /* m is now error in rtt est */
663 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
664 if (m < 0) {
665 m = -m; /* m is now abs(error) */
666 m -= (tp->mdev >> 2); /* similar update on mdev */
667 /* This is similar to one of Eifel findings.
668 * Eifel blocks mdev updates when rtt decreases.
669 * This solution is a bit different: we use finer gain
670 * for mdev in this case (alpha*beta).
671 * Like Eifel it also prevents growth of rto,
672 * but also it limits too fast rto decreases,
673 * happening in pure Eifel.
674 */
675 if (m > 0)
676 m >>= 3;
677 } else {
678 m -= (tp->mdev >> 2); /* similar update on mdev */
679 }
680 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
681 if (tp->mdev > tp->mdev_max) {
682 tp->mdev_max = tp->mdev;
683 if (tp->mdev_max > tp->rttvar)
684 tp->rttvar = tp->mdev_max;
685 }
686 if (after(tp->snd_una, tp->rtt_seq)) {
687 if (tp->mdev_max < tp->rttvar)
688 tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2;
689 tp->rtt_seq = tp->snd_nxt;
690 tp->mdev_max = tcp_rto_min(sk);
691 }
692 } else {
693 /* no previous measure. */
694 tp->srtt = m << 3; /* take the measured time to be rtt */
695 tp->mdev = m << 1; /* make sure rto = 3*rtt */
696 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
697 tp->rtt_seq = tp->snd_nxt;
698 }
699}
700
701/* Calculate rto without backoff. This is the second half of Van Jacobson's
702 * routine referred to above.
703 */
704static inline void tcp_set_rto(struct sock *sk)
705{
706 const struct tcp_sock *tp = tcp_sk(sk);
707 /* Old crap is replaced with new one. 8)
708 *
709 * More seriously:
710 * 1. If rtt variance happened to be less 50msec, it is hallucination.
711 * It cannot be less due to utterly erratic ACK generation made
712 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
713 * to do with delayed acks, because at cwnd>2 true delack timeout
714 * is invisible. Actually, Linux-2.4 also generates erratic
715 * ACKs in some circumstances.
716 */
717 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
718
719 /* 2. Fixups made earlier cannot be right.
720 * If we do not estimate RTO correctly without them,
721 * all the algo is pure shit and should be replaced
722 * with correct one. It is exactly, which we pretend to do.
723 */
724
725 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
726 * guarantees that rto is higher.
727 */
728 tcp_bound_rto(sk);
729}
730
731/* Save metrics learned by this TCP session.
732 This function is called only, when TCP finishes successfully
733 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
734 */
735void tcp_update_metrics(struct sock *sk)
736{
737 struct tcp_sock *tp = tcp_sk(sk);
738 struct dst_entry *dst = __sk_dst_get(sk);
739
740 if (sysctl_tcp_nometrics_save)
741 return;
742
743 dst_confirm(dst);
744
745 if (dst && (dst->flags & DST_HOST)) {
746 const struct inet_connection_sock *icsk = inet_csk(sk);
747 int m;
748 unsigned long rtt;
749
750 if (icsk->icsk_backoff || !tp->srtt) {
751 /* This session failed to estimate rtt. Why?
752 * Probably, no packets returned in time.
753 * Reset our results.
754 */
755 if (!(dst_metric_locked(dst, RTAX_RTT)))
756 dst_metric_set(dst, RTAX_RTT, 0);
757 return;
758 }
759
760 rtt = dst_metric_rtt(dst, RTAX_RTT);
761 m = rtt - tp->srtt;
762
763 /* If newly calculated rtt larger than stored one,
764 * store new one. Otherwise, use EWMA. Remember,
765 * rtt overestimation is always better than underestimation.
766 */
767 if (!(dst_metric_locked(dst, RTAX_RTT))) {
768 if (m <= 0)
769 set_dst_metric_rtt(dst, RTAX_RTT, tp->srtt);
770 else
771 set_dst_metric_rtt(dst, RTAX_RTT, rtt - (m >> 3));
772 }
773
774 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
775 unsigned long var;
776 if (m < 0)
777 m = -m;
778
779 /* Scale deviation to rttvar fixed point */
780 m >>= 1;
781 if (m < tp->mdev)
782 m = tp->mdev;
783
784 var = dst_metric_rtt(dst, RTAX_RTTVAR);
785 if (m >= var)
786 var = m;
787 else
788 var -= (var - m) >> 2;
789
790 set_dst_metric_rtt(dst, RTAX_RTTVAR, var);
791 }
792
793 if (tcp_in_initial_slowstart(tp)) {
794 /* Slow start still did not finish. */
795 if (dst_metric(dst, RTAX_SSTHRESH) &&
796 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
797 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
798 dst_metric_set(dst, RTAX_SSTHRESH, tp->snd_cwnd >> 1);
799 if (!dst_metric_locked(dst, RTAX_CWND) &&
800 tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
801 dst_metric_set(dst, RTAX_CWND, tp->snd_cwnd);
802 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
803 icsk->icsk_ca_state == TCP_CA_Open) {
804 /* Cong. avoidance phase, cwnd is reliable. */
805 if (!dst_metric_locked(dst, RTAX_SSTHRESH))
806 dst_metric_set(dst, RTAX_SSTHRESH,
807 max(tp->snd_cwnd >> 1, tp->snd_ssthresh));
808 if (!dst_metric_locked(dst, RTAX_CWND))
809 dst_metric_set(dst, RTAX_CWND,
810 (dst_metric(dst, RTAX_CWND) +
811 tp->snd_cwnd) >> 1);
812 } else {
813 /* Else slow start did not finish, cwnd is non-sense,
814 ssthresh may be also invalid.
815 */
816 if (!dst_metric_locked(dst, RTAX_CWND))
817 dst_metric_set(dst, RTAX_CWND,
818 (dst_metric(dst, RTAX_CWND) +
819 tp->snd_ssthresh) >> 1);
820 if (dst_metric(dst, RTAX_SSTHRESH) &&
821 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
822 tp->snd_ssthresh > dst_metric(dst, RTAX_SSTHRESH))
823 dst_metric_set(dst, RTAX_SSTHRESH, tp->snd_ssthresh);
824 }
825
826 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
827 if (dst_metric(dst, RTAX_REORDERING) < tp->reordering &&
828 tp->reordering != sysctl_tcp_reordering)
829 dst_metric_set(dst, RTAX_REORDERING, tp->reordering);
830 }
831 }
832}
833
834__u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
835{
836 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
837
838 if (!cwnd)
839 cwnd = TCP_INIT_CWND;
840 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
841}
842
843/* Set slow start threshold and cwnd not falling to slow start */
844void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
845{
846 struct tcp_sock *tp = tcp_sk(sk);
847 const struct inet_connection_sock *icsk = inet_csk(sk);
848
849 tp->prior_ssthresh = 0;
850 tp->bytes_acked = 0;
851 if (icsk->icsk_ca_state < TCP_CA_CWR) {
852 tp->undo_marker = 0;
853 if (set_ssthresh)
854 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
855 tp->snd_cwnd = min(tp->snd_cwnd,
856 tcp_packets_in_flight(tp) + 1U);
857 tp->snd_cwnd_cnt = 0;
858 tp->high_seq = tp->snd_nxt;
859 tp->snd_cwnd_stamp = tcp_time_stamp;
860 TCP_ECN_queue_cwr(tp);
861
862 tcp_set_ca_state(sk, TCP_CA_CWR);
863 }
864}
865
866/*
867 * Packet counting of FACK is based on in-order assumptions, therefore TCP
868 * disables it when reordering is detected
869 */
870static void tcp_disable_fack(struct tcp_sock *tp)
871{
872 /* RFC3517 uses different metric in lost marker => reset on change */
873 if (tcp_is_fack(tp))
874 tp->lost_skb_hint = NULL;
875 tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
876}
877
878/* Take a notice that peer is sending D-SACKs */
879static void tcp_dsack_seen(struct tcp_sock *tp)
880{
881 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
882}
883
884/* Initialize metrics on socket. */
885
886static void tcp_init_metrics(struct sock *sk)
887{
888 struct tcp_sock *tp = tcp_sk(sk);
889 struct dst_entry *dst = __sk_dst_get(sk);
890
891 if (dst == NULL)
892 goto reset;
893
894 dst_confirm(dst);
895
896 if (dst_metric_locked(dst, RTAX_CWND))
897 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
898 if (dst_metric(dst, RTAX_SSTHRESH)) {
899 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
900 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
901 tp->snd_ssthresh = tp->snd_cwnd_clamp;
902 } else {
903 /* ssthresh may have been reduced unnecessarily during.
904 * 3WHS. Restore it back to its initial default.
905 */
906 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
907 }
908 if (dst_metric(dst, RTAX_REORDERING) &&
909 tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
910 tcp_disable_fack(tp);
911 tcp_disable_early_retrans(tp);
912 tp->reordering = dst_metric(dst, RTAX_REORDERING);
913 }
914
915 if (dst_metric(dst, RTAX_RTT) == 0 || tp->srtt == 0)
916 goto reset;
917
918 /* Initial rtt is determined from SYN,SYN-ACK.
919 * The segment is small and rtt may appear much
920 * less than real one. Use per-dst memory
921 * to make it more realistic.
922 *
923 * A bit of theory. RTT is time passed after "normal" sized packet
924 * is sent until it is ACKed. In normal circumstances sending small
925 * packets force peer to delay ACKs and calculation is correct too.
926 * The algorithm is adaptive and, provided we follow specs, it
927 * NEVER underestimate RTT. BUT! If peer tries to make some clever
928 * tricks sort of "quick acks" for time long enough to decrease RTT
929 * to low value, and then abruptly stops to do it and starts to delay
930 * ACKs, wait for troubles.
931 */
932 if (dst_metric_rtt(dst, RTAX_RTT) > tp->srtt) {
933 tp->srtt = dst_metric_rtt(dst, RTAX_RTT);
934 tp->rtt_seq = tp->snd_nxt;
935 }
936 if (dst_metric_rtt(dst, RTAX_RTTVAR) > tp->mdev) {
937 tp->mdev = dst_metric_rtt(dst, RTAX_RTTVAR);
938 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
939 }
940 tcp_set_rto(sk);
941reset:
942 if (tp->srtt == 0) {
943 /* RFC6298: 5.7 We've failed to get a valid RTT sample from
944 * 3WHS. This is most likely due to retransmission,
945 * including spurious one. Reset the RTO back to 3secs
946 * from the more aggressive 1sec to avoid more spurious
947 * retransmission.
948 */
949 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_FALLBACK;
950 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
951 }
952 /* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
953 * retransmitted. In light of RFC6298 more aggressive 1sec
954 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
955 * retransmission has occurred.
956 */
957 if (tp->total_retrans > 1)
958 tp->snd_cwnd = 1;
959 else
960 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
961 tp->snd_cwnd_stamp = tcp_time_stamp;
962}
963
964static void tcp_update_reordering(struct sock *sk, const int metric,
965 const int ts)
966{
967 struct tcp_sock *tp = tcp_sk(sk);
968 if (metric > tp->reordering) {
969 int mib_idx;
970
971 tp->reordering = min(TCP_MAX_REORDERING, metric);
972
973 /* This exciting event is worth to be remembered. 8) */
974 if (ts)
975 mib_idx = LINUX_MIB_TCPTSREORDER;
976 else if (tcp_is_reno(tp))
977 mib_idx = LINUX_MIB_TCPRENOREORDER;
978 else if (tcp_is_fack(tp))
979 mib_idx = LINUX_MIB_TCPFACKREORDER;
980 else
981 mib_idx = LINUX_MIB_TCPSACKREORDER;
982
983 NET_INC_STATS_BH(sock_net(sk), mib_idx);
984#if FASTRETRANS_DEBUG > 1
985 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
986 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
987 tp->reordering,
988 tp->fackets_out,
989 tp->sacked_out,
990 tp->undo_marker ? tp->undo_retrans : 0);
991#endif
992 tcp_disable_fack(tp);
993 }
994
995 if (metric > 0)
996 tcp_disable_early_retrans(tp);
997}
998
999/* This must be called before lost_out is incremented */
1000static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
1001{
1002 if ((tp->retransmit_skb_hint == NULL) ||
1003 before(TCP_SKB_CB(skb)->seq,
1004 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
1005 tp->retransmit_skb_hint = skb;
1006
1007 if (!tp->lost_out ||
1008 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
1009 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
1010}
1011
1012static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
1013{
1014 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1015 tcp_verify_retransmit_hint(tp, skb);
1016
1017 tp->lost_out += tcp_skb_pcount(skb);
1018 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1019 }
1020}
1021
1022static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
1023 struct sk_buff *skb)
1024{
1025 tcp_verify_retransmit_hint(tp, skb);
1026
1027 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1028 tp->lost_out += tcp_skb_pcount(skb);
1029 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1030 }
1031}
1032
1033/* This procedure tags the retransmission queue when SACKs arrive.
1034 *
1035 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1036 * Packets in queue with these bits set are counted in variables
1037 * sacked_out, retrans_out and lost_out, correspondingly.
1038 *
1039 * Valid combinations are:
1040 * Tag InFlight Description
1041 * 0 1 - orig segment is in flight.
1042 * S 0 - nothing flies, orig reached receiver.
1043 * L 0 - nothing flies, orig lost by net.
1044 * R 2 - both orig and retransmit are in flight.
1045 * L|R 1 - orig is lost, retransmit is in flight.
1046 * S|R 1 - orig reached receiver, retrans is still in flight.
1047 * (L|S|R is logically valid, it could occur when L|R is sacked,
1048 * but it is equivalent to plain S and code short-curcuits it to S.
1049 * L|S is logically invalid, it would mean -1 packet in flight 8))
1050 *
1051 * These 6 states form finite state machine, controlled by the following events:
1052 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1053 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1054 * 3. Loss detection event of two flavors:
1055 * A. Scoreboard estimator decided the packet is lost.
1056 * A'. Reno "three dupacks" marks head of queue lost.
1057 * A''. Its FACK modification, head until snd.fack is lost.
1058 * B. SACK arrives sacking SND.NXT at the moment, when the
1059 * segment was retransmitted.
1060 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1061 *
1062 * It is pleasant to note, that state diagram turns out to be commutative,
1063 * so that we are allowed not to be bothered by order of our actions,
1064 * when multiple events arrive simultaneously. (see the function below).
1065 *
1066 * Reordering detection.
1067 * --------------------
1068 * Reordering metric is maximal distance, which a packet can be displaced
1069 * in packet stream. With SACKs we can estimate it:
1070 *
1071 * 1. SACK fills old hole and the corresponding segment was not
1072 * ever retransmitted -> reordering. Alas, we cannot use it
1073 * when segment was retransmitted.
1074 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1075 * for retransmitted and already SACKed segment -> reordering..
1076 * Both of these heuristics are not used in Loss state, when we cannot
1077 * account for retransmits accurately.
1078 *
1079 * SACK block validation.
1080 * ----------------------
1081 *
1082 * SACK block range validation checks that the received SACK block fits to
1083 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1084 * Note that SND.UNA is not included to the range though being valid because
1085 * it means that the receiver is rather inconsistent with itself reporting
1086 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1087 * perfectly valid, however, in light of RFC2018 which explicitly states
1088 * that "SACK block MUST reflect the newest segment. Even if the newest
1089 * segment is going to be discarded ...", not that it looks very clever
1090 * in case of head skb. Due to potentional receiver driven attacks, we
1091 * choose to avoid immediate execution of a walk in write queue due to
1092 * reneging and defer head skb's loss recovery to standard loss recovery
1093 * procedure that will eventually trigger (nothing forbids us doing this).
1094 *
1095 * Implements also blockage to start_seq wrap-around. Problem lies in the
1096 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1097 * there's no guarantee that it will be before snd_nxt (n). The problem
1098 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1099 * wrap (s_w):
1100 *
1101 * <- outs wnd -> <- wrapzone ->
1102 * u e n u_w e_w s n_w
1103 * | | | | | | |
1104 * |<------------+------+----- TCP seqno space --------------+---------->|
1105 * ...-- <2^31 ->| |<--------...
1106 * ...---- >2^31 ------>| |<--------...
1107 *
1108 * Current code wouldn't be vulnerable but it's better still to discard such
1109 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1110 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1111 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1112 * equal to the ideal case (infinite seqno space without wrap caused issues).
1113 *
1114 * With D-SACK the lower bound is extended to cover sequence space below
1115 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1116 * again, D-SACK block must not to go across snd_una (for the same reason as
1117 * for the normal SACK blocks, explained above). But there all simplicity
1118 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1119 * fully below undo_marker they do not affect behavior in anyway and can
1120 * therefore be safely ignored. In rare cases (which are more or less
1121 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1122 * fragmentation and packet reordering past skb's retransmission. To consider
1123 * them correctly, the acceptable range must be extended even more though
1124 * the exact amount is rather hard to quantify. However, tp->max_window can
1125 * be used as an exaggerated estimate.
1126 */
1127static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1128 u32 start_seq, u32 end_seq)
1129{
1130 /* Too far in future, or reversed (interpretation is ambiguous) */
1131 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1132 return false;
1133
1134 /* Nasty start_seq wrap-around check (see comments above) */
1135 if (!before(start_seq, tp->snd_nxt))
1136 return false;
1137
1138 /* In outstanding window? ...This is valid exit for D-SACKs too.
1139 * start_seq == snd_una is non-sensical (see comments above)
1140 */
1141 if (after(start_seq, tp->snd_una))
1142 return true;
1143
1144 if (!is_dsack || !tp->undo_marker)
1145 return false;
1146
1147 /* ...Then it's D-SACK, and must reside below snd_una completely */
1148 if (after(end_seq, tp->snd_una))
1149 return false;
1150
1151 if (!before(start_seq, tp->undo_marker))
1152 return true;
1153
1154 /* Too old */
1155 if (!after(end_seq, tp->undo_marker))
1156 return false;
1157
1158 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1159 * start_seq < undo_marker and end_seq >= undo_marker.
1160 */
1161 return !before(start_seq, end_seq - tp->max_window);
1162}
1163
1164/* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1165 * Event "B". Later note: FACK people cheated me again 8), we have to account
1166 * for reordering! Ugly, but should help.
1167 *
1168 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1169 * less than what is now known to be received by the other end (derived from
1170 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1171 * retransmitted skbs to avoid some costly processing per ACKs.
1172 */
1173static void tcp_mark_lost_retrans(struct sock *sk)
1174{
1175 const struct inet_connection_sock *icsk = inet_csk(sk);
1176 struct tcp_sock *tp = tcp_sk(sk);
1177 struct sk_buff *skb;
1178 int cnt = 0;
1179 u32 new_low_seq = tp->snd_nxt;
1180 u32 received_upto = tcp_highest_sack_seq(tp);
1181
1182 if (!tcp_is_fack(tp) || !tp->retrans_out ||
1183 !after(received_upto, tp->lost_retrans_low) ||
1184 icsk->icsk_ca_state != TCP_CA_Recovery)
1185 return;
1186
1187 tcp_for_write_queue(skb, sk) {
1188 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1189
1190 if (skb == tcp_send_head(sk))
1191 break;
1192 if (cnt == tp->retrans_out)
1193 break;
1194 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1195 continue;
1196
1197 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1198 continue;
1199
1200 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1201 * constraint here (see above) but figuring out that at
1202 * least tp->reordering SACK blocks reside between ack_seq
1203 * and received_upto is not easy task to do cheaply with
1204 * the available datastructures.
1205 *
1206 * Whether FACK should check here for tp->reordering segs
1207 * in-between one could argue for either way (it would be
1208 * rather simple to implement as we could count fack_count
1209 * during the walk and do tp->fackets_out - fack_count).
1210 */
1211 if (after(received_upto, ack_seq)) {
1212 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1213 tp->retrans_out -= tcp_skb_pcount(skb);
1214
1215 tcp_skb_mark_lost_uncond_verify(tp, skb);
1216 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1217 } else {
1218 if (before(ack_seq, new_low_seq))
1219 new_low_seq = ack_seq;
1220 cnt += tcp_skb_pcount(skb);
1221 }
1222 }
1223
1224 if (tp->retrans_out)
1225 tp->lost_retrans_low = new_low_seq;
1226}
1227
1228static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1229 struct tcp_sack_block_wire *sp, int num_sacks,
1230 u32 prior_snd_una)
1231{
1232 struct tcp_sock *tp = tcp_sk(sk);
1233 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1234 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1235 bool dup_sack = false;
1236
1237 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1238 dup_sack = true;
1239 tcp_dsack_seen(tp);
1240 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1241 } else if (num_sacks > 1) {
1242 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1243 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1244
1245 if (!after(end_seq_0, end_seq_1) &&
1246 !before(start_seq_0, start_seq_1)) {
1247 dup_sack = true;
1248 tcp_dsack_seen(tp);
1249 NET_INC_STATS_BH(sock_net(sk),
1250 LINUX_MIB_TCPDSACKOFORECV);
1251 }
1252 }
1253
1254 /* D-SACK for already forgotten data... Do dumb counting. */
1255 if (dup_sack && tp->undo_marker && tp->undo_retrans &&
1256 !after(end_seq_0, prior_snd_una) &&
1257 after(end_seq_0, tp->undo_marker))
1258 tp->undo_retrans--;
1259
1260 return dup_sack;
1261}
1262
1263struct tcp_sacktag_state {
1264 int reord;
1265 int fack_count;
1266 int flag;
1267};
1268
1269/* Check if skb is fully within the SACK block. In presence of GSO skbs,
1270 * the incoming SACK may not exactly match but we can find smaller MSS
1271 * aligned portion of it that matches. Therefore we might need to fragment
1272 * which may fail and creates some hassle (caller must handle error case
1273 * returns).
1274 *
1275 * FIXME: this could be merged to shift decision code
1276 */
1277static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1278 u32 start_seq, u32 end_seq)
1279{
1280 int err;
1281 bool in_sack;
1282 unsigned int pkt_len;
1283 unsigned int mss;
1284
1285 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1286 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1287
1288 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1289 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1290 mss = tcp_skb_mss(skb);
1291 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1292
1293 if (!in_sack) {
1294 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1295 if (pkt_len < mss)
1296 pkt_len = mss;
1297 } else {
1298 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1299 if (pkt_len < mss)
1300 return -EINVAL;
1301 }
1302
1303 /* Round if necessary so that SACKs cover only full MSSes
1304 * and/or the remaining small portion (if present)
1305 */
1306 if (pkt_len > mss) {
1307 unsigned int new_len = (pkt_len / mss) * mss;
1308 if (!in_sack && new_len < pkt_len) {
1309 new_len += mss;
1310 if (new_len > skb->len)
1311 return 0;
1312 }
1313 pkt_len = new_len;
1314 }
1315 err = tcp_fragment(sk, skb, pkt_len, mss);
1316 if (err < 0)
1317 return err;
1318 }
1319
1320 return in_sack;
1321}
1322
1323/* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1324static u8 tcp_sacktag_one(struct sock *sk,
1325 struct tcp_sacktag_state *state, u8 sacked,
1326 u32 start_seq, u32 end_seq,
1327 bool dup_sack, int pcount)
1328{
1329 struct tcp_sock *tp = tcp_sk(sk);
1330 int fack_count = state->fack_count;
1331
1332 /* Account D-SACK for retransmitted packet. */
1333 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1334 if (tp->undo_marker && tp->undo_retrans &&
1335 after(end_seq, tp->undo_marker))
1336 tp->undo_retrans--;
1337 if (sacked & TCPCB_SACKED_ACKED)
1338 state->reord = min(fack_count, state->reord);
1339 }
1340
1341 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1342 if (!after(end_seq, tp->snd_una))
1343 return sacked;
1344
1345 if (!(sacked & TCPCB_SACKED_ACKED)) {
1346 if (sacked & TCPCB_SACKED_RETRANS) {
1347 /* If the segment is not tagged as lost,
1348 * we do not clear RETRANS, believing
1349 * that retransmission is still in flight.
1350 */
1351 if (sacked & TCPCB_LOST) {
1352 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1353 tp->lost_out -= pcount;
1354 tp->retrans_out -= pcount;
1355 }
1356 } else {
1357 if (!(sacked & TCPCB_RETRANS)) {
1358 /* New sack for not retransmitted frame,
1359 * which was in hole. It is reordering.
1360 */
1361 if (before(start_seq,
1362 tcp_highest_sack_seq(tp)))
1363 state->reord = min(fack_count,
1364 state->reord);
1365
1366 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1367 if (!after(end_seq, tp->frto_highmark))
1368 state->flag |= FLAG_ONLY_ORIG_SACKED;
1369 }
1370
1371 if (sacked & TCPCB_LOST) {
1372 sacked &= ~TCPCB_LOST;
1373 tp->lost_out -= pcount;
1374 }
1375 }
1376
1377 sacked |= TCPCB_SACKED_ACKED;
1378 state->flag |= FLAG_DATA_SACKED;
1379 tp->sacked_out += pcount;
1380
1381 fack_count += pcount;
1382
1383 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1384 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1385 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1386 tp->lost_cnt_hint += pcount;
1387
1388 if (fack_count > tp->fackets_out)
1389 tp->fackets_out = fack_count;
1390 }
1391
1392 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1393 * frames and clear it. undo_retrans is decreased above, L|R frames
1394 * are accounted above as well.
1395 */
1396 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1397 sacked &= ~TCPCB_SACKED_RETRANS;
1398 tp->retrans_out -= pcount;
1399 }
1400
1401 return sacked;
1402}
1403
1404/* Shift newly-SACKed bytes from this skb to the immediately previous
1405 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1406 */
1407static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1408 struct tcp_sacktag_state *state,
1409 unsigned int pcount, int shifted, int mss,
1410 bool dup_sack)
1411{
1412 struct tcp_sock *tp = tcp_sk(sk);
1413 struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1414 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1415 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1416
1417 BUG_ON(!pcount);
1418
1419 /* Adjust counters and hints for the newly sacked sequence
1420 * range but discard the return value since prev is already
1421 * marked. We must tag the range first because the seq
1422 * advancement below implicitly advances
1423 * tcp_highest_sack_seq() when skb is highest_sack.
1424 */
1425 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1426 start_seq, end_seq, dup_sack, pcount);
1427
1428 if (skb == tp->lost_skb_hint)
1429 tp->lost_cnt_hint += pcount;
1430
1431 TCP_SKB_CB(prev)->end_seq += shifted;
1432 TCP_SKB_CB(skb)->seq += shifted;
1433
1434 skb_shinfo(prev)->gso_segs += pcount;
1435 BUG_ON(skb_shinfo(skb)->gso_segs < pcount);
1436 skb_shinfo(skb)->gso_segs -= pcount;
1437
1438 /* When we're adding to gso_segs == 1, gso_size will be zero,
1439 * in theory this shouldn't be necessary but as long as DSACK
1440 * code can come after this skb later on it's better to keep
1441 * setting gso_size to something.
1442 */
1443 if (!skb_shinfo(prev)->gso_size) {
1444 skb_shinfo(prev)->gso_size = mss;
1445 skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1446 }
1447
1448 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1449 if (skb_shinfo(skb)->gso_segs <= 1) {
1450 skb_shinfo(skb)->gso_size = 0;
1451 skb_shinfo(skb)->gso_type = 0;
1452 }
1453
1454 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1455 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1456
1457 if (skb->len > 0) {
1458 BUG_ON(!tcp_skb_pcount(skb));
1459 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1460 return false;
1461 }
1462
1463 /* Whole SKB was eaten :-) */
1464
1465 if (skb == tp->retransmit_skb_hint)
1466 tp->retransmit_skb_hint = prev;
1467 if (skb == tp->scoreboard_skb_hint)
1468 tp->scoreboard_skb_hint = prev;
1469 if (skb == tp->lost_skb_hint) {
1470 tp->lost_skb_hint = prev;
1471 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1472 }
1473
1474 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(prev)->tcp_flags;
1475 if (skb == tcp_highest_sack(sk))
1476 tcp_advance_highest_sack(sk, skb);
1477
1478 tcp_unlink_write_queue(skb, sk);
1479 sk_wmem_free_skb(sk, skb);
1480
1481 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1482
1483 return true;
1484}
1485
1486/* I wish gso_size would have a bit more sane initialization than
1487 * something-or-zero which complicates things
1488 */
1489static int tcp_skb_seglen(const struct sk_buff *skb)
1490{
1491 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1492}
1493
1494/* Shifting pages past head area doesn't work */
1495static int skb_can_shift(const struct sk_buff *skb)
1496{
1497 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1498}
1499
1500/* Try collapsing SACK blocks spanning across multiple skbs to a single
1501 * skb.
1502 */
1503static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1504 struct tcp_sacktag_state *state,
1505 u32 start_seq, u32 end_seq,
1506 bool dup_sack)
1507{
1508 struct tcp_sock *tp = tcp_sk(sk);
1509 struct sk_buff *prev;
1510 int mss;
1511 int pcount = 0;
1512 int len;
1513 int in_sack;
1514
1515 if (!sk_can_gso(sk))
1516 goto fallback;
1517
1518 /* Normally R but no L won't result in plain S */
1519 if (!dup_sack &&
1520 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1521 goto fallback;
1522 if (!skb_can_shift(skb))
1523 goto fallback;
1524 /* This frame is about to be dropped (was ACKed). */
1525 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1526 goto fallback;
1527
1528 /* Can only happen with delayed DSACK + discard craziness */
1529 if (unlikely(skb == tcp_write_queue_head(sk)))
1530 goto fallback;
1531 prev = tcp_write_queue_prev(sk, skb);
1532
1533 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1534 goto fallback;
1535
1536 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1537 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1538
1539 if (in_sack) {
1540 len = skb->len;
1541 pcount = tcp_skb_pcount(skb);
1542 mss = tcp_skb_seglen(skb);
1543
1544 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1545 * drop this restriction as unnecessary
1546 */
1547 if (mss != tcp_skb_seglen(prev))
1548 goto fallback;
1549 } else {
1550 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1551 goto noop;
1552 /* CHECKME: This is non-MSS split case only?, this will
1553 * cause skipped skbs due to advancing loop btw, original
1554 * has that feature too
1555 */
1556 if (tcp_skb_pcount(skb) <= 1)
1557 goto noop;
1558
1559 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1560 if (!in_sack) {
1561 /* TODO: head merge to next could be attempted here
1562 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1563 * though it might not be worth of the additional hassle
1564 *
1565 * ...we can probably just fallback to what was done
1566 * previously. We could try merging non-SACKed ones
1567 * as well but it probably isn't going to buy off
1568 * because later SACKs might again split them, and
1569 * it would make skb timestamp tracking considerably
1570 * harder problem.
1571 */
1572 goto fallback;
1573 }
1574
1575 len = end_seq - TCP_SKB_CB(skb)->seq;
1576 BUG_ON(len < 0);
1577 BUG_ON(len > skb->len);
1578
1579 /* MSS boundaries should be honoured or else pcount will
1580 * severely break even though it makes things bit trickier.
1581 * Optimize common case to avoid most of the divides
1582 */
1583 mss = tcp_skb_mss(skb);
1584
1585 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1586 * drop this restriction as unnecessary
1587 */
1588 if (mss != tcp_skb_seglen(prev))
1589 goto fallback;
1590
1591 if (len == mss) {
1592 pcount = 1;
1593 } else if (len < mss) {
1594 goto noop;
1595 } else {
1596 pcount = len / mss;
1597 len = pcount * mss;
1598 }
1599 }
1600
1601 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1602 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1603 goto fallback;
1604
1605 if (!skb_shift(prev, skb, len))
1606 goto fallback;
1607 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1608 goto out;
1609
1610 /* Hole filled allows collapsing with the next as well, this is very
1611 * useful when hole on every nth skb pattern happens
1612 */
1613 if (prev == tcp_write_queue_tail(sk))
1614 goto out;
1615 skb = tcp_write_queue_next(sk, prev);
1616
1617 if (!skb_can_shift(skb) ||
1618 (skb == tcp_send_head(sk)) ||
1619 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1620 (mss != tcp_skb_seglen(skb)))
1621 goto out;
1622
1623 len = skb->len;
1624 if (skb_shift(prev, skb, len)) {
1625 pcount += tcp_skb_pcount(skb);
1626 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1627 }
1628
1629out:
1630 state->fack_count += pcount;
1631 return prev;
1632
1633noop:
1634 return skb;
1635
1636fallback:
1637 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1638 return NULL;
1639}
1640
1641static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1642 struct tcp_sack_block *next_dup,
1643 struct tcp_sacktag_state *state,
1644 u32 start_seq, u32 end_seq,
1645 bool dup_sack_in)
1646{
1647 struct tcp_sock *tp = tcp_sk(sk);
1648 struct sk_buff *tmp;
1649
1650 tcp_for_write_queue_from(skb, sk) {
1651 int in_sack = 0;
1652 bool dup_sack = dup_sack_in;
1653
1654 if (skb == tcp_send_head(sk))
1655 break;
1656
1657 /* queue is in-order => we can short-circuit the walk early */
1658 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1659 break;
1660
1661 if ((next_dup != NULL) &&
1662 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1663 in_sack = tcp_match_skb_to_sack(sk, skb,
1664 next_dup->start_seq,
1665 next_dup->end_seq);
1666 if (in_sack > 0)
1667 dup_sack = true;
1668 }
1669
1670 /* skb reference here is a bit tricky to get right, since
1671 * shifting can eat and free both this skb and the next,
1672 * so not even _safe variant of the loop is enough.
1673 */
1674 if (in_sack <= 0) {
1675 tmp = tcp_shift_skb_data(sk, skb, state,
1676 start_seq, end_seq, dup_sack);
1677 if (tmp != NULL) {
1678 if (tmp != skb) {
1679 skb = tmp;
1680 continue;
1681 }
1682
1683 in_sack = 0;
1684 } else {
1685 in_sack = tcp_match_skb_to_sack(sk, skb,
1686 start_seq,
1687 end_seq);
1688 }
1689 }
1690
1691 if (unlikely(in_sack < 0))
1692 break;
1693
1694 if (in_sack) {
1695 TCP_SKB_CB(skb)->sacked =
1696 tcp_sacktag_one(sk,
1697 state,
1698 TCP_SKB_CB(skb)->sacked,
1699 TCP_SKB_CB(skb)->seq,
1700 TCP_SKB_CB(skb)->end_seq,
1701 dup_sack,
1702 tcp_skb_pcount(skb));
1703
1704 if (!before(TCP_SKB_CB(skb)->seq,
1705 tcp_highest_sack_seq(tp)))
1706 tcp_advance_highest_sack(sk, skb);
1707 }
1708
1709 state->fack_count += tcp_skb_pcount(skb);
1710 }
1711 return skb;
1712}
1713
1714/* Avoid all extra work that is being done by sacktag while walking in
1715 * a normal way
1716 */
1717static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1718 struct tcp_sacktag_state *state,
1719 u32 skip_to_seq)
1720{
1721 tcp_for_write_queue_from(skb, sk) {
1722 if (skb == tcp_send_head(sk))
1723 break;
1724
1725 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1726 break;
1727
1728 state->fack_count += tcp_skb_pcount(skb);
1729 }
1730 return skb;
1731}
1732
1733static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1734 struct sock *sk,
1735 struct tcp_sack_block *next_dup,
1736 struct tcp_sacktag_state *state,
1737 u32 skip_to_seq)
1738{
1739 if (next_dup == NULL)
1740 return skb;
1741
1742 if (before(next_dup->start_seq, skip_to_seq)) {
1743 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1744 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1745 next_dup->start_seq, next_dup->end_seq,
1746 1);
1747 }
1748
1749 return skb;
1750}
1751
1752static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1753{
1754 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1755}
1756
1757static int
1758tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1759 u32 prior_snd_una)
1760{
1761 const struct inet_connection_sock *icsk = inet_csk(sk);
1762 struct tcp_sock *tp = tcp_sk(sk);
1763 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1764 TCP_SKB_CB(ack_skb)->sacked);
1765 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1766 struct tcp_sack_block sp[TCP_NUM_SACKS];
1767 struct tcp_sack_block *cache;
1768 struct tcp_sacktag_state state;
1769 struct sk_buff *skb;
1770 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1771 int used_sacks;
1772 bool found_dup_sack = false;
1773 int i, j;
1774 int first_sack_index;
1775
1776 state.flag = 0;
1777 state.reord = tp->packets_out;
1778
1779 if (!tp->sacked_out) {
1780 if (WARN_ON(tp->fackets_out))
1781 tp->fackets_out = 0;
1782 tcp_highest_sack_reset(sk);
1783 }
1784
1785 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1786 num_sacks, prior_snd_una);
1787 if (found_dup_sack)
1788 state.flag |= FLAG_DSACKING_ACK;
1789
1790 /* Eliminate too old ACKs, but take into
1791 * account more or less fresh ones, they can
1792 * contain valid SACK info.
1793 */
1794 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1795 return 0;
1796
1797 if (!tp->packets_out)
1798 goto out;
1799
1800 used_sacks = 0;
1801 first_sack_index = 0;
1802 for (i = 0; i < num_sacks; i++) {
1803 bool dup_sack = !i && found_dup_sack;
1804
1805 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1806 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1807
1808 if (!tcp_is_sackblock_valid(tp, dup_sack,
1809 sp[used_sacks].start_seq,
1810 sp[used_sacks].end_seq)) {
1811 int mib_idx;
1812
1813 if (dup_sack) {
1814 if (!tp->undo_marker)
1815 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1816 else
1817 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1818 } else {
1819 /* Don't count olds caused by ACK reordering */
1820 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1821 !after(sp[used_sacks].end_seq, tp->snd_una))
1822 continue;
1823 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1824 }
1825
1826 NET_INC_STATS_BH(sock_net(sk), mib_idx);
1827 if (i == 0)
1828 first_sack_index = -1;
1829 continue;
1830 }
1831
1832 /* Ignore very old stuff early */
1833 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1834 continue;
1835
1836 used_sacks++;
1837 }
1838
1839 /* order SACK blocks to allow in order walk of the retrans queue */
1840 for (i = used_sacks - 1; i > 0; i--) {
1841 for (j = 0; j < i; j++) {
1842 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1843 swap(sp[j], sp[j + 1]);
1844
1845 /* Track where the first SACK block goes to */
1846 if (j == first_sack_index)
1847 first_sack_index = j + 1;
1848 }
1849 }
1850 }
1851
1852 skb = tcp_write_queue_head(sk);
1853 state.fack_count = 0;
1854 i = 0;
1855
1856 if (!tp->sacked_out) {
1857 /* It's already past, so skip checking against it */
1858 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1859 } else {
1860 cache = tp->recv_sack_cache;
1861 /* Skip empty blocks in at head of the cache */
1862 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1863 !cache->end_seq)
1864 cache++;
1865 }
1866
1867 while (i < used_sacks) {
1868 u32 start_seq = sp[i].start_seq;
1869 u32 end_seq = sp[i].end_seq;
1870 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1871 struct tcp_sack_block *next_dup = NULL;
1872
1873 if (found_dup_sack && ((i + 1) == first_sack_index))
1874 next_dup = &sp[i + 1];
1875
1876 /* Skip too early cached blocks */
1877 while (tcp_sack_cache_ok(tp, cache) &&
1878 !before(start_seq, cache->end_seq))
1879 cache++;
1880
1881 /* Can skip some work by looking recv_sack_cache? */
1882 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1883 after(end_seq, cache->start_seq)) {
1884
1885 /* Head todo? */
1886 if (before(start_seq, cache->start_seq)) {
1887 skb = tcp_sacktag_skip(skb, sk, &state,
1888 start_seq);
1889 skb = tcp_sacktag_walk(skb, sk, next_dup,
1890 &state,
1891 start_seq,
1892 cache->start_seq,
1893 dup_sack);
1894 }
1895
1896 /* Rest of the block already fully processed? */
1897 if (!after(end_seq, cache->end_seq))
1898 goto advance_sp;
1899
1900 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1901 &state,
1902 cache->end_seq);
1903
1904 /* ...tail remains todo... */
1905 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1906 /* ...but better entrypoint exists! */
1907 skb = tcp_highest_sack(sk);
1908 if (skb == NULL)
1909 break;
1910 state.fack_count = tp->fackets_out;
1911 cache++;
1912 goto walk;
1913 }
1914
1915 skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
1916 /* Check overlap against next cached too (past this one already) */
1917 cache++;
1918 continue;
1919 }
1920
1921 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1922 skb = tcp_highest_sack(sk);
1923 if (skb == NULL)
1924 break;
1925 state.fack_count = tp->fackets_out;
1926 }
1927 skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
1928
1929walk:
1930 skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1931 start_seq, end_seq, dup_sack);
1932
1933advance_sp:
1934 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1935 * due to in-order walk
1936 */
1937 if (after(end_seq, tp->frto_highmark))
1938 state.flag &= ~FLAG_ONLY_ORIG_SACKED;
1939
1940 i++;
1941 }
1942
1943 /* Clear the head of the cache sack blocks so we can skip it next time */
1944 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1945 tp->recv_sack_cache[i].start_seq = 0;
1946 tp->recv_sack_cache[i].end_seq = 0;
1947 }
1948 for (j = 0; j < used_sacks; j++)
1949 tp->recv_sack_cache[i++] = sp[j];
1950
1951 tcp_mark_lost_retrans(sk);
1952
1953 tcp_verify_left_out(tp);
1954
1955 if ((state.reord < tp->fackets_out) &&
1956 ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) &&
1957 (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1958 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
1959
1960out:
1961
1962#if FASTRETRANS_DEBUG > 0
1963 WARN_ON((int)tp->sacked_out < 0);
1964 WARN_ON((int)tp->lost_out < 0);
1965 WARN_ON((int)tp->retrans_out < 0);
1966 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1967#endif
1968 return state.flag;
1969}
1970
1971/* Limits sacked_out so that sum with lost_out isn't ever larger than
1972 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1973 */
1974static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1975{
1976 u32 holes;
1977
1978 holes = max(tp->lost_out, 1U);
1979 holes = min(holes, tp->packets_out);
1980
1981 if ((tp->sacked_out + holes) > tp->packets_out) {
1982 tp->sacked_out = tp->packets_out - holes;
1983 return true;
1984 }
1985 return false;
1986}
1987
1988/* If we receive more dupacks than we expected counting segments
1989 * in assumption of absent reordering, interpret this as reordering.
1990 * The only another reason could be bug in receiver TCP.
1991 */
1992static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1993{
1994 struct tcp_sock *tp = tcp_sk(sk);
1995 if (tcp_limit_reno_sacked(tp))
1996 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1997}
1998
1999/* Emulate SACKs for SACKless connection: account for a new dupack. */
2000
2001static void tcp_add_reno_sack(struct sock *sk)
2002{
2003 struct tcp_sock *tp = tcp_sk(sk);
2004 tp->sacked_out++;
2005 tcp_check_reno_reordering(sk, 0);
2006 tcp_verify_left_out(tp);
2007}
2008
2009/* Account for ACK, ACKing some data in Reno Recovery phase. */
2010
2011static void tcp_remove_reno_sacks(struct sock *sk, int acked)
2012{
2013 struct tcp_sock *tp = tcp_sk(sk);
2014
2015 if (acked > 0) {
2016 /* One ACK acked hole. The rest eat duplicate ACKs. */
2017 if (acked - 1 >= tp->sacked_out)
2018 tp->sacked_out = 0;
2019 else
2020 tp->sacked_out -= acked - 1;
2021 }
2022 tcp_check_reno_reordering(sk, acked);
2023 tcp_verify_left_out(tp);
2024}
2025
2026static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
2027{
2028 tp->sacked_out = 0;
2029}
2030
2031static int tcp_is_sackfrto(const struct tcp_sock *tp)
2032{
2033 return (sysctl_tcp_frto == 0x2) && !tcp_is_reno(tp);
2034}
2035
2036/* F-RTO can only be used if TCP has never retransmitted anything other than
2037 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
2038 */
2039bool tcp_use_frto(struct sock *sk)
2040{
2041 const struct tcp_sock *tp = tcp_sk(sk);
2042 const struct inet_connection_sock *icsk = inet_csk(sk);
2043 struct sk_buff *skb;
2044
2045 if (!sysctl_tcp_frto)
2046 return false;
2047
2048 /* MTU probe and F-RTO won't really play nicely along currently */
2049 if (icsk->icsk_mtup.probe_size)
2050 return false;
2051
2052 if (tcp_is_sackfrto(tp))
2053 return true;
2054
2055 /* Avoid expensive walking of rexmit queue if possible */
2056 if (tp->retrans_out > 1)
2057 return false;
2058
2059 skb = tcp_write_queue_head(sk);
2060 if (tcp_skb_is_last(sk, skb))
2061 return true;
2062 skb = tcp_write_queue_next(sk, skb); /* Skips head */
2063 tcp_for_write_queue_from(skb, sk) {
2064 if (skb == tcp_send_head(sk))
2065 break;
2066 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2067 return false;
2068 /* Short-circuit when first non-SACKed skb has been checked */
2069 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2070 break;
2071 }
2072 return true;
2073}
2074
2075/* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
2076 * recovery a bit and use heuristics in tcp_process_frto() to detect if
2077 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
2078 * keep retrans_out counting accurate (with SACK F-RTO, other than head
2079 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
2080 * bits are handled if the Loss state is really to be entered (in
2081 * tcp_enter_frto_loss).
2082 *
2083 * Do like tcp_enter_loss() would; when RTO expires the second time it
2084 * does:
2085 * "Reduce ssthresh if it has not yet been made inside this window."
2086 */
2087void tcp_enter_frto(struct sock *sk)
2088{
2089 const struct inet_connection_sock *icsk = inet_csk(sk);
2090 struct tcp_sock *tp = tcp_sk(sk);
2091 struct sk_buff *skb;
2092
2093 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
2094 tp->snd_una == tp->high_seq ||
2095 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
2096 !icsk->icsk_retransmits)) {
2097 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2098 /* Our state is too optimistic in ssthresh() call because cwnd
2099 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
2100 * recovery has not yet completed. Pattern would be this: RTO,
2101 * Cumulative ACK, RTO (2xRTO for the same segment does not end
2102 * up here twice).
2103 * RFC4138 should be more specific on what to do, even though
2104 * RTO is quite unlikely to occur after the first Cumulative ACK
2105 * due to back-off and complexity of triggering events ...
2106 */
2107 if (tp->frto_counter) {
2108 u32 stored_cwnd;
2109 stored_cwnd = tp->snd_cwnd;
2110 tp->snd_cwnd = 2;
2111 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2112 tp->snd_cwnd = stored_cwnd;
2113 } else {
2114 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2115 }
2116 /* ... in theory, cong.control module could do "any tricks" in
2117 * ssthresh(), which means that ca_state, lost bits and lost_out
2118 * counter would have to be faked before the call occurs. We
2119 * consider that too expensive, unlikely and hacky, so modules
2120 * using these in ssthresh() must deal these incompatibility
2121 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
2122 */
2123 tcp_ca_event(sk, CA_EVENT_FRTO);
2124 }
2125
2126 tp->undo_marker = tp->snd_una;
2127 tp->undo_retrans = 0;
2128
2129 skb = tcp_write_queue_head(sk);
2130 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2131 tp->undo_marker = 0;
2132 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2133 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2134 tp->retrans_out -= tcp_skb_pcount(skb);
2135 }
2136 tcp_verify_left_out(tp);
2137
2138 /* Too bad if TCP was application limited */
2139 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2140
2141 /* Earlier loss recovery underway (see RFC4138; Appendix B).
2142 * The last condition is necessary at least in tp->frto_counter case.
2143 */
2144 if (tcp_is_sackfrto(tp) && (tp->frto_counter ||
2145 ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
2146 after(tp->high_seq, tp->snd_una)) {
2147 tp->frto_highmark = tp->high_seq;
2148 } else {
2149 tp->frto_highmark = tp->snd_nxt;
2150 }
2151 tcp_set_ca_state(sk, TCP_CA_Disorder);
2152 tp->high_seq = tp->snd_nxt;
2153 tp->frto_counter = 1;
2154}
2155
2156/* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
2157 * which indicates that we should follow the traditional RTO recovery,
2158 * i.e. mark everything lost and do go-back-N retransmission.
2159 */
2160static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
2161{
2162 struct tcp_sock *tp = tcp_sk(sk);
2163 struct sk_buff *skb;
2164
2165 tp->lost_out = 0;
2166 tp->retrans_out = 0;
2167 if (tcp_is_reno(tp))
2168 tcp_reset_reno_sack(tp);
2169
2170 tcp_for_write_queue(skb, sk) {
2171 if (skb == tcp_send_head(sk))
2172 break;
2173
2174 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2175 /*
2176 * Count the retransmission made on RTO correctly (only when
2177 * waiting for the first ACK and did not get it)...
2178 */
2179 if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) {
2180 /* For some reason this R-bit might get cleared? */
2181 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
2182 tp->retrans_out += tcp_skb_pcount(skb);
2183 /* ...enter this if branch just for the first segment */
2184 flag |= FLAG_DATA_ACKED;
2185 } else {
2186 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2187 tp->undo_marker = 0;
2188 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2189 }
2190
2191 /* Marking forward transmissions that were made after RTO lost
2192 * can cause unnecessary retransmissions in some scenarios,
2193 * SACK blocks will mitigate that in some but not in all cases.
2194 * We used to not mark them but it was causing break-ups with
2195 * receivers that do only in-order receival.
2196 *
2197 * TODO: we could detect presence of such receiver and select
2198 * different behavior per flow.
2199 */
2200 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2201 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2202 tp->lost_out += tcp_skb_pcount(skb);
2203 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2204 }
2205 }
2206 tcp_verify_left_out(tp);
2207
2208 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
2209 tp->snd_cwnd_cnt = 0;
2210 tp->snd_cwnd_stamp = tcp_time_stamp;
2211 tp->frto_counter = 0;
2212 tp->bytes_acked = 0;
2213
2214 tp->reordering = min_t(unsigned int, tp->reordering,
2215 sysctl_tcp_reordering);
2216 tcp_set_ca_state(sk, TCP_CA_Loss);
2217 tp->high_seq = tp->snd_nxt;
2218 TCP_ECN_queue_cwr(tp);
2219
2220 tcp_clear_all_retrans_hints(tp);
2221}
2222
2223static void tcp_clear_retrans_partial(struct tcp_sock *tp)
2224{
2225 tp->retrans_out = 0;
2226 tp->lost_out = 0;
2227
2228 tp->undo_marker = 0;
2229 tp->undo_retrans = 0;
2230}
2231
2232void tcp_clear_retrans(struct tcp_sock *tp)
2233{
2234 tcp_clear_retrans_partial(tp);
2235
2236 tp->fackets_out = 0;
2237 tp->sacked_out = 0;
2238}
2239
2240/* Enter Loss state. If "how" is not zero, forget all SACK information
2241 * and reset tags completely, otherwise preserve SACKs. If receiver
2242 * dropped its ofo queue, we will know this due to reneging detection.
2243 */
2244void tcp_enter_loss(struct sock *sk, int how)
2245{
2246 const struct inet_connection_sock *icsk = inet_csk(sk);
2247 struct tcp_sock *tp = tcp_sk(sk);
2248 struct sk_buff *skb;
2249
2250 /* Reduce ssthresh if it has not yet been made inside this window. */
2251 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
2252 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2253 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2254 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2255 tcp_ca_event(sk, CA_EVENT_LOSS);
2256 }
2257 tp->snd_cwnd = 1;
2258 tp->snd_cwnd_cnt = 0;
2259 tp->snd_cwnd_stamp = tcp_time_stamp;
2260
2261 tp->bytes_acked = 0;
2262 tcp_clear_retrans_partial(tp);
2263
2264 if (tcp_is_reno(tp))
2265 tcp_reset_reno_sack(tp);
2266
2267 if (!how) {
2268 /* Push undo marker, if it was plain RTO and nothing
2269 * was retransmitted. */
2270 tp->undo_marker = tp->snd_una;
2271 } else {
2272 tp->sacked_out = 0;
2273 tp->fackets_out = 0;
2274 }
2275 tcp_clear_all_retrans_hints(tp);
2276
2277 tcp_for_write_queue(skb, sk) {
2278 if (skb == tcp_send_head(sk))
2279 break;
2280
2281 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2282 tp->undo_marker = 0;
2283 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
2284 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
2285 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2286 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2287 tp->lost_out += tcp_skb_pcount(skb);
2288 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
2289 }
2290 }
2291 tcp_verify_left_out(tp);
2292
2293 tp->reordering = min_t(unsigned int, tp->reordering,
2294 sysctl_tcp_reordering);
2295 tcp_set_ca_state(sk, TCP_CA_Loss);
2296 tp->high_seq = tp->snd_nxt;
2297 TCP_ECN_queue_cwr(tp);
2298 /* Abort F-RTO algorithm if one is in progress */
2299 tp->frto_counter = 0;
2300}
2301
2302/* If ACK arrived pointing to a remembered SACK, it means that our
2303 * remembered SACKs do not reflect real state of receiver i.e.
2304 * receiver _host_ is heavily congested (or buggy).
2305 *
2306 * Do processing similar to RTO timeout.
2307 */
2308static bool tcp_check_sack_reneging(struct sock *sk, int flag)
2309{
2310 if (flag & FLAG_SACK_RENEGING) {
2311 struct inet_connection_sock *icsk = inet_csk(sk);
2312 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2313
2314 tcp_enter_loss(sk, 1);
2315 icsk->icsk_retransmits++;
2316 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
2317 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2318 icsk->icsk_rto, TCP_RTO_MAX);
2319 return true;
2320 }
2321 return false;
2322}
2323
2324static inline int tcp_fackets_out(const struct tcp_sock *tp)
2325{
2326 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2327}
2328
2329/* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2330 * counter when SACK is enabled (without SACK, sacked_out is used for
2331 * that purpose).
2332 *
2333 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2334 * segments up to the highest received SACK block so far and holes in
2335 * between them.
2336 *
2337 * With reordering, holes may still be in flight, so RFC3517 recovery
2338 * uses pure sacked_out (total number of SACKed segments) even though
2339 * it violates the RFC that uses duplicate ACKs, often these are equal
2340 * but when e.g. out-of-window ACKs or packet duplication occurs,
2341 * they differ. Since neither occurs due to loss, TCP should really
2342 * ignore them.
2343 */
2344static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2345{
2346 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2347}
2348
2349static bool tcp_pause_early_retransmit(struct sock *sk, int flag)
2350{
2351 struct tcp_sock *tp = tcp_sk(sk);
2352 unsigned long delay;
2353
2354 /* Delay early retransmit and entering fast recovery for
2355 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
2356 * available, or RTO is scheduled to fire first.
2357 */
2358 if (sysctl_tcp_early_retrans < 2 || (flag & FLAG_ECE) || !tp->srtt)
2359 return false;
2360
2361 delay = max_t(unsigned long, (tp->srtt >> 5), msecs_to_jiffies(2));
2362 if (!time_after(inet_csk(sk)->icsk_timeout, (jiffies + delay)))
2363 return false;
2364
2365 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, delay, TCP_RTO_MAX);
2366 tp->early_retrans_delayed = 1;
2367 return true;
2368}
2369
2370static inline int tcp_skb_timedout(const struct sock *sk,
2371 const struct sk_buff *skb)
2372{
2373 return tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto;
2374}
2375
2376static inline int tcp_head_timedout(const struct sock *sk)
2377{
2378 const struct tcp_sock *tp = tcp_sk(sk);
2379
2380 return tp->packets_out &&
2381 tcp_skb_timedout(sk, tcp_write_queue_head(sk));
2382}
2383
2384/* Linux NewReno/SACK/FACK/ECN state machine.
2385 * --------------------------------------
2386 *
2387 * "Open" Normal state, no dubious events, fast path.
2388 * "Disorder" In all the respects it is "Open",
2389 * but requires a bit more attention. It is entered when
2390 * we see some SACKs or dupacks. It is split of "Open"
2391 * mainly to move some processing from fast path to slow one.
2392 * "CWR" CWND was reduced due to some Congestion Notification event.
2393 * It can be ECN, ICMP source quench, local device congestion.
2394 * "Recovery" CWND was reduced, we are fast-retransmitting.
2395 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2396 *
2397 * tcp_fastretrans_alert() is entered:
2398 * - each incoming ACK, if state is not "Open"
2399 * - when arrived ACK is unusual, namely:
2400 * * SACK
2401 * * Duplicate ACK.
2402 * * ECN ECE.
2403 *
2404 * Counting packets in flight is pretty simple.
2405 *
2406 * in_flight = packets_out - left_out + retrans_out
2407 *
2408 * packets_out is SND.NXT-SND.UNA counted in packets.
2409 *
2410 * retrans_out is number of retransmitted segments.
2411 *
2412 * left_out is number of segments left network, but not ACKed yet.
2413 *
2414 * left_out = sacked_out + lost_out
2415 *
2416 * sacked_out: Packets, which arrived to receiver out of order
2417 * and hence not ACKed. With SACKs this number is simply
2418 * amount of SACKed data. Even without SACKs
2419 * it is easy to give pretty reliable estimate of this number,
2420 * counting duplicate ACKs.
2421 *
2422 * lost_out: Packets lost by network. TCP has no explicit
2423 * "loss notification" feedback from network (for now).
2424 * It means that this number can be only _guessed_.
2425 * Actually, it is the heuristics to predict lossage that
2426 * distinguishes different algorithms.
2427 *
2428 * F.e. after RTO, when all the queue is considered as lost,
2429 * lost_out = packets_out and in_flight = retrans_out.
2430 *
2431 * Essentially, we have now two algorithms counting
2432 * lost packets.
2433 *
2434 * FACK: It is the simplest heuristics. As soon as we decided
2435 * that something is lost, we decide that _all_ not SACKed
2436 * packets until the most forward SACK are lost. I.e.
2437 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2438 * It is absolutely correct estimate, if network does not reorder
2439 * packets. And it loses any connection to reality when reordering
2440 * takes place. We use FACK by default until reordering
2441 * is suspected on the path to this destination.
2442 *
2443 * NewReno: when Recovery is entered, we assume that one segment
2444 * is lost (classic Reno). While we are in Recovery and
2445 * a partial ACK arrives, we assume that one more packet
2446 * is lost (NewReno). This heuristics are the same in NewReno
2447 * and SACK.
2448 *
2449 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2450 * deflation etc. CWND is real congestion window, never inflated, changes
2451 * only according to classic VJ rules.
2452 *
2453 * Really tricky (and requiring careful tuning) part of algorithm
2454 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2455 * The first determines the moment _when_ we should reduce CWND and,
2456 * hence, slow down forward transmission. In fact, it determines the moment
2457 * when we decide that hole is caused by loss, rather than by a reorder.
2458 *
2459 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2460 * holes, caused by lost packets.
2461 *
2462 * And the most logically complicated part of algorithm is undo
2463 * heuristics. We detect false retransmits due to both too early
2464 * fast retransmit (reordering) and underestimated RTO, analyzing
2465 * timestamps and D-SACKs. When we detect that some segments were
2466 * retransmitted by mistake and CWND reduction was wrong, we undo
2467 * window reduction and abort recovery phase. This logic is hidden
2468 * inside several functions named tcp_try_undo_<something>.
2469 */
2470
2471/* This function decides, when we should leave Disordered state
2472 * and enter Recovery phase, reducing congestion window.
2473 *
2474 * Main question: may we further continue forward transmission
2475 * with the same cwnd?
2476 */
2477static bool tcp_time_to_recover(struct sock *sk, int flag)
2478{
2479 struct tcp_sock *tp = tcp_sk(sk);
2480 __u32 packets_out;
2481
2482 /* Do not perform any recovery during F-RTO algorithm */
2483 if (tp->frto_counter)
2484 return false;
2485
2486 /* Trick#1: The loss is proven. */
2487 if (tp->lost_out)
2488 return true;
2489
2490 /* Not-A-Trick#2 : Classic rule... */
2491 if (tcp_dupack_heuristics(tp) > tp->reordering)
2492 return true;
2493
2494 /* Trick#3 : when we use RFC2988 timer restart, fast
2495 * retransmit can be triggered by timeout of queue head.
2496 */
2497 if (tcp_is_fack(tp) && tcp_head_timedout(sk))
2498 return true;
2499
2500 /* Trick#4: It is still not OK... But will it be useful to delay
2501 * recovery more?
2502 */
2503 packets_out = tp->packets_out;
2504 if (packets_out <= tp->reordering &&
2505 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2506 !tcp_may_send_now(sk)) {
2507 /* We have nothing to send. This connection is limited
2508 * either by receiver window or by application.
2509 */
2510 return true;
2511 }
2512
2513 /* If a thin stream is detected, retransmit after first
2514 * received dupack. Employ only if SACK is supported in order
2515 * to avoid possible corner-case series of spurious retransmissions
2516 * Use only if there are no unsent data.
2517 */
2518 if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2519 tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2520 tcp_is_sack(tp) && !tcp_send_head(sk))
2521 return true;
2522
2523 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious
2524 * retransmissions due to small network reorderings, we implement
2525 * Mitigation A.3 in the RFC and delay the retransmission for a short
2526 * interval if appropriate.
2527 */
2528 if (tp->do_early_retrans && !tp->retrans_out && tp->sacked_out &&
2529 (tp->packets_out == (tp->sacked_out + 1) && tp->packets_out < 4) &&
2530 !tcp_may_send_now(sk))
2531 return !tcp_pause_early_retransmit(sk, flag);
2532
2533 return false;
2534}
2535
2536/* New heuristics: it is possible only after we switched to restart timer
2537 * each time when something is ACKed. Hence, we can detect timed out packets
2538 * during fast retransmit without falling to slow start.
2539 *
2540 * Usefulness of this as is very questionable, since we should know which of
2541 * the segments is the next to timeout which is relatively expensive to find
2542 * in general case unless we add some data structure just for that. The
2543 * current approach certainly won't find the right one too often and when it
2544 * finally does find _something_ it usually marks large part of the window
2545 * right away (because a retransmission with a larger timestamp blocks the
2546 * loop from advancing). -ij
2547 */
2548static void tcp_timeout_skbs(struct sock *sk)
2549{
2550 struct tcp_sock *tp = tcp_sk(sk);
2551 struct sk_buff *skb;
2552
2553 if (!tcp_is_fack(tp) || !tcp_head_timedout(sk))
2554 return;
2555
2556 skb = tp->scoreboard_skb_hint;
2557 if (tp->scoreboard_skb_hint == NULL)
2558 skb = tcp_write_queue_head(sk);
2559
2560 tcp_for_write_queue_from(skb, sk) {
2561 if (skb == tcp_send_head(sk))
2562 break;
2563 if (!tcp_skb_timedout(sk, skb))
2564 break;
2565
2566 tcp_skb_mark_lost(tp, skb);
2567 }
2568
2569 tp->scoreboard_skb_hint = skb;
2570
2571 tcp_verify_left_out(tp);
2572}
2573
2574/* Detect loss in event "A" above by marking head of queue up as lost.
2575 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2576 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2577 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2578 * the maximum SACKed segments to pass before reaching this limit.
2579 */
2580static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2581{
2582 struct tcp_sock *tp = tcp_sk(sk);
2583 struct sk_buff *skb;
2584 int cnt, oldcnt;
2585 int err;
2586 unsigned int mss;
2587 /* Use SACK to deduce losses of new sequences sent during recovery */
2588 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
2589
2590 WARN_ON(packets > tp->packets_out);
2591 if (tp->lost_skb_hint) {
2592 skb = tp->lost_skb_hint;
2593 cnt = tp->lost_cnt_hint;
2594 /* Head already handled? */
2595 if (mark_head && skb != tcp_write_queue_head(sk))
2596 return;
2597 } else {
2598 skb = tcp_write_queue_head(sk);
2599 cnt = 0;
2600 }
2601
2602 tcp_for_write_queue_from(skb, sk) {
2603 if (skb == tcp_send_head(sk))
2604 break;
2605 /* TODO: do this better */
2606 /* this is not the most efficient way to do this... */
2607 tp->lost_skb_hint = skb;
2608 tp->lost_cnt_hint = cnt;
2609
2610 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2611 break;
2612
2613 oldcnt = cnt;
2614 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2615 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2616 cnt += tcp_skb_pcount(skb);
2617
2618 if (cnt > packets) {
2619 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2620 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2621 (oldcnt >= packets))
2622 break;
2623
2624 mss = skb_shinfo(skb)->gso_size;
2625 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2626 if (err < 0)
2627 break;
2628 cnt = packets;
2629 }
2630
2631 tcp_skb_mark_lost(tp, skb);
2632
2633 if (mark_head)
2634 break;
2635 }
2636 tcp_verify_left_out(tp);
2637}
2638
2639/* Account newly detected lost packet(s) */
2640
2641static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2642{
2643 struct tcp_sock *tp = tcp_sk(sk);
2644
2645 if (tcp_is_reno(tp)) {
2646 tcp_mark_head_lost(sk, 1, 1);
2647 } else if (tcp_is_fack(tp)) {
2648 int lost = tp->fackets_out - tp->reordering;
2649 if (lost <= 0)
2650 lost = 1;
2651 tcp_mark_head_lost(sk, lost, 0);
2652 } else {
2653 int sacked_upto = tp->sacked_out - tp->reordering;
2654 if (sacked_upto >= 0)
2655 tcp_mark_head_lost(sk, sacked_upto, 0);
2656 else if (fast_rexmit)
2657 tcp_mark_head_lost(sk, 1, 1);
2658 }
2659
2660 tcp_timeout_skbs(sk);
2661}
2662
2663/* CWND moderation, preventing bursts due to too big ACKs
2664 * in dubious situations.
2665 */
2666static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2667{
2668 tp->snd_cwnd = min(tp->snd_cwnd,
2669 tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2670 tp->snd_cwnd_stamp = tcp_time_stamp;
2671}
2672
2673/* Lower bound on congestion window is slow start threshold
2674 * unless congestion avoidance choice decides to overide it.
2675 */
2676static inline u32 tcp_cwnd_min(const struct sock *sk)
2677{
2678 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2679
2680 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
2681}
2682
2683/* Decrease cwnd each second ack. */
2684static void tcp_cwnd_down(struct sock *sk, int flag)
2685{
2686 struct tcp_sock *tp = tcp_sk(sk);
2687 int decr = tp->snd_cwnd_cnt + 1;
2688
2689 if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK)) ||
2690 (tcp_is_reno(tp) && !(flag & FLAG_NOT_DUP))) {
2691 tp->snd_cwnd_cnt = decr & 1;
2692 decr >>= 1;
2693
2694 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
2695 tp->snd_cwnd -= decr;
2696
2697 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2698 tp->snd_cwnd_stamp = tcp_time_stamp;
2699 }
2700}
2701
2702/* Nothing was retransmitted or returned timestamp is less
2703 * than timestamp of the first retransmission.
2704 */
2705static inline int tcp_packet_delayed(const struct tcp_sock *tp)
2706{
2707 return !tp->retrans_stamp ||
2708 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2709 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
2710}
2711
2712/* Undo procedures. */
2713
2714#if FASTRETRANS_DEBUG > 1
2715static void DBGUNDO(struct sock *sk, const char *msg)
2716{
2717 struct tcp_sock *tp = tcp_sk(sk);
2718 struct inet_sock *inet = inet_sk(sk);
2719
2720 if (sk->sk_family == AF_INET) {
2721 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2722 msg,
2723 &inet->inet_daddr, ntohs(inet->inet_dport),
2724 tp->snd_cwnd, tcp_left_out(tp),
2725 tp->snd_ssthresh, tp->prior_ssthresh,
2726 tp->packets_out);
2727 }
2728#if IS_ENABLED(CONFIG_IPV6)
2729 else if (sk->sk_family == AF_INET6) {
2730 struct ipv6_pinfo *np = inet6_sk(sk);
2731 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2732 msg,
2733 &np->daddr, ntohs(inet->inet_dport),
2734 tp->snd_cwnd, tcp_left_out(tp),
2735 tp->snd_ssthresh, tp->prior_ssthresh,
2736 tp->packets_out);
2737 }
2738#endif
2739}
2740#else
2741#define DBGUNDO(x...) do { } while (0)
2742#endif
2743
2744static void tcp_undo_cwr(struct sock *sk, const bool undo_ssthresh)
2745{
2746 struct tcp_sock *tp = tcp_sk(sk);
2747
2748 if (tp->prior_ssthresh) {
2749 const struct inet_connection_sock *icsk = inet_csk(sk);
2750
2751 if (icsk->icsk_ca_ops->undo_cwnd)
2752 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2753 else
2754 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2755
2756 if (undo_ssthresh && tp->prior_ssthresh > tp->snd_ssthresh) {
2757 tp->snd_ssthresh = tp->prior_ssthresh;
2758 TCP_ECN_withdraw_cwr(tp);
2759 }
2760 } else {
2761 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2762 }
2763 tp->snd_cwnd_stamp = tcp_time_stamp;
2764}
2765
2766static inline int tcp_may_undo(const struct tcp_sock *tp)
2767{
2768 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2769}
2770
2771/* People celebrate: "We love our President!" */
2772static bool tcp_try_undo_recovery(struct sock *sk)
2773{
2774 struct tcp_sock *tp = tcp_sk(sk);
2775
2776 if (tcp_may_undo(tp)) {
2777 int mib_idx;
2778
2779 /* Happy end! We did not retransmit anything
2780 * or our original transmission succeeded.
2781 */
2782 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2783 tcp_undo_cwr(sk, true);
2784 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2785 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2786 else
2787 mib_idx = LINUX_MIB_TCPFULLUNDO;
2788
2789 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2790 tp->undo_marker = 0;
2791 }
2792 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2793 /* Hold old state until something *above* high_seq
2794 * is ACKed. For Reno it is MUST to prevent false
2795 * fast retransmits (RFC2582). SACK TCP is safe. */
2796 tcp_moderate_cwnd(tp);
2797 return true;
2798 }
2799 tcp_set_ca_state(sk, TCP_CA_Open);
2800 return false;
2801}
2802
2803/* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2804static void tcp_try_undo_dsack(struct sock *sk)
2805{
2806 struct tcp_sock *tp = tcp_sk(sk);
2807
2808 if (tp->undo_marker && !tp->undo_retrans) {
2809 DBGUNDO(sk, "D-SACK");
2810 tcp_undo_cwr(sk, true);
2811 tp->undo_marker = 0;
2812 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2813 }
2814}
2815
2816/* We can clear retrans_stamp when there are no retransmissions in the
2817 * window. It would seem that it is trivially available for us in
2818 * tp->retrans_out, however, that kind of assumptions doesn't consider
2819 * what will happen if errors occur when sending retransmission for the
2820 * second time. ...It could the that such segment has only
2821 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2822 * the head skb is enough except for some reneging corner cases that
2823 * are not worth the effort.
2824 *
2825 * Main reason for all this complexity is the fact that connection dying
2826 * time now depends on the validity of the retrans_stamp, in particular,
2827 * that successive retransmissions of a segment must not advance
2828 * retrans_stamp under any conditions.
2829 */
2830static bool tcp_any_retrans_done(const struct sock *sk)
2831{
2832 const struct tcp_sock *tp = tcp_sk(sk);
2833 struct sk_buff *skb;
2834
2835 if (tp->retrans_out)
2836 return true;
2837
2838 skb = tcp_write_queue_head(sk);
2839 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2840 return true;
2841
2842 return false;
2843}
2844
2845/* Undo during fast recovery after partial ACK. */
2846
2847static int tcp_try_undo_partial(struct sock *sk, int acked)
2848{
2849 struct tcp_sock *tp = tcp_sk(sk);
2850 /* Partial ACK arrived. Force Hoe's retransmit. */
2851 int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering);
2852
2853 if (tcp_may_undo(tp)) {
2854 /* Plain luck! Hole if filled with delayed
2855 * packet, rather than with a retransmit.
2856 */
2857 if (!tcp_any_retrans_done(sk))
2858 tp->retrans_stamp = 0;
2859
2860 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2861
2862 DBGUNDO(sk, "Hoe");
2863 tcp_undo_cwr(sk, false);
2864 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2865
2866 /* So... Do not make Hoe's retransmit yet.
2867 * If the first packet was delayed, the rest
2868 * ones are most probably delayed as well.
2869 */
2870 failed = 0;
2871 }
2872 return failed;
2873}
2874
2875/* Undo during loss recovery after partial ACK. */
2876static bool tcp_try_undo_loss(struct sock *sk)
2877{
2878 struct tcp_sock *tp = tcp_sk(sk);
2879
2880 if (tcp_may_undo(tp)) {
2881 struct sk_buff *skb;
2882 tcp_for_write_queue(skb, sk) {
2883 if (skb == tcp_send_head(sk))
2884 break;
2885 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2886 }
2887
2888 tcp_clear_all_retrans_hints(tp);
2889
2890 DBGUNDO(sk, "partial loss");
2891 tp->lost_out = 0;
2892 tcp_undo_cwr(sk, true);
2893 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2894 inet_csk(sk)->icsk_retransmits = 0;
2895 tp->undo_marker = 0;
2896 if (tcp_is_sack(tp))
2897 tcp_set_ca_state(sk, TCP_CA_Open);
2898 return true;
2899 }
2900 return false;
2901}
2902
2903static inline void tcp_complete_cwr(struct sock *sk)
2904{
2905 struct tcp_sock *tp = tcp_sk(sk);
2906
2907 /* Do not moderate cwnd if it's already undone in cwr or recovery. */
2908 if (tp->undo_marker) {
2909 if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR) {
2910 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2911 tp->snd_cwnd_stamp = tcp_time_stamp;
2912 } else if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH) {
2913 /* PRR algorithm. */
2914 tp->snd_cwnd = tp->snd_ssthresh;
2915 tp->snd_cwnd_stamp = tcp_time_stamp;
2916 }
2917 }
2918 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2919}
2920
2921static void tcp_try_keep_open(struct sock *sk)
2922{
2923 struct tcp_sock *tp = tcp_sk(sk);
2924 int state = TCP_CA_Open;
2925
2926 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2927 state = TCP_CA_Disorder;
2928
2929 if (inet_csk(sk)->icsk_ca_state != state) {
2930 tcp_set_ca_state(sk, state);
2931 tp->high_seq = tp->snd_nxt;
2932 }
2933}
2934
2935static void tcp_try_to_open(struct sock *sk, int flag)
2936{
2937 struct tcp_sock *tp = tcp_sk(sk);
2938
2939 tcp_verify_left_out(tp);
2940
2941 if (!tp->frto_counter && !tcp_any_retrans_done(sk))
2942 tp->retrans_stamp = 0;
2943
2944 if (flag & FLAG_ECE)
2945 tcp_enter_cwr(sk, 1);
2946
2947 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2948 tcp_try_keep_open(sk);
2949 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
2950 tcp_moderate_cwnd(tp);
2951 } else {
2952 tcp_cwnd_down(sk, flag);
2953 }
2954}
2955
2956static void tcp_mtup_probe_failed(struct sock *sk)
2957{
2958 struct inet_connection_sock *icsk = inet_csk(sk);
2959
2960 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2961 icsk->icsk_mtup.probe_size = 0;
2962}
2963
2964static void tcp_mtup_probe_success(struct sock *sk)
2965{
2966 struct tcp_sock *tp = tcp_sk(sk);
2967 struct inet_connection_sock *icsk = inet_csk(sk);
2968
2969 /* FIXME: breaks with very large cwnd */
2970 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2971 tp->snd_cwnd = tp->snd_cwnd *
2972 tcp_mss_to_mtu(sk, tp->mss_cache) /
2973 icsk->icsk_mtup.probe_size;
2974 tp->snd_cwnd_cnt = 0;
2975 tp->snd_cwnd_stamp = tcp_time_stamp;
2976 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2977
2978 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2979 icsk->icsk_mtup.probe_size = 0;
2980 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2981}
2982
2983/* Do a simple retransmit without using the backoff mechanisms in
2984 * tcp_timer. This is used for path mtu discovery.
2985 * The socket is already locked here.
2986 */
2987void tcp_simple_retransmit(struct sock *sk)
2988{
2989 const struct inet_connection_sock *icsk = inet_csk(sk);
2990 struct tcp_sock *tp = tcp_sk(sk);
2991 struct sk_buff *skb;
2992 unsigned int mss = tcp_current_mss(sk);
2993 u32 prior_lost = tp->lost_out;
2994
2995 tcp_for_write_queue(skb, sk) {
2996 if (skb == tcp_send_head(sk))
2997 break;
2998 if (tcp_skb_seglen(skb) > mss &&
2999 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
3000 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
3001 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
3002 tp->retrans_out -= tcp_skb_pcount(skb);
3003 }
3004 tcp_skb_mark_lost_uncond_verify(tp, skb);
3005 }
3006 }
3007
3008 tcp_clear_retrans_hints_partial(tp);
3009
3010 if (prior_lost == tp->lost_out)
3011 return;
3012
3013 if (tcp_is_reno(tp))
3014 tcp_limit_reno_sacked(tp);
3015
3016 tcp_verify_left_out(tp);
3017
3018 /* Don't muck with the congestion window here.
3019 * Reason is that we do not increase amount of _data_
3020 * in network, but units changed and effective
3021 * cwnd/ssthresh really reduced now.
3022 */
3023 if (icsk->icsk_ca_state != TCP_CA_Loss) {
3024 tp->high_seq = tp->snd_nxt;
3025 tp->snd_ssthresh = tcp_current_ssthresh(sk);
3026 tp->prior_ssthresh = 0;
3027 tp->undo_marker = 0;
3028 tcp_set_ca_state(sk, TCP_CA_Loss);
3029 }
3030 tcp_xmit_retransmit_queue(sk);
3031}
3032EXPORT_SYMBOL(tcp_simple_retransmit);
3033
3034/* This function implements the PRR algorithm, specifcally the PRR-SSRB
3035 * (proportional rate reduction with slow start reduction bound) as described in
3036 * http://www.ietf.org/id/draft-mathis-tcpm-proportional-rate-reduction-01.txt.
3037 * It computes the number of packets to send (sndcnt) based on packets newly
3038 * delivered:
3039 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
3040 * cwnd reductions across a full RTT.
3041 * 2) If packets in flight is lower than ssthresh (such as due to excess
3042 * losses and/or application stalls), do not perform any further cwnd
3043 * reductions, but instead slow start up to ssthresh.
3044 */
3045static void tcp_update_cwnd_in_recovery(struct sock *sk, int newly_acked_sacked,
3046 int fast_rexmit, int flag)
3047{
3048 struct tcp_sock *tp = tcp_sk(sk);
3049 int sndcnt = 0;
3050 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
3051
3052 if (tcp_packets_in_flight(tp) > tp->snd_ssthresh) {
3053 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
3054 tp->prior_cwnd - 1;
3055 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
3056 } else {
3057 sndcnt = min_t(int, delta,
3058 max_t(int, tp->prr_delivered - tp->prr_out,
3059 newly_acked_sacked) + 1);
3060 }
3061
3062 sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0));
3063 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
3064}
3065
3066static void tcp_enter_recovery(struct sock *sk, bool ece_ack)
3067{
3068 struct tcp_sock *tp = tcp_sk(sk);
3069 int mib_idx;
3070
3071 if (tcp_is_reno(tp))
3072 mib_idx = LINUX_MIB_TCPRENORECOVERY;
3073 else
3074 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
3075
3076 NET_INC_STATS_BH(sock_net(sk), mib_idx);
3077
3078 tp->high_seq = tp->snd_nxt;
3079 tp->prior_ssthresh = 0;
3080 tp->undo_marker = tp->snd_una;
3081 tp->undo_retrans = tp->retrans_out;
3082
3083 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
3084 if (!ece_ack)
3085 tp->prior_ssthresh = tcp_current_ssthresh(sk);
3086 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
3087 TCP_ECN_queue_cwr(tp);
3088 }
3089
3090 tp->bytes_acked = 0;
3091 tp->snd_cwnd_cnt = 0;
3092 tp->prior_cwnd = tp->snd_cwnd;
3093 tp->prr_delivered = 0;
3094 tp->prr_out = 0;
3095 tcp_set_ca_state(sk, TCP_CA_Recovery);
3096}
3097
3098/* Process an event, which can update packets-in-flight not trivially.
3099 * Main goal of this function is to calculate new estimate for left_out,
3100 * taking into account both packets sitting in receiver's buffer and
3101 * packets lost by network.
3102 *
3103 * Besides that it does CWND reduction, when packet loss is detected
3104 * and changes state of machine.
3105 *
3106 * It does _not_ decide what to send, it is made in function
3107 * tcp_xmit_retransmit_queue().
3108 */
3109static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked,
3110 int prior_sacked, bool is_dupack,
3111 int flag)
3112{
3113 struct inet_connection_sock *icsk = inet_csk(sk);
3114 struct tcp_sock *tp = tcp_sk(sk);
3115 int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
3116 (tcp_fackets_out(tp) > tp->reordering));
3117 int newly_acked_sacked = 0;
3118 int fast_rexmit = 0;
3119
3120 if (WARN_ON(!tp->packets_out && tp->sacked_out))
3121 tp->sacked_out = 0;
3122 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
3123 tp->fackets_out = 0;
3124
3125 /* Now state machine starts.
3126 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
3127 if (flag & FLAG_ECE)
3128 tp->prior_ssthresh = 0;
3129
3130 /* B. In all the states check for reneging SACKs. */
3131 if (tcp_check_sack_reneging(sk, flag))
3132 return;
3133
3134 /* C. Check consistency of the current state. */
3135 tcp_verify_left_out(tp);
3136
3137 /* D. Check state exit conditions. State can be terminated
3138 * when high_seq is ACKed. */
3139 if (icsk->icsk_ca_state == TCP_CA_Open) {
3140 WARN_ON(tp->retrans_out != 0);
3141 tp->retrans_stamp = 0;
3142 } else if (!before(tp->snd_una, tp->high_seq)) {
3143 switch (icsk->icsk_ca_state) {
3144 case TCP_CA_Loss:
3145 icsk->icsk_retransmits = 0;
3146 if (tcp_try_undo_recovery(sk))
3147 return;
3148 break;
3149
3150 case TCP_CA_CWR:
3151 /* CWR is to be held something *above* high_seq
3152 * is ACKed for CWR bit to reach receiver. */
3153 if (tp->snd_una != tp->high_seq) {
3154 tcp_complete_cwr(sk);
3155 tcp_set_ca_state(sk, TCP_CA_Open);
3156 }
3157 break;
3158
3159 case TCP_CA_Recovery:
3160 if (tcp_is_reno(tp))
3161 tcp_reset_reno_sack(tp);
3162 if (tcp_try_undo_recovery(sk))
3163 return;
3164 tcp_complete_cwr(sk);
3165 break;
3166 }
3167 }
3168
3169 /* E. Process state. */
3170 switch (icsk->icsk_ca_state) {
3171 case TCP_CA_Recovery:
3172 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
3173 if (tcp_is_reno(tp) && is_dupack)
3174 tcp_add_reno_sack(sk);
3175 } else
3176 do_lost = tcp_try_undo_partial(sk, pkts_acked);
3177 newly_acked_sacked = pkts_acked + tp->sacked_out - prior_sacked;
3178 break;
3179 case TCP_CA_Loss:
3180 if (flag & FLAG_DATA_ACKED)
3181 icsk->icsk_retransmits = 0;
3182 if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED)
3183 tcp_reset_reno_sack(tp);
3184 if (!tcp_try_undo_loss(sk)) {
3185 tcp_moderate_cwnd(tp);
3186 tcp_xmit_retransmit_queue(sk);
3187 return;
3188 }
3189 if (icsk->icsk_ca_state != TCP_CA_Open)
3190 return;
3191 /* Loss is undone; fall through to processing in Open state. */
3192 default:
3193 if (tcp_is_reno(tp)) {
3194 if (flag & FLAG_SND_UNA_ADVANCED)
3195 tcp_reset_reno_sack(tp);
3196 if (is_dupack)
3197 tcp_add_reno_sack(sk);
3198 }
3199 newly_acked_sacked = pkts_acked + tp->sacked_out - prior_sacked;
3200
3201 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
3202 tcp_try_undo_dsack(sk);
3203
3204 if (!tcp_time_to_recover(sk, flag)) {
3205 tcp_try_to_open(sk, flag);
3206 return;
3207 }
3208
3209 /* MTU probe failure: don't reduce cwnd */
3210 if (icsk->icsk_ca_state < TCP_CA_CWR &&
3211 icsk->icsk_mtup.probe_size &&
3212 tp->snd_una == tp->mtu_probe.probe_seq_start) {
3213 tcp_mtup_probe_failed(sk);
3214 /* Restores the reduction we did in tcp_mtup_probe() */
3215 tp->snd_cwnd++;
3216 tcp_simple_retransmit(sk);
3217 return;
3218 }
3219
3220 /* Otherwise enter Recovery state */
3221 tcp_enter_recovery(sk, (flag & FLAG_ECE));
3222 fast_rexmit = 1;
3223 }
3224
3225 if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk)))
3226 tcp_update_scoreboard(sk, fast_rexmit);
3227 tp->prr_delivered += newly_acked_sacked;
3228 tcp_update_cwnd_in_recovery(sk, newly_acked_sacked, fast_rexmit, flag);
3229 tcp_xmit_retransmit_queue(sk);
3230}
3231
3232void tcp_valid_rtt_meas(struct sock *sk, u32 seq_rtt)
3233{
3234 tcp_rtt_estimator(sk, seq_rtt);
3235 tcp_set_rto(sk);
3236 inet_csk(sk)->icsk_backoff = 0;
3237}
3238EXPORT_SYMBOL(tcp_valid_rtt_meas);
3239
3240/* Read draft-ietf-tcplw-high-performance before mucking
3241 * with this code. (Supersedes RFC1323)
3242 */
3243static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
3244{
3245 /* RTTM Rule: A TSecr value received in a segment is used to
3246 * update the averaged RTT measurement only if the segment
3247 * acknowledges some new data, i.e., only if it advances the
3248 * left edge of the send window.
3249 *
3250 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3251 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
3252 *
3253 * Changed: reset backoff as soon as we see the first valid sample.
3254 * If we do not, we get strongly overestimated rto. With timestamps
3255 * samples are accepted even from very old segments: f.e., when rtt=1
3256 * increases to 8, we retransmit 5 times and after 8 seconds delayed
3257 * answer arrives rto becomes 120 seconds! If at least one of segments
3258 * in window is lost... Voila. --ANK (010210)
3259 */
3260 struct tcp_sock *tp = tcp_sk(sk);
3261
3262 tcp_valid_rtt_meas(sk, tcp_time_stamp - tp->rx_opt.rcv_tsecr);
3263}
3264
3265static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
3266{
3267 /* We don't have a timestamp. Can only use
3268 * packets that are not retransmitted to determine
3269 * rtt estimates. Also, we must not reset the
3270 * backoff for rto until we get a non-retransmitted
3271 * packet. This allows us to deal with a situation
3272 * where the network delay has increased suddenly.
3273 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
3274 */
3275
3276 if (flag & FLAG_RETRANS_DATA_ACKED)
3277 return;
3278
3279 tcp_valid_rtt_meas(sk, seq_rtt);
3280}
3281
3282static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
3283 const s32 seq_rtt)
3284{
3285 const struct tcp_sock *tp = tcp_sk(sk);
3286 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
3287 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3288 tcp_ack_saw_tstamp(sk, flag);
3289 else if (seq_rtt >= 0)
3290 tcp_ack_no_tstamp(sk, seq_rtt, flag);
3291}
3292
3293static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
3294{
3295 const struct inet_connection_sock *icsk = inet_csk(sk);
3296 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
3297 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
3298}
3299
3300/* Restart timer after forward progress on connection.
3301 * RFC2988 recommends to restart timer to now+rto.
3302 */
3303void tcp_rearm_rto(struct sock *sk)
3304{
3305 struct tcp_sock *tp = tcp_sk(sk);
3306
3307 if (!tp->packets_out) {
3308 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3309 } else {
3310 u32 rto = inet_csk(sk)->icsk_rto;
3311 /* Offset the time elapsed after installing regular RTO */
3312 if (tp->early_retrans_delayed) {
3313 struct sk_buff *skb = tcp_write_queue_head(sk);
3314 const u32 rto_time_stamp = TCP_SKB_CB(skb)->when + rto;
3315 s32 delta = (s32)(rto_time_stamp - tcp_time_stamp);
3316 /* delta may not be positive if the socket is locked
3317 * when the delayed ER timer fires and is rescheduled.
3318 */
3319 if (delta > 0)
3320 rto = delta;
3321 }
3322 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3323 TCP_RTO_MAX);
3324 }
3325 tp->early_retrans_delayed = 0;
3326}
3327
3328/* This function is called when the delayed ER timer fires. TCP enters
3329 * fast recovery and performs fast-retransmit.
3330 */
3331void tcp_resume_early_retransmit(struct sock *sk)
3332{
3333 struct tcp_sock *tp = tcp_sk(sk);
3334
3335 tcp_rearm_rto(sk);
3336
3337 /* Stop if ER is disabled after the delayed ER timer is scheduled */
3338 if (!tp->do_early_retrans)
3339 return;
3340
3341 tcp_enter_recovery(sk, false);
3342 tcp_update_scoreboard(sk, 1);
3343 tcp_xmit_retransmit_queue(sk);
3344}
3345
3346/* If we get here, the whole TSO packet has not been acked. */
3347static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3348{
3349 struct tcp_sock *tp = tcp_sk(sk);
3350 u32 packets_acked;
3351
3352 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3353
3354 packets_acked = tcp_skb_pcount(skb);
3355 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3356 return 0;
3357 packets_acked -= tcp_skb_pcount(skb);
3358
3359 if (packets_acked) {
3360 BUG_ON(tcp_skb_pcount(skb) == 0);
3361 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3362 }
3363
3364 return packets_acked;
3365}
3366
3367/* Remove acknowledged frames from the retransmission queue. If our packet
3368 * is before the ack sequence we can discard it as it's confirmed to have
3369 * arrived at the other end.
3370 */
3371static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3372 u32 prior_snd_una)
3373{
3374 struct tcp_sock *tp = tcp_sk(sk);
3375 const struct inet_connection_sock *icsk = inet_csk(sk);
3376 struct sk_buff *skb;
3377 u32 now = tcp_time_stamp;
3378 int fully_acked = true;
3379 int flag = 0;
3380 u32 pkts_acked = 0;
3381 u32 reord = tp->packets_out;
3382 u32 prior_sacked = tp->sacked_out;
3383 s32 seq_rtt = -1;
3384 s32 ca_seq_rtt = -1;
3385 ktime_t last_ackt = net_invalid_timestamp();
3386
3387 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3388 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3389 u32 acked_pcount;
3390 u8 sacked = scb->sacked;
3391
3392 /* Determine how many packets and what bytes were acked, tso and else */
3393 if (after(scb->end_seq, tp->snd_una)) {
3394 if (tcp_skb_pcount(skb) == 1 ||
3395 !after(tp->snd_una, scb->seq))
3396 break;
3397
3398 acked_pcount = tcp_tso_acked(sk, skb);
3399 if (!acked_pcount)
3400 break;
3401
3402 fully_acked = false;
3403 } else {
3404 acked_pcount = tcp_skb_pcount(skb);
3405 }
3406
3407 if (sacked & TCPCB_RETRANS) {
3408 if (sacked & TCPCB_SACKED_RETRANS)
3409 tp->retrans_out -= acked_pcount;
3410 flag |= FLAG_RETRANS_DATA_ACKED;
3411 ca_seq_rtt = -1;
3412 seq_rtt = -1;
3413 if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1))
3414 flag |= FLAG_NONHEAD_RETRANS_ACKED;
3415 } else {
3416 ca_seq_rtt = now - scb->when;
3417 last_ackt = skb->tstamp;
3418 if (seq_rtt < 0) {
3419 seq_rtt = ca_seq_rtt;
3420 }
3421 if (!(sacked & TCPCB_SACKED_ACKED))
3422 reord = min(pkts_acked, reord);
3423 }
3424
3425 if (sacked & TCPCB_SACKED_ACKED)
3426 tp->sacked_out -= acked_pcount;
3427 if (sacked & TCPCB_LOST)
3428 tp->lost_out -= acked_pcount;
3429
3430 tp->packets_out -= acked_pcount;
3431 pkts_acked += acked_pcount;
3432
3433 /* Initial outgoing SYN's get put onto the write_queue
3434 * just like anything else we transmit. It is not
3435 * true data, and if we misinform our callers that
3436 * this ACK acks real data, we will erroneously exit
3437 * connection startup slow start one packet too
3438 * quickly. This is severely frowned upon behavior.
3439 */
3440 if (!(scb->tcp_flags & TCPHDR_SYN)) {
3441 flag |= FLAG_DATA_ACKED;
3442 } else {
3443 flag |= FLAG_SYN_ACKED;
3444 tp->retrans_stamp = 0;
3445 }
3446
3447 if (!fully_acked)
3448 break;
3449
3450 tcp_unlink_write_queue(skb, sk);
3451 sk_wmem_free_skb(sk, skb);
3452 tp->scoreboard_skb_hint = NULL;
3453 if (skb == tp->retransmit_skb_hint)
3454 tp->retransmit_skb_hint = NULL;
3455 if (skb == tp->lost_skb_hint)
3456 tp->lost_skb_hint = NULL;
3457 }
3458
3459 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3460 tp->snd_up = tp->snd_una;
3461
3462 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3463 flag |= FLAG_SACK_RENEGING;
3464
3465 if (flag & FLAG_ACKED) {
3466 const struct tcp_congestion_ops *ca_ops
3467 = inet_csk(sk)->icsk_ca_ops;
3468
3469 if (unlikely(icsk->icsk_mtup.probe_size &&
3470 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3471 tcp_mtup_probe_success(sk);
3472 }
3473
3474 tcp_ack_update_rtt(sk, flag, seq_rtt);
3475 tcp_rearm_rto(sk);
3476
3477 if (tcp_is_reno(tp)) {
3478 tcp_remove_reno_sacks(sk, pkts_acked);
3479 } else {
3480 int delta;
3481
3482 /* Non-retransmitted hole got filled? That's reordering */
3483 if (reord < prior_fackets)
3484 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3485
3486 delta = tcp_is_fack(tp) ? pkts_acked :
3487 prior_sacked - tp->sacked_out;
3488 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3489 }
3490
3491 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3492
3493 if (ca_ops->pkts_acked) {
3494 s32 rtt_us = -1;
3495
3496 /* Is the ACK triggering packet unambiguous? */
3497 if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
3498 /* High resolution needed and available? */
3499 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
3500 !ktime_equal(last_ackt,
3501 net_invalid_timestamp()))
3502 rtt_us = ktime_us_delta(ktime_get_real(),
3503 last_ackt);
3504 else if (ca_seq_rtt >= 0)
3505 rtt_us = jiffies_to_usecs(ca_seq_rtt);
3506 }
3507
3508 ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
3509 }
3510 }
3511
3512#if FASTRETRANS_DEBUG > 0
3513 WARN_ON((int)tp->sacked_out < 0);
3514 WARN_ON((int)tp->lost_out < 0);
3515 WARN_ON((int)tp->retrans_out < 0);
3516 if (!tp->packets_out && tcp_is_sack(tp)) {
3517 icsk = inet_csk(sk);
3518 if (tp->lost_out) {
3519 pr_debug("Leak l=%u %d\n",
3520 tp->lost_out, icsk->icsk_ca_state);
3521 tp->lost_out = 0;
3522 }
3523 if (tp->sacked_out) {
3524 pr_debug("Leak s=%u %d\n",
3525 tp->sacked_out, icsk->icsk_ca_state);
3526 tp->sacked_out = 0;
3527 }
3528 if (tp->retrans_out) {
3529 pr_debug("Leak r=%u %d\n",
3530 tp->retrans_out, icsk->icsk_ca_state);
3531 tp->retrans_out = 0;
3532 }
3533 }
3534#endif
3535 return flag;
3536}
3537
3538static void tcp_ack_probe(struct sock *sk)
3539{
3540 const struct tcp_sock *tp = tcp_sk(sk);
3541 struct inet_connection_sock *icsk = inet_csk(sk);
3542
3543 /* Was it a usable window open? */
3544
3545 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3546 icsk->icsk_backoff = 0;
3547 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3548 /* Socket must be waked up by subsequent tcp_data_snd_check().
3549 * This function is not for random using!
3550 */
3551 } else {
3552 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3553 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3554 TCP_RTO_MAX);
3555 }
3556}
3557
3558static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
3559{
3560 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3561 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3562}
3563
3564static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3565{
3566 const struct tcp_sock *tp = tcp_sk(sk);
3567 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
3568 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
3569}
3570
3571/* Check that window update is acceptable.
3572 * The function assumes that snd_una<=ack<=snd_next.
3573 */
3574static inline int tcp_may_update_window(const struct tcp_sock *tp,
3575 const u32 ack, const u32 ack_seq,
3576 const u32 nwin)
3577{
3578 return after(ack, tp->snd_una) ||
3579 after(ack_seq, tp->snd_wl1) ||
3580 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3581}
3582
3583/* Update our send window.
3584 *
3585 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3586 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3587 */
3588static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3589 u32 ack_seq)
3590{
3591 struct tcp_sock *tp = tcp_sk(sk);
3592 int flag = 0;
3593 u32 nwin = ntohs(tcp_hdr(skb)->window);
3594
3595 if (likely(!tcp_hdr(skb)->syn))
3596 nwin <<= tp->rx_opt.snd_wscale;
3597
3598 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3599 flag |= FLAG_WIN_UPDATE;
3600 tcp_update_wl(tp, ack_seq);
3601
3602 if (tp->snd_wnd != nwin) {
3603 tp->snd_wnd = nwin;
3604
3605 /* Note, it is the only place, where
3606 * fast path is recovered for sending TCP.
3607 */
3608 tp->pred_flags = 0;
3609 tcp_fast_path_check(sk);
3610
3611 if (nwin > tp->max_window) {
3612 tp->max_window = nwin;
3613 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3614 }
3615 }
3616 }
3617
3618 tp->snd_una = ack;
3619
3620 return flag;
3621}
3622
3623/* A very conservative spurious RTO response algorithm: reduce cwnd and
3624 * continue in congestion avoidance.
3625 */
3626static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
3627{
3628 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
3629 tp->snd_cwnd_cnt = 0;
3630 tp->bytes_acked = 0;
3631 TCP_ECN_queue_cwr(tp);
3632 tcp_moderate_cwnd(tp);
3633}
3634
3635/* A conservative spurious RTO response algorithm: reduce cwnd using
3636 * rate halving and continue in congestion avoidance.
3637 */
3638static void tcp_ratehalving_spur_to_response(struct sock *sk)
3639{
3640 tcp_enter_cwr(sk, 0);
3641}
3642
3643static void tcp_undo_spur_to_response(struct sock *sk, int flag)
3644{
3645 if (flag & FLAG_ECE)
3646 tcp_ratehalving_spur_to_response(sk);
3647 else
3648 tcp_undo_cwr(sk, true);
3649}
3650
3651/* F-RTO spurious RTO detection algorithm (RFC4138)
3652 *
3653 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3654 * comments). State (ACK number) is kept in frto_counter. When ACK advances
3655 * window (but not to or beyond highest sequence sent before RTO):
3656 * On First ACK, send two new segments out.
3657 * On Second ACK, RTO was likely spurious. Do spurious response (response
3658 * algorithm is not part of the F-RTO detection algorithm
3659 * given in RFC4138 but can be selected separately).
3660 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
3661 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3662 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3663 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
3664 *
3665 * Rationale: if the RTO was spurious, new ACKs should arrive from the
3666 * original window even after we transmit two new data segments.
3667 *
3668 * SACK version:
3669 * on first step, wait until first cumulative ACK arrives, then move to
3670 * the second step. In second step, the next ACK decides.
3671 *
3672 * F-RTO is implemented (mainly) in four functions:
3673 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
3674 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3675 * called when tcp_use_frto() showed green light
3676 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3677 * - tcp_enter_frto_loss() is called if there is not enough evidence
3678 * to prove that the RTO is indeed spurious. It transfers the control
3679 * from F-RTO to the conventional RTO recovery
3680 */
3681static bool tcp_process_frto(struct sock *sk, int flag)
3682{
3683 struct tcp_sock *tp = tcp_sk(sk);
3684
3685 tcp_verify_left_out(tp);
3686
3687 /* Duplicate the behavior from Loss state (fastretrans_alert) */
3688 if (flag & FLAG_DATA_ACKED)
3689 inet_csk(sk)->icsk_retransmits = 0;
3690
3691 if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||
3692 ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
3693 tp->undo_marker = 0;
3694
3695 if (!before(tp->snd_una, tp->frto_highmark)) {
3696 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
3697 return true;
3698 }
3699
3700 if (!tcp_is_sackfrto(tp)) {
3701 /* RFC4138 shortcoming in step 2; should also have case c):
3702 * ACK isn't duplicate nor advances window, e.g., opposite dir
3703 * data, winupdate
3704 */
3705 if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP))
3706 return true;
3707
3708 if (!(flag & FLAG_DATA_ACKED)) {
3709 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
3710 flag);
3711 return true;
3712 }
3713 } else {
3714 if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
3715 /* Prevent sending of new data. */
3716 tp->snd_cwnd = min(tp->snd_cwnd,
3717 tcp_packets_in_flight(tp));
3718 return true;
3719 }
3720
3721 if ((tp->frto_counter >= 2) &&
3722 (!(flag & FLAG_FORWARD_PROGRESS) ||
3723 ((flag & FLAG_DATA_SACKED) &&
3724 !(flag & FLAG_ONLY_ORIG_SACKED)))) {
3725 /* RFC4138 shortcoming (see comment above) */
3726 if (!(flag & FLAG_FORWARD_PROGRESS) &&
3727 (flag & FLAG_NOT_DUP))
3728 return true;
3729
3730 tcp_enter_frto_loss(sk, 3, flag);
3731 return true;
3732 }
3733 }
3734
3735 if (tp->frto_counter == 1) {
3736 /* tcp_may_send_now needs to see updated state */
3737 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
3738 tp->frto_counter = 2;
3739
3740 if (!tcp_may_send_now(sk))
3741 tcp_enter_frto_loss(sk, 2, flag);
3742
3743 return true;
3744 } else {
3745 switch (sysctl_tcp_frto_response) {
3746 case 2:
3747 tcp_undo_spur_to_response(sk, flag);
3748 break;
3749 case 1:
3750 tcp_conservative_spur_to_response(tp);
3751 break;
3752 default:
3753 tcp_ratehalving_spur_to_response(sk);
3754 break;
3755 }
3756 tp->frto_counter = 0;
3757 tp->undo_marker = 0;
3758 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSPURIOUSRTOS);
3759 }
3760 return false;
3761}
3762
3763/* This routine deals with incoming acks, but not outgoing ones. */
3764static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3765{
3766 struct inet_connection_sock *icsk = inet_csk(sk);
3767 struct tcp_sock *tp = tcp_sk(sk);
3768 u32 prior_snd_una = tp->snd_una;
3769 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3770 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3771 bool is_dupack = false;
3772 u32 prior_in_flight;
3773 u32 prior_fackets;
3774 int prior_packets;
3775 int prior_sacked = tp->sacked_out;
3776 int pkts_acked = 0;
3777 bool frto_cwnd = false;
3778
3779 /* If the ack is older than previous acks
3780 * then we can probably ignore it.
3781 */
3782 if (before(ack, prior_snd_una))
3783 goto old_ack;
3784
3785 /* If the ack includes data we haven't sent yet, discard
3786 * this segment (RFC793 Section 3.9).
3787 */
3788 if (after(ack, tp->snd_nxt))
3789 goto invalid_ack;
3790
3791 if (tp->early_retrans_delayed)
3792 tcp_rearm_rto(sk);
3793
3794 if (after(ack, prior_snd_una))
3795 flag |= FLAG_SND_UNA_ADVANCED;
3796
3797 if (sysctl_tcp_abc) {
3798 if (icsk->icsk_ca_state < TCP_CA_CWR)
3799 tp->bytes_acked += ack - prior_snd_una;
3800 else if (icsk->icsk_ca_state == TCP_CA_Loss)
3801 /* we assume just one segment left network */
3802 tp->bytes_acked += min(ack - prior_snd_una,
3803 tp->mss_cache);
3804 }
3805
3806 prior_fackets = tp->fackets_out;
3807 prior_in_flight = tcp_packets_in_flight(tp);
3808
3809 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3810 /* Window is constant, pure forward advance.
3811 * No more checks are required.
3812 * Note, we use the fact that SND.UNA>=SND.WL2.
3813 */
3814 tcp_update_wl(tp, ack_seq);
3815 tp->snd_una = ack;
3816 flag |= FLAG_WIN_UPDATE;
3817
3818 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3819
3820 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3821 } else {
3822 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3823 flag |= FLAG_DATA;
3824 else
3825 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3826
3827 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3828
3829 if (TCP_SKB_CB(skb)->sacked)
3830 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3831
3832 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3833 flag |= FLAG_ECE;
3834
3835 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
3836 }
3837
3838 /* We passed data and got it acked, remove any soft error
3839 * log. Something worked...
3840 */
3841 sk->sk_err_soft = 0;
3842 icsk->icsk_probes_out = 0;
3843 tp->rcv_tstamp = tcp_time_stamp;
3844 prior_packets = tp->packets_out;
3845 if (!prior_packets)
3846 goto no_queue;
3847
3848 /* See if we can take anything off of the retransmit queue. */
3849 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una);
3850
3851 pkts_acked = prior_packets - tp->packets_out;
3852
3853 if (tp->frto_counter)
3854 frto_cwnd = tcp_process_frto(sk, flag);
3855 /* Guarantee sacktag reordering detection against wrap-arounds */
3856 if (before(tp->frto_highmark, tp->snd_una))
3857 tp->frto_highmark = 0;
3858
3859 if (tcp_ack_is_dubious(sk, flag)) {
3860 /* Advance CWND, if state allows this. */
3861 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
3862 tcp_may_raise_cwnd(sk, flag))
3863 tcp_cong_avoid(sk, ack, prior_in_flight);
3864 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3865 tcp_fastretrans_alert(sk, pkts_acked, prior_sacked,
3866 is_dupack, flag);
3867 } else {
3868 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
3869 tcp_cong_avoid(sk, ack, prior_in_flight);
3870 }
3871
3872 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3873 dst_confirm(__sk_dst_get(sk));
3874
3875 return 1;
3876
3877no_queue:
3878 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3879 if (flag & FLAG_DSACKING_ACK)
3880 tcp_fastretrans_alert(sk, pkts_acked, prior_sacked,
3881 is_dupack, flag);
3882 /* If this ack opens up a zero window, clear backoff. It was
3883 * being used to time the probes, and is probably far higher than
3884 * it needs to be for normal retransmission.
3885 */
3886 if (tcp_send_head(sk))
3887 tcp_ack_probe(sk);
3888 return 1;
3889
3890invalid_ack:
3891 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3892 return -1;
3893
3894old_ack:
3895 /* If data was SACKed, tag it and see if we should send more data.
3896 * If data was DSACKed, see if we can undo a cwnd reduction.
3897 */
3898 if (TCP_SKB_CB(skb)->sacked) {
3899 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3900 tcp_fastretrans_alert(sk, pkts_acked, prior_sacked,
3901 is_dupack, flag);
3902 }
3903
3904 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3905 return 0;
3906}
3907
3908/* Look for tcp options. Normally only called on SYN and SYNACK packets.
3909 * But, this can also be called on packets in the established flow when
3910 * the fast version below fails.
3911 */
3912void tcp_parse_options(const struct sk_buff *skb, struct tcp_options_received *opt_rx,
3913 const u8 **hvpp, int estab)
3914{
3915 const unsigned char *ptr;
3916 const struct tcphdr *th = tcp_hdr(skb);
3917 int length = (th->doff * 4) - sizeof(struct tcphdr);
3918
3919 ptr = (const unsigned char *)(th + 1);
3920 opt_rx->saw_tstamp = 0;
3921
3922 while (length > 0) {
3923 int opcode = *ptr++;
3924 int opsize;
3925
3926 switch (opcode) {
3927 case TCPOPT_EOL:
3928 return;
3929 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3930 length--;
3931 continue;
3932 default:
3933 opsize = *ptr++;
3934 if (opsize < 2) /* "silly options" */
3935 return;
3936 if (opsize > length)
3937 return; /* don't parse partial options */
3938 switch (opcode) {
3939 case TCPOPT_MSS:
3940 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3941 u16 in_mss = get_unaligned_be16(ptr);
3942 if (in_mss) {
3943 if (opt_rx->user_mss &&
3944 opt_rx->user_mss < in_mss)
3945 in_mss = opt_rx->user_mss;
3946 opt_rx->mss_clamp = in_mss;
3947 }
3948 }
3949 break;
3950 case TCPOPT_WINDOW:
3951 if (opsize == TCPOLEN_WINDOW && th->syn &&
3952 !estab && sysctl_tcp_window_scaling) {
3953 __u8 snd_wscale = *(__u8 *)ptr;
3954 opt_rx->wscale_ok = 1;
3955 if (snd_wscale > 14) {
3956 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3957 __func__,
3958 snd_wscale);
3959 snd_wscale = 14;
3960 }
3961 opt_rx->snd_wscale = snd_wscale;
3962 }
3963 break;
3964 case TCPOPT_TIMESTAMP:
3965 if ((opsize == TCPOLEN_TIMESTAMP) &&
3966 ((estab && opt_rx->tstamp_ok) ||
3967 (!estab && sysctl_tcp_timestamps))) {
3968 opt_rx->saw_tstamp = 1;
3969 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3970 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3971 }
3972 break;
3973 case TCPOPT_SACK_PERM:
3974 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3975 !estab && sysctl_tcp_sack) {
3976 opt_rx->sack_ok = TCP_SACK_SEEN;
3977 tcp_sack_reset(opt_rx);
3978 }
3979 break;
3980
3981 case TCPOPT_SACK:
3982 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3983 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3984 opt_rx->sack_ok) {
3985 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3986 }
3987 break;
3988#ifdef CONFIG_TCP_MD5SIG
3989 case TCPOPT_MD5SIG:
3990 /*
3991 * The MD5 Hash has already been
3992 * checked (see tcp_v{4,6}_do_rcv()).
3993 */
3994 break;
3995#endif
3996 case TCPOPT_COOKIE:
3997 /* This option is variable length.
3998 */
3999 switch (opsize) {
4000 case TCPOLEN_COOKIE_BASE:
4001 /* not yet implemented */
4002 break;
4003 case TCPOLEN_COOKIE_PAIR:
4004 /* not yet implemented */
4005 break;
4006 case TCPOLEN_COOKIE_MIN+0:
4007 case TCPOLEN_COOKIE_MIN+2:
4008 case TCPOLEN_COOKIE_MIN+4:
4009 case TCPOLEN_COOKIE_MIN+6:
4010 case TCPOLEN_COOKIE_MAX:
4011 /* 16-bit multiple */
4012 opt_rx->cookie_plus = opsize;
4013 *hvpp = ptr;
4014 break;
4015 default:
4016 /* ignore option */
4017 break;
4018 }
4019 break;
4020 }
4021
4022 ptr += opsize-2;
4023 length -= opsize;
4024 }
4025 }
4026}
4027EXPORT_SYMBOL(tcp_parse_options);
4028
4029static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
4030{
4031 const __be32 *ptr = (const __be32 *)(th + 1);
4032
4033 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4034 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
4035 tp->rx_opt.saw_tstamp = 1;
4036 ++ptr;
4037 tp->rx_opt.rcv_tsval = ntohl(*ptr);
4038 ++ptr;
4039 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
4040 return true;
4041 }
4042 return false;
4043}
4044
4045/* Fast parse options. This hopes to only see timestamps.
4046 * If it is wrong it falls back on tcp_parse_options().
4047 */
4048static bool tcp_fast_parse_options(const struct sk_buff *skb,
4049 const struct tcphdr *th,
4050 struct tcp_sock *tp, const u8 **hvpp)
4051{
4052 /* In the spirit of fast parsing, compare doff directly to constant
4053 * values. Because equality is used, short doff can be ignored here.
4054 */
4055 if (th->doff == (sizeof(*th) / 4)) {
4056 tp->rx_opt.saw_tstamp = 0;
4057 return false;
4058 } else if (tp->rx_opt.tstamp_ok &&
4059 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
4060 if (tcp_parse_aligned_timestamp(tp, th))
4061 return true;
4062 }
4063 tcp_parse_options(skb, &tp->rx_opt, hvpp, 1);
4064 return true;
4065}
4066
4067#ifdef CONFIG_TCP_MD5SIG
4068/*
4069 * Parse MD5 Signature option
4070 */
4071const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
4072{
4073 int length = (th->doff << 2) - sizeof(*th);
4074 const u8 *ptr = (const u8 *)(th + 1);
4075
4076 /* If the TCP option is too short, we can short cut */
4077 if (length < TCPOLEN_MD5SIG)
4078 return NULL;
4079
4080 while (length > 0) {
4081 int opcode = *ptr++;
4082 int opsize;
4083
4084 switch(opcode) {
4085 case TCPOPT_EOL:
4086 return NULL;
4087 case TCPOPT_NOP:
4088 length--;
4089 continue;
4090 default:
4091 opsize = *ptr++;
4092 if (opsize < 2 || opsize > length)
4093 return NULL;
4094 if (opcode == TCPOPT_MD5SIG)
4095 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
4096 }
4097 ptr += opsize - 2;
4098 length -= opsize;
4099 }
4100 return NULL;
4101}
4102EXPORT_SYMBOL(tcp_parse_md5sig_option);
4103#endif
4104
4105static inline void tcp_store_ts_recent(struct tcp_sock *tp)
4106{
4107 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
4108 tp->rx_opt.ts_recent_stamp = get_seconds();
4109}
4110
4111static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
4112{
4113 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
4114 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
4115 * extra check below makes sure this can only happen
4116 * for pure ACK frames. -DaveM
4117 *
4118 * Not only, also it occurs for expired timestamps.
4119 */
4120
4121 if (tcp_paws_check(&tp->rx_opt, 0))
4122 tcp_store_ts_recent(tp);
4123 }
4124}
4125
4126/* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
4127 *
4128 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
4129 * it can pass through stack. So, the following predicate verifies that
4130 * this segment is not used for anything but congestion avoidance or
4131 * fast retransmit. Moreover, we even are able to eliminate most of such
4132 * second order effects, if we apply some small "replay" window (~RTO)
4133 * to timestamp space.
4134 *
4135 * All these measures still do not guarantee that we reject wrapped ACKs
4136 * on networks with high bandwidth, when sequence space is recycled fastly,
4137 * but it guarantees that such events will be very rare and do not affect
4138 * connection seriously. This doesn't look nice, but alas, PAWS is really
4139 * buggy extension.
4140 *
4141 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
4142 * states that events when retransmit arrives after original data are rare.
4143 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
4144 * the biggest problem on large power networks even with minor reordering.
4145 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
4146 * up to bandwidth of 18Gigabit/sec. 8) ]
4147 */
4148
4149static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
4150{
4151 const struct tcp_sock *tp = tcp_sk(sk);
4152 const struct tcphdr *th = tcp_hdr(skb);
4153 u32 seq = TCP_SKB_CB(skb)->seq;
4154 u32 ack = TCP_SKB_CB(skb)->ack_seq;
4155
4156 return (/* 1. Pure ACK with correct sequence number. */
4157 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
4158
4159 /* 2. ... and duplicate ACK. */
4160 ack == tp->snd_una &&
4161
4162 /* 3. ... and does not update window. */
4163 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4164
4165 /* 4. ... and sits in replay window. */
4166 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
4167}
4168
4169static inline int tcp_paws_discard(const struct sock *sk,
4170 const struct sk_buff *skb)
4171{
4172 const struct tcp_sock *tp = tcp_sk(sk);
4173
4174 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4175 !tcp_disordered_ack(sk, skb);
4176}
4177
4178/* Check segment sequence number for validity.
4179 *
4180 * Segment controls are considered valid, if the segment
4181 * fits to the window after truncation to the window. Acceptability
4182 * of data (and SYN, FIN, of course) is checked separately.
4183 * See tcp_data_queue(), for example.
4184 *
4185 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4186 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4187 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4188 * (borrowed from freebsd)
4189 */
4190
4191static inline int tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
4192{
4193 return !before(end_seq, tp->rcv_wup) &&
4194 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4195}
4196
4197/* When we get a reset we do this. */
4198static void tcp_reset(struct sock *sk)
4199{
4200 /* We want the right error as BSD sees it (and indeed as we do). */
4201 switch (sk->sk_state) {
4202 case TCP_SYN_SENT:
4203 sk->sk_err = ECONNREFUSED;
4204 break;
4205 case TCP_CLOSE_WAIT:
4206 sk->sk_err = EPIPE;
4207 break;
4208 case TCP_CLOSE:
4209 return;
4210 default:
4211 sk->sk_err = ECONNRESET;
4212 }
4213 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4214 smp_wmb();
4215
4216 if (!sock_flag(sk, SOCK_DEAD))
4217 sk->sk_error_report(sk);
4218
4219 tcp_done(sk);
4220}
4221
4222/*
4223 * Process the FIN bit. This now behaves as it is supposed to work
4224 * and the FIN takes effect when it is validly part of sequence
4225 * space. Not before when we get holes.
4226 *
4227 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4228 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4229 * TIME-WAIT)
4230 *
4231 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4232 * close and we go into CLOSING (and later onto TIME-WAIT)
4233 *
4234 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4235 */
4236static void tcp_fin(struct sock *sk)
4237{
4238 struct tcp_sock *tp = tcp_sk(sk);
4239
4240 inet_csk_schedule_ack(sk);
4241
4242 sk->sk_shutdown |= RCV_SHUTDOWN;
4243 sock_set_flag(sk, SOCK_DONE);
4244
4245 switch (sk->sk_state) {
4246 case TCP_SYN_RECV:
4247 case TCP_ESTABLISHED:
4248 /* Move to CLOSE_WAIT */
4249 tcp_set_state(sk, TCP_CLOSE_WAIT);
4250 inet_csk(sk)->icsk_ack.pingpong = 1;
4251 break;
4252
4253 case TCP_CLOSE_WAIT:
4254 case TCP_CLOSING:
4255 /* Received a retransmission of the FIN, do
4256 * nothing.
4257 */
4258 break;
4259 case TCP_LAST_ACK:
4260 /* RFC793: Remain in the LAST-ACK state. */
4261 break;
4262
4263 case TCP_FIN_WAIT1:
4264 /* This case occurs when a simultaneous close
4265 * happens, we must ack the received FIN and
4266 * enter the CLOSING state.
4267 */
4268 tcp_send_ack(sk);
4269 tcp_set_state(sk, TCP_CLOSING);
4270 break;
4271 case TCP_FIN_WAIT2:
4272 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4273 tcp_send_ack(sk);
4274 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4275 break;
4276 default:
4277 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4278 * cases we should never reach this piece of code.
4279 */
4280 pr_err("%s: Impossible, sk->sk_state=%d\n",
4281 __func__, sk->sk_state);
4282 break;
4283 }
4284
4285 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4286 * Probably, we should reset in this case. For now drop them.
4287 */
4288 __skb_queue_purge(&tp->out_of_order_queue);
4289 if (tcp_is_sack(tp))
4290 tcp_sack_reset(&tp->rx_opt);
4291 sk_mem_reclaim(sk);
4292
4293 if (!sock_flag(sk, SOCK_DEAD)) {
4294 sk->sk_state_change(sk);
4295
4296 /* Do not send POLL_HUP for half duplex close. */
4297 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4298 sk->sk_state == TCP_CLOSE)
4299 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4300 else
4301 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4302 }
4303}
4304
4305static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4306 u32 end_seq)
4307{
4308 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4309 if (before(seq, sp->start_seq))
4310 sp->start_seq = seq;
4311 if (after(end_seq, sp->end_seq))
4312 sp->end_seq = end_seq;
4313 return true;
4314 }
4315 return false;
4316}
4317
4318static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4319{
4320 struct tcp_sock *tp = tcp_sk(sk);
4321
4322 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4323 int mib_idx;
4324
4325 if (before(seq, tp->rcv_nxt))
4326 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4327 else
4328 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4329
4330 NET_INC_STATS_BH(sock_net(sk), mib_idx);
4331
4332 tp->rx_opt.dsack = 1;
4333 tp->duplicate_sack[0].start_seq = seq;
4334 tp->duplicate_sack[0].end_seq = end_seq;
4335 }
4336}
4337
4338static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4339{
4340 struct tcp_sock *tp = tcp_sk(sk);
4341
4342 if (!tp->rx_opt.dsack)
4343 tcp_dsack_set(sk, seq, end_seq);
4344 else
4345 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4346}
4347
4348static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4349{
4350 struct tcp_sock *tp = tcp_sk(sk);
4351
4352 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4353 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4354 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4355 tcp_enter_quickack_mode(sk);
4356
4357 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4358 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4359
4360 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4361 end_seq = tp->rcv_nxt;
4362 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4363 }
4364 }
4365
4366 tcp_send_ack(sk);
4367}
4368
4369/* These routines update the SACK block as out-of-order packets arrive or
4370 * in-order packets close up the sequence space.
4371 */
4372static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4373{
4374 int this_sack;
4375 struct tcp_sack_block *sp = &tp->selective_acks[0];
4376 struct tcp_sack_block *swalk = sp + 1;
4377
4378 /* See if the recent change to the first SACK eats into
4379 * or hits the sequence space of other SACK blocks, if so coalesce.
4380 */
4381 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4382 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4383 int i;
4384
4385 /* Zap SWALK, by moving every further SACK up by one slot.
4386 * Decrease num_sacks.
4387 */
4388 tp->rx_opt.num_sacks--;
4389 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4390 sp[i] = sp[i + 1];
4391 continue;
4392 }
4393 this_sack++, swalk++;
4394 }
4395}
4396
4397static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4398{
4399 struct tcp_sock *tp = tcp_sk(sk);
4400 struct tcp_sack_block *sp = &tp->selective_acks[0];
4401 int cur_sacks = tp->rx_opt.num_sacks;
4402 int this_sack;
4403
4404 if (!cur_sacks)
4405 goto new_sack;
4406
4407 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4408 if (tcp_sack_extend(sp, seq, end_seq)) {
4409 /* Rotate this_sack to the first one. */
4410 for (; this_sack > 0; this_sack--, sp--)
4411 swap(*sp, *(sp - 1));
4412 if (cur_sacks > 1)
4413 tcp_sack_maybe_coalesce(tp);
4414 return;
4415 }
4416 }
4417
4418 /* Could not find an adjacent existing SACK, build a new one,
4419 * put it at the front, and shift everyone else down. We
4420 * always know there is at least one SACK present already here.
4421 *
4422 * If the sack array is full, forget about the last one.
4423 */
4424 if (this_sack >= TCP_NUM_SACKS) {
4425 this_sack--;
4426 tp->rx_opt.num_sacks--;
4427 sp--;
4428 }
4429 for (; this_sack > 0; this_sack--, sp--)
4430 *sp = *(sp - 1);
4431
4432new_sack:
4433 /* Build the new head SACK, and we're done. */
4434 sp->start_seq = seq;
4435 sp->end_seq = end_seq;
4436 tp->rx_opt.num_sacks++;
4437}
4438
4439/* RCV.NXT advances, some SACKs should be eaten. */
4440
4441static void tcp_sack_remove(struct tcp_sock *tp)
4442{
4443 struct tcp_sack_block *sp = &tp->selective_acks[0];
4444 int num_sacks = tp->rx_opt.num_sacks;
4445 int this_sack;
4446
4447 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4448 if (skb_queue_empty(&tp->out_of_order_queue)) {
4449 tp->rx_opt.num_sacks = 0;
4450 return;
4451 }
4452
4453 for (this_sack = 0; this_sack < num_sacks;) {
4454 /* Check if the start of the sack is covered by RCV.NXT. */
4455 if (!before(tp->rcv_nxt, sp->start_seq)) {
4456 int i;
4457
4458 /* RCV.NXT must cover all the block! */
4459 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4460
4461 /* Zap this SACK, by moving forward any other SACKS. */
4462 for (i=this_sack+1; i < num_sacks; i++)
4463 tp->selective_acks[i-1] = tp->selective_acks[i];
4464 num_sacks--;
4465 continue;
4466 }
4467 this_sack++;
4468 sp++;
4469 }
4470 tp->rx_opt.num_sacks = num_sacks;
4471}
4472
4473/* This one checks to see if we can put data from the
4474 * out_of_order queue into the receive_queue.
4475 */
4476static void tcp_ofo_queue(struct sock *sk)
4477{
4478 struct tcp_sock *tp = tcp_sk(sk);
4479 __u32 dsack_high = tp->rcv_nxt;
4480 struct sk_buff *skb;
4481
4482 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4483 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4484 break;
4485
4486 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4487 __u32 dsack = dsack_high;
4488 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4489 dsack_high = TCP_SKB_CB(skb)->end_seq;
4490 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4491 }
4492
4493 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4494 SOCK_DEBUG(sk, "ofo packet was already received\n");
4495 __skb_unlink(skb, &tp->out_of_order_queue);
4496 __kfree_skb(skb);
4497 continue;
4498 }
4499 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4500 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4501 TCP_SKB_CB(skb)->end_seq);
4502
4503 __skb_unlink(skb, &tp->out_of_order_queue);
4504 __skb_queue_tail(&sk->sk_receive_queue, skb);
4505 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4506 if (tcp_hdr(skb)->fin)
4507 tcp_fin(sk);
4508 }
4509}
4510
4511static bool tcp_prune_ofo_queue(struct sock *sk);
4512static int tcp_prune_queue(struct sock *sk);
4513
4514static int tcp_try_rmem_schedule(struct sock *sk, unsigned int size)
4515{
4516 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4517 !sk_rmem_schedule(sk, size)) {
4518
4519 if (tcp_prune_queue(sk) < 0)
4520 return -1;
4521
4522 if (!sk_rmem_schedule(sk, size)) {
4523 if (!tcp_prune_ofo_queue(sk))
4524 return -1;
4525
4526 if (!sk_rmem_schedule(sk, size))
4527 return -1;
4528 }
4529 }
4530 return 0;
4531}
4532
4533/**
4534 * tcp_try_coalesce - try to merge skb to prior one
4535 * @sk: socket
4536 * @to: prior buffer
4537 * @from: buffer to add in queue
4538 * @fragstolen: pointer to boolean
4539 *
4540 * Before queueing skb @from after @to, try to merge them
4541 * to reduce overall memory use and queue lengths, if cost is small.
4542 * Packets in ofo or receive queues can stay a long time.
4543 * Better try to coalesce them right now to avoid future collapses.
4544 * Returns true if caller should free @from instead of queueing it
4545 */
4546static bool tcp_try_coalesce(struct sock *sk,
4547 struct sk_buff *to,
4548 struct sk_buff *from,
4549 bool *fragstolen)
4550{
4551 int delta;
4552
4553 *fragstolen = false;
4554
4555 if (tcp_hdr(from)->fin)
4556 return false;
4557
4558 /* Its possible this segment overlaps with prior segment in queue */
4559 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4560 return false;
4561
4562 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4563 return false;
4564
4565 atomic_add(delta, &sk->sk_rmem_alloc);
4566 sk_mem_charge(sk, delta);
4567 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4568 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4569 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4570 return true;
4571}
4572
4573static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4574{
4575 struct tcp_sock *tp = tcp_sk(sk);
4576 struct sk_buff *skb1;
4577 u32 seq, end_seq;
4578
4579 TCP_ECN_check_ce(tp, skb);
4580
4581 if (tcp_try_rmem_schedule(sk, skb->truesize)) {
4582 /* TODO: should increment a counter */
4583 __kfree_skb(skb);
4584 return;
4585 }
4586
4587 /* Disable header prediction. */
4588 tp->pred_flags = 0;
4589 inet_csk_schedule_ack(sk);
4590
4591 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4592 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4593
4594 skb1 = skb_peek_tail(&tp->out_of_order_queue);
4595 if (!skb1) {
4596 /* Initial out of order segment, build 1 SACK. */
4597 if (tcp_is_sack(tp)) {
4598 tp->rx_opt.num_sacks = 1;
4599 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4600 tp->selective_acks[0].end_seq =
4601 TCP_SKB_CB(skb)->end_seq;
4602 }
4603 __skb_queue_head(&tp->out_of_order_queue, skb);
4604 goto end;
4605 }
4606
4607 seq = TCP_SKB_CB(skb)->seq;
4608 end_seq = TCP_SKB_CB(skb)->end_seq;
4609
4610 if (seq == TCP_SKB_CB(skb1)->end_seq) {
4611 bool fragstolen;
4612
4613 if (!tcp_try_coalesce(sk, skb1, skb, &fragstolen)) {
4614 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4615 } else {
4616 kfree_skb_partial(skb, fragstolen);
4617 skb = NULL;
4618 }
4619
4620 if (!tp->rx_opt.num_sacks ||
4621 tp->selective_acks[0].end_seq != seq)
4622 goto add_sack;
4623
4624 /* Common case: data arrive in order after hole. */
4625 tp->selective_acks[0].end_seq = end_seq;
4626 goto end;
4627 }
4628
4629 /* Find place to insert this segment. */
4630 while (1) {
4631 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4632 break;
4633 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4634 skb1 = NULL;
4635 break;
4636 }
4637 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4638 }
4639
4640 /* Do skb overlap to previous one? */
4641 if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4642 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4643 /* All the bits are present. Drop. */
4644 __kfree_skb(skb);
4645 skb = NULL;
4646 tcp_dsack_set(sk, seq, end_seq);
4647 goto add_sack;
4648 }
4649 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4650 /* Partial overlap. */
4651 tcp_dsack_set(sk, seq,
4652 TCP_SKB_CB(skb1)->end_seq);
4653 } else {
4654 if (skb_queue_is_first(&tp->out_of_order_queue,
4655 skb1))
4656 skb1 = NULL;
4657 else
4658 skb1 = skb_queue_prev(
4659 &tp->out_of_order_queue,
4660 skb1);
4661 }
4662 }
4663 if (!skb1)
4664 __skb_queue_head(&tp->out_of_order_queue, skb);
4665 else
4666 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4667
4668 /* And clean segments covered by new one as whole. */
4669 while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4670 skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4671
4672 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4673 break;
4674 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4675 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4676 end_seq);
4677 break;
4678 }
4679 __skb_unlink(skb1, &tp->out_of_order_queue);
4680 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4681 TCP_SKB_CB(skb1)->end_seq);
4682 __kfree_skb(skb1);
4683 }
4684
4685add_sack:
4686 if (tcp_is_sack(tp))
4687 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4688end:
4689 if (skb)
4690 skb_set_owner_r(skb, sk);
4691}
4692
4693static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4694 bool *fragstolen)
4695{
4696 int eaten;
4697 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4698
4699 __skb_pull(skb, hdrlen);
4700 eaten = (tail &&
4701 tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0;
4702 tcp_sk(sk)->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4703 if (!eaten) {
4704 __skb_queue_tail(&sk->sk_receive_queue, skb);
4705 skb_set_owner_r(skb, sk);
4706 }
4707 return eaten;
4708}
4709
4710int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4711{
4712 struct sk_buff *skb;
4713 struct tcphdr *th;
4714 bool fragstolen;
4715
4716 if (tcp_try_rmem_schedule(sk, size + sizeof(*th)))
4717 goto err;
4718
4719 skb = alloc_skb(size + sizeof(*th), sk->sk_allocation);
4720 if (!skb)
4721 goto err;
4722
4723 th = (struct tcphdr *)skb_put(skb, sizeof(*th));
4724 skb_reset_transport_header(skb);
4725 memset(th, 0, sizeof(*th));
4726
4727 if (memcpy_fromiovec(skb_put(skb, size), msg->msg_iov, size))
4728 goto err_free;
4729
4730 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4731 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4732 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4733
4734 if (tcp_queue_rcv(sk, skb, sizeof(*th), &fragstolen)) {
4735 WARN_ON_ONCE(fragstolen); /* should not happen */
4736 __kfree_skb(skb);
4737 }
4738 return size;
4739
4740err_free:
4741 kfree_skb(skb);
4742err:
4743 return -ENOMEM;
4744}
4745
4746static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4747{
4748 const struct tcphdr *th = tcp_hdr(skb);
4749 struct tcp_sock *tp = tcp_sk(sk);
4750 int eaten = -1;
4751 bool fragstolen = false;
4752
4753 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4754 goto drop;
4755
4756 skb_dst_drop(skb);
4757 __skb_pull(skb, th->doff * 4);
4758
4759 TCP_ECN_accept_cwr(tp, skb);
4760
4761 tp->rx_opt.dsack = 0;
4762
4763 /* Queue data for delivery to the user.
4764 * Packets in sequence go to the receive queue.
4765 * Out of sequence packets to the out_of_order_queue.
4766 */
4767 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4768 if (tcp_receive_window(tp) == 0)
4769 goto out_of_window;
4770
4771 /* Ok. In sequence. In window. */
4772 if (tp->ucopy.task == current &&
4773 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4774 sock_owned_by_user(sk) && !tp->urg_data) {
4775 int chunk = min_t(unsigned int, skb->len,
4776 tp->ucopy.len);
4777
4778 __set_current_state(TASK_RUNNING);
4779
4780 local_bh_enable();
4781 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
4782 tp->ucopy.len -= chunk;
4783 tp->copied_seq += chunk;
4784 eaten = (chunk == skb->len);
4785 tcp_rcv_space_adjust(sk);
4786 }
4787 local_bh_disable();
4788 }
4789
4790 if (eaten <= 0) {
4791queue_and_out:
4792 if (eaten < 0 &&
4793 tcp_try_rmem_schedule(sk, skb->truesize))
4794 goto drop;
4795
4796 eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4797 }
4798 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4799 if (skb->len)
4800 tcp_event_data_recv(sk, skb);
4801 if (th->fin)
4802 tcp_fin(sk);
4803
4804 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4805 tcp_ofo_queue(sk);
4806
4807 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4808 * gap in queue is filled.
4809 */
4810 if (skb_queue_empty(&tp->out_of_order_queue))
4811 inet_csk(sk)->icsk_ack.pingpong = 0;
4812 }
4813
4814 if (tp->rx_opt.num_sacks)
4815 tcp_sack_remove(tp);
4816
4817 tcp_fast_path_check(sk);
4818
4819 if (eaten > 0)
4820 kfree_skb_partial(skb, fragstolen);
4821 else if (!sock_flag(sk, SOCK_DEAD))
4822 sk->sk_data_ready(sk, 0);
4823 return;
4824 }
4825
4826 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4827 /* A retransmit, 2nd most common case. Force an immediate ack. */
4828 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4829 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4830
4831out_of_window:
4832 tcp_enter_quickack_mode(sk);
4833 inet_csk_schedule_ack(sk);
4834drop:
4835 __kfree_skb(skb);
4836 return;
4837 }
4838
4839 /* Out of window. F.e. zero window probe. */
4840 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4841 goto out_of_window;
4842
4843 tcp_enter_quickack_mode(sk);
4844
4845 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4846 /* Partial packet, seq < rcv_next < end_seq */
4847 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4848 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4849 TCP_SKB_CB(skb)->end_seq);
4850
4851 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4852
4853 /* If window is closed, drop tail of packet. But after
4854 * remembering D-SACK for its head made in previous line.
4855 */
4856 if (!tcp_receive_window(tp))
4857 goto out_of_window;
4858 goto queue_and_out;
4859 }
4860
4861 tcp_data_queue_ofo(sk, skb);
4862}
4863
4864static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4865 struct sk_buff_head *list)
4866{
4867 struct sk_buff *next = NULL;
4868
4869 if (!skb_queue_is_last(list, skb))
4870 next = skb_queue_next(list, skb);
4871
4872 __skb_unlink(skb, list);
4873 __kfree_skb(skb);
4874 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4875
4876 return next;
4877}
4878
4879/* Collapse contiguous sequence of skbs head..tail with
4880 * sequence numbers start..end.
4881 *
4882 * If tail is NULL, this means until the end of the list.
4883 *
4884 * Segments with FIN/SYN are not collapsed (only because this
4885 * simplifies code)
4886 */
4887static void
4888tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4889 struct sk_buff *head, struct sk_buff *tail,
4890 u32 start, u32 end)
4891{
4892 struct sk_buff *skb, *n;
4893 bool end_of_skbs;
4894
4895 /* First, check that queue is collapsible and find
4896 * the point where collapsing can be useful. */
4897 skb = head;
4898restart:
4899 end_of_skbs = true;
4900 skb_queue_walk_from_safe(list, skb, n) {
4901 if (skb == tail)
4902 break;
4903 /* No new bits? It is possible on ofo queue. */
4904 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4905 skb = tcp_collapse_one(sk, skb, list);
4906 if (!skb)
4907 break;
4908 goto restart;
4909 }
4910
4911 /* The first skb to collapse is:
4912 * - not SYN/FIN and
4913 * - bloated or contains data before "start" or
4914 * overlaps to the next one.
4915 */
4916 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
4917 (tcp_win_from_space(skb->truesize) > skb->len ||
4918 before(TCP_SKB_CB(skb)->seq, start))) {
4919 end_of_skbs = false;
4920 break;
4921 }
4922
4923 if (!skb_queue_is_last(list, skb)) {
4924 struct sk_buff *next = skb_queue_next(list, skb);
4925 if (next != tail &&
4926 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4927 end_of_skbs = false;
4928 break;
4929 }
4930 }
4931
4932 /* Decided to skip this, advance start seq. */
4933 start = TCP_SKB_CB(skb)->end_seq;
4934 }
4935 if (end_of_skbs || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
4936 return;
4937
4938 while (before(start, end)) {
4939 struct sk_buff *nskb;
4940 unsigned int header = skb_headroom(skb);
4941 int copy = SKB_MAX_ORDER(header, 0);
4942
4943 /* Too big header? This can happen with IPv6. */
4944 if (copy < 0)
4945 return;
4946 if (end - start < copy)
4947 copy = end - start;
4948 nskb = alloc_skb(copy + header, GFP_ATOMIC);
4949 if (!nskb)
4950 return;
4951
4952 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
4953 skb_set_network_header(nskb, (skb_network_header(skb) -
4954 skb->head));
4955 skb_set_transport_header(nskb, (skb_transport_header(skb) -
4956 skb->head));
4957 skb_reserve(nskb, header);
4958 memcpy(nskb->head, skb->head, header);
4959 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4960 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4961 __skb_queue_before(list, skb, nskb);
4962 skb_set_owner_r(nskb, sk);
4963
4964 /* Copy data, releasing collapsed skbs. */
4965 while (copy > 0) {
4966 int offset = start - TCP_SKB_CB(skb)->seq;
4967 int size = TCP_SKB_CB(skb)->end_seq - start;
4968
4969 BUG_ON(offset < 0);
4970 if (size > 0) {
4971 size = min(copy, size);
4972 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4973 BUG();
4974 TCP_SKB_CB(nskb)->end_seq += size;
4975 copy -= size;
4976 start += size;
4977 }
4978 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4979 skb = tcp_collapse_one(sk, skb, list);
4980 if (!skb ||
4981 skb == tail ||
4982 tcp_hdr(skb)->syn ||
4983 tcp_hdr(skb)->fin)
4984 return;
4985 }
4986 }
4987 }
4988}
4989
4990/* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4991 * and tcp_collapse() them until all the queue is collapsed.
4992 */
4993static void tcp_collapse_ofo_queue(struct sock *sk)
4994{
4995 struct tcp_sock *tp = tcp_sk(sk);
4996 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4997 struct sk_buff *head;
4998 u32 start, end;
4999
5000 if (skb == NULL)
5001 return;
5002
5003 start = TCP_SKB_CB(skb)->seq;
5004 end = TCP_SKB_CB(skb)->end_seq;
5005 head = skb;
5006
5007 for (;;) {
5008 struct sk_buff *next = NULL;
5009
5010 if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
5011 next = skb_queue_next(&tp->out_of_order_queue, skb);
5012 skb = next;
5013
5014 /* Segment is terminated when we see gap or when
5015 * we are at the end of all the queue. */
5016 if (!skb ||
5017 after(TCP_SKB_CB(skb)->seq, end) ||
5018 before(TCP_SKB_CB(skb)->end_seq, start)) {
5019 tcp_collapse(sk, &tp->out_of_order_queue,
5020 head, skb, start, end);
5021 head = skb;
5022 if (!skb)
5023 break;
5024 /* Start new segment */
5025 start = TCP_SKB_CB(skb)->seq;
5026 end = TCP_SKB_CB(skb)->end_seq;
5027 } else {
5028 if (before(TCP_SKB_CB(skb)->seq, start))
5029 start = TCP_SKB_CB(skb)->seq;
5030 if (after(TCP_SKB_CB(skb)->end_seq, end))
5031 end = TCP_SKB_CB(skb)->end_seq;
5032 }
5033 }
5034}
5035
5036/*
5037 * Purge the out-of-order queue.
5038 * Return true if queue was pruned.
5039 */
5040static bool tcp_prune_ofo_queue(struct sock *sk)
5041{
5042 struct tcp_sock *tp = tcp_sk(sk);
5043 bool res = false;
5044
5045 if (!skb_queue_empty(&tp->out_of_order_queue)) {
5046 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
5047 __skb_queue_purge(&tp->out_of_order_queue);
5048
5049 /* Reset SACK state. A conforming SACK implementation will
5050 * do the same at a timeout based retransmit. When a connection
5051 * is in a sad state like this, we care only about integrity
5052 * of the connection not performance.
5053 */
5054 if (tp->rx_opt.sack_ok)
5055 tcp_sack_reset(&tp->rx_opt);
5056 sk_mem_reclaim(sk);
5057 res = true;
5058 }
5059 return res;
5060}
5061
5062/* Reduce allocated memory if we can, trying to get
5063 * the socket within its memory limits again.
5064 *
5065 * Return less than zero if we should start dropping frames
5066 * until the socket owning process reads some of the data
5067 * to stabilize the situation.
5068 */
5069static int tcp_prune_queue(struct sock *sk)
5070{
5071 struct tcp_sock *tp = tcp_sk(sk);
5072
5073 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
5074
5075 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
5076
5077 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
5078 tcp_clamp_window(sk);
5079 else if (sk_under_memory_pressure(sk))
5080 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
5081
5082 tcp_collapse_ofo_queue(sk);
5083 if (!skb_queue_empty(&sk->sk_receive_queue))
5084 tcp_collapse(sk, &sk->sk_receive_queue,
5085 skb_peek(&sk->sk_receive_queue),
5086 NULL,
5087 tp->copied_seq, tp->rcv_nxt);
5088 sk_mem_reclaim(sk);
5089
5090 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5091 return 0;
5092
5093 /* Collapsing did not help, destructive actions follow.
5094 * This must not ever occur. */
5095
5096 tcp_prune_ofo_queue(sk);
5097
5098 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5099 return 0;
5100
5101 /* If we are really being abused, tell the caller to silently
5102 * drop receive data on the floor. It will get retransmitted
5103 * and hopefully then we'll have sufficient space.
5104 */
5105 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
5106
5107 /* Massive buffer overcommit. */
5108 tp->pred_flags = 0;
5109 return -1;
5110}
5111
5112/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
5113 * As additional protections, we do not touch cwnd in retransmission phases,
5114 * and if application hit its sndbuf limit recently.
5115 */
5116void tcp_cwnd_application_limited(struct sock *sk)
5117{
5118 struct tcp_sock *tp = tcp_sk(sk);
5119
5120 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
5121 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
5122 /* Limited by application or receiver window. */
5123 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
5124 u32 win_used = max(tp->snd_cwnd_used, init_win);
5125 if (win_used < tp->snd_cwnd) {
5126 tp->snd_ssthresh = tcp_current_ssthresh(sk);
5127 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
5128 }
5129 tp->snd_cwnd_used = 0;
5130 }
5131 tp->snd_cwnd_stamp = tcp_time_stamp;
5132}
5133
5134static bool tcp_should_expand_sndbuf(const struct sock *sk)
5135{
5136 const struct tcp_sock *tp = tcp_sk(sk);
5137
5138 /* If the user specified a specific send buffer setting, do
5139 * not modify it.
5140 */
5141 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
5142 return false;
5143
5144 /* If we are under global TCP memory pressure, do not expand. */
5145 if (sk_under_memory_pressure(sk))
5146 return false;
5147
5148 /* If we are under soft global TCP memory pressure, do not expand. */
5149 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
5150 return false;
5151
5152 /* If we filled the congestion window, do not expand. */
5153 if (tp->packets_out >= tp->snd_cwnd)
5154 return false;
5155
5156 return true;
5157}
5158
5159/* When incoming ACK allowed to free some skb from write_queue,
5160 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
5161 * on the exit from tcp input handler.
5162 *
5163 * PROBLEM: sndbuf expansion does not work well with largesend.
5164 */
5165static void tcp_new_space(struct sock *sk)
5166{
5167 struct tcp_sock *tp = tcp_sk(sk);
5168
5169 if (tcp_should_expand_sndbuf(sk)) {
5170 int sndmem = SKB_TRUESIZE(max_t(u32,
5171 tp->rx_opt.mss_clamp,
5172 tp->mss_cache) +
5173 MAX_TCP_HEADER);
5174 int demanded = max_t(unsigned int, tp->snd_cwnd,
5175 tp->reordering + 1);
5176 sndmem *= 2 * demanded;
5177 if (sndmem > sk->sk_sndbuf)
5178 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
5179 tp->snd_cwnd_stamp = tcp_time_stamp;
5180 }
5181
5182 sk->sk_write_space(sk);
5183}
5184
5185static void tcp_check_space(struct sock *sk)
5186{
5187 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
5188 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
5189 if (sk->sk_socket &&
5190 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5191 tcp_new_space(sk);
5192 }
5193}
5194
5195static inline void tcp_data_snd_check(struct sock *sk)
5196{
5197 tcp_push_pending_frames(sk);
5198 tcp_check_space(sk);
5199}
5200
5201/*
5202 * Check if sending an ack is needed.
5203 */
5204static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5205{
5206 struct tcp_sock *tp = tcp_sk(sk);
5207
5208 /* More than one full frame received... */
5209 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5210 /* ... and right edge of window advances far enough.
5211 * (tcp_recvmsg() will send ACK otherwise). Or...
5212 */
5213 __tcp_select_window(sk) >= tp->rcv_wnd) ||
5214 /* We ACK each frame or... */
5215 tcp_in_quickack_mode(sk) ||
5216 /* We have out of order data. */
5217 (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
5218 /* Then ack it now */
5219 tcp_send_ack(sk);
5220 } else {
5221 /* Else, send delayed ack. */
5222 tcp_send_delayed_ack(sk);
5223 }
5224}
5225
5226static inline void tcp_ack_snd_check(struct sock *sk)
5227{
5228 if (!inet_csk_ack_scheduled(sk)) {
5229 /* We sent a data segment already. */
5230 return;
5231 }
5232 __tcp_ack_snd_check(sk, 1);
5233}
5234
5235/*
5236 * This routine is only called when we have urgent data
5237 * signaled. Its the 'slow' part of tcp_urg. It could be
5238 * moved inline now as tcp_urg is only called from one
5239 * place. We handle URGent data wrong. We have to - as
5240 * BSD still doesn't use the correction from RFC961.
5241 * For 1003.1g we should support a new option TCP_STDURG to permit
5242 * either form (or just set the sysctl tcp_stdurg).
5243 */
5244
5245static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5246{
5247 struct tcp_sock *tp = tcp_sk(sk);
5248 u32 ptr = ntohs(th->urg_ptr);
5249
5250 if (ptr && !sysctl_tcp_stdurg)
5251 ptr--;
5252 ptr += ntohl(th->seq);
5253
5254 /* Ignore urgent data that we've already seen and read. */
5255 if (after(tp->copied_seq, ptr))
5256 return;
5257
5258 /* Do not replay urg ptr.
5259 *
5260 * NOTE: interesting situation not covered by specs.
5261 * Misbehaving sender may send urg ptr, pointing to segment,
5262 * which we already have in ofo queue. We are not able to fetch
5263 * such data and will stay in TCP_URG_NOTYET until will be eaten
5264 * by recvmsg(). Seems, we are not obliged to handle such wicked
5265 * situations. But it is worth to think about possibility of some
5266 * DoSes using some hypothetical application level deadlock.
5267 */
5268 if (before(ptr, tp->rcv_nxt))
5269 return;
5270
5271 /* Do we already have a newer (or duplicate) urgent pointer? */
5272 if (tp->urg_data && !after(ptr, tp->urg_seq))
5273 return;
5274
5275 /* Tell the world about our new urgent pointer. */
5276 sk_send_sigurg(sk);
5277
5278 /* We may be adding urgent data when the last byte read was
5279 * urgent. To do this requires some care. We cannot just ignore
5280 * tp->copied_seq since we would read the last urgent byte again
5281 * as data, nor can we alter copied_seq until this data arrives
5282 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5283 *
5284 * NOTE. Double Dutch. Rendering to plain English: author of comment
5285 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5286 * and expect that both A and B disappear from stream. This is _wrong_.
5287 * Though this happens in BSD with high probability, this is occasional.
5288 * Any application relying on this is buggy. Note also, that fix "works"
5289 * only in this artificial test. Insert some normal data between A and B and we will
5290 * decline of BSD again. Verdict: it is better to remove to trap
5291 * buggy users.
5292 */
5293 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5294 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5295 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5296 tp->copied_seq++;
5297 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5298 __skb_unlink(skb, &sk->sk_receive_queue);
5299 __kfree_skb(skb);
5300 }
5301 }
5302
5303 tp->urg_data = TCP_URG_NOTYET;
5304 tp->urg_seq = ptr;
5305
5306 /* Disable header prediction. */
5307 tp->pred_flags = 0;
5308}
5309
5310/* This is the 'fast' part of urgent handling. */
5311static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5312{
5313 struct tcp_sock *tp = tcp_sk(sk);
5314
5315 /* Check if we get a new urgent pointer - normally not. */
5316 if (th->urg)
5317 tcp_check_urg(sk, th);
5318
5319 /* Do we wait for any urgent data? - normally not... */
5320 if (tp->urg_data == TCP_URG_NOTYET) {
5321 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5322 th->syn;
5323
5324 /* Is the urgent pointer pointing into this packet? */
5325 if (ptr < skb->len) {
5326 u8 tmp;
5327 if (skb_copy_bits(skb, ptr, &tmp, 1))
5328 BUG();
5329 tp->urg_data = TCP_URG_VALID | tmp;
5330 if (!sock_flag(sk, SOCK_DEAD))
5331 sk->sk_data_ready(sk, 0);
5332 }
5333 }
5334}
5335
5336static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
5337{
5338 struct tcp_sock *tp = tcp_sk(sk);
5339 int chunk = skb->len - hlen;
5340 int err;
5341
5342 local_bh_enable();
5343 if (skb_csum_unnecessary(skb))
5344 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
5345 else
5346 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
5347 tp->ucopy.iov);
5348
5349 if (!err) {
5350 tp->ucopy.len -= chunk;
5351 tp->copied_seq += chunk;
5352 tcp_rcv_space_adjust(sk);
5353 }
5354
5355 local_bh_disable();
5356 return err;
5357}
5358
5359static __sum16 __tcp_checksum_complete_user(struct sock *sk,
5360 struct sk_buff *skb)
5361{
5362 __sum16 result;
5363
5364 if (sock_owned_by_user(sk)) {
5365 local_bh_enable();
5366 result = __tcp_checksum_complete(skb);
5367 local_bh_disable();
5368 } else {
5369 result = __tcp_checksum_complete(skb);
5370 }
5371 return result;
5372}
5373
5374static inline int tcp_checksum_complete_user(struct sock *sk,
5375 struct sk_buff *skb)
5376{
5377 return !skb_csum_unnecessary(skb) &&
5378 __tcp_checksum_complete_user(sk, skb);
5379}
5380
5381#ifdef CONFIG_NET_DMA
5382static bool tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb,
5383 int hlen)
5384{
5385 struct tcp_sock *tp = tcp_sk(sk);
5386 int chunk = skb->len - hlen;
5387 int dma_cookie;
5388 bool copied_early = false;
5389
5390 if (tp->ucopy.wakeup)
5391 return false;
5392
5393 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
5394 tp->ucopy.dma_chan = net_dma_find_channel();
5395
5396 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
5397
5398 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
5399 skb, hlen,
5400 tp->ucopy.iov, chunk,
5401 tp->ucopy.pinned_list);
5402
5403 if (dma_cookie < 0)
5404 goto out;
5405
5406 tp->ucopy.dma_cookie = dma_cookie;
5407 copied_early = true;
5408
5409 tp->ucopy.len -= chunk;
5410 tp->copied_seq += chunk;
5411 tcp_rcv_space_adjust(sk);
5412
5413 if ((tp->ucopy.len == 0) ||
5414 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
5415 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
5416 tp->ucopy.wakeup = 1;
5417 sk->sk_data_ready(sk, 0);
5418 }
5419 } else if (chunk > 0) {
5420 tp->ucopy.wakeup = 1;
5421 sk->sk_data_ready(sk, 0);
5422 }
5423out:
5424 return copied_early;
5425}
5426#endif /* CONFIG_NET_DMA */
5427
5428/* Does PAWS and seqno based validation of an incoming segment, flags will
5429 * play significant role here.
5430 */
5431static int tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5432 const struct tcphdr *th, int syn_inerr)
5433{
5434 const u8 *hash_location;
5435 struct tcp_sock *tp = tcp_sk(sk);
5436
5437 /* RFC1323: H1. Apply PAWS check first. */
5438 if (tcp_fast_parse_options(skb, th, tp, &hash_location) &&
5439 tp->rx_opt.saw_tstamp &&
5440 tcp_paws_discard(sk, skb)) {
5441 if (!th->rst) {
5442 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5443 tcp_send_dupack(sk, skb);
5444 goto discard;
5445 }
5446 /* Reset is accepted even if it did not pass PAWS. */
5447 }
5448
5449 /* Step 1: check sequence number */
5450 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5451 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5452 * (RST) segments are validated by checking their SEQ-fields."
5453 * And page 69: "If an incoming segment is not acceptable,
5454 * an acknowledgment should be sent in reply (unless the RST
5455 * bit is set, if so drop the segment and return)".
5456 */
5457 if (!th->rst)
5458 tcp_send_dupack(sk, skb);
5459 goto discard;
5460 }
5461
5462 /* Step 2: check RST bit */
5463 if (th->rst) {
5464 tcp_reset(sk);
5465 goto discard;
5466 }
5467
5468 /* ts_recent update must be made after we are sure that the packet
5469 * is in window.
5470 */
5471 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
5472
5473 /* step 3: check security and precedence [ignored] */
5474
5475 /* step 4: Check for a SYN in window. */
5476 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5477 if (syn_inerr)
5478 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5479 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONSYN);
5480 tcp_reset(sk);
5481 return -1;
5482 }
5483
5484 return 1;
5485
5486discard:
5487 __kfree_skb(skb);
5488 return 0;
5489}
5490
5491/*
5492 * TCP receive function for the ESTABLISHED state.
5493 *
5494 * It is split into a fast path and a slow path. The fast path is
5495 * disabled when:
5496 * - A zero window was announced from us - zero window probing
5497 * is only handled properly in the slow path.
5498 * - Out of order segments arrived.
5499 * - Urgent data is expected.
5500 * - There is no buffer space left
5501 * - Unexpected TCP flags/window values/header lengths are received
5502 * (detected by checking the TCP header against pred_flags)
5503 * - Data is sent in both directions. Fast path only supports pure senders
5504 * or pure receivers (this means either the sequence number or the ack
5505 * value must stay constant)
5506 * - Unexpected TCP option.
5507 *
5508 * When these conditions are not satisfied it drops into a standard
5509 * receive procedure patterned after RFC793 to handle all cases.
5510 * The first three cases are guaranteed by proper pred_flags setting,
5511 * the rest is checked inline. Fast processing is turned on in
5512 * tcp_data_queue when everything is OK.
5513 */
5514int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5515 const struct tcphdr *th, unsigned int len)
5516{
5517 struct tcp_sock *tp = tcp_sk(sk);
5518 int res;
5519
5520 /*
5521 * Header prediction.
5522 * The code loosely follows the one in the famous
5523 * "30 instruction TCP receive" Van Jacobson mail.
5524 *
5525 * Van's trick is to deposit buffers into socket queue
5526 * on a device interrupt, to call tcp_recv function
5527 * on the receive process context and checksum and copy
5528 * the buffer to user space. smart...
5529 *
5530 * Our current scheme is not silly either but we take the
5531 * extra cost of the net_bh soft interrupt processing...
5532 * We do checksum and copy also but from device to kernel.
5533 */
5534
5535 tp->rx_opt.saw_tstamp = 0;
5536
5537 /* pred_flags is 0xS?10 << 16 + snd_wnd
5538 * if header_prediction is to be made
5539 * 'S' will always be tp->tcp_header_len >> 2
5540 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5541 * turn it off (when there are holes in the receive
5542 * space for instance)
5543 * PSH flag is ignored.
5544 */
5545
5546 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5547 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5548 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5549 int tcp_header_len = tp->tcp_header_len;
5550
5551 /* Timestamp header prediction: tcp_header_len
5552 * is automatically equal to th->doff*4 due to pred_flags
5553 * match.
5554 */
5555
5556 /* Check timestamp */
5557 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5558 /* No? Slow path! */
5559 if (!tcp_parse_aligned_timestamp(tp, th))
5560 goto slow_path;
5561
5562 /* If PAWS failed, check it more carefully in slow path */
5563 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5564 goto slow_path;
5565
5566 /* DO NOT update ts_recent here, if checksum fails
5567 * and timestamp was corrupted part, it will result
5568 * in a hung connection since we will drop all
5569 * future packets due to the PAWS test.
5570 */
5571 }
5572
5573 if (len <= tcp_header_len) {
5574 /* Bulk data transfer: sender */
5575 if (len == tcp_header_len) {
5576 /* Predicted packet is in window by definition.
5577 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5578 * Hence, check seq<=rcv_wup reduces to:
5579 */
5580 if (tcp_header_len ==
5581 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5582 tp->rcv_nxt == tp->rcv_wup)
5583 tcp_store_ts_recent(tp);
5584
5585 /* We know that such packets are checksummed
5586 * on entry.
5587 */
5588 tcp_ack(sk, skb, 0);
5589 __kfree_skb(skb);
5590 tcp_data_snd_check(sk);
5591 return 0;
5592 } else { /* Header too small */
5593 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5594 goto discard;
5595 }
5596 } else {
5597 int eaten = 0;
5598 int copied_early = 0;
5599 bool fragstolen = false;
5600
5601 if (tp->copied_seq == tp->rcv_nxt &&
5602 len - tcp_header_len <= tp->ucopy.len) {
5603#ifdef CONFIG_NET_DMA
5604 if (tp->ucopy.task == current &&
5605 sock_owned_by_user(sk) &&
5606 tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
5607 copied_early = 1;
5608 eaten = 1;
5609 }
5610#endif
5611 if (tp->ucopy.task == current &&
5612 sock_owned_by_user(sk) && !copied_early) {
5613 __set_current_state(TASK_RUNNING);
5614
5615 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
5616 eaten = 1;
5617 }
5618 if (eaten) {
5619 /* Predicted packet is in window by definition.
5620 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5621 * Hence, check seq<=rcv_wup reduces to:
5622 */
5623 if (tcp_header_len ==
5624 (sizeof(struct tcphdr) +
5625 TCPOLEN_TSTAMP_ALIGNED) &&
5626 tp->rcv_nxt == tp->rcv_wup)
5627 tcp_store_ts_recent(tp);
5628
5629 tcp_rcv_rtt_measure_ts(sk, skb);
5630
5631 __skb_pull(skb, tcp_header_len);
5632 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5633 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5634 }
5635 if (copied_early)
5636 tcp_cleanup_rbuf(sk, skb->len);
5637 }
5638 if (!eaten) {
5639 if (tcp_checksum_complete_user(sk, skb))
5640 goto csum_error;
5641
5642 /* Predicted packet is in window by definition.
5643 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5644 * Hence, check seq<=rcv_wup reduces to:
5645 */
5646 if (tcp_header_len ==
5647 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5648 tp->rcv_nxt == tp->rcv_wup)
5649 tcp_store_ts_recent(tp);
5650
5651 tcp_rcv_rtt_measure_ts(sk, skb);
5652
5653 if ((int)skb->truesize > sk->sk_forward_alloc)
5654 goto step5;
5655
5656 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5657
5658 /* Bulk data transfer: receiver */
5659 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5660 &fragstolen);
5661 }
5662
5663 tcp_event_data_recv(sk, skb);
5664
5665 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5666 /* Well, only one small jumplet in fast path... */
5667 tcp_ack(sk, skb, FLAG_DATA);
5668 tcp_data_snd_check(sk);
5669 if (!inet_csk_ack_scheduled(sk))
5670 goto no_ack;
5671 }
5672
5673 if (!copied_early || tp->rcv_nxt != tp->rcv_wup)
5674 __tcp_ack_snd_check(sk, 0);
5675no_ack:
5676#ifdef CONFIG_NET_DMA
5677 if (copied_early)
5678 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
5679 else
5680#endif
5681 if (eaten)
5682 kfree_skb_partial(skb, fragstolen);
5683 else
5684 sk->sk_data_ready(sk, 0);
5685 return 0;
5686 }
5687 }
5688
5689slow_path:
5690 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5691 goto csum_error;
5692
5693 /*
5694 * Standard slow path.
5695 */
5696
5697 res = tcp_validate_incoming(sk, skb, th, 1);
5698 if (res <= 0)
5699 return -res;
5700
5701step5:
5702 if (th->ack && tcp_ack(sk, skb, FLAG_SLOWPATH) < 0)
5703 goto discard;
5704
5705 tcp_rcv_rtt_measure_ts(sk, skb);
5706
5707 /* Process urgent data. */
5708 tcp_urg(sk, skb, th);
5709
5710 /* step 7: process the segment text */
5711 tcp_data_queue(sk, skb);
5712
5713 tcp_data_snd_check(sk);
5714 tcp_ack_snd_check(sk);
5715 return 0;
5716
5717csum_error:
5718 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5719
5720discard:
5721 __kfree_skb(skb);
5722 return 0;
5723}
5724EXPORT_SYMBOL(tcp_rcv_established);
5725
5726void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5727{
5728 struct tcp_sock *tp = tcp_sk(sk);
5729 struct inet_connection_sock *icsk = inet_csk(sk);
5730
5731 tcp_set_state(sk, TCP_ESTABLISHED);
5732
5733 if (skb != NULL)
5734 security_inet_conn_established(sk, skb);
5735
5736 /* Make sure socket is routed, for correct metrics. */
5737 icsk->icsk_af_ops->rebuild_header(sk);
5738
5739 tcp_init_metrics(sk);
5740
5741 tcp_init_congestion_control(sk);
5742
5743 /* Prevent spurious tcp_cwnd_restart() on first data
5744 * packet.
5745 */
5746 tp->lsndtime = tcp_time_stamp;
5747
5748 tcp_init_buffer_space(sk);
5749
5750 if (sock_flag(sk, SOCK_KEEPOPEN))
5751 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5752
5753 if (!tp->rx_opt.snd_wscale)
5754 __tcp_fast_path_on(tp, tp->snd_wnd);
5755 else
5756 tp->pred_flags = 0;
5757
5758 if (!sock_flag(sk, SOCK_DEAD)) {
5759 sk->sk_state_change(sk);
5760 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5761 }
5762}
5763
5764static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5765 const struct tcphdr *th, unsigned int len)
5766{
5767 const u8 *hash_location;
5768 struct inet_connection_sock *icsk = inet_csk(sk);
5769 struct tcp_sock *tp = tcp_sk(sk);
5770 struct tcp_cookie_values *cvp = tp->cookie_values;
5771 int saved_clamp = tp->rx_opt.mss_clamp;
5772
5773 tcp_parse_options(skb, &tp->rx_opt, &hash_location, 0);
5774
5775 if (th->ack) {
5776 /* rfc793:
5777 * "If the state is SYN-SENT then
5778 * first check the ACK bit
5779 * If the ACK bit is set
5780 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5781 * a reset (unless the RST bit is set, if so drop
5782 * the segment and return)"
5783 *
5784 * We do not send data with SYN, so that RFC-correct
5785 * test reduces to:
5786 */
5787 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
5788 goto reset_and_undo;
5789
5790 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5791 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5792 tcp_time_stamp)) {
5793 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5794 goto reset_and_undo;
5795 }
5796
5797 /* Now ACK is acceptable.
5798 *
5799 * "If the RST bit is set
5800 * If the ACK was acceptable then signal the user "error:
5801 * connection reset", drop the segment, enter CLOSED state,
5802 * delete TCB, and return."
5803 */
5804
5805 if (th->rst) {
5806 tcp_reset(sk);
5807 goto discard;
5808 }
5809
5810 /* rfc793:
5811 * "fifth, if neither of the SYN or RST bits is set then
5812 * drop the segment and return."
5813 *
5814 * See note below!
5815 * --ANK(990513)
5816 */
5817 if (!th->syn)
5818 goto discard_and_undo;
5819
5820 /* rfc793:
5821 * "If the SYN bit is on ...
5822 * are acceptable then ...
5823 * (our SYN has been ACKed), change the connection
5824 * state to ESTABLISHED..."
5825 */
5826
5827 TCP_ECN_rcv_synack(tp, th);
5828
5829 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5830 tcp_ack(sk, skb, FLAG_SLOWPATH);
5831
5832 /* Ok.. it's good. Set up sequence numbers and
5833 * move to established.
5834 */
5835 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5836 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5837
5838 /* RFC1323: The window in SYN & SYN/ACK segments is
5839 * never scaled.
5840 */
5841 tp->snd_wnd = ntohs(th->window);
5842 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5843
5844 if (!tp->rx_opt.wscale_ok) {
5845 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5846 tp->window_clamp = min(tp->window_clamp, 65535U);
5847 }
5848
5849 if (tp->rx_opt.saw_tstamp) {
5850 tp->rx_opt.tstamp_ok = 1;
5851 tp->tcp_header_len =
5852 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5853 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5854 tcp_store_ts_recent(tp);
5855 } else {
5856 tp->tcp_header_len = sizeof(struct tcphdr);
5857 }
5858
5859 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5860 tcp_enable_fack(tp);
5861
5862 tcp_mtup_init(sk);
5863 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5864 tcp_initialize_rcv_mss(sk);
5865
5866 /* Remember, tcp_poll() does not lock socket!
5867 * Change state from SYN-SENT only after copied_seq
5868 * is initialized. */
5869 tp->copied_seq = tp->rcv_nxt;
5870
5871 if (cvp != NULL &&
5872 cvp->cookie_pair_size > 0 &&
5873 tp->rx_opt.cookie_plus > 0) {
5874 int cookie_size = tp->rx_opt.cookie_plus
5875 - TCPOLEN_COOKIE_BASE;
5876 int cookie_pair_size = cookie_size
5877 + cvp->cookie_desired;
5878
5879 /* A cookie extension option was sent and returned.
5880 * Note that each incoming SYNACK replaces the
5881 * Responder cookie. The initial exchange is most
5882 * fragile, as protection against spoofing relies
5883 * entirely upon the sequence and timestamp (above).
5884 * This replacement strategy allows the correct pair to
5885 * pass through, while any others will be filtered via
5886 * Responder verification later.
5887 */
5888 if (sizeof(cvp->cookie_pair) >= cookie_pair_size) {
5889 memcpy(&cvp->cookie_pair[cvp->cookie_desired],
5890 hash_location, cookie_size);
5891 cvp->cookie_pair_size = cookie_pair_size;
5892 }
5893 }
5894
5895 smp_mb();
5896
5897 tcp_finish_connect(sk, skb);
5898
5899 if (sk->sk_write_pending ||
5900 icsk->icsk_accept_queue.rskq_defer_accept ||
5901 icsk->icsk_ack.pingpong) {
5902 /* Save one ACK. Data will be ready after
5903 * several ticks, if write_pending is set.
5904 *
5905 * It may be deleted, but with this feature tcpdumps
5906 * look so _wonderfully_ clever, that I was not able
5907 * to stand against the temptation 8) --ANK
5908 */
5909 inet_csk_schedule_ack(sk);
5910 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5911 tcp_enter_quickack_mode(sk);
5912 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5913 TCP_DELACK_MAX, TCP_RTO_MAX);
5914
5915discard:
5916 __kfree_skb(skb);
5917 return 0;
5918 } else {
5919 tcp_send_ack(sk);
5920 }
5921 return -1;
5922 }
5923
5924 /* No ACK in the segment */
5925
5926 if (th->rst) {
5927 /* rfc793:
5928 * "If the RST bit is set
5929 *
5930 * Otherwise (no ACK) drop the segment and return."
5931 */
5932
5933 goto discard_and_undo;
5934 }
5935
5936 /* PAWS check. */
5937 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5938 tcp_paws_reject(&tp->rx_opt, 0))
5939 goto discard_and_undo;
5940
5941 if (th->syn) {
5942 /* We see SYN without ACK. It is attempt of
5943 * simultaneous connect with crossed SYNs.
5944 * Particularly, it can be connect to self.
5945 */
5946 tcp_set_state(sk, TCP_SYN_RECV);
5947
5948 if (tp->rx_opt.saw_tstamp) {
5949 tp->rx_opt.tstamp_ok = 1;
5950 tcp_store_ts_recent(tp);
5951 tp->tcp_header_len =
5952 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5953 } else {
5954 tp->tcp_header_len = sizeof(struct tcphdr);
5955 }
5956
5957 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5958 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5959
5960 /* RFC1323: The window in SYN & SYN/ACK segments is
5961 * never scaled.
5962 */
5963 tp->snd_wnd = ntohs(th->window);
5964 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5965 tp->max_window = tp->snd_wnd;
5966
5967 TCP_ECN_rcv_syn(tp, th);
5968
5969 tcp_mtup_init(sk);
5970 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5971 tcp_initialize_rcv_mss(sk);
5972
5973 tcp_send_synack(sk);
5974#if 0
5975 /* Note, we could accept data and URG from this segment.
5976 * There are no obstacles to make this.
5977 *
5978 * However, if we ignore data in ACKless segments sometimes,
5979 * we have no reasons to accept it sometimes.
5980 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5981 * is not flawless. So, discard packet for sanity.
5982 * Uncomment this return to process the data.
5983 */
5984 return -1;
5985#else
5986 goto discard;
5987#endif
5988 }
5989 /* "fifth, if neither of the SYN or RST bits is set then
5990 * drop the segment and return."
5991 */
5992
5993discard_and_undo:
5994 tcp_clear_options(&tp->rx_opt);
5995 tp->rx_opt.mss_clamp = saved_clamp;
5996 goto discard;
5997
5998reset_and_undo:
5999 tcp_clear_options(&tp->rx_opt);
6000 tp->rx_opt.mss_clamp = saved_clamp;
6001 return 1;
6002}
6003
6004/*
6005 * This function implements the receiving procedure of RFC 793 for
6006 * all states except ESTABLISHED and TIME_WAIT.
6007 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
6008 * address independent.
6009 */
6010
6011int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
6012 const struct tcphdr *th, unsigned int len)
6013{
6014 struct tcp_sock *tp = tcp_sk(sk);
6015 struct inet_connection_sock *icsk = inet_csk(sk);
6016 int queued = 0;
6017 int res;
6018
6019 tp->rx_opt.saw_tstamp = 0;
6020
6021 switch (sk->sk_state) {
6022 case TCP_CLOSE:
6023 goto discard;
6024
6025 case TCP_LISTEN:
6026 if (th->ack)
6027 return 1;
6028
6029 if (th->rst)
6030 goto discard;
6031
6032 if (th->syn) {
6033 if (th->fin)
6034 goto discard;
6035 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
6036 return 1;
6037
6038 /* Now we have several options: In theory there is
6039 * nothing else in the frame. KA9Q has an option to
6040 * send data with the syn, BSD accepts data with the
6041 * syn up to the [to be] advertised window and
6042 * Solaris 2.1 gives you a protocol error. For now
6043 * we just ignore it, that fits the spec precisely
6044 * and avoids incompatibilities. It would be nice in
6045 * future to drop through and process the data.
6046 *
6047 * Now that TTCP is starting to be used we ought to
6048 * queue this data.
6049 * But, this leaves one open to an easy denial of
6050 * service attack, and SYN cookies can't defend
6051 * against this problem. So, we drop the data
6052 * in the interest of security over speed unless
6053 * it's still in use.
6054 */
6055 kfree_skb(skb);
6056 return 0;
6057 }
6058 goto discard;
6059
6060 case TCP_SYN_SENT:
6061 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
6062 if (queued >= 0)
6063 return queued;
6064
6065 /* Do step6 onward by hand. */
6066 tcp_urg(sk, skb, th);
6067 __kfree_skb(skb);
6068 tcp_data_snd_check(sk);
6069 return 0;
6070 }
6071
6072 res = tcp_validate_incoming(sk, skb, th, 0);
6073 if (res <= 0)
6074 return -res;
6075
6076 /* step 5: check the ACK field */
6077 if (th->ack) {
6078 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH) > 0;
6079
6080 switch (sk->sk_state) {
6081 case TCP_SYN_RECV:
6082 if (acceptable) {
6083 tp->copied_seq = tp->rcv_nxt;
6084 smp_mb();
6085 tcp_set_state(sk, TCP_ESTABLISHED);
6086 sk->sk_state_change(sk);
6087
6088 /* Note, that this wakeup is only for marginal
6089 * crossed SYN case. Passively open sockets
6090 * are not waked up, because sk->sk_sleep ==
6091 * NULL and sk->sk_socket == NULL.
6092 */
6093 if (sk->sk_socket)
6094 sk_wake_async(sk,
6095 SOCK_WAKE_IO, POLL_OUT);
6096
6097 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
6098 tp->snd_wnd = ntohs(th->window) <<
6099 tp->rx_opt.snd_wscale;
6100 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6101
6102 if (tp->rx_opt.tstamp_ok)
6103 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6104
6105 /* Make sure socket is routed, for
6106 * correct metrics.
6107 */
6108 icsk->icsk_af_ops->rebuild_header(sk);
6109
6110 tcp_init_metrics(sk);
6111
6112 tcp_init_congestion_control(sk);
6113
6114 /* Prevent spurious tcp_cwnd_restart() on
6115 * first data packet.
6116 */
6117 tp->lsndtime = tcp_time_stamp;
6118
6119 tcp_mtup_init(sk);
6120 tcp_initialize_rcv_mss(sk);
6121 tcp_init_buffer_space(sk);
6122 tcp_fast_path_on(tp);
6123 } else {
6124 return 1;
6125 }
6126 break;
6127
6128 case TCP_FIN_WAIT1:
6129 if (tp->snd_una == tp->write_seq) {
6130 tcp_set_state(sk, TCP_FIN_WAIT2);
6131 sk->sk_shutdown |= SEND_SHUTDOWN;
6132 dst_confirm(__sk_dst_get(sk));
6133
6134 if (!sock_flag(sk, SOCK_DEAD))
6135 /* Wake up lingering close() */
6136 sk->sk_state_change(sk);
6137 else {
6138 int tmo;
6139
6140 if (tp->linger2 < 0 ||
6141 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6142 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
6143 tcp_done(sk);
6144 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6145 return 1;
6146 }
6147
6148 tmo = tcp_fin_time(sk);
6149 if (tmo > TCP_TIMEWAIT_LEN) {
6150 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
6151 } else if (th->fin || sock_owned_by_user(sk)) {
6152 /* Bad case. We could lose such FIN otherwise.
6153 * It is not a big problem, but it looks confusing
6154 * and not so rare event. We still can lose it now,
6155 * if it spins in bh_lock_sock(), but it is really
6156 * marginal case.
6157 */
6158 inet_csk_reset_keepalive_timer(sk, tmo);
6159 } else {
6160 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6161 goto discard;
6162 }
6163 }
6164 }
6165 break;
6166
6167 case TCP_CLOSING:
6168 if (tp->snd_una == tp->write_seq) {
6169 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6170 goto discard;
6171 }
6172 break;
6173
6174 case TCP_LAST_ACK:
6175 if (tp->snd_una == tp->write_seq) {
6176 tcp_update_metrics(sk);
6177 tcp_done(sk);
6178 goto discard;
6179 }
6180 break;
6181 }
6182 } else
6183 goto discard;
6184
6185 /* step 6: check the URG bit */
6186 tcp_urg(sk, skb, th);
6187
6188 /* step 7: process the segment text */
6189 switch (sk->sk_state) {
6190 case TCP_CLOSE_WAIT:
6191 case TCP_CLOSING:
6192 case TCP_LAST_ACK:
6193 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
6194 break;
6195 case TCP_FIN_WAIT1:
6196 case TCP_FIN_WAIT2:
6197 /* RFC 793 says to queue data in these states,
6198 * RFC 1122 says we MUST send a reset.
6199 * BSD 4.4 also does reset.
6200 */
6201 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6202 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6203 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6204 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6205 tcp_reset(sk);
6206 return 1;
6207 }
6208 }
6209 /* Fall through */
6210 case TCP_ESTABLISHED:
6211 tcp_data_queue(sk, skb);
6212 queued = 1;
6213 break;
6214 }
6215
6216 /* tcp_data could move socket to TIME-WAIT */
6217 if (sk->sk_state != TCP_CLOSE) {
6218 tcp_data_snd_check(sk);
6219 tcp_ack_snd_check(sk);
6220 }
6221
6222 if (!queued) {
6223discard:
6224 __kfree_skb(skb);
6225 }
6226 return 0;
6227}
6228EXPORT_SYMBOL(tcp_rcv_state_process);
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 <linux/prefetch.h>
72#include <net/dst.h>
73#include <net/tcp.h>
74#include <net/inet_common.h>
75#include <linux/ipsec.h>
76#include <asm/unaligned.h>
77#include <linux/errqueue.h>
78
79int sysctl_tcp_timestamps __read_mostly = 1;
80int sysctl_tcp_window_scaling __read_mostly = 1;
81int sysctl_tcp_sack __read_mostly = 1;
82int sysctl_tcp_fack __read_mostly = 1;
83int sysctl_tcp_max_reordering __read_mostly = 300;
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);
88EXPORT_SYMBOL(sysctl_tcp_timestamps);
89
90/* rfc5961 challenge ack rate limiting */
91int sysctl_tcp_challenge_ack_limit = 1000;
92
93int sysctl_tcp_stdurg __read_mostly;
94int sysctl_tcp_rfc1337 __read_mostly;
95int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
96int sysctl_tcp_frto __read_mostly = 2;
97int sysctl_tcp_min_rtt_wlen __read_mostly = 300;
98
99int sysctl_tcp_thin_dupack __read_mostly;
100
101int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
102int sysctl_tcp_early_retrans __read_mostly = 3;
103int sysctl_tcp_invalid_ratelimit __read_mostly = HZ/2;
104
105#define FLAG_DATA 0x01 /* Incoming frame contained data. */
106#define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
107#define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
108#define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
109#define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
110#define FLAG_DATA_SACKED 0x20 /* New SACK. */
111#define FLAG_ECE 0x40 /* ECE in this ACK */
112#define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
113#define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
114#define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
115#define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
116#define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
117#define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
118#define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
119
120#define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
121#define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
122#define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
123#define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
124
125#define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
126#define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
127
128#define REXMIT_NONE 0 /* no loss recovery to do */
129#define REXMIT_LOST 1 /* retransmit packets marked lost */
130#define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */
131
132static void tcp_gro_dev_warn(struct sock *sk, const struct sk_buff *skb)
133{
134 static bool __once __read_mostly;
135
136 if (!__once) {
137 struct net_device *dev;
138
139 __once = true;
140
141 rcu_read_lock();
142 dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif);
143 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n",
144 dev ? dev->name : "Unknown driver");
145 rcu_read_unlock();
146 }
147}
148
149/* Adapt the MSS value used to make delayed ack decision to the
150 * real world.
151 */
152static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
153{
154 struct inet_connection_sock *icsk = inet_csk(sk);
155 const unsigned int lss = icsk->icsk_ack.last_seg_size;
156 unsigned int len;
157
158 icsk->icsk_ack.last_seg_size = 0;
159
160 /* skb->len may jitter because of SACKs, even if peer
161 * sends good full-sized frames.
162 */
163 len = skb_shinfo(skb)->gso_size ? : skb->len;
164 if (len >= icsk->icsk_ack.rcv_mss) {
165 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len,
166 tcp_sk(sk)->advmss);
167 if (unlikely(icsk->icsk_ack.rcv_mss != len))
168 tcp_gro_dev_warn(sk, skb);
169 } else {
170 /* Otherwise, we make more careful check taking into account,
171 * that SACKs block is variable.
172 *
173 * "len" is invariant segment length, including TCP header.
174 */
175 len += skb->data - skb_transport_header(skb);
176 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
177 /* If PSH is not set, packet should be
178 * full sized, provided peer TCP is not badly broken.
179 * This observation (if it is correct 8)) allows
180 * to handle super-low mtu links fairly.
181 */
182 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
183 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
184 /* Subtract also invariant (if peer is RFC compliant),
185 * tcp header plus fixed timestamp option length.
186 * Resulting "len" is MSS free of SACK jitter.
187 */
188 len -= tcp_sk(sk)->tcp_header_len;
189 icsk->icsk_ack.last_seg_size = len;
190 if (len == lss) {
191 icsk->icsk_ack.rcv_mss = len;
192 return;
193 }
194 }
195 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
196 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
197 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
198 }
199}
200
201static void tcp_incr_quickack(struct sock *sk)
202{
203 struct inet_connection_sock *icsk = inet_csk(sk);
204 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
205
206 if (quickacks == 0)
207 quickacks = 2;
208 if (quickacks > icsk->icsk_ack.quick)
209 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
210}
211
212static void tcp_enter_quickack_mode(struct sock *sk)
213{
214 struct inet_connection_sock *icsk = inet_csk(sk);
215 tcp_incr_quickack(sk);
216 icsk->icsk_ack.pingpong = 0;
217 icsk->icsk_ack.ato = TCP_ATO_MIN;
218}
219
220/* Send ACKs quickly, if "quick" count is not exhausted
221 * and the session is not interactive.
222 */
223
224static bool tcp_in_quickack_mode(struct sock *sk)
225{
226 const struct inet_connection_sock *icsk = inet_csk(sk);
227 const struct dst_entry *dst = __sk_dst_get(sk);
228
229 return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
230 (icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong);
231}
232
233static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
234{
235 if (tp->ecn_flags & TCP_ECN_OK)
236 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
237}
238
239static void tcp_ecn_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
240{
241 if (tcp_hdr(skb)->cwr)
242 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
243}
244
245static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
246{
247 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
248}
249
250static void __tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
251{
252 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
253 case INET_ECN_NOT_ECT:
254 /* Funny extension: if ECT is not set on a segment,
255 * and we already seen ECT on a previous segment,
256 * it is probably a retransmit.
257 */
258 if (tp->ecn_flags & TCP_ECN_SEEN)
259 tcp_enter_quickack_mode((struct sock *)tp);
260 break;
261 case INET_ECN_CE:
262 if (tcp_ca_needs_ecn((struct sock *)tp))
263 tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_IS_CE);
264
265 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
266 /* Better not delay acks, sender can have a very low cwnd */
267 tcp_enter_quickack_mode((struct sock *)tp);
268 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
269 }
270 tp->ecn_flags |= TCP_ECN_SEEN;
271 break;
272 default:
273 if (tcp_ca_needs_ecn((struct sock *)tp))
274 tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_NO_CE);
275 tp->ecn_flags |= TCP_ECN_SEEN;
276 break;
277 }
278}
279
280static void tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
281{
282 if (tp->ecn_flags & TCP_ECN_OK)
283 __tcp_ecn_check_ce(tp, skb);
284}
285
286static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
287{
288 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
289 tp->ecn_flags &= ~TCP_ECN_OK;
290}
291
292static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
293{
294 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
295 tp->ecn_flags &= ~TCP_ECN_OK;
296}
297
298static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
299{
300 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
301 return true;
302 return false;
303}
304
305/* Buffer size and advertised window tuning.
306 *
307 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
308 */
309
310static void tcp_sndbuf_expand(struct sock *sk)
311{
312 const struct tcp_sock *tp = tcp_sk(sk);
313 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
314 int sndmem, per_mss;
315 u32 nr_segs;
316
317 /* Worst case is non GSO/TSO : each frame consumes one skb
318 * and skb->head is kmalloced using power of two area of memory
319 */
320 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
321 MAX_TCP_HEADER +
322 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
323
324 per_mss = roundup_pow_of_two(per_mss) +
325 SKB_DATA_ALIGN(sizeof(struct sk_buff));
326
327 nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
328 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
329
330 /* Fast Recovery (RFC 5681 3.2) :
331 * Cubic needs 1.7 factor, rounded to 2 to include
332 * extra cushion (application might react slowly to POLLOUT)
333 */
334 sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2;
335 sndmem *= nr_segs * per_mss;
336
337 if (sk->sk_sndbuf < sndmem)
338 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
339}
340
341/* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
342 *
343 * All tcp_full_space() is split to two parts: "network" buffer, allocated
344 * forward and advertised in receiver window (tp->rcv_wnd) and
345 * "application buffer", required to isolate scheduling/application
346 * latencies from network.
347 * window_clamp is maximal advertised window. It can be less than
348 * tcp_full_space(), in this case tcp_full_space() - window_clamp
349 * is reserved for "application" buffer. The less window_clamp is
350 * the smoother our behaviour from viewpoint of network, but the lower
351 * throughput and the higher sensitivity of the connection to losses. 8)
352 *
353 * rcv_ssthresh is more strict window_clamp used at "slow start"
354 * phase to predict further behaviour of this connection.
355 * It is used for two goals:
356 * - to enforce header prediction at sender, even when application
357 * requires some significant "application buffer". It is check #1.
358 * - to prevent pruning of receive queue because of misprediction
359 * of receiver window. Check #2.
360 *
361 * The scheme does not work when sender sends good segments opening
362 * window and then starts to feed us spaghetti. But it should work
363 * in common situations. Otherwise, we have to rely on queue collapsing.
364 */
365
366/* Slow part of check#2. */
367static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
368{
369 struct tcp_sock *tp = tcp_sk(sk);
370 /* Optimize this! */
371 int truesize = tcp_win_from_space(skb->truesize) >> 1;
372 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
373
374 while (tp->rcv_ssthresh <= window) {
375 if (truesize <= skb->len)
376 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
377
378 truesize >>= 1;
379 window >>= 1;
380 }
381 return 0;
382}
383
384static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
385{
386 struct tcp_sock *tp = tcp_sk(sk);
387
388 /* Check #1 */
389 if (tp->rcv_ssthresh < tp->window_clamp &&
390 (int)tp->rcv_ssthresh < tcp_space(sk) &&
391 !tcp_under_memory_pressure(sk)) {
392 int incr;
393
394 /* Check #2. Increase window, if skb with such overhead
395 * will fit to rcvbuf in future.
396 */
397 if (tcp_win_from_space(skb->truesize) <= skb->len)
398 incr = 2 * tp->advmss;
399 else
400 incr = __tcp_grow_window(sk, skb);
401
402 if (incr) {
403 incr = max_t(int, incr, 2 * skb->len);
404 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
405 tp->window_clamp);
406 inet_csk(sk)->icsk_ack.quick |= 1;
407 }
408 }
409}
410
411/* 3. Tuning rcvbuf, when connection enters established state. */
412static void tcp_fixup_rcvbuf(struct sock *sk)
413{
414 u32 mss = tcp_sk(sk)->advmss;
415 int rcvmem;
416
417 rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) *
418 tcp_default_init_rwnd(mss);
419
420 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
421 * Allow enough cushion so that sender is not limited by our window
422 */
423 if (sysctl_tcp_moderate_rcvbuf)
424 rcvmem <<= 2;
425
426 if (sk->sk_rcvbuf < rcvmem)
427 sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
428}
429
430/* 4. Try to fixup all. It is made immediately after connection enters
431 * established state.
432 */
433void tcp_init_buffer_space(struct sock *sk)
434{
435 struct tcp_sock *tp = tcp_sk(sk);
436 int maxwin;
437
438 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
439 tcp_fixup_rcvbuf(sk);
440 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
441 tcp_sndbuf_expand(sk);
442
443 tp->rcvq_space.space = tp->rcv_wnd;
444 tp->rcvq_space.time = tcp_time_stamp;
445 tp->rcvq_space.seq = tp->copied_seq;
446
447 maxwin = tcp_full_space(sk);
448
449 if (tp->window_clamp >= maxwin) {
450 tp->window_clamp = maxwin;
451
452 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
453 tp->window_clamp = max(maxwin -
454 (maxwin >> sysctl_tcp_app_win),
455 4 * tp->advmss);
456 }
457
458 /* Force reservation of one segment. */
459 if (sysctl_tcp_app_win &&
460 tp->window_clamp > 2 * tp->advmss &&
461 tp->window_clamp + tp->advmss > maxwin)
462 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
463
464 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
465 tp->snd_cwnd_stamp = tcp_time_stamp;
466}
467
468/* 5. Recalculate window clamp after socket hit its memory bounds. */
469static void tcp_clamp_window(struct sock *sk)
470{
471 struct tcp_sock *tp = tcp_sk(sk);
472 struct inet_connection_sock *icsk = inet_csk(sk);
473
474 icsk->icsk_ack.quick = 0;
475
476 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
477 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
478 !tcp_under_memory_pressure(sk) &&
479 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
480 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
481 sysctl_tcp_rmem[2]);
482 }
483 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
484 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
485}
486
487/* Initialize RCV_MSS value.
488 * RCV_MSS is an our guess about MSS used by the peer.
489 * We haven't any direct information about the MSS.
490 * It's better to underestimate the RCV_MSS rather than overestimate.
491 * Overestimations make us ACKing less frequently than needed.
492 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
493 */
494void tcp_initialize_rcv_mss(struct sock *sk)
495{
496 const struct tcp_sock *tp = tcp_sk(sk);
497 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
498
499 hint = min(hint, tp->rcv_wnd / 2);
500 hint = min(hint, TCP_MSS_DEFAULT);
501 hint = max(hint, TCP_MIN_MSS);
502
503 inet_csk(sk)->icsk_ack.rcv_mss = hint;
504}
505EXPORT_SYMBOL(tcp_initialize_rcv_mss);
506
507/* Receiver "autotuning" code.
508 *
509 * The algorithm for RTT estimation w/o timestamps is based on
510 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
511 * <http://public.lanl.gov/radiant/pubs.html#DRS>
512 *
513 * More detail on this code can be found at
514 * <http://staff.psc.edu/jheffner/>,
515 * though this reference is out of date. A new paper
516 * is pending.
517 */
518static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
519{
520 u32 new_sample = tp->rcv_rtt_est.rtt;
521 long m = sample;
522
523 if (m == 0)
524 m = 1;
525
526 if (new_sample != 0) {
527 /* If we sample in larger samples in the non-timestamp
528 * case, we could grossly overestimate the RTT especially
529 * with chatty applications or bulk transfer apps which
530 * are stalled on filesystem I/O.
531 *
532 * Also, since we are only going for a minimum in the
533 * non-timestamp case, we do not smooth things out
534 * else with timestamps disabled convergence takes too
535 * long.
536 */
537 if (!win_dep) {
538 m -= (new_sample >> 3);
539 new_sample += m;
540 } else {
541 m <<= 3;
542 if (m < new_sample)
543 new_sample = m;
544 }
545 } else {
546 /* No previous measure. */
547 new_sample = m << 3;
548 }
549
550 if (tp->rcv_rtt_est.rtt != new_sample)
551 tp->rcv_rtt_est.rtt = new_sample;
552}
553
554static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
555{
556 if (tp->rcv_rtt_est.time == 0)
557 goto new_measure;
558 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
559 return;
560 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1);
561
562new_measure:
563 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
564 tp->rcv_rtt_est.time = tcp_time_stamp;
565}
566
567static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
568 const struct sk_buff *skb)
569{
570 struct tcp_sock *tp = tcp_sk(sk);
571 if (tp->rx_opt.rcv_tsecr &&
572 (TCP_SKB_CB(skb)->end_seq -
573 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
574 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
575}
576
577/*
578 * This function should be called every time data is copied to user space.
579 * It calculates the appropriate TCP receive buffer space.
580 */
581void tcp_rcv_space_adjust(struct sock *sk)
582{
583 struct tcp_sock *tp = tcp_sk(sk);
584 int time;
585 int copied;
586
587 time = tcp_time_stamp - tp->rcvq_space.time;
588 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
589 return;
590
591 /* Number of bytes copied to user in last RTT */
592 copied = tp->copied_seq - tp->rcvq_space.seq;
593 if (copied <= tp->rcvq_space.space)
594 goto new_measure;
595
596 /* A bit of theory :
597 * copied = bytes received in previous RTT, our base window
598 * To cope with packet losses, we need a 2x factor
599 * To cope with slow start, and sender growing its cwin by 100 %
600 * every RTT, we need a 4x factor, because the ACK we are sending
601 * now is for the next RTT, not the current one :
602 * <prev RTT . ><current RTT .. ><next RTT .... >
603 */
604
605 if (sysctl_tcp_moderate_rcvbuf &&
606 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
607 int rcvwin, rcvmem, rcvbuf;
608
609 /* minimal window to cope with packet losses, assuming
610 * steady state. Add some cushion because of small variations.
611 */
612 rcvwin = (copied << 1) + 16 * tp->advmss;
613
614 /* If rate increased by 25%,
615 * assume slow start, rcvwin = 3 * copied
616 * If rate increased by 50%,
617 * assume sender can use 2x growth, rcvwin = 4 * copied
618 */
619 if (copied >=
620 tp->rcvq_space.space + (tp->rcvq_space.space >> 2)) {
621 if (copied >=
622 tp->rcvq_space.space + (tp->rcvq_space.space >> 1))
623 rcvwin <<= 1;
624 else
625 rcvwin += (rcvwin >> 1);
626 }
627
628 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
629 while (tcp_win_from_space(rcvmem) < tp->advmss)
630 rcvmem += 128;
631
632 rcvbuf = min(rcvwin / tp->advmss * rcvmem, sysctl_tcp_rmem[2]);
633 if (rcvbuf > sk->sk_rcvbuf) {
634 sk->sk_rcvbuf = rcvbuf;
635
636 /* Make the window clamp follow along. */
637 tp->window_clamp = rcvwin;
638 }
639 }
640 tp->rcvq_space.space = copied;
641
642new_measure:
643 tp->rcvq_space.seq = tp->copied_seq;
644 tp->rcvq_space.time = tcp_time_stamp;
645}
646
647/* There is something which you must keep in mind when you analyze the
648 * behavior of the tp->ato delayed ack timeout interval. When a
649 * connection starts up, we want to ack as quickly as possible. The
650 * problem is that "good" TCP's do slow start at the beginning of data
651 * transmission. The means that until we send the first few ACK's the
652 * sender will sit on his end and only queue most of his data, because
653 * he can only send snd_cwnd unacked packets at any given time. For
654 * each ACK we send, he increments snd_cwnd and transmits more of his
655 * queue. -DaveM
656 */
657static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
658{
659 struct tcp_sock *tp = tcp_sk(sk);
660 struct inet_connection_sock *icsk = inet_csk(sk);
661 u32 now;
662
663 inet_csk_schedule_ack(sk);
664
665 tcp_measure_rcv_mss(sk, skb);
666
667 tcp_rcv_rtt_measure(tp);
668
669 now = tcp_time_stamp;
670
671 if (!icsk->icsk_ack.ato) {
672 /* The _first_ data packet received, initialize
673 * delayed ACK engine.
674 */
675 tcp_incr_quickack(sk);
676 icsk->icsk_ack.ato = TCP_ATO_MIN;
677 } else {
678 int m = now - icsk->icsk_ack.lrcvtime;
679
680 if (m <= TCP_ATO_MIN / 2) {
681 /* The fastest case is the first. */
682 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
683 } else if (m < icsk->icsk_ack.ato) {
684 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
685 if (icsk->icsk_ack.ato > icsk->icsk_rto)
686 icsk->icsk_ack.ato = icsk->icsk_rto;
687 } else if (m > icsk->icsk_rto) {
688 /* Too long gap. Apparently sender failed to
689 * restart window, so that we send ACKs quickly.
690 */
691 tcp_incr_quickack(sk);
692 sk_mem_reclaim(sk);
693 }
694 }
695 icsk->icsk_ack.lrcvtime = now;
696
697 tcp_ecn_check_ce(tp, skb);
698
699 if (skb->len >= 128)
700 tcp_grow_window(sk, skb);
701}
702
703/* Called to compute a smoothed rtt estimate. The data fed to this
704 * routine either comes from timestamps, or from segments that were
705 * known _not_ to have been retransmitted [see Karn/Partridge
706 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
707 * piece by Van Jacobson.
708 * NOTE: the next three routines used to be one big routine.
709 * To save cycles in the RFC 1323 implementation it was better to break
710 * it up into three procedures. -- erics
711 */
712static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
713{
714 struct tcp_sock *tp = tcp_sk(sk);
715 long m = mrtt_us; /* RTT */
716 u32 srtt = tp->srtt_us;
717
718 /* The following amusing code comes from Jacobson's
719 * article in SIGCOMM '88. Note that rtt and mdev
720 * are scaled versions of rtt and mean deviation.
721 * This is designed to be as fast as possible
722 * m stands for "measurement".
723 *
724 * On a 1990 paper the rto value is changed to:
725 * RTO = rtt + 4 * mdev
726 *
727 * Funny. This algorithm seems to be very broken.
728 * These formulae increase RTO, when it should be decreased, increase
729 * too slowly, when it should be increased quickly, decrease too quickly
730 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
731 * does not matter how to _calculate_ it. Seems, it was trap
732 * that VJ failed to avoid. 8)
733 */
734 if (srtt != 0) {
735 m -= (srtt >> 3); /* m is now error in rtt est */
736 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
737 if (m < 0) {
738 m = -m; /* m is now abs(error) */
739 m -= (tp->mdev_us >> 2); /* similar update on mdev */
740 /* This is similar to one of Eifel findings.
741 * Eifel blocks mdev updates when rtt decreases.
742 * This solution is a bit different: we use finer gain
743 * for mdev in this case (alpha*beta).
744 * Like Eifel it also prevents growth of rto,
745 * but also it limits too fast rto decreases,
746 * happening in pure Eifel.
747 */
748 if (m > 0)
749 m >>= 3;
750 } else {
751 m -= (tp->mdev_us >> 2); /* similar update on mdev */
752 }
753 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
754 if (tp->mdev_us > tp->mdev_max_us) {
755 tp->mdev_max_us = tp->mdev_us;
756 if (tp->mdev_max_us > tp->rttvar_us)
757 tp->rttvar_us = tp->mdev_max_us;
758 }
759 if (after(tp->snd_una, tp->rtt_seq)) {
760 if (tp->mdev_max_us < tp->rttvar_us)
761 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
762 tp->rtt_seq = tp->snd_nxt;
763 tp->mdev_max_us = tcp_rto_min_us(sk);
764 }
765 } else {
766 /* no previous measure. */
767 srtt = m << 3; /* take the measured time to be rtt */
768 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
769 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
770 tp->mdev_max_us = tp->rttvar_us;
771 tp->rtt_seq = tp->snd_nxt;
772 }
773 tp->srtt_us = max(1U, srtt);
774}
775
776/* Set the sk_pacing_rate to allow proper sizing of TSO packets.
777 * Note: TCP stack does not yet implement pacing.
778 * FQ packet scheduler can be used to implement cheap but effective
779 * TCP pacing, to smooth the burst on large writes when packets
780 * in flight is significantly lower than cwnd (or rwin)
781 */
782int sysctl_tcp_pacing_ss_ratio __read_mostly = 200;
783int sysctl_tcp_pacing_ca_ratio __read_mostly = 120;
784
785static void tcp_update_pacing_rate(struct sock *sk)
786{
787 const struct tcp_sock *tp = tcp_sk(sk);
788 u64 rate;
789
790 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
791 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
792
793 /* current rate is (cwnd * mss) / srtt
794 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
795 * In Congestion Avoidance phase, set it to 120 % the current rate.
796 *
797 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
798 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
799 * end of slow start and should slow down.
800 */
801 if (tp->snd_cwnd < tp->snd_ssthresh / 2)
802 rate *= sysctl_tcp_pacing_ss_ratio;
803 else
804 rate *= sysctl_tcp_pacing_ca_ratio;
805
806 rate *= max(tp->snd_cwnd, tp->packets_out);
807
808 if (likely(tp->srtt_us))
809 do_div(rate, tp->srtt_us);
810
811 /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate
812 * without any lock. We want to make sure compiler wont store
813 * intermediate values in this location.
814 */
815 ACCESS_ONCE(sk->sk_pacing_rate) = min_t(u64, rate,
816 sk->sk_max_pacing_rate);
817}
818
819/* Calculate rto without backoff. This is the second half of Van Jacobson's
820 * routine referred to above.
821 */
822static void tcp_set_rto(struct sock *sk)
823{
824 const struct tcp_sock *tp = tcp_sk(sk);
825 /* Old crap is replaced with new one. 8)
826 *
827 * More seriously:
828 * 1. If rtt variance happened to be less 50msec, it is hallucination.
829 * It cannot be less due to utterly erratic ACK generation made
830 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
831 * to do with delayed acks, because at cwnd>2 true delack timeout
832 * is invisible. Actually, Linux-2.4 also generates erratic
833 * ACKs in some circumstances.
834 */
835 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
836
837 /* 2. Fixups made earlier cannot be right.
838 * If we do not estimate RTO correctly without them,
839 * all the algo is pure shit and should be replaced
840 * with correct one. It is exactly, which we pretend to do.
841 */
842
843 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
844 * guarantees that rto is higher.
845 */
846 tcp_bound_rto(sk);
847}
848
849__u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
850{
851 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
852
853 if (!cwnd)
854 cwnd = TCP_INIT_CWND;
855 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
856}
857
858/*
859 * Packet counting of FACK is based on in-order assumptions, therefore TCP
860 * disables it when reordering is detected
861 */
862void tcp_disable_fack(struct tcp_sock *tp)
863{
864 /* RFC3517 uses different metric in lost marker => reset on change */
865 if (tcp_is_fack(tp))
866 tp->lost_skb_hint = NULL;
867 tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
868}
869
870/* Take a notice that peer is sending D-SACKs */
871static void tcp_dsack_seen(struct tcp_sock *tp)
872{
873 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
874}
875
876static void tcp_update_reordering(struct sock *sk, const int metric,
877 const int ts)
878{
879 struct tcp_sock *tp = tcp_sk(sk);
880 if (metric > tp->reordering) {
881 int mib_idx;
882
883 tp->reordering = min(sysctl_tcp_max_reordering, metric);
884
885 /* This exciting event is worth to be remembered. 8) */
886 if (ts)
887 mib_idx = LINUX_MIB_TCPTSREORDER;
888 else if (tcp_is_reno(tp))
889 mib_idx = LINUX_MIB_TCPRENOREORDER;
890 else if (tcp_is_fack(tp))
891 mib_idx = LINUX_MIB_TCPFACKREORDER;
892 else
893 mib_idx = LINUX_MIB_TCPSACKREORDER;
894
895 NET_INC_STATS(sock_net(sk), mib_idx);
896#if FASTRETRANS_DEBUG > 1
897 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
898 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
899 tp->reordering,
900 tp->fackets_out,
901 tp->sacked_out,
902 tp->undo_marker ? tp->undo_retrans : 0);
903#endif
904 tcp_disable_fack(tp);
905 }
906
907 if (metric > 0)
908 tcp_disable_early_retrans(tp);
909 tp->rack.reord = 1;
910}
911
912/* This must be called before lost_out is incremented */
913static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
914{
915 if (!tp->retransmit_skb_hint ||
916 before(TCP_SKB_CB(skb)->seq,
917 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
918 tp->retransmit_skb_hint = skb;
919
920 if (!tp->lost_out ||
921 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
922 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
923}
924
925/* Sum the number of packets on the wire we have marked as lost.
926 * There are two cases we care about here:
927 * a) Packet hasn't been marked lost (nor retransmitted),
928 * and this is the first loss.
929 * b) Packet has been marked both lost and retransmitted,
930 * and this means we think it was lost again.
931 */
932static void tcp_sum_lost(struct tcp_sock *tp, struct sk_buff *skb)
933{
934 __u8 sacked = TCP_SKB_CB(skb)->sacked;
935
936 if (!(sacked & TCPCB_LOST) ||
937 ((sacked & TCPCB_LOST) && (sacked & TCPCB_SACKED_RETRANS)))
938 tp->lost += tcp_skb_pcount(skb);
939}
940
941static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
942{
943 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
944 tcp_verify_retransmit_hint(tp, skb);
945
946 tp->lost_out += tcp_skb_pcount(skb);
947 tcp_sum_lost(tp, skb);
948 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
949 }
950}
951
952void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb)
953{
954 tcp_verify_retransmit_hint(tp, skb);
955
956 tcp_sum_lost(tp, skb);
957 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
958 tp->lost_out += tcp_skb_pcount(skb);
959 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
960 }
961}
962
963/* This procedure tags the retransmission queue when SACKs arrive.
964 *
965 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
966 * Packets in queue with these bits set are counted in variables
967 * sacked_out, retrans_out and lost_out, correspondingly.
968 *
969 * Valid combinations are:
970 * Tag InFlight Description
971 * 0 1 - orig segment is in flight.
972 * S 0 - nothing flies, orig reached receiver.
973 * L 0 - nothing flies, orig lost by net.
974 * R 2 - both orig and retransmit are in flight.
975 * L|R 1 - orig is lost, retransmit is in flight.
976 * S|R 1 - orig reached receiver, retrans is still in flight.
977 * (L|S|R is logically valid, it could occur when L|R is sacked,
978 * but it is equivalent to plain S and code short-curcuits it to S.
979 * L|S is logically invalid, it would mean -1 packet in flight 8))
980 *
981 * These 6 states form finite state machine, controlled by the following events:
982 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
983 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
984 * 3. Loss detection event of two flavors:
985 * A. Scoreboard estimator decided the packet is lost.
986 * A'. Reno "three dupacks" marks head of queue lost.
987 * A''. Its FACK modification, head until snd.fack is lost.
988 * B. SACK arrives sacking SND.NXT at the moment, when the
989 * segment was retransmitted.
990 * 4. D-SACK added new rule: D-SACK changes any tag to S.
991 *
992 * It is pleasant to note, that state diagram turns out to be commutative,
993 * so that we are allowed not to be bothered by order of our actions,
994 * when multiple events arrive simultaneously. (see the function below).
995 *
996 * Reordering detection.
997 * --------------------
998 * Reordering metric is maximal distance, which a packet can be displaced
999 * in packet stream. With SACKs we can estimate it:
1000 *
1001 * 1. SACK fills old hole and the corresponding segment was not
1002 * ever retransmitted -> reordering. Alas, we cannot use it
1003 * when segment was retransmitted.
1004 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1005 * for retransmitted and already SACKed segment -> reordering..
1006 * Both of these heuristics are not used in Loss state, when we cannot
1007 * account for retransmits accurately.
1008 *
1009 * SACK block validation.
1010 * ----------------------
1011 *
1012 * SACK block range validation checks that the received SACK block fits to
1013 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1014 * Note that SND.UNA is not included to the range though being valid because
1015 * it means that the receiver is rather inconsistent with itself reporting
1016 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1017 * perfectly valid, however, in light of RFC2018 which explicitly states
1018 * that "SACK block MUST reflect the newest segment. Even if the newest
1019 * segment is going to be discarded ...", not that it looks very clever
1020 * in case of head skb. Due to potentional receiver driven attacks, we
1021 * choose to avoid immediate execution of a walk in write queue due to
1022 * reneging and defer head skb's loss recovery to standard loss recovery
1023 * procedure that will eventually trigger (nothing forbids us doing this).
1024 *
1025 * Implements also blockage to start_seq wrap-around. Problem lies in the
1026 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1027 * there's no guarantee that it will be before snd_nxt (n). The problem
1028 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1029 * wrap (s_w):
1030 *
1031 * <- outs wnd -> <- wrapzone ->
1032 * u e n u_w e_w s n_w
1033 * | | | | | | |
1034 * |<------------+------+----- TCP seqno space --------------+---------->|
1035 * ...-- <2^31 ->| |<--------...
1036 * ...---- >2^31 ------>| |<--------...
1037 *
1038 * Current code wouldn't be vulnerable but it's better still to discard such
1039 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1040 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1041 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1042 * equal to the ideal case (infinite seqno space without wrap caused issues).
1043 *
1044 * With D-SACK the lower bound is extended to cover sequence space below
1045 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1046 * again, D-SACK block must not to go across snd_una (for the same reason as
1047 * for the normal SACK blocks, explained above). But there all simplicity
1048 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1049 * fully below undo_marker they do not affect behavior in anyway and can
1050 * therefore be safely ignored. In rare cases (which are more or less
1051 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1052 * fragmentation and packet reordering past skb's retransmission. To consider
1053 * them correctly, the acceptable range must be extended even more though
1054 * the exact amount is rather hard to quantify. However, tp->max_window can
1055 * be used as an exaggerated estimate.
1056 */
1057static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1058 u32 start_seq, u32 end_seq)
1059{
1060 /* Too far in future, or reversed (interpretation is ambiguous) */
1061 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1062 return false;
1063
1064 /* Nasty start_seq wrap-around check (see comments above) */
1065 if (!before(start_seq, tp->snd_nxt))
1066 return false;
1067
1068 /* In outstanding window? ...This is valid exit for D-SACKs too.
1069 * start_seq == snd_una is non-sensical (see comments above)
1070 */
1071 if (after(start_seq, tp->snd_una))
1072 return true;
1073
1074 if (!is_dsack || !tp->undo_marker)
1075 return false;
1076
1077 /* ...Then it's D-SACK, and must reside below snd_una completely */
1078 if (after(end_seq, tp->snd_una))
1079 return false;
1080
1081 if (!before(start_seq, tp->undo_marker))
1082 return true;
1083
1084 /* Too old */
1085 if (!after(end_seq, tp->undo_marker))
1086 return false;
1087
1088 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1089 * start_seq < undo_marker and end_seq >= undo_marker.
1090 */
1091 return !before(start_seq, end_seq - tp->max_window);
1092}
1093
1094static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1095 struct tcp_sack_block_wire *sp, int num_sacks,
1096 u32 prior_snd_una)
1097{
1098 struct tcp_sock *tp = tcp_sk(sk);
1099 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1100 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1101 bool dup_sack = false;
1102
1103 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1104 dup_sack = true;
1105 tcp_dsack_seen(tp);
1106 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1107 } else if (num_sacks > 1) {
1108 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1109 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1110
1111 if (!after(end_seq_0, end_seq_1) &&
1112 !before(start_seq_0, start_seq_1)) {
1113 dup_sack = true;
1114 tcp_dsack_seen(tp);
1115 NET_INC_STATS(sock_net(sk),
1116 LINUX_MIB_TCPDSACKOFORECV);
1117 }
1118 }
1119
1120 /* D-SACK for already forgotten data... Do dumb counting. */
1121 if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
1122 !after(end_seq_0, prior_snd_una) &&
1123 after(end_seq_0, tp->undo_marker))
1124 tp->undo_retrans--;
1125
1126 return dup_sack;
1127}
1128
1129struct tcp_sacktag_state {
1130 int reord;
1131 int fack_count;
1132 /* Timestamps for earliest and latest never-retransmitted segment
1133 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1134 * but congestion control should still get an accurate delay signal.
1135 */
1136 struct skb_mstamp first_sackt;
1137 struct skb_mstamp last_sackt;
1138 struct rate_sample *rate;
1139 int flag;
1140};
1141
1142/* Check if skb is fully within the SACK block. In presence of GSO skbs,
1143 * the incoming SACK may not exactly match but we can find smaller MSS
1144 * aligned portion of it that matches. Therefore we might need to fragment
1145 * which may fail and creates some hassle (caller must handle error case
1146 * returns).
1147 *
1148 * FIXME: this could be merged to shift decision code
1149 */
1150static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1151 u32 start_seq, u32 end_seq)
1152{
1153 int err;
1154 bool in_sack;
1155 unsigned int pkt_len;
1156 unsigned int mss;
1157
1158 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1159 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1160
1161 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1162 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1163 mss = tcp_skb_mss(skb);
1164 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1165
1166 if (!in_sack) {
1167 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1168 if (pkt_len < mss)
1169 pkt_len = mss;
1170 } else {
1171 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1172 if (pkt_len < mss)
1173 return -EINVAL;
1174 }
1175
1176 /* Round if necessary so that SACKs cover only full MSSes
1177 * and/or the remaining small portion (if present)
1178 */
1179 if (pkt_len > mss) {
1180 unsigned int new_len = (pkt_len / mss) * mss;
1181 if (!in_sack && new_len < pkt_len) {
1182 new_len += mss;
1183 if (new_len >= skb->len)
1184 return 0;
1185 }
1186 pkt_len = new_len;
1187 }
1188 err = tcp_fragment(sk, skb, pkt_len, mss, GFP_ATOMIC);
1189 if (err < 0)
1190 return err;
1191 }
1192
1193 return in_sack;
1194}
1195
1196/* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1197static u8 tcp_sacktag_one(struct sock *sk,
1198 struct tcp_sacktag_state *state, u8 sacked,
1199 u32 start_seq, u32 end_seq,
1200 int dup_sack, int pcount,
1201 const struct skb_mstamp *xmit_time)
1202{
1203 struct tcp_sock *tp = tcp_sk(sk);
1204 int fack_count = state->fack_count;
1205
1206 /* Account D-SACK for retransmitted packet. */
1207 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1208 if (tp->undo_marker && tp->undo_retrans > 0 &&
1209 after(end_seq, tp->undo_marker))
1210 tp->undo_retrans--;
1211 if (sacked & TCPCB_SACKED_ACKED)
1212 state->reord = min(fack_count, state->reord);
1213 }
1214
1215 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1216 if (!after(end_seq, tp->snd_una))
1217 return sacked;
1218
1219 if (!(sacked & TCPCB_SACKED_ACKED)) {
1220 tcp_rack_advance(tp, xmit_time, sacked);
1221
1222 if (sacked & TCPCB_SACKED_RETRANS) {
1223 /* If the segment is not tagged as lost,
1224 * we do not clear RETRANS, believing
1225 * that retransmission is still in flight.
1226 */
1227 if (sacked & TCPCB_LOST) {
1228 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1229 tp->lost_out -= pcount;
1230 tp->retrans_out -= pcount;
1231 }
1232 } else {
1233 if (!(sacked & TCPCB_RETRANS)) {
1234 /* New sack for not retransmitted frame,
1235 * which was in hole. It is reordering.
1236 */
1237 if (before(start_seq,
1238 tcp_highest_sack_seq(tp)))
1239 state->reord = min(fack_count,
1240 state->reord);
1241 if (!after(end_seq, tp->high_seq))
1242 state->flag |= FLAG_ORIG_SACK_ACKED;
1243 if (state->first_sackt.v64 == 0)
1244 state->first_sackt = *xmit_time;
1245 state->last_sackt = *xmit_time;
1246 }
1247
1248 if (sacked & TCPCB_LOST) {
1249 sacked &= ~TCPCB_LOST;
1250 tp->lost_out -= pcount;
1251 }
1252 }
1253
1254 sacked |= TCPCB_SACKED_ACKED;
1255 state->flag |= FLAG_DATA_SACKED;
1256 tp->sacked_out += pcount;
1257 tp->delivered += pcount; /* Out-of-order packets delivered */
1258
1259 fack_count += pcount;
1260
1261 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1262 if (!tcp_is_fack(tp) && tp->lost_skb_hint &&
1263 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1264 tp->lost_cnt_hint += pcount;
1265
1266 if (fack_count > tp->fackets_out)
1267 tp->fackets_out = fack_count;
1268 }
1269
1270 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1271 * frames and clear it. undo_retrans is decreased above, L|R frames
1272 * are accounted above as well.
1273 */
1274 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1275 sacked &= ~TCPCB_SACKED_RETRANS;
1276 tp->retrans_out -= pcount;
1277 }
1278
1279 return sacked;
1280}
1281
1282/* Shift newly-SACKed bytes from this skb to the immediately previous
1283 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1284 */
1285static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1286 struct tcp_sacktag_state *state,
1287 unsigned int pcount, int shifted, int mss,
1288 bool dup_sack)
1289{
1290 struct tcp_sock *tp = tcp_sk(sk);
1291 struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1292 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1293 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1294
1295 BUG_ON(!pcount);
1296
1297 /* Adjust counters and hints for the newly sacked sequence
1298 * range but discard the return value since prev is already
1299 * marked. We must tag the range first because the seq
1300 * advancement below implicitly advances
1301 * tcp_highest_sack_seq() when skb is highest_sack.
1302 */
1303 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1304 start_seq, end_seq, dup_sack, pcount,
1305 &skb->skb_mstamp);
1306 tcp_rate_skb_delivered(sk, skb, state->rate);
1307
1308 if (skb == tp->lost_skb_hint)
1309 tp->lost_cnt_hint += pcount;
1310
1311 TCP_SKB_CB(prev)->end_seq += shifted;
1312 TCP_SKB_CB(skb)->seq += shifted;
1313
1314 tcp_skb_pcount_add(prev, pcount);
1315 BUG_ON(tcp_skb_pcount(skb) < pcount);
1316 tcp_skb_pcount_add(skb, -pcount);
1317
1318 /* When we're adding to gso_segs == 1, gso_size will be zero,
1319 * in theory this shouldn't be necessary but as long as DSACK
1320 * code can come after this skb later on it's better to keep
1321 * setting gso_size to something.
1322 */
1323 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1324 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1325
1326 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1327 if (tcp_skb_pcount(skb) <= 1)
1328 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1329
1330 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1331 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1332
1333 if (skb->len > 0) {
1334 BUG_ON(!tcp_skb_pcount(skb));
1335 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1336 return false;
1337 }
1338
1339 /* Whole SKB was eaten :-) */
1340
1341 if (skb == tp->retransmit_skb_hint)
1342 tp->retransmit_skb_hint = prev;
1343 if (skb == tp->lost_skb_hint) {
1344 tp->lost_skb_hint = prev;
1345 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1346 }
1347
1348 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1349 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
1350 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1351 TCP_SKB_CB(prev)->end_seq++;
1352
1353 if (skb == tcp_highest_sack(sk))
1354 tcp_advance_highest_sack(sk, skb);
1355
1356 tcp_skb_collapse_tstamp(prev, skb);
1357 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp.v64))
1358 TCP_SKB_CB(prev)->tx.delivered_mstamp.v64 = 0;
1359
1360 tcp_unlink_write_queue(skb, sk);
1361 sk_wmem_free_skb(sk, skb);
1362
1363 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
1364
1365 return true;
1366}
1367
1368/* I wish gso_size would have a bit more sane initialization than
1369 * something-or-zero which complicates things
1370 */
1371static int tcp_skb_seglen(const struct sk_buff *skb)
1372{
1373 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1374}
1375
1376/* Shifting pages past head area doesn't work */
1377static int skb_can_shift(const struct sk_buff *skb)
1378{
1379 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1380}
1381
1382/* Try collapsing SACK blocks spanning across multiple skbs to a single
1383 * skb.
1384 */
1385static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1386 struct tcp_sacktag_state *state,
1387 u32 start_seq, u32 end_seq,
1388 bool dup_sack)
1389{
1390 struct tcp_sock *tp = tcp_sk(sk);
1391 struct sk_buff *prev;
1392 int mss;
1393 int pcount = 0;
1394 int len;
1395 int in_sack;
1396
1397 if (!sk_can_gso(sk))
1398 goto fallback;
1399
1400 /* Normally R but no L won't result in plain S */
1401 if (!dup_sack &&
1402 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1403 goto fallback;
1404 if (!skb_can_shift(skb))
1405 goto fallback;
1406 /* This frame is about to be dropped (was ACKed). */
1407 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1408 goto fallback;
1409
1410 /* Can only happen with delayed DSACK + discard craziness */
1411 if (unlikely(skb == tcp_write_queue_head(sk)))
1412 goto fallback;
1413 prev = tcp_write_queue_prev(sk, skb);
1414
1415 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1416 goto fallback;
1417
1418 if (!tcp_skb_can_collapse_to(prev))
1419 goto fallback;
1420
1421 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1422 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1423
1424 if (in_sack) {
1425 len = skb->len;
1426 pcount = tcp_skb_pcount(skb);
1427 mss = tcp_skb_seglen(skb);
1428
1429 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1430 * drop this restriction as unnecessary
1431 */
1432 if (mss != tcp_skb_seglen(prev))
1433 goto fallback;
1434 } else {
1435 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1436 goto noop;
1437 /* CHECKME: This is non-MSS split case only?, this will
1438 * cause skipped skbs due to advancing loop btw, original
1439 * has that feature too
1440 */
1441 if (tcp_skb_pcount(skb) <= 1)
1442 goto noop;
1443
1444 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1445 if (!in_sack) {
1446 /* TODO: head merge to next could be attempted here
1447 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1448 * though it might not be worth of the additional hassle
1449 *
1450 * ...we can probably just fallback to what was done
1451 * previously. We could try merging non-SACKed ones
1452 * as well but it probably isn't going to buy off
1453 * because later SACKs might again split them, and
1454 * it would make skb timestamp tracking considerably
1455 * harder problem.
1456 */
1457 goto fallback;
1458 }
1459
1460 len = end_seq - TCP_SKB_CB(skb)->seq;
1461 BUG_ON(len < 0);
1462 BUG_ON(len > skb->len);
1463
1464 /* MSS boundaries should be honoured or else pcount will
1465 * severely break even though it makes things bit trickier.
1466 * Optimize common case to avoid most of the divides
1467 */
1468 mss = tcp_skb_mss(skb);
1469
1470 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1471 * drop this restriction as unnecessary
1472 */
1473 if (mss != tcp_skb_seglen(prev))
1474 goto fallback;
1475
1476 if (len == mss) {
1477 pcount = 1;
1478 } else if (len < mss) {
1479 goto noop;
1480 } else {
1481 pcount = len / mss;
1482 len = pcount * mss;
1483 }
1484 }
1485
1486 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1487 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1488 goto fallback;
1489
1490 if (!skb_shift(prev, skb, len))
1491 goto fallback;
1492 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1493 goto out;
1494
1495 /* Hole filled allows collapsing with the next as well, this is very
1496 * useful when hole on every nth skb pattern happens
1497 */
1498 if (prev == tcp_write_queue_tail(sk))
1499 goto out;
1500 skb = tcp_write_queue_next(sk, prev);
1501
1502 if (!skb_can_shift(skb) ||
1503 (skb == tcp_send_head(sk)) ||
1504 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1505 (mss != tcp_skb_seglen(skb)))
1506 goto out;
1507
1508 len = skb->len;
1509 if (skb_shift(prev, skb, len)) {
1510 pcount += tcp_skb_pcount(skb);
1511 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1512 }
1513
1514out:
1515 state->fack_count += pcount;
1516 return prev;
1517
1518noop:
1519 return skb;
1520
1521fallback:
1522 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1523 return NULL;
1524}
1525
1526static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1527 struct tcp_sack_block *next_dup,
1528 struct tcp_sacktag_state *state,
1529 u32 start_seq, u32 end_seq,
1530 bool dup_sack_in)
1531{
1532 struct tcp_sock *tp = tcp_sk(sk);
1533 struct sk_buff *tmp;
1534
1535 tcp_for_write_queue_from(skb, sk) {
1536 int in_sack = 0;
1537 bool dup_sack = dup_sack_in;
1538
1539 if (skb == tcp_send_head(sk))
1540 break;
1541
1542 /* queue is in-order => we can short-circuit the walk early */
1543 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1544 break;
1545
1546 if (next_dup &&
1547 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1548 in_sack = tcp_match_skb_to_sack(sk, skb,
1549 next_dup->start_seq,
1550 next_dup->end_seq);
1551 if (in_sack > 0)
1552 dup_sack = true;
1553 }
1554
1555 /* skb reference here is a bit tricky to get right, since
1556 * shifting can eat and free both this skb and the next,
1557 * so not even _safe variant of the loop is enough.
1558 */
1559 if (in_sack <= 0) {
1560 tmp = tcp_shift_skb_data(sk, skb, state,
1561 start_seq, end_seq, dup_sack);
1562 if (tmp) {
1563 if (tmp != skb) {
1564 skb = tmp;
1565 continue;
1566 }
1567
1568 in_sack = 0;
1569 } else {
1570 in_sack = tcp_match_skb_to_sack(sk, skb,
1571 start_seq,
1572 end_seq);
1573 }
1574 }
1575
1576 if (unlikely(in_sack < 0))
1577 break;
1578
1579 if (in_sack) {
1580 TCP_SKB_CB(skb)->sacked =
1581 tcp_sacktag_one(sk,
1582 state,
1583 TCP_SKB_CB(skb)->sacked,
1584 TCP_SKB_CB(skb)->seq,
1585 TCP_SKB_CB(skb)->end_seq,
1586 dup_sack,
1587 tcp_skb_pcount(skb),
1588 &skb->skb_mstamp);
1589 tcp_rate_skb_delivered(sk, skb, state->rate);
1590
1591 if (!before(TCP_SKB_CB(skb)->seq,
1592 tcp_highest_sack_seq(tp)))
1593 tcp_advance_highest_sack(sk, skb);
1594 }
1595
1596 state->fack_count += tcp_skb_pcount(skb);
1597 }
1598 return skb;
1599}
1600
1601/* Avoid all extra work that is being done by sacktag while walking in
1602 * a normal way
1603 */
1604static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1605 struct tcp_sacktag_state *state,
1606 u32 skip_to_seq)
1607{
1608 tcp_for_write_queue_from(skb, sk) {
1609 if (skb == tcp_send_head(sk))
1610 break;
1611
1612 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1613 break;
1614
1615 state->fack_count += tcp_skb_pcount(skb);
1616 }
1617 return skb;
1618}
1619
1620static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1621 struct sock *sk,
1622 struct tcp_sack_block *next_dup,
1623 struct tcp_sacktag_state *state,
1624 u32 skip_to_seq)
1625{
1626 if (!next_dup)
1627 return skb;
1628
1629 if (before(next_dup->start_seq, skip_to_seq)) {
1630 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1631 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1632 next_dup->start_seq, next_dup->end_seq,
1633 1);
1634 }
1635
1636 return skb;
1637}
1638
1639static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1640{
1641 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1642}
1643
1644static int
1645tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1646 u32 prior_snd_una, struct tcp_sacktag_state *state)
1647{
1648 struct tcp_sock *tp = tcp_sk(sk);
1649 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1650 TCP_SKB_CB(ack_skb)->sacked);
1651 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1652 struct tcp_sack_block sp[TCP_NUM_SACKS];
1653 struct tcp_sack_block *cache;
1654 struct sk_buff *skb;
1655 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1656 int used_sacks;
1657 bool found_dup_sack = false;
1658 int i, j;
1659 int first_sack_index;
1660
1661 state->flag = 0;
1662 state->reord = tp->packets_out;
1663
1664 if (!tp->sacked_out) {
1665 if (WARN_ON(tp->fackets_out))
1666 tp->fackets_out = 0;
1667 tcp_highest_sack_reset(sk);
1668 }
1669
1670 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1671 num_sacks, prior_snd_una);
1672 if (found_dup_sack) {
1673 state->flag |= FLAG_DSACKING_ACK;
1674 tp->delivered++; /* A spurious retransmission is delivered */
1675 }
1676
1677 /* Eliminate too old ACKs, but take into
1678 * account more or less fresh ones, they can
1679 * contain valid SACK info.
1680 */
1681 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1682 return 0;
1683
1684 if (!tp->packets_out)
1685 goto out;
1686
1687 used_sacks = 0;
1688 first_sack_index = 0;
1689 for (i = 0; i < num_sacks; i++) {
1690 bool dup_sack = !i && found_dup_sack;
1691
1692 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1693 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1694
1695 if (!tcp_is_sackblock_valid(tp, dup_sack,
1696 sp[used_sacks].start_seq,
1697 sp[used_sacks].end_seq)) {
1698 int mib_idx;
1699
1700 if (dup_sack) {
1701 if (!tp->undo_marker)
1702 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1703 else
1704 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1705 } else {
1706 /* Don't count olds caused by ACK reordering */
1707 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1708 !after(sp[used_sacks].end_seq, tp->snd_una))
1709 continue;
1710 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1711 }
1712
1713 NET_INC_STATS(sock_net(sk), mib_idx);
1714 if (i == 0)
1715 first_sack_index = -1;
1716 continue;
1717 }
1718
1719 /* Ignore very old stuff early */
1720 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1721 continue;
1722
1723 used_sacks++;
1724 }
1725
1726 /* order SACK blocks to allow in order walk of the retrans queue */
1727 for (i = used_sacks - 1; i > 0; i--) {
1728 for (j = 0; j < i; j++) {
1729 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1730 swap(sp[j], sp[j + 1]);
1731
1732 /* Track where the first SACK block goes to */
1733 if (j == first_sack_index)
1734 first_sack_index = j + 1;
1735 }
1736 }
1737 }
1738
1739 skb = tcp_write_queue_head(sk);
1740 state->fack_count = 0;
1741 i = 0;
1742
1743 if (!tp->sacked_out) {
1744 /* It's already past, so skip checking against it */
1745 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1746 } else {
1747 cache = tp->recv_sack_cache;
1748 /* Skip empty blocks in at head of the cache */
1749 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1750 !cache->end_seq)
1751 cache++;
1752 }
1753
1754 while (i < used_sacks) {
1755 u32 start_seq = sp[i].start_seq;
1756 u32 end_seq = sp[i].end_seq;
1757 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1758 struct tcp_sack_block *next_dup = NULL;
1759
1760 if (found_dup_sack && ((i + 1) == first_sack_index))
1761 next_dup = &sp[i + 1];
1762
1763 /* Skip too early cached blocks */
1764 while (tcp_sack_cache_ok(tp, cache) &&
1765 !before(start_seq, cache->end_seq))
1766 cache++;
1767
1768 /* Can skip some work by looking recv_sack_cache? */
1769 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1770 after(end_seq, cache->start_seq)) {
1771
1772 /* Head todo? */
1773 if (before(start_seq, cache->start_seq)) {
1774 skb = tcp_sacktag_skip(skb, sk, state,
1775 start_seq);
1776 skb = tcp_sacktag_walk(skb, sk, next_dup,
1777 state,
1778 start_seq,
1779 cache->start_seq,
1780 dup_sack);
1781 }
1782
1783 /* Rest of the block already fully processed? */
1784 if (!after(end_seq, cache->end_seq))
1785 goto advance_sp;
1786
1787 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1788 state,
1789 cache->end_seq);
1790
1791 /* ...tail remains todo... */
1792 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1793 /* ...but better entrypoint exists! */
1794 skb = tcp_highest_sack(sk);
1795 if (!skb)
1796 break;
1797 state->fack_count = tp->fackets_out;
1798 cache++;
1799 goto walk;
1800 }
1801
1802 skb = tcp_sacktag_skip(skb, sk, state, cache->end_seq);
1803 /* Check overlap against next cached too (past this one already) */
1804 cache++;
1805 continue;
1806 }
1807
1808 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1809 skb = tcp_highest_sack(sk);
1810 if (!skb)
1811 break;
1812 state->fack_count = tp->fackets_out;
1813 }
1814 skb = tcp_sacktag_skip(skb, sk, state, start_seq);
1815
1816walk:
1817 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
1818 start_seq, end_seq, dup_sack);
1819
1820advance_sp:
1821 i++;
1822 }
1823
1824 /* Clear the head of the cache sack blocks so we can skip it next time */
1825 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1826 tp->recv_sack_cache[i].start_seq = 0;
1827 tp->recv_sack_cache[i].end_seq = 0;
1828 }
1829 for (j = 0; j < used_sacks; j++)
1830 tp->recv_sack_cache[i++] = sp[j];
1831
1832 if ((state->reord < tp->fackets_out) &&
1833 ((inet_csk(sk)->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker))
1834 tcp_update_reordering(sk, tp->fackets_out - state->reord, 0);
1835
1836 tcp_verify_left_out(tp);
1837out:
1838
1839#if FASTRETRANS_DEBUG > 0
1840 WARN_ON((int)tp->sacked_out < 0);
1841 WARN_ON((int)tp->lost_out < 0);
1842 WARN_ON((int)tp->retrans_out < 0);
1843 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1844#endif
1845 return state->flag;
1846}
1847
1848/* Limits sacked_out so that sum with lost_out isn't ever larger than
1849 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1850 */
1851static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1852{
1853 u32 holes;
1854
1855 holes = max(tp->lost_out, 1U);
1856 holes = min(holes, tp->packets_out);
1857
1858 if ((tp->sacked_out + holes) > tp->packets_out) {
1859 tp->sacked_out = tp->packets_out - holes;
1860 return true;
1861 }
1862 return false;
1863}
1864
1865/* If we receive more dupacks than we expected counting segments
1866 * in assumption of absent reordering, interpret this as reordering.
1867 * The only another reason could be bug in receiver TCP.
1868 */
1869static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1870{
1871 struct tcp_sock *tp = tcp_sk(sk);
1872 if (tcp_limit_reno_sacked(tp))
1873 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1874}
1875
1876/* Emulate SACKs for SACKless connection: account for a new dupack. */
1877
1878static void tcp_add_reno_sack(struct sock *sk)
1879{
1880 struct tcp_sock *tp = tcp_sk(sk);
1881 u32 prior_sacked = tp->sacked_out;
1882
1883 tp->sacked_out++;
1884 tcp_check_reno_reordering(sk, 0);
1885 if (tp->sacked_out > prior_sacked)
1886 tp->delivered++; /* Some out-of-order packet is delivered */
1887 tcp_verify_left_out(tp);
1888}
1889
1890/* Account for ACK, ACKing some data in Reno Recovery phase. */
1891
1892static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1893{
1894 struct tcp_sock *tp = tcp_sk(sk);
1895
1896 if (acked > 0) {
1897 /* One ACK acked hole. The rest eat duplicate ACKs. */
1898 tp->delivered += max_t(int, acked - tp->sacked_out, 1);
1899 if (acked - 1 >= tp->sacked_out)
1900 tp->sacked_out = 0;
1901 else
1902 tp->sacked_out -= acked - 1;
1903 }
1904 tcp_check_reno_reordering(sk, acked);
1905 tcp_verify_left_out(tp);
1906}
1907
1908static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1909{
1910 tp->sacked_out = 0;
1911}
1912
1913void tcp_clear_retrans(struct tcp_sock *tp)
1914{
1915 tp->retrans_out = 0;
1916 tp->lost_out = 0;
1917 tp->undo_marker = 0;
1918 tp->undo_retrans = -1;
1919 tp->fackets_out = 0;
1920 tp->sacked_out = 0;
1921}
1922
1923static inline void tcp_init_undo(struct tcp_sock *tp)
1924{
1925 tp->undo_marker = tp->snd_una;
1926 /* Retransmission still in flight may cause DSACKs later. */
1927 tp->undo_retrans = tp->retrans_out ? : -1;
1928}
1929
1930/* Enter Loss state. If we detect SACK reneging, forget all SACK information
1931 * and reset tags completely, otherwise preserve SACKs. If receiver
1932 * dropped its ofo queue, we will know this due to reneging detection.
1933 */
1934void tcp_enter_loss(struct sock *sk)
1935{
1936 const struct inet_connection_sock *icsk = inet_csk(sk);
1937 struct tcp_sock *tp = tcp_sk(sk);
1938 struct net *net = sock_net(sk);
1939 struct sk_buff *skb;
1940 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
1941 bool is_reneg; /* is receiver reneging on SACKs? */
1942 bool mark_lost;
1943
1944 /* Reduce ssthresh if it has not yet been made inside this window. */
1945 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1946 !after(tp->high_seq, tp->snd_una) ||
1947 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1948 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1949 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1950 tcp_ca_event(sk, CA_EVENT_LOSS);
1951 tcp_init_undo(tp);
1952 }
1953 tp->snd_cwnd = 1;
1954 tp->snd_cwnd_cnt = 0;
1955 tp->snd_cwnd_stamp = tcp_time_stamp;
1956
1957 tp->retrans_out = 0;
1958 tp->lost_out = 0;
1959
1960 if (tcp_is_reno(tp))
1961 tcp_reset_reno_sack(tp);
1962
1963 skb = tcp_write_queue_head(sk);
1964 is_reneg = skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED);
1965 if (is_reneg) {
1966 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1967 tp->sacked_out = 0;
1968 tp->fackets_out = 0;
1969 }
1970 tcp_clear_all_retrans_hints(tp);
1971
1972 tcp_for_write_queue(skb, sk) {
1973 if (skb == tcp_send_head(sk))
1974 break;
1975
1976 mark_lost = (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
1977 is_reneg);
1978 if (mark_lost)
1979 tcp_sum_lost(tp, skb);
1980 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1981 if (mark_lost) {
1982 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1983 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1984 tp->lost_out += tcp_skb_pcount(skb);
1985 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
1986 }
1987 }
1988 tcp_verify_left_out(tp);
1989
1990 /* Timeout in disordered state after receiving substantial DUPACKs
1991 * suggests that the degree of reordering is over-estimated.
1992 */
1993 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
1994 tp->sacked_out >= net->ipv4.sysctl_tcp_reordering)
1995 tp->reordering = min_t(unsigned int, tp->reordering,
1996 net->ipv4.sysctl_tcp_reordering);
1997 tcp_set_ca_state(sk, TCP_CA_Loss);
1998 tp->high_seq = tp->snd_nxt;
1999 tcp_ecn_queue_cwr(tp);
2000
2001 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
2002 * loss recovery is underway except recurring timeout(s) on
2003 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
2004 */
2005 tp->frto = sysctl_tcp_frto &&
2006 (new_recovery || icsk->icsk_retransmits) &&
2007 !inet_csk(sk)->icsk_mtup.probe_size;
2008}
2009
2010/* If ACK arrived pointing to a remembered SACK, it means that our
2011 * remembered SACKs do not reflect real state of receiver i.e.
2012 * receiver _host_ is heavily congested (or buggy).
2013 *
2014 * To avoid big spurious retransmission bursts due to transient SACK
2015 * scoreboard oddities that look like reneging, we give the receiver a
2016 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2017 * restore sanity to the SACK scoreboard. If the apparent reneging
2018 * persists until this RTO then we'll clear the SACK scoreboard.
2019 */
2020static bool tcp_check_sack_reneging(struct sock *sk, int flag)
2021{
2022 if (flag & FLAG_SACK_RENEGING) {
2023 struct tcp_sock *tp = tcp_sk(sk);
2024 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2025 msecs_to_jiffies(10));
2026
2027 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2028 delay, TCP_RTO_MAX);
2029 return true;
2030 }
2031 return false;
2032}
2033
2034static inline int tcp_fackets_out(const struct tcp_sock *tp)
2035{
2036 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2037}
2038
2039/* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2040 * counter when SACK is enabled (without SACK, sacked_out is used for
2041 * that purpose).
2042 *
2043 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2044 * segments up to the highest received SACK block so far and holes in
2045 * between them.
2046 *
2047 * With reordering, holes may still be in flight, so RFC3517 recovery
2048 * uses pure sacked_out (total number of SACKed segments) even though
2049 * it violates the RFC that uses duplicate ACKs, often these are equal
2050 * but when e.g. out-of-window ACKs or packet duplication occurs,
2051 * they differ. Since neither occurs due to loss, TCP should really
2052 * ignore them.
2053 */
2054static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2055{
2056 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2057}
2058
2059static bool tcp_pause_early_retransmit(struct sock *sk, int flag)
2060{
2061 struct tcp_sock *tp = tcp_sk(sk);
2062 unsigned long delay;
2063
2064 /* Delay early retransmit and entering fast recovery for
2065 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
2066 * available, or RTO is scheduled to fire first.
2067 */
2068 if (sysctl_tcp_early_retrans < 2 || sysctl_tcp_early_retrans > 3 ||
2069 (flag & FLAG_ECE) || !tp->srtt_us)
2070 return false;
2071
2072 delay = max(usecs_to_jiffies(tp->srtt_us >> 5),
2073 msecs_to_jiffies(2));
2074
2075 if (!time_after(inet_csk(sk)->icsk_timeout, (jiffies + delay)))
2076 return false;
2077
2078 inet_csk_reset_xmit_timer(sk, ICSK_TIME_EARLY_RETRANS, delay,
2079 TCP_RTO_MAX);
2080 return true;
2081}
2082
2083/* Linux NewReno/SACK/FACK/ECN state machine.
2084 * --------------------------------------
2085 *
2086 * "Open" Normal state, no dubious events, fast path.
2087 * "Disorder" In all the respects it is "Open",
2088 * but requires a bit more attention. It is entered when
2089 * we see some SACKs or dupacks. It is split of "Open"
2090 * mainly to move some processing from fast path to slow one.
2091 * "CWR" CWND was reduced due to some Congestion Notification event.
2092 * It can be ECN, ICMP source quench, local device congestion.
2093 * "Recovery" CWND was reduced, we are fast-retransmitting.
2094 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2095 *
2096 * tcp_fastretrans_alert() is entered:
2097 * - each incoming ACK, if state is not "Open"
2098 * - when arrived ACK is unusual, namely:
2099 * * SACK
2100 * * Duplicate ACK.
2101 * * ECN ECE.
2102 *
2103 * Counting packets in flight is pretty simple.
2104 *
2105 * in_flight = packets_out - left_out + retrans_out
2106 *
2107 * packets_out is SND.NXT-SND.UNA counted in packets.
2108 *
2109 * retrans_out is number of retransmitted segments.
2110 *
2111 * left_out is number of segments left network, but not ACKed yet.
2112 *
2113 * left_out = sacked_out + lost_out
2114 *
2115 * sacked_out: Packets, which arrived to receiver out of order
2116 * and hence not ACKed. With SACKs this number is simply
2117 * amount of SACKed data. Even without SACKs
2118 * it is easy to give pretty reliable estimate of this number,
2119 * counting duplicate ACKs.
2120 *
2121 * lost_out: Packets lost by network. TCP has no explicit
2122 * "loss notification" feedback from network (for now).
2123 * It means that this number can be only _guessed_.
2124 * Actually, it is the heuristics to predict lossage that
2125 * distinguishes different algorithms.
2126 *
2127 * F.e. after RTO, when all the queue is considered as lost,
2128 * lost_out = packets_out and in_flight = retrans_out.
2129 *
2130 * Essentially, we have now two algorithms counting
2131 * lost packets.
2132 *
2133 * FACK: It is the simplest heuristics. As soon as we decided
2134 * that something is lost, we decide that _all_ not SACKed
2135 * packets until the most forward SACK are lost. I.e.
2136 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2137 * It is absolutely correct estimate, if network does not reorder
2138 * packets. And it loses any connection to reality when reordering
2139 * takes place. We use FACK by default until reordering
2140 * is suspected on the path to this destination.
2141 *
2142 * NewReno: when Recovery is entered, we assume that one segment
2143 * is lost (classic Reno). While we are in Recovery and
2144 * a partial ACK arrives, we assume that one more packet
2145 * is lost (NewReno). This heuristics are the same in NewReno
2146 * and SACK.
2147 *
2148 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2149 * deflation etc. CWND is real congestion window, never inflated, changes
2150 * only according to classic VJ rules.
2151 *
2152 * Really tricky (and requiring careful tuning) part of algorithm
2153 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2154 * The first determines the moment _when_ we should reduce CWND and,
2155 * hence, slow down forward transmission. In fact, it determines the moment
2156 * when we decide that hole is caused by loss, rather than by a reorder.
2157 *
2158 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2159 * holes, caused by lost packets.
2160 *
2161 * And the most logically complicated part of algorithm is undo
2162 * heuristics. We detect false retransmits due to both too early
2163 * fast retransmit (reordering) and underestimated RTO, analyzing
2164 * timestamps and D-SACKs. When we detect that some segments were
2165 * retransmitted by mistake and CWND reduction was wrong, we undo
2166 * window reduction and abort recovery phase. This logic is hidden
2167 * inside several functions named tcp_try_undo_<something>.
2168 */
2169
2170/* This function decides, when we should leave Disordered state
2171 * and enter Recovery phase, reducing congestion window.
2172 *
2173 * Main question: may we further continue forward transmission
2174 * with the same cwnd?
2175 */
2176static bool tcp_time_to_recover(struct sock *sk, int flag)
2177{
2178 struct tcp_sock *tp = tcp_sk(sk);
2179 __u32 packets_out;
2180 int tcp_reordering = sock_net(sk)->ipv4.sysctl_tcp_reordering;
2181
2182 /* Trick#1: The loss is proven. */
2183 if (tp->lost_out)
2184 return true;
2185
2186 /* Not-A-Trick#2 : Classic rule... */
2187 if (tcp_dupack_heuristics(tp) > tp->reordering)
2188 return true;
2189
2190 /* Trick#4: It is still not OK... But will it be useful to delay
2191 * recovery more?
2192 */
2193 packets_out = tp->packets_out;
2194 if (packets_out <= tp->reordering &&
2195 tp->sacked_out >= max_t(__u32, packets_out/2, tcp_reordering) &&
2196 !tcp_may_send_now(sk)) {
2197 /* We have nothing to send. This connection is limited
2198 * either by receiver window or by application.
2199 */
2200 return true;
2201 }
2202
2203 /* If a thin stream is detected, retransmit after first
2204 * received dupack. Employ only if SACK is supported in order
2205 * to avoid possible corner-case series of spurious retransmissions
2206 * Use only if there are no unsent data.
2207 */
2208 if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2209 tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2210 tcp_is_sack(tp) && !tcp_send_head(sk))
2211 return true;
2212
2213 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious
2214 * retransmissions due to small network reorderings, we implement
2215 * Mitigation A.3 in the RFC and delay the retransmission for a short
2216 * interval if appropriate.
2217 */
2218 if (tp->do_early_retrans && !tp->retrans_out && tp->sacked_out &&
2219 (tp->packets_out >= (tp->sacked_out + 1) && tp->packets_out < 4) &&
2220 !tcp_may_send_now(sk))
2221 return !tcp_pause_early_retransmit(sk, flag);
2222
2223 return false;
2224}
2225
2226/* Detect loss in event "A" above by marking head of queue up as lost.
2227 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2228 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2229 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2230 * the maximum SACKed segments to pass before reaching this limit.
2231 */
2232static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2233{
2234 struct tcp_sock *tp = tcp_sk(sk);
2235 struct sk_buff *skb;
2236 int cnt, oldcnt, lost;
2237 unsigned int mss;
2238 /* Use SACK to deduce losses of new sequences sent during recovery */
2239 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
2240
2241 WARN_ON(packets > tp->packets_out);
2242 if (tp->lost_skb_hint) {
2243 skb = tp->lost_skb_hint;
2244 cnt = tp->lost_cnt_hint;
2245 /* Head already handled? */
2246 if (mark_head && skb != tcp_write_queue_head(sk))
2247 return;
2248 } else {
2249 skb = tcp_write_queue_head(sk);
2250 cnt = 0;
2251 }
2252
2253 tcp_for_write_queue_from(skb, sk) {
2254 if (skb == tcp_send_head(sk))
2255 break;
2256 /* TODO: do this better */
2257 /* this is not the most efficient way to do this... */
2258 tp->lost_skb_hint = skb;
2259 tp->lost_cnt_hint = cnt;
2260
2261 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2262 break;
2263
2264 oldcnt = cnt;
2265 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2266 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2267 cnt += tcp_skb_pcount(skb);
2268
2269 if (cnt > packets) {
2270 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2271 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2272 (oldcnt >= packets))
2273 break;
2274
2275 mss = tcp_skb_mss(skb);
2276 /* If needed, chop off the prefix to mark as lost. */
2277 lost = (packets - oldcnt) * mss;
2278 if (lost < skb->len &&
2279 tcp_fragment(sk, skb, lost, mss, GFP_ATOMIC) < 0)
2280 break;
2281 cnt = packets;
2282 }
2283
2284 tcp_skb_mark_lost(tp, skb);
2285
2286 if (mark_head)
2287 break;
2288 }
2289 tcp_verify_left_out(tp);
2290}
2291
2292/* Account newly detected lost packet(s) */
2293
2294static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2295{
2296 struct tcp_sock *tp = tcp_sk(sk);
2297
2298 if (tcp_is_reno(tp)) {
2299 tcp_mark_head_lost(sk, 1, 1);
2300 } else if (tcp_is_fack(tp)) {
2301 int lost = tp->fackets_out - tp->reordering;
2302 if (lost <= 0)
2303 lost = 1;
2304 tcp_mark_head_lost(sk, lost, 0);
2305 } else {
2306 int sacked_upto = tp->sacked_out - tp->reordering;
2307 if (sacked_upto >= 0)
2308 tcp_mark_head_lost(sk, sacked_upto, 0);
2309 else if (fast_rexmit)
2310 tcp_mark_head_lost(sk, 1, 1);
2311 }
2312}
2313
2314static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2315{
2316 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2317 before(tp->rx_opt.rcv_tsecr, when);
2318}
2319
2320/* skb is spurious retransmitted if the returned timestamp echo
2321 * reply is prior to the skb transmission time
2322 */
2323static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2324 const struct sk_buff *skb)
2325{
2326 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2327 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb));
2328}
2329
2330/* Nothing was retransmitted or returned timestamp is less
2331 * than timestamp of the first retransmission.
2332 */
2333static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2334{
2335 return !tp->retrans_stamp ||
2336 tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
2337}
2338
2339/* Undo procedures. */
2340
2341/* We can clear retrans_stamp when there are no retransmissions in the
2342 * window. It would seem that it is trivially available for us in
2343 * tp->retrans_out, however, that kind of assumptions doesn't consider
2344 * what will happen if errors occur when sending retransmission for the
2345 * second time. ...It could the that such segment has only
2346 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2347 * the head skb is enough except for some reneging corner cases that
2348 * are not worth the effort.
2349 *
2350 * Main reason for all this complexity is the fact that connection dying
2351 * time now depends on the validity of the retrans_stamp, in particular,
2352 * that successive retransmissions of a segment must not advance
2353 * retrans_stamp under any conditions.
2354 */
2355static bool tcp_any_retrans_done(const struct sock *sk)
2356{
2357 const struct tcp_sock *tp = tcp_sk(sk);
2358 struct sk_buff *skb;
2359
2360 if (tp->retrans_out)
2361 return true;
2362
2363 skb = tcp_write_queue_head(sk);
2364 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2365 return true;
2366
2367 return false;
2368}
2369
2370#if FASTRETRANS_DEBUG > 1
2371static void DBGUNDO(struct sock *sk, const char *msg)
2372{
2373 struct tcp_sock *tp = tcp_sk(sk);
2374 struct inet_sock *inet = inet_sk(sk);
2375
2376 if (sk->sk_family == AF_INET) {
2377 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2378 msg,
2379 &inet->inet_daddr, ntohs(inet->inet_dport),
2380 tp->snd_cwnd, tcp_left_out(tp),
2381 tp->snd_ssthresh, tp->prior_ssthresh,
2382 tp->packets_out);
2383 }
2384#if IS_ENABLED(CONFIG_IPV6)
2385 else if (sk->sk_family == AF_INET6) {
2386 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2387 msg,
2388 &sk->sk_v6_daddr, ntohs(inet->inet_dport),
2389 tp->snd_cwnd, tcp_left_out(tp),
2390 tp->snd_ssthresh, tp->prior_ssthresh,
2391 tp->packets_out);
2392 }
2393#endif
2394}
2395#else
2396#define DBGUNDO(x...) do { } while (0)
2397#endif
2398
2399static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2400{
2401 struct tcp_sock *tp = tcp_sk(sk);
2402
2403 if (unmark_loss) {
2404 struct sk_buff *skb;
2405
2406 tcp_for_write_queue(skb, sk) {
2407 if (skb == tcp_send_head(sk))
2408 break;
2409 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2410 }
2411 tp->lost_out = 0;
2412 tcp_clear_all_retrans_hints(tp);
2413 }
2414
2415 if (tp->prior_ssthresh) {
2416 const struct inet_connection_sock *icsk = inet_csk(sk);
2417
2418 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2419
2420 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2421 tp->snd_ssthresh = tp->prior_ssthresh;
2422 tcp_ecn_withdraw_cwr(tp);
2423 }
2424 }
2425 tp->snd_cwnd_stamp = tcp_time_stamp;
2426 tp->undo_marker = 0;
2427}
2428
2429static inline bool tcp_may_undo(const struct tcp_sock *tp)
2430{
2431 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2432}
2433
2434/* People celebrate: "We love our President!" */
2435static bool tcp_try_undo_recovery(struct sock *sk)
2436{
2437 struct tcp_sock *tp = tcp_sk(sk);
2438
2439 if (tcp_may_undo(tp)) {
2440 int mib_idx;
2441
2442 /* Happy end! We did not retransmit anything
2443 * or our original transmission succeeded.
2444 */
2445 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2446 tcp_undo_cwnd_reduction(sk, false);
2447 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2448 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2449 else
2450 mib_idx = LINUX_MIB_TCPFULLUNDO;
2451
2452 NET_INC_STATS(sock_net(sk), mib_idx);
2453 }
2454 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2455 /* Hold old state until something *above* high_seq
2456 * is ACKed. For Reno it is MUST to prevent false
2457 * fast retransmits (RFC2582). SACK TCP is safe. */
2458 if (!tcp_any_retrans_done(sk))
2459 tp->retrans_stamp = 0;
2460 return true;
2461 }
2462 tcp_set_ca_state(sk, TCP_CA_Open);
2463 return false;
2464}
2465
2466/* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2467static bool tcp_try_undo_dsack(struct sock *sk)
2468{
2469 struct tcp_sock *tp = tcp_sk(sk);
2470
2471 if (tp->undo_marker && !tp->undo_retrans) {
2472 DBGUNDO(sk, "D-SACK");
2473 tcp_undo_cwnd_reduction(sk, false);
2474 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2475 return true;
2476 }
2477 return false;
2478}
2479
2480/* Undo during loss recovery after partial ACK or using F-RTO. */
2481static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2482{
2483 struct tcp_sock *tp = tcp_sk(sk);
2484
2485 if (frto_undo || tcp_may_undo(tp)) {
2486 tcp_undo_cwnd_reduction(sk, true);
2487
2488 DBGUNDO(sk, "partial loss");
2489 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2490 if (frto_undo)
2491 NET_INC_STATS(sock_net(sk),
2492 LINUX_MIB_TCPSPURIOUSRTOS);
2493 inet_csk(sk)->icsk_retransmits = 0;
2494 if (frto_undo || tcp_is_sack(tp))
2495 tcp_set_ca_state(sk, TCP_CA_Open);
2496 return true;
2497 }
2498 return false;
2499}
2500
2501/* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2502 * It computes the number of packets to send (sndcnt) based on packets newly
2503 * delivered:
2504 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2505 * cwnd reductions across a full RTT.
2506 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2507 * But when the retransmits are acked without further losses, PRR
2508 * slow starts cwnd up to ssthresh to speed up the recovery.
2509 */
2510static void tcp_init_cwnd_reduction(struct sock *sk)
2511{
2512 struct tcp_sock *tp = tcp_sk(sk);
2513
2514 tp->high_seq = tp->snd_nxt;
2515 tp->tlp_high_seq = 0;
2516 tp->snd_cwnd_cnt = 0;
2517 tp->prior_cwnd = tp->snd_cwnd;
2518 tp->prr_delivered = 0;
2519 tp->prr_out = 0;
2520 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2521 tcp_ecn_queue_cwr(tp);
2522}
2523
2524static void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked,
2525 int flag)
2526{
2527 struct tcp_sock *tp = tcp_sk(sk);
2528 int sndcnt = 0;
2529 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2530
2531 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2532 return;
2533
2534 tp->prr_delivered += newly_acked_sacked;
2535 if (delta < 0) {
2536 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2537 tp->prior_cwnd - 1;
2538 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2539 } else if ((flag & FLAG_RETRANS_DATA_ACKED) &&
2540 !(flag & FLAG_LOST_RETRANS)) {
2541 sndcnt = min_t(int, delta,
2542 max_t(int, tp->prr_delivered - tp->prr_out,
2543 newly_acked_sacked) + 1);
2544 } else {
2545 sndcnt = min(delta, newly_acked_sacked);
2546 }
2547 /* Force a fast retransmit upon entering fast recovery */
2548 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2549 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2550}
2551
2552static inline void tcp_end_cwnd_reduction(struct sock *sk)
2553{
2554 struct tcp_sock *tp = tcp_sk(sk);
2555
2556 if (inet_csk(sk)->icsk_ca_ops->cong_control)
2557 return;
2558
2559 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2560 if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR ||
2561 (tp->undo_marker && tp->snd_ssthresh < TCP_INFINITE_SSTHRESH)) {
2562 tp->snd_cwnd = tp->snd_ssthresh;
2563 tp->snd_cwnd_stamp = tcp_time_stamp;
2564 }
2565 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2566}
2567
2568/* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2569void tcp_enter_cwr(struct sock *sk)
2570{
2571 struct tcp_sock *tp = tcp_sk(sk);
2572
2573 tp->prior_ssthresh = 0;
2574 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2575 tp->undo_marker = 0;
2576 tcp_init_cwnd_reduction(sk);
2577 tcp_set_ca_state(sk, TCP_CA_CWR);
2578 }
2579}
2580EXPORT_SYMBOL(tcp_enter_cwr);
2581
2582static void tcp_try_keep_open(struct sock *sk)
2583{
2584 struct tcp_sock *tp = tcp_sk(sk);
2585 int state = TCP_CA_Open;
2586
2587 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2588 state = TCP_CA_Disorder;
2589
2590 if (inet_csk(sk)->icsk_ca_state != state) {
2591 tcp_set_ca_state(sk, state);
2592 tp->high_seq = tp->snd_nxt;
2593 }
2594}
2595
2596static void tcp_try_to_open(struct sock *sk, int flag)
2597{
2598 struct tcp_sock *tp = tcp_sk(sk);
2599
2600 tcp_verify_left_out(tp);
2601
2602 if (!tcp_any_retrans_done(sk))
2603 tp->retrans_stamp = 0;
2604
2605 if (flag & FLAG_ECE)
2606 tcp_enter_cwr(sk);
2607
2608 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2609 tcp_try_keep_open(sk);
2610 }
2611}
2612
2613static void tcp_mtup_probe_failed(struct sock *sk)
2614{
2615 struct inet_connection_sock *icsk = inet_csk(sk);
2616
2617 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2618 icsk->icsk_mtup.probe_size = 0;
2619 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2620}
2621
2622static void tcp_mtup_probe_success(struct sock *sk)
2623{
2624 struct tcp_sock *tp = tcp_sk(sk);
2625 struct inet_connection_sock *icsk = inet_csk(sk);
2626
2627 /* FIXME: breaks with very large cwnd */
2628 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2629 tp->snd_cwnd = tp->snd_cwnd *
2630 tcp_mss_to_mtu(sk, tp->mss_cache) /
2631 icsk->icsk_mtup.probe_size;
2632 tp->snd_cwnd_cnt = 0;
2633 tp->snd_cwnd_stamp = tcp_time_stamp;
2634 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2635
2636 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2637 icsk->icsk_mtup.probe_size = 0;
2638 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2639 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2640}
2641
2642/* Do a simple retransmit without using the backoff mechanisms in
2643 * tcp_timer. This is used for path mtu discovery.
2644 * The socket is already locked here.
2645 */
2646void tcp_simple_retransmit(struct sock *sk)
2647{
2648 const struct inet_connection_sock *icsk = inet_csk(sk);
2649 struct tcp_sock *tp = tcp_sk(sk);
2650 struct sk_buff *skb;
2651 unsigned int mss = tcp_current_mss(sk);
2652 u32 prior_lost = tp->lost_out;
2653
2654 tcp_for_write_queue(skb, sk) {
2655 if (skb == tcp_send_head(sk))
2656 break;
2657 if (tcp_skb_seglen(skb) > mss &&
2658 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2659 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2660 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2661 tp->retrans_out -= tcp_skb_pcount(skb);
2662 }
2663 tcp_skb_mark_lost_uncond_verify(tp, skb);
2664 }
2665 }
2666
2667 tcp_clear_retrans_hints_partial(tp);
2668
2669 if (prior_lost == tp->lost_out)
2670 return;
2671
2672 if (tcp_is_reno(tp))
2673 tcp_limit_reno_sacked(tp);
2674
2675 tcp_verify_left_out(tp);
2676
2677 /* Don't muck with the congestion window here.
2678 * Reason is that we do not increase amount of _data_
2679 * in network, but units changed and effective
2680 * cwnd/ssthresh really reduced now.
2681 */
2682 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2683 tp->high_seq = tp->snd_nxt;
2684 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2685 tp->prior_ssthresh = 0;
2686 tp->undo_marker = 0;
2687 tcp_set_ca_state(sk, TCP_CA_Loss);
2688 }
2689 tcp_xmit_retransmit_queue(sk);
2690}
2691EXPORT_SYMBOL(tcp_simple_retransmit);
2692
2693static void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2694{
2695 struct tcp_sock *tp = tcp_sk(sk);
2696 int mib_idx;
2697
2698 if (tcp_is_reno(tp))
2699 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2700 else
2701 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2702
2703 NET_INC_STATS(sock_net(sk), mib_idx);
2704
2705 tp->prior_ssthresh = 0;
2706 tcp_init_undo(tp);
2707
2708 if (!tcp_in_cwnd_reduction(sk)) {
2709 if (!ece_ack)
2710 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2711 tcp_init_cwnd_reduction(sk);
2712 }
2713 tcp_set_ca_state(sk, TCP_CA_Recovery);
2714}
2715
2716/* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2717 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2718 */
2719static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack,
2720 int *rexmit)
2721{
2722 struct tcp_sock *tp = tcp_sk(sk);
2723 bool recovered = !before(tp->snd_una, tp->high_seq);
2724
2725 if ((flag & FLAG_SND_UNA_ADVANCED) &&
2726 tcp_try_undo_loss(sk, false))
2727 return;
2728
2729 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2730 /* Step 3.b. A timeout is spurious if not all data are
2731 * lost, i.e., never-retransmitted data are (s)acked.
2732 */
2733 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2734 tcp_try_undo_loss(sk, true))
2735 return;
2736
2737 if (after(tp->snd_nxt, tp->high_seq)) {
2738 if (flag & FLAG_DATA_SACKED || is_dupack)
2739 tp->frto = 0; /* Step 3.a. loss was real */
2740 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2741 tp->high_seq = tp->snd_nxt;
2742 /* Step 2.b. Try send new data (but deferred until cwnd
2743 * is updated in tcp_ack()). Otherwise fall back to
2744 * the conventional recovery.
2745 */
2746 if (tcp_send_head(sk) &&
2747 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2748 *rexmit = REXMIT_NEW;
2749 return;
2750 }
2751 tp->frto = 0;
2752 }
2753 }
2754
2755 if (recovered) {
2756 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2757 tcp_try_undo_recovery(sk);
2758 return;
2759 }
2760 if (tcp_is_reno(tp)) {
2761 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2762 * delivered. Lower inflight to clock out (re)tranmissions.
2763 */
2764 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2765 tcp_add_reno_sack(sk);
2766 else if (flag & FLAG_SND_UNA_ADVANCED)
2767 tcp_reset_reno_sack(tp);
2768 }
2769 *rexmit = REXMIT_LOST;
2770}
2771
2772/* Undo during fast recovery after partial ACK. */
2773static bool tcp_try_undo_partial(struct sock *sk, const int acked)
2774{
2775 struct tcp_sock *tp = tcp_sk(sk);
2776
2777 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2778 /* Plain luck! Hole if filled with delayed
2779 * packet, rather than with a retransmit.
2780 */
2781 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2782
2783 /* We are getting evidence that the reordering degree is higher
2784 * than we realized. If there are no retransmits out then we
2785 * can undo. Otherwise we clock out new packets but do not
2786 * mark more packets lost or retransmit more.
2787 */
2788 if (tp->retrans_out)
2789 return true;
2790
2791 if (!tcp_any_retrans_done(sk))
2792 tp->retrans_stamp = 0;
2793
2794 DBGUNDO(sk, "partial recovery");
2795 tcp_undo_cwnd_reduction(sk, true);
2796 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2797 tcp_try_keep_open(sk);
2798 return true;
2799 }
2800 return false;
2801}
2802
2803/* Process an event, which can update packets-in-flight not trivially.
2804 * Main goal of this function is to calculate new estimate for left_out,
2805 * taking into account both packets sitting in receiver's buffer and
2806 * packets lost by network.
2807 *
2808 * Besides that it updates the congestion state when packet loss or ECN
2809 * is detected. But it does not reduce the cwnd, it is done by the
2810 * congestion control later.
2811 *
2812 * It does _not_ decide what to send, it is made in function
2813 * tcp_xmit_retransmit_queue().
2814 */
2815static void tcp_fastretrans_alert(struct sock *sk, const int acked,
2816 bool is_dupack, int *ack_flag, int *rexmit)
2817{
2818 struct inet_connection_sock *icsk = inet_csk(sk);
2819 struct tcp_sock *tp = tcp_sk(sk);
2820 int fast_rexmit = 0, flag = *ack_flag;
2821 bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2822 (tcp_fackets_out(tp) > tp->reordering));
2823
2824 if (WARN_ON(!tp->packets_out && tp->sacked_out))
2825 tp->sacked_out = 0;
2826 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2827 tp->fackets_out = 0;
2828
2829 /* Now state machine starts.
2830 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2831 if (flag & FLAG_ECE)
2832 tp->prior_ssthresh = 0;
2833
2834 /* B. In all the states check for reneging SACKs. */
2835 if (tcp_check_sack_reneging(sk, flag))
2836 return;
2837
2838 /* C. Check consistency of the current state. */
2839 tcp_verify_left_out(tp);
2840
2841 /* D. Check state exit conditions. State can be terminated
2842 * when high_seq is ACKed. */
2843 if (icsk->icsk_ca_state == TCP_CA_Open) {
2844 WARN_ON(tp->retrans_out != 0);
2845 tp->retrans_stamp = 0;
2846 } else if (!before(tp->snd_una, tp->high_seq)) {
2847 switch (icsk->icsk_ca_state) {
2848 case TCP_CA_CWR:
2849 /* CWR is to be held something *above* high_seq
2850 * is ACKed for CWR bit to reach receiver. */
2851 if (tp->snd_una != tp->high_seq) {
2852 tcp_end_cwnd_reduction(sk);
2853 tcp_set_ca_state(sk, TCP_CA_Open);
2854 }
2855 break;
2856
2857 case TCP_CA_Recovery:
2858 if (tcp_is_reno(tp))
2859 tcp_reset_reno_sack(tp);
2860 if (tcp_try_undo_recovery(sk))
2861 return;
2862 tcp_end_cwnd_reduction(sk);
2863 break;
2864 }
2865 }
2866
2867 /* Use RACK to detect loss */
2868 if (sysctl_tcp_recovery & TCP_RACK_LOST_RETRANS &&
2869 tcp_rack_mark_lost(sk)) {
2870 flag |= FLAG_LOST_RETRANS;
2871 *ack_flag |= FLAG_LOST_RETRANS;
2872 }
2873
2874 /* E. Process state. */
2875 switch (icsk->icsk_ca_state) {
2876 case TCP_CA_Recovery:
2877 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2878 if (tcp_is_reno(tp) && is_dupack)
2879 tcp_add_reno_sack(sk);
2880 } else {
2881 if (tcp_try_undo_partial(sk, acked))
2882 return;
2883 /* Partial ACK arrived. Force fast retransmit. */
2884 do_lost = tcp_is_reno(tp) ||
2885 tcp_fackets_out(tp) > tp->reordering;
2886 }
2887 if (tcp_try_undo_dsack(sk)) {
2888 tcp_try_keep_open(sk);
2889 return;
2890 }
2891 break;
2892 case TCP_CA_Loss:
2893 tcp_process_loss(sk, flag, is_dupack, rexmit);
2894 if (icsk->icsk_ca_state != TCP_CA_Open &&
2895 !(flag & FLAG_LOST_RETRANS))
2896 return;
2897 /* Change state if cwnd is undone or retransmits are lost */
2898 default:
2899 if (tcp_is_reno(tp)) {
2900 if (flag & FLAG_SND_UNA_ADVANCED)
2901 tcp_reset_reno_sack(tp);
2902 if (is_dupack)
2903 tcp_add_reno_sack(sk);
2904 }
2905
2906 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2907 tcp_try_undo_dsack(sk);
2908
2909 if (!tcp_time_to_recover(sk, flag)) {
2910 tcp_try_to_open(sk, flag);
2911 return;
2912 }
2913
2914 /* MTU probe failure: don't reduce cwnd */
2915 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2916 icsk->icsk_mtup.probe_size &&
2917 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2918 tcp_mtup_probe_failed(sk);
2919 /* Restores the reduction we did in tcp_mtup_probe() */
2920 tp->snd_cwnd++;
2921 tcp_simple_retransmit(sk);
2922 return;
2923 }
2924
2925 /* Otherwise enter Recovery state */
2926 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2927 fast_rexmit = 1;
2928 }
2929
2930 if (do_lost)
2931 tcp_update_scoreboard(sk, fast_rexmit);
2932 *rexmit = REXMIT_LOST;
2933}
2934
2935static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us)
2936{
2937 struct tcp_sock *tp = tcp_sk(sk);
2938 u32 wlen = sysctl_tcp_min_rtt_wlen * HZ;
2939
2940 minmax_running_min(&tp->rtt_min, wlen, tcp_time_stamp,
2941 rtt_us ? : jiffies_to_usecs(1));
2942}
2943
2944static inline bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2945 long seq_rtt_us, long sack_rtt_us,
2946 long ca_rtt_us)
2947{
2948 const struct tcp_sock *tp = tcp_sk(sk);
2949
2950 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2951 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2952 * Karn's algorithm forbids taking RTT if some retransmitted data
2953 * is acked (RFC6298).
2954 */
2955 if (seq_rtt_us < 0)
2956 seq_rtt_us = sack_rtt_us;
2957
2958 /* RTTM Rule: A TSecr value received in a segment is used to
2959 * update the averaged RTT measurement only if the segment
2960 * acknowledges some new data, i.e., only if it advances the
2961 * left edge of the send window.
2962 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2963 */
2964 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2965 flag & FLAG_ACKED)
2966 seq_rtt_us = ca_rtt_us = jiffies_to_usecs(tcp_time_stamp -
2967 tp->rx_opt.rcv_tsecr);
2968 if (seq_rtt_us < 0)
2969 return false;
2970
2971 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2972 * always taken together with ACK, SACK, or TS-opts. Any negative
2973 * values will be skipped with the seq_rtt_us < 0 check above.
2974 */
2975 tcp_update_rtt_min(sk, ca_rtt_us);
2976 tcp_rtt_estimator(sk, seq_rtt_us);
2977 tcp_set_rto(sk);
2978
2979 /* RFC6298: only reset backoff on valid RTT measurement. */
2980 inet_csk(sk)->icsk_backoff = 0;
2981 return true;
2982}
2983
2984/* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2985void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
2986{
2987 long rtt_us = -1L;
2988
2989 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack.v64) {
2990 struct skb_mstamp now;
2991
2992 skb_mstamp_get(&now);
2993 rtt_us = skb_mstamp_us_delta(&now, &tcp_rsk(req)->snt_synack);
2994 }
2995
2996 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us);
2997}
2998
2999
3000static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
3001{
3002 const struct inet_connection_sock *icsk = inet_csk(sk);
3003
3004 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
3005 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
3006}
3007
3008/* Restart timer after forward progress on connection.
3009 * RFC2988 recommends to restart timer to now+rto.
3010 */
3011void tcp_rearm_rto(struct sock *sk)
3012{
3013 const struct inet_connection_sock *icsk = inet_csk(sk);
3014 struct tcp_sock *tp = tcp_sk(sk);
3015
3016 /* If the retrans timer is currently being used by Fast Open
3017 * for SYN-ACK retrans purpose, stay put.
3018 */
3019 if (tp->fastopen_rsk)
3020 return;
3021
3022 if (!tp->packets_out) {
3023 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3024 } else {
3025 u32 rto = inet_csk(sk)->icsk_rto;
3026 /* Offset the time elapsed after installing regular RTO */
3027 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
3028 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
3029 struct sk_buff *skb = tcp_write_queue_head(sk);
3030 const u32 rto_time_stamp =
3031 tcp_skb_timestamp(skb) + rto;
3032 s32 delta = (s32)(rto_time_stamp - tcp_time_stamp);
3033 /* delta may not be positive if the socket is locked
3034 * when the retrans timer fires and is rescheduled.
3035 */
3036 if (delta > 0)
3037 rto = delta;
3038 }
3039 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3040 TCP_RTO_MAX);
3041 }
3042}
3043
3044/* This function is called when the delayed ER timer fires. TCP enters
3045 * fast recovery and performs fast-retransmit.
3046 */
3047void tcp_resume_early_retransmit(struct sock *sk)
3048{
3049 struct tcp_sock *tp = tcp_sk(sk);
3050
3051 tcp_rearm_rto(sk);
3052
3053 /* Stop if ER is disabled after the delayed ER timer is scheduled */
3054 if (!tp->do_early_retrans)
3055 return;
3056
3057 tcp_enter_recovery(sk, false);
3058 tcp_update_scoreboard(sk, 1);
3059 tcp_xmit_retransmit_queue(sk);
3060}
3061
3062/* If we get here, the whole TSO packet has not been acked. */
3063static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3064{
3065 struct tcp_sock *tp = tcp_sk(sk);
3066 u32 packets_acked;
3067
3068 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3069
3070 packets_acked = tcp_skb_pcount(skb);
3071 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3072 return 0;
3073 packets_acked -= tcp_skb_pcount(skb);
3074
3075 if (packets_acked) {
3076 BUG_ON(tcp_skb_pcount(skb) == 0);
3077 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3078 }
3079
3080 return packets_acked;
3081}
3082
3083static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3084 u32 prior_snd_una)
3085{
3086 const struct skb_shared_info *shinfo;
3087
3088 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3089 if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
3090 return;
3091
3092 shinfo = skb_shinfo(skb);
3093 if (!before(shinfo->tskey, prior_snd_una) &&
3094 before(shinfo->tskey, tcp_sk(sk)->snd_una))
3095 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3096}
3097
3098/* Remove acknowledged frames from the retransmission queue. If our packet
3099 * is before the ack sequence we can discard it as it's confirmed to have
3100 * arrived at the other end.
3101 */
3102static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3103 u32 prior_snd_una, int *acked,
3104 struct tcp_sacktag_state *sack,
3105 struct skb_mstamp *now)
3106{
3107 const struct inet_connection_sock *icsk = inet_csk(sk);
3108 struct skb_mstamp first_ackt, last_ackt;
3109 struct tcp_sock *tp = tcp_sk(sk);
3110 u32 prior_sacked = tp->sacked_out;
3111 u32 reord = tp->packets_out;
3112 bool fully_acked = true;
3113 long sack_rtt_us = -1L;
3114 long seq_rtt_us = -1L;
3115 long ca_rtt_us = -1L;
3116 struct sk_buff *skb;
3117 u32 pkts_acked = 0;
3118 u32 last_in_flight = 0;
3119 bool rtt_update;
3120 int flag = 0;
3121
3122 first_ackt.v64 = 0;
3123
3124 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3125 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3126 u8 sacked = scb->sacked;
3127 u32 acked_pcount;
3128
3129 tcp_ack_tstamp(sk, skb, prior_snd_una);
3130
3131 /* Determine how many packets and what bytes were acked, tso and else */
3132 if (after(scb->end_seq, tp->snd_una)) {
3133 if (tcp_skb_pcount(skb) == 1 ||
3134 !after(tp->snd_una, scb->seq))
3135 break;
3136
3137 acked_pcount = tcp_tso_acked(sk, skb);
3138 if (!acked_pcount)
3139 break;
3140 fully_acked = false;
3141 } else {
3142 /* Speedup tcp_unlink_write_queue() and next loop */
3143 prefetchw(skb->next);
3144 acked_pcount = tcp_skb_pcount(skb);
3145 }
3146
3147 if (unlikely(sacked & TCPCB_RETRANS)) {
3148 if (sacked & TCPCB_SACKED_RETRANS)
3149 tp->retrans_out -= acked_pcount;
3150 flag |= FLAG_RETRANS_DATA_ACKED;
3151 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3152 last_ackt = skb->skb_mstamp;
3153 WARN_ON_ONCE(last_ackt.v64 == 0);
3154 if (!first_ackt.v64)
3155 first_ackt = last_ackt;
3156
3157 last_in_flight = TCP_SKB_CB(skb)->tx.in_flight;
3158 reord = min(pkts_acked, reord);
3159 if (!after(scb->end_seq, tp->high_seq))
3160 flag |= FLAG_ORIG_SACK_ACKED;
3161 }
3162
3163 if (sacked & TCPCB_SACKED_ACKED) {
3164 tp->sacked_out -= acked_pcount;
3165 } else if (tcp_is_sack(tp)) {
3166 tp->delivered += acked_pcount;
3167 if (!tcp_skb_spurious_retrans(tp, skb))
3168 tcp_rack_advance(tp, &skb->skb_mstamp, sacked);
3169 }
3170 if (sacked & TCPCB_LOST)
3171 tp->lost_out -= acked_pcount;
3172
3173 tp->packets_out -= acked_pcount;
3174 pkts_acked += acked_pcount;
3175 tcp_rate_skb_delivered(sk, skb, sack->rate);
3176
3177 /* Initial outgoing SYN's get put onto the write_queue
3178 * just like anything else we transmit. It is not
3179 * true data, and if we misinform our callers that
3180 * this ACK acks real data, we will erroneously exit
3181 * connection startup slow start one packet too
3182 * quickly. This is severely frowned upon behavior.
3183 */
3184 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3185 flag |= FLAG_DATA_ACKED;
3186 } else {
3187 flag |= FLAG_SYN_ACKED;
3188 tp->retrans_stamp = 0;
3189 }
3190
3191 if (!fully_acked)
3192 break;
3193
3194 tcp_unlink_write_queue(skb, sk);
3195 sk_wmem_free_skb(sk, skb);
3196 if (unlikely(skb == tp->retransmit_skb_hint))
3197 tp->retransmit_skb_hint = NULL;
3198 if (unlikely(skb == tp->lost_skb_hint))
3199 tp->lost_skb_hint = NULL;
3200 }
3201
3202 if (!skb)
3203 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3204
3205 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3206 tp->snd_up = tp->snd_una;
3207
3208 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3209 flag |= FLAG_SACK_RENEGING;
3210
3211 if (likely(first_ackt.v64) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3212 seq_rtt_us = skb_mstamp_us_delta(now, &first_ackt);
3213 ca_rtt_us = skb_mstamp_us_delta(now, &last_ackt);
3214 }
3215 if (sack->first_sackt.v64) {
3216 sack_rtt_us = skb_mstamp_us_delta(now, &sack->first_sackt);
3217 ca_rtt_us = skb_mstamp_us_delta(now, &sack->last_sackt);
3218 }
3219 sack->rate->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet, or -1 */
3220 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3221 ca_rtt_us);
3222
3223 if (flag & FLAG_ACKED) {
3224 tcp_rearm_rto(sk);
3225 if (unlikely(icsk->icsk_mtup.probe_size &&
3226 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3227 tcp_mtup_probe_success(sk);
3228 }
3229
3230 if (tcp_is_reno(tp)) {
3231 tcp_remove_reno_sacks(sk, pkts_acked);
3232 } else {
3233 int delta;
3234
3235 /* Non-retransmitted hole got filled? That's reordering */
3236 if (reord < prior_fackets)
3237 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3238
3239 delta = tcp_is_fack(tp) ? pkts_acked :
3240 prior_sacked - tp->sacked_out;
3241 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3242 }
3243
3244 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3245
3246 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3247 sack_rtt_us > skb_mstamp_us_delta(now, &skb->skb_mstamp)) {
3248 /* Do not re-arm RTO if the sack RTT is measured from data sent
3249 * after when the head was last (re)transmitted. Otherwise the
3250 * timeout may continue to extend in loss recovery.
3251 */
3252 tcp_rearm_rto(sk);
3253 }
3254
3255 if (icsk->icsk_ca_ops->pkts_acked) {
3256 struct ack_sample sample = { .pkts_acked = pkts_acked,
3257 .rtt_us = ca_rtt_us,
3258 .in_flight = last_in_flight };
3259
3260 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3261 }
3262
3263#if FASTRETRANS_DEBUG > 0
3264 WARN_ON((int)tp->sacked_out < 0);
3265 WARN_ON((int)tp->lost_out < 0);
3266 WARN_ON((int)tp->retrans_out < 0);
3267 if (!tp->packets_out && tcp_is_sack(tp)) {
3268 icsk = inet_csk(sk);
3269 if (tp->lost_out) {
3270 pr_debug("Leak l=%u %d\n",
3271 tp->lost_out, icsk->icsk_ca_state);
3272 tp->lost_out = 0;
3273 }
3274 if (tp->sacked_out) {
3275 pr_debug("Leak s=%u %d\n",
3276 tp->sacked_out, icsk->icsk_ca_state);
3277 tp->sacked_out = 0;
3278 }
3279 if (tp->retrans_out) {
3280 pr_debug("Leak r=%u %d\n",
3281 tp->retrans_out, icsk->icsk_ca_state);
3282 tp->retrans_out = 0;
3283 }
3284 }
3285#endif
3286 *acked = pkts_acked;
3287 return flag;
3288}
3289
3290static void tcp_ack_probe(struct sock *sk)
3291{
3292 const struct tcp_sock *tp = tcp_sk(sk);
3293 struct inet_connection_sock *icsk = inet_csk(sk);
3294
3295 /* Was it a usable window open? */
3296
3297 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3298 icsk->icsk_backoff = 0;
3299 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3300 /* Socket must be waked up by subsequent tcp_data_snd_check().
3301 * This function is not for random using!
3302 */
3303 } else {
3304 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3305
3306 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3307 when, TCP_RTO_MAX);
3308 }
3309}
3310
3311static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3312{
3313 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3314 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3315}
3316
3317/* Decide wheather to run the increase function of congestion control. */
3318static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3319{
3320 /* If reordering is high then always grow cwnd whenever data is
3321 * delivered regardless of its ordering. Otherwise stay conservative
3322 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3323 * new SACK or ECE mark may first advance cwnd here and later reduce
3324 * cwnd in tcp_fastretrans_alert() based on more states.
3325 */
3326 if (tcp_sk(sk)->reordering > sock_net(sk)->ipv4.sysctl_tcp_reordering)
3327 return flag & FLAG_FORWARD_PROGRESS;
3328
3329 return flag & FLAG_DATA_ACKED;
3330}
3331
3332/* The "ultimate" congestion control function that aims to replace the rigid
3333 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3334 * It's called toward the end of processing an ACK with precise rate
3335 * information. All transmission or retransmission are delayed afterwards.
3336 */
3337static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3338 int flag, const struct rate_sample *rs)
3339{
3340 const struct inet_connection_sock *icsk = inet_csk(sk);
3341
3342 if (icsk->icsk_ca_ops->cong_control) {
3343 icsk->icsk_ca_ops->cong_control(sk, rs);
3344 return;
3345 }
3346
3347 if (tcp_in_cwnd_reduction(sk)) {
3348 /* Reduce cwnd if state mandates */
3349 tcp_cwnd_reduction(sk, acked_sacked, flag);
3350 } else if (tcp_may_raise_cwnd(sk, flag)) {
3351 /* Advance cwnd if state allows */
3352 tcp_cong_avoid(sk, ack, acked_sacked);
3353 }
3354 tcp_update_pacing_rate(sk);
3355}
3356
3357/* Check that window update is acceptable.
3358 * The function assumes that snd_una<=ack<=snd_next.
3359 */
3360static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3361 const u32 ack, const u32 ack_seq,
3362 const u32 nwin)
3363{
3364 return after(ack, tp->snd_una) ||
3365 after(ack_seq, tp->snd_wl1) ||
3366 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3367}
3368
3369/* If we update tp->snd_una, also update tp->bytes_acked */
3370static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3371{
3372 u32 delta = ack - tp->snd_una;
3373
3374 sock_owned_by_me((struct sock *)tp);
3375 tp->bytes_acked += delta;
3376 tp->snd_una = ack;
3377}
3378
3379/* If we update tp->rcv_nxt, also update tp->bytes_received */
3380static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3381{
3382 u32 delta = seq - tp->rcv_nxt;
3383
3384 sock_owned_by_me((struct sock *)tp);
3385 tp->bytes_received += delta;
3386 tp->rcv_nxt = seq;
3387}
3388
3389/* Update our send window.
3390 *
3391 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3392 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3393 */
3394static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3395 u32 ack_seq)
3396{
3397 struct tcp_sock *tp = tcp_sk(sk);
3398 int flag = 0;
3399 u32 nwin = ntohs(tcp_hdr(skb)->window);
3400
3401 if (likely(!tcp_hdr(skb)->syn))
3402 nwin <<= tp->rx_opt.snd_wscale;
3403
3404 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3405 flag |= FLAG_WIN_UPDATE;
3406 tcp_update_wl(tp, ack_seq);
3407
3408 if (tp->snd_wnd != nwin) {
3409 tp->snd_wnd = nwin;
3410
3411 /* Note, it is the only place, where
3412 * fast path is recovered for sending TCP.
3413 */
3414 tp->pred_flags = 0;
3415 tcp_fast_path_check(sk);
3416
3417 if (tcp_send_head(sk))
3418 tcp_slow_start_after_idle_check(sk);
3419
3420 if (nwin > tp->max_window) {
3421 tp->max_window = nwin;
3422 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3423 }
3424 }
3425 }
3426
3427 tcp_snd_una_update(tp, ack);
3428
3429 return flag;
3430}
3431
3432static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3433 u32 *last_oow_ack_time)
3434{
3435 if (*last_oow_ack_time) {
3436 s32 elapsed = (s32)(tcp_time_stamp - *last_oow_ack_time);
3437
3438 if (0 <= elapsed && elapsed < sysctl_tcp_invalid_ratelimit) {
3439 NET_INC_STATS(net, mib_idx);
3440 return true; /* rate-limited: don't send yet! */
3441 }
3442 }
3443
3444 *last_oow_ack_time = tcp_time_stamp;
3445
3446 return false; /* not rate-limited: go ahead, send dupack now! */
3447}
3448
3449/* Return true if we're currently rate-limiting out-of-window ACKs and
3450 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3451 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3452 * attacks that send repeated SYNs or ACKs for the same connection. To
3453 * do this, we do not send a duplicate SYNACK or ACK if the remote
3454 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3455 */
3456bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3457 int mib_idx, u32 *last_oow_ack_time)
3458{
3459 /* Data packets without SYNs are not likely part of an ACK loop. */
3460 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3461 !tcp_hdr(skb)->syn)
3462 return false;
3463
3464 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3465}
3466
3467/* RFC 5961 7 [ACK Throttling] */
3468static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
3469{
3470 /* unprotected vars, we dont care of overwrites */
3471 static u32 challenge_timestamp;
3472 static unsigned int challenge_count;
3473 struct tcp_sock *tp = tcp_sk(sk);
3474 u32 count, now;
3475
3476 /* First check our per-socket dupack rate limit. */
3477 if (__tcp_oow_rate_limited(sock_net(sk),
3478 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3479 &tp->last_oow_ack_time))
3480 return;
3481
3482 /* Then check host-wide RFC 5961 rate limit. */
3483 now = jiffies / HZ;
3484 if (now != challenge_timestamp) {
3485 u32 half = (sysctl_tcp_challenge_ack_limit + 1) >> 1;
3486
3487 challenge_timestamp = now;
3488 WRITE_ONCE(challenge_count, half +
3489 prandom_u32_max(sysctl_tcp_challenge_ack_limit));
3490 }
3491 count = READ_ONCE(challenge_count);
3492 if (count > 0) {
3493 WRITE_ONCE(challenge_count, count - 1);
3494 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK);
3495 tcp_send_ack(sk);
3496 }
3497}
3498
3499static void tcp_store_ts_recent(struct tcp_sock *tp)
3500{
3501 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3502 tp->rx_opt.ts_recent_stamp = get_seconds();
3503}
3504
3505static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3506{
3507 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3508 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3509 * extra check below makes sure this can only happen
3510 * for pure ACK frames. -DaveM
3511 *
3512 * Not only, also it occurs for expired timestamps.
3513 */
3514
3515 if (tcp_paws_check(&tp->rx_opt, 0))
3516 tcp_store_ts_recent(tp);
3517 }
3518}
3519
3520/* This routine deals with acks during a TLP episode.
3521 * We mark the end of a TLP episode on receiving TLP dupack or when
3522 * ack is after tlp_high_seq.
3523 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3524 */
3525static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3526{
3527 struct tcp_sock *tp = tcp_sk(sk);
3528
3529 if (before(ack, tp->tlp_high_seq))
3530 return;
3531
3532 if (flag & FLAG_DSACKING_ACK) {
3533 /* This DSACK means original and TLP probe arrived; no loss */
3534 tp->tlp_high_seq = 0;
3535 } else if (after(ack, tp->tlp_high_seq)) {
3536 /* ACK advances: there was a loss, so reduce cwnd. Reset
3537 * tlp_high_seq in tcp_init_cwnd_reduction()
3538 */
3539 tcp_init_cwnd_reduction(sk);
3540 tcp_set_ca_state(sk, TCP_CA_CWR);
3541 tcp_end_cwnd_reduction(sk);
3542 tcp_try_keep_open(sk);
3543 NET_INC_STATS(sock_net(sk),
3544 LINUX_MIB_TCPLOSSPROBERECOVERY);
3545 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3546 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3547 /* Pure dupack: original and TLP probe arrived; no loss */
3548 tp->tlp_high_seq = 0;
3549 }
3550}
3551
3552static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3553{
3554 const struct inet_connection_sock *icsk = inet_csk(sk);
3555
3556 if (icsk->icsk_ca_ops->in_ack_event)
3557 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3558}
3559
3560/* Congestion control has updated the cwnd already. So if we're in
3561 * loss recovery then now we do any new sends (for FRTO) or
3562 * retransmits (for CA_Loss or CA_recovery) that make sense.
3563 */
3564static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3565{
3566 struct tcp_sock *tp = tcp_sk(sk);
3567
3568 if (rexmit == REXMIT_NONE)
3569 return;
3570
3571 if (unlikely(rexmit == 2)) {
3572 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3573 TCP_NAGLE_OFF);
3574 if (after(tp->snd_nxt, tp->high_seq))
3575 return;
3576 tp->frto = 0;
3577 }
3578 tcp_xmit_retransmit_queue(sk);
3579}
3580
3581/* This routine deals with incoming acks, but not outgoing ones. */
3582static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3583{
3584 struct inet_connection_sock *icsk = inet_csk(sk);
3585 struct tcp_sock *tp = tcp_sk(sk);
3586 struct tcp_sacktag_state sack_state;
3587 struct rate_sample rs = { .prior_delivered = 0 };
3588 u32 prior_snd_una = tp->snd_una;
3589 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3590 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3591 bool is_dupack = false;
3592 u32 prior_fackets;
3593 int prior_packets = tp->packets_out;
3594 u32 delivered = tp->delivered;
3595 u32 lost = tp->lost;
3596 int acked = 0; /* Number of packets newly acked */
3597 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3598 struct skb_mstamp now;
3599
3600 sack_state.first_sackt.v64 = 0;
3601 sack_state.rate = &rs;
3602
3603 /* We very likely will need to access write queue head. */
3604 prefetchw(sk->sk_write_queue.next);
3605
3606 /* If the ack is older than previous acks
3607 * then we can probably ignore it.
3608 */
3609 if (before(ack, prior_snd_una)) {
3610 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3611 if (before(ack, prior_snd_una - tp->max_window)) {
3612 tcp_send_challenge_ack(sk, skb);
3613 return -1;
3614 }
3615 goto old_ack;
3616 }
3617
3618 /* If the ack includes data we haven't sent yet, discard
3619 * this segment (RFC793 Section 3.9).
3620 */
3621 if (after(ack, tp->snd_nxt))
3622 goto invalid_ack;
3623
3624 skb_mstamp_get(&now);
3625
3626 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
3627 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
3628 tcp_rearm_rto(sk);
3629
3630 if (after(ack, prior_snd_una)) {
3631 flag |= FLAG_SND_UNA_ADVANCED;
3632 icsk->icsk_retransmits = 0;
3633 }
3634
3635 prior_fackets = tp->fackets_out;
3636 rs.prior_in_flight = tcp_packets_in_flight(tp);
3637
3638 /* ts_recent update must be made after we are sure that the packet
3639 * is in window.
3640 */
3641 if (flag & FLAG_UPDATE_TS_RECENT)
3642 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3643
3644 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3645 /* Window is constant, pure forward advance.
3646 * No more checks are required.
3647 * Note, we use the fact that SND.UNA>=SND.WL2.
3648 */
3649 tcp_update_wl(tp, ack_seq);
3650 tcp_snd_una_update(tp, ack);
3651 flag |= FLAG_WIN_UPDATE;
3652
3653 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3654
3655 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3656 } else {
3657 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3658
3659 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3660 flag |= FLAG_DATA;
3661 else
3662 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3663
3664 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3665
3666 if (TCP_SKB_CB(skb)->sacked)
3667 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3668 &sack_state);
3669
3670 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3671 flag |= FLAG_ECE;
3672 ack_ev_flags |= CA_ACK_ECE;
3673 }
3674
3675 if (flag & FLAG_WIN_UPDATE)
3676 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3677
3678 tcp_in_ack_event(sk, ack_ev_flags);
3679 }
3680
3681 /* We passed data and got it acked, remove any soft error
3682 * log. Something worked...
3683 */
3684 sk->sk_err_soft = 0;
3685 icsk->icsk_probes_out = 0;
3686 tp->rcv_tstamp = tcp_time_stamp;
3687 if (!prior_packets)
3688 goto no_queue;
3689
3690 /* See if we can take anything off of the retransmit queue. */
3691 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una, &acked,
3692 &sack_state, &now);
3693
3694 if (tcp_ack_is_dubious(sk, flag)) {
3695 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3696 tcp_fastretrans_alert(sk, acked, is_dupack, &flag, &rexmit);
3697 }
3698 if (tp->tlp_high_seq)
3699 tcp_process_tlp_ack(sk, ack, flag);
3700
3701 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) {
3702 struct dst_entry *dst = __sk_dst_get(sk);
3703 if (dst)
3704 dst_confirm(dst);
3705 }
3706
3707 if (icsk->icsk_pending == ICSK_TIME_RETRANS)
3708 tcp_schedule_loss_probe(sk);
3709 delivered = tp->delivered - delivered; /* freshly ACKed or SACKed */
3710 lost = tp->lost - lost; /* freshly marked lost */
3711 tcp_rate_gen(sk, delivered, lost, &now, &rs);
3712 tcp_cong_control(sk, ack, delivered, flag, &rs);
3713 tcp_xmit_recovery(sk, rexmit);
3714 return 1;
3715
3716no_queue:
3717 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3718 if (flag & FLAG_DSACKING_ACK)
3719 tcp_fastretrans_alert(sk, acked, is_dupack, &flag, &rexmit);
3720 /* If this ack opens up a zero window, clear backoff. It was
3721 * being used to time the probes, and is probably far higher than
3722 * it needs to be for normal retransmission.
3723 */
3724 if (tcp_send_head(sk))
3725 tcp_ack_probe(sk);
3726
3727 if (tp->tlp_high_seq)
3728 tcp_process_tlp_ack(sk, ack, flag);
3729 return 1;
3730
3731invalid_ack:
3732 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3733 return -1;
3734
3735old_ack:
3736 /* If data was SACKed, tag it and see if we should send more data.
3737 * If data was DSACKed, see if we can undo a cwnd reduction.
3738 */
3739 if (TCP_SKB_CB(skb)->sacked) {
3740 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3741 &sack_state);
3742 tcp_fastretrans_alert(sk, acked, is_dupack, &flag, &rexmit);
3743 tcp_xmit_recovery(sk, rexmit);
3744 }
3745
3746 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3747 return 0;
3748}
3749
3750static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3751 bool syn, struct tcp_fastopen_cookie *foc,
3752 bool exp_opt)
3753{
3754 /* Valid only in SYN or SYN-ACK with an even length. */
3755 if (!foc || !syn || len < 0 || (len & 1))
3756 return;
3757
3758 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3759 len <= TCP_FASTOPEN_COOKIE_MAX)
3760 memcpy(foc->val, cookie, len);
3761 else if (len != 0)
3762 len = -1;
3763 foc->len = len;
3764 foc->exp = exp_opt;
3765}
3766
3767/* Look for tcp options. Normally only called on SYN and SYNACK packets.
3768 * But, this can also be called on packets in the established flow when
3769 * the fast version below fails.
3770 */
3771void tcp_parse_options(const struct sk_buff *skb,
3772 struct tcp_options_received *opt_rx, int estab,
3773 struct tcp_fastopen_cookie *foc)
3774{
3775 const unsigned char *ptr;
3776 const struct tcphdr *th = tcp_hdr(skb);
3777 int length = (th->doff * 4) - sizeof(struct tcphdr);
3778
3779 ptr = (const unsigned char *)(th + 1);
3780 opt_rx->saw_tstamp = 0;
3781
3782 while (length > 0) {
3783 int opcode = *ptr++;
3784 int opsize;
3785
3786 switch (opcode) {
3787 case TCPOPT_EOL:
3788 return;
3789 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3790 length--;
3791 continue;
3792 default:
3793 opsize = *ptr++;
3794 if (opsize < 2) /* "silly options" */
3795 return;
3796 if (opsize > length)
3797 return; /* don't parse partial options */
3798 switch (opcode) {
3799 case TCPOPT_MSS:
3800 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3801 u16 in_mss = get_unaligned_be16(ptr);
3802 if (in_mss) {
3803 if (opt_rx->user_mss &&
3804 opt_rx->user_mss < in_mss)
3805 in_mss = opt_rx->user_mss;
3806 opt_rx->mss_clamp = in_mss;
3807 }
3808 }
3809 break;
3810 case TCPOPT_WINDOW:
3811 if (opsize == TCPOLEN_WINDOW && th->syn &&
3812 !estab && sysctl_tcp_window_scaling) {
3813 __u8 snd_wscale = *(__u8 *)ptr;
3814 opt_rx->wscale_ok = 1;
3815 if (snd_wscale > 14) {
3816 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3817 __func__,
3818 snd_wscale);
3819 snd_wscale = 14;
3820 }
3821 opt_rx->snd_wscale = snd_wscale;
3822 }
3823 break;
3824 case TCPOPT_TIMESTAMP:
3825 if ((opsize == TCPOLEN_TIMESTAMP) &&
3826 ((estab && opt_rx->tstamp_ok) ||
3827 (!estab && sysctl_tcp_timestamps))) {
3828 opt_rx->saw_tstamp = 1;
3829 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3830 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3831 }
3832 break;
3833 case TCPOPT_SACK_PERM:
3834 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3835 !estab && sysctl_tcp_sack) {
3836 opt_rx->sack_ok = TCP_SACK_SEEN;
3837 tcp_sack_reset(opt_rx);
3838 }
3839 break;
3840
3841 case TCPOPT_SACK:
3842 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3843 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3844 opt_rx->sack_ok) {
3845 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3846 }
3847 break;
3848#ifdef CONFIG_TCP_MD5SIG
3849 case TCPOPT_MD5SIG:
3850 /*
3851 * The MD5 Hash has already been
3852 * checked (see tcp_v{4,6}_do_rcv()).
3853 */
3854 break;
3855#endif
3856 case TCPOPT_FASTOPEN:
3857 tcp_parse_fastopen_option(
3858 opsize - TCPOLEN_FASTOPEN_BASE,
3859 ptr, th->syn, foc, false);
3860 break;
3861
3862 case TCPOPT_EXP:
3863 /* Fast Open option shares code 254 using a
3864 * 16 bits magic number.
3865 */
3866 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
3867 get_unaligned_be16(ptr) ==
3868 TCPOPT_FASTOPEN_MAGIC)
3869 tcp_parse_fastopen_option(opsize -
3870 TCPOLEN_EXP_FASTOPEN_BASE,
3871 ptr + 2, th->syn, foc, true);
3872 break;
3873
3874 }
3875 ptr += opsize-2;
3876 length -= opsize;
3877 }
3878 }
3879}
3880EXPORT_SYMBOL(tcp_parse_options);
3881
3882static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3883{
3884 const __be32 *ptr = (const __be32 *)(th + 1);
3885
3886 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3887 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3888 tp->rx_opt.saw_tstamp = 1;
3889 ++ptr;
3890 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3891 ++ptr;
3892 if (*ptr)
3893 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3894 else
3895 tp->rx_opt.rcv_tsecr = 0;
3896 return true;
3897 }
3898 return false;
3899}
3900
3901/* Fast parse options. This hopes to only see timestamps.
3902 * If it is wrong it falls back on tcp_parse_options().
3903 */
3904static bool tcp_fast_parse_options(const struct sk_buff *skb,
3905 const struct tcphdr *th, struct tcp_sock *tp)
3906{
3907 /* In the spirit of fast parsing, compare doff directly to constant
3908 * values. Because equality is used, short doff can be ignored here.
3909 */
3910 if (th->doff == (sizeof(*th) / 4)) {
3911 tp->rx_opt.saw_tstamp = 0;
3912 return false;
3913 } else if (tp->rx_opt.tstamp_ok &&
3914 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3915 if (tcp_parse_aligned_timestamp(tp, th))
3916 return true;
3917 }
3918
3919 tcp_parse_options(skb, &tp->rx_opt, 1, NULL);
3920 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3921 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3922
3923 return true;
3924}
3925
3926#ifdef CONFIG_TCP_MD5SIG
3927/*
3928 * Parse MD5 Signature option
3929 */
3930const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3931{
3932 int length = (th->doff << 2) - sizeof(*th);
3933 const u8 *ptr = (const u8 *)(th + 1);
3934
3935 /* If the TCP option is too short, we can short cut */
3936 if (length < TCPOLEN_MD5SIG)
3937 return NULL;
3938
3939 while (length > 0) {
3940 int opcode = *ptr++;
3941 int opsize;
3942
3943 switch (opcode) {
3944 case TCPOPT_EOL:
3945 return NULL;
3946 case TCPOPT_NOP:
3947 length--;
3948 continue;
3949 default:
3950 opsize = *ptr++;
3951 if (opsize < 2 || opsize > length)
3952 return NULL;
3953 if (opcode == TCPOPT_MD5SIG)
3954 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3955 }
3956 ptr += opsize - 2;
3957 length -= opsize;
3958 }
3959 return NULL;
3960}
3961EXPORT_SYMBOL(tcp_parse_md5sig_option);
3962#endif
3963
3964/* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3965 *
3966 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3967 * it can pass through stack. So, the following predicate verifies that
3968 * this segment is not used for anything but congestion avoidance or
3969 * fast retransmit. Moreover, we even are able to eliminate most of such
3970 * second order effects, if we apply some small "replay" window (~RTO)
3971 * to timestamp space.
3972 *
3973 * All these measures still do not guarantee that we reject wrapped ACKs
3974 * on networks with high bandwidth, when sequence space is recycled fastly,
3975 * but it guarantees that such events will be very rare and do not affect
3976 * connection seriously. This doesn't look nice, but alas, PAWS is really
3977 * buggy extension.
3978 *
3979 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3980 * states that events when retransmit arrives after original data are rare.
3981 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3982 * the biggest problem on large power networks even with minor reordering.
3983 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3984 * up to bandwidth of 18Gigabit/sec. 8) ]
3985 */
3986
3987static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3988{
3989 const struct tcp_sock *tp = tcp_sk(sk);
3990 const struct tcphdr *th = tcp_hdr(skb);
3991 u32 seq = TCP_SKB_CB(skb)->seq;
3992 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3993
3994 return (/* 1. Pure ACK with correct sequence number. */
3995 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3996
3997 /* 2. ... and duplicate ACK. */
3998 ack == tp->snd_una &&
3999
4000 /* 3. ... and does not update window. */
4001 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4002
4003 /* 4. ... and sits in replay window. */
4004 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
4005}
4006
4007static inline bool tcp_paws_discard(const struct sock *sk,
4008 const struct sk_buff *skb)
4009{
4010 const struct tcp_sock *tp = tcp_sk(sk);
4011
4012 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4013 !tcp_disordered_ack(sk, skb);
4014}
4015
4016/* Check segment sequence number for validity.
4017 *
4018 * Segment controls are considered valid, if the segment
4019 * fits to the window after truncation to the window. Acceptability
4020 * of data (and SYN, FIN, of course) is checked separately.
4021 * See tcp_data_queue(), for example.
4022 *
4023 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4024 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4025 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4026 * (borrowed from freebsd)
4027 */
4028
4029static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
4030{
4031 return !before(end_seq, tp->rcv_wup) &&
4032 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4033}
4034
4035/* When we get a reset we do this. */
4036void tcp_reset(struct sock *sk)
4037{
4038 /* We want the right error as BSD sees it (and indeed as we do). */
4039 switch (sk->sk_state) {
4040 case TCP_SYN_SENT:
4041 sk->sk_err = ECONNREFUSED;
4042 break;
4043 case TCP_CLOSE_WAIT:
4044 sk->sk_err = EPIPE;
4045 break;
4046 case TCP_CLOSE:
4047 return;
4048 default:
4049 sk->sk_err = ECONNRESET;
4050 }
4051 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4052 smp_wmb();
4053
4054 if (!sock_flag(sk, SOCK_DEAD))
4055 sk->sk_error_report(sk);
4056
4057 tcp_done(sk);
4058}
4059
4060/*
4061 * Process the FIN bit. This now behaves as it is supposed to work
4062 * and the FIN takes effect when it is validly part of sequence
4063 * space. Not before when we get holes.
4064 *
4065 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4066 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4067 * TIME-WAIT)
4068 *
4069 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4070 * close and we go into CLOSING (and later onto TIME-WAIT)
4071 *
4072 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4073 */
4074void tcp_fin(struct sock *sk)
4075{
4076 struct tcp_sock *tp = tcp_sk(sk);
4077
4078 inet_csk_schedule_ack(sk);
4079
4080 sk->sk_shutdown |= RCV_SHUTDOWN;
4081 sock_set_flag(sk, SOCK_DONE);
4082
4083 switch (sk->sk_state) {
4084 case TCP_SYN_RECV:
4085 case TCP_ESTABLISHED:
4086 /* Move to CLOSE_WAIT */
4087 tcp_set_state(sk, TCP_CLOSE_WAIT);
4088 inet_csk(sk)->icsk_ack.pingpong = 1;
4089 break;
4090
4091 case TCP_CLOSE_WAIT:
4092 case TCP_CLOSING:
4093 /* Received a retransmission of the FIN, do
4094 * nothing.
4095 */
4096 break;
4097 case TCP_LAST_ACK:
4098 /* RFC793: Remain in the LAST-ACK state. */
4099 break;
4100
4101 case TCP_FIN_WAIT1:
4102 /* This case occurs when a simultaneous close
4103 * happens, we must ack the received FIN and
4104 * enter the CLOSING state.
4105 */
4106 tcp_send_ack(sk);
4107 tcp_set_state(sk, TCP_CLOSING);
4108 break;
4109 case TCP_FIN_WAIT2:
4110 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4111 tcp_send_ack(sk);
4112 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4113 break;
4114 default:
4115 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4116 * cases we should never reach this piece of code.
4117 */
4118 pr_err("%s: Impossible, sk->sk_state=%d\n",
4119 __func__, sk->sk_state);
4120 break;
4121 }
4122
4123 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4124 * Probably, we should reset in this case. For now drop them.
4125 */
4126 skb_rbtree_purge(&tp->out_of_order_queue);
4127 if (tcp_is_sack(tp))
4128 tcp_sack_reset(&tp->rx_opt);
4129 sk_mem_reclaim(sk);
4130
4131 if (!sock_flag(sk, SOCK_DEAD)) {
4132 sk->sk_state_change(sk);
4133
4134 /* Do not send POLL_HUP for half duplex close. */
4135 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4136 sk->sk_state == TCP_CLOSE)
4137 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4138 else
4139 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4140 }
4141}
4142
4143static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4144 u32 end_seq)
4145{
4146 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4147 if (before(seq, sp->start_seq))
4148 sp->start_seq = seq;
4149 if (after(end_seq, sp->end_seq))
4150 sp->end_seq = end_seq;
4151 return true;
4152 }
4153 return false;
4154}
4155
4156static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4157{
4158 struct tcp_sock *tp = tcp_sk(sk);
4159
4160 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4161 int mib_idx;
4162
4163 if (before(seq, tp->rcv_nxt))
4164 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4165 else
4166 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4167
4168 NET_INC_STATS(sock_net(sk), mib_idx);
4169
4170 tp->rx_opt.dsack = 1;
4171 tp->duplicate_sack[0].start_seq = seq;
4172 tp->duplicate_sack[0].end_seq = end_seq;
4173 }
4174}
4175
4176static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4177{
4178 struct tcp_sock *tp = tcp_sk(sk);
4179
4180 if (!tp->rx_opt.dsack)
4181 tcp_dsack_set(sk, seq, end_seq);
4182 else
4183 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4184}
4185
4186static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4187{
4188 struct tcp_sock *tp = tcp_sk(sk);
4189
4190 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4191 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4192 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4193 tcp_enter_quickack_mode(sk);
4194
4195 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4196 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4197
4198 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4199 end_seq = tp->rcv_nxt;
4200 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4201 }
4202 }
4203
4204 tcp_send_ack(sk);
4205}
4206
4207/* These routines update the SACK block as out-of-order packets arrive or
4208 * in-order packets close up the sequence space.
4209 */
4210static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4211{
4212 int this_sack;
4213 struct tcp_sack_block *sp = &tp->selective_acks[0];
4214 struct tcp_sack_block *swalk = sp + 1;
4215
4216 /* See if the recent change to the first SACK eats into
4217 * or hits the sequence space of other SACK blocks, if so coalesce.
4218 */
4219 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4220 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4221 int i;
4222
4223 /* Zap SWALK, by moving every further SACK up by one slot.
4224 * Decrease num_sacks.
4225 */
4226 tp->rx_opt.num_sacks--;
4227 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4228 sp[i] = sp[i + 1];
4229 continue;
4230 }
4231 this_sack++, swalk++;
4232 }
4233}
4234
4235static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4236{
4237 struct tcp_sock *tp = tcp_sk(sk);
4238 struct tcp_sack_block *sp = &tp->selective_acks[0];
4239 int cur_sacks = tp->rx_opt.num_sacks;
4240 int this_sack;
4241
4242 if (!cur_sacks)
4243 goto new_sack;
4244
4245 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4246 if (tcp_sack_extend(sp, seq, end_seq)) {
4247 /* Rotate this_sack to the first one. */
4248 for (; this_sack > 0; this_sack--, sp--)
4249 swap(*sp, *(sp - 1));
4250 if (cur_sacks > 1)
4251 tcp_sack_maybe_coalesce(tp);
4252 return;
4253 }
4254 }
4255
4256 /* Could not find an adjacent existing SACK, build a new one,
4257 * put it at the front, and shift everyone else down. We
4258 * always know there is at least one SACK present already here.
4259 *
4260 * If the sack array is full, forget about the last one.
4261 */
4262 if (this_sack >= TCP_NUM_SACKS) {
4263 this_sack--;
4264 tp->rx_opt.num_sacks--;
4265 sp--;
4266 }
4267 for (; this_sack > 0; this_sack--, sp--)
4268 *sp = *(sp - 1);
4269
4270new_sack:
4271 /* Build the new head SACK, and we're done. */
4272 sp->start_seq = seq;
4273 sp->end_seq = end_seq;
4274 tp->rx_opt.num_sacks++;
4275}
4276
4277/* RCV.NXT advances, some SACKs should be eaten. */
4278
4279static void tcp_sack_remove(struct tcp_sock *tp)
4280{
4281 struct tcp_sack_block *sp = &tp->selective_acks[0];
4282 int num_sacks = tp->rx_opt.num_sacks;
4283 int this_sack;
4284
4285 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4286 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4287 tp->rx_opt.num_sacks = 0;
4288 return;
4289 }
4290
4291 for (this_sack = 0; this_sack < num_sacks;) {
4292 /* Check if the start of the sack is covered by RCV.NXT. */
4293 if (!before(tp->rcv_nxt, sp->start_seq)) {
4294 int i;
4295
4296 /* RCV.NXT must cover all the block! */
4297 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4298
4299 /* Zap this SACK, by moving forward any other SACKS. */
4300 for (i = this_sack+1; i < num_sacks; i++)
4301 tp->selective_acks[i-1] = tp->selective_acks[i];
4302 num_sacks--;
4303 continue;
4304 }
4305 this_sack++;
4306 sp++;
4307 }
4308 tp->rx_opt.num_sacks = num_sacks;
4309}
4310
4311/**
4312 * tcp_try_coalesce - try to merge skb to prior one
4313 * @sk: socket
4314 * @to: prior buffer
4315 * @from: buffer to add in queue
4316 * @fragstolen: pointer to boolean
4317 *
4318 * Before queueing skb @from after @to, try to merge them
4319 * to reduce overall memory use and queue lengths, if cost is small.
4320 * Packets in ofo or receive queues can stay a long time.
4321 * Better try to coalesce them right now to avoid future collapses.
4322 * Returns true if caller should free @from instead of queueing it
4323 */
4324static bool tcp_try_coalesce(struct sock *sk,
4325 struct sk_buff *to,
4326 struct sk_buff *from,
4327 bool *fragstolen)
4328{
4329 int delta;
4330
4331 *fragstolen = false;
4332
4333 /* Its possible this segment overlaps with prior segment in queue */
4334 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4335 return false;
4336
4337 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4338 return false;
4339
4340 atomic_add(delta, &sk->sk_rmem_alloc);
4341 sk_mem_charge(sk, delta);
4342 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4343 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4344 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4345 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4346 return true;
4347}
4348
4349static void tcp_drop(struct sock *sk, struct sk_buff *skb)
4350{
4351 sk_drops_add(sk, skb);
4352 __kfree_skb(skb);
4353}
4354
4355/* This one checks to see if we can put data from the
4356 * out_of_order queue into the receive_queue.
4357 */
4358static void tcp_ofo_queue(struct sock *sk)
4359{
4360 struct tcp_sock *tp = tcp_sk(sk);
4361 __u32 dsack_high = tp->rcv_nxt;
4362 bool fin, fragstolen, eaten;
4363 struct sk_buff *skb, *tail;
4364 struct rb_node *p;
4365
4366 p = rb_first(&tp->out_of_order_queue);
4367 while (p) {
4368 skb = rb_entry(p, struct sk_buff, rbnode);
4369 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4370 break;
4371
4372 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4373 __u32 dsack = dsack_high;
4374 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4375 dsack_high = TCP_SKB_CB(skb)->end_seq;
4376 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4377 }
4378 p = rb_next(p);
4379 rb_erase(&skb->rbnode, &tp->out_of_order_queue);
4380
4381 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
4382 SOCK_DEBUG(sk, "ofo packet was already received\n");
4383 tcp_drop(sk, skb);
4384 continue;
4385 }
4386 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4387 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4388 TCP_SKB_CB(skb)->end_seq);
4389
4390 tail = skb_peek_tail(&sk->sk_receive_queue);
4391 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4392 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4393 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
4394 if (!eaten)
4395 __skb_queue_tail(&sk->sk_receive_queue, skb);
4396 else
4397 kfree_skb_partial(skb, fragstolen);
4398
4399 if (unlikely(fin)) {
4400 tcp_fin(sk);
4401 /* tcp_fin() purges tp->out_of_order_queue,
4402 * so we must end this loop right now.
4403 */
4404 break;
4405 }
4406 }
4407}
4408
4409static bool tcp_prune_ofo_queue(struct sock *sk);
4410static int tcp_prune_queue(struct sock *sk);
4411
4412static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4413 unsigned int size)
4414{
4415 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4416 !sk_rmem_schedule(sk, skb, size)) {
4417
4418 if (tcp_prune_queue(sk) < 0)
4419 return -1;
4420
4421 while (!sk_rmem_schedule(sk, skb, size)) {
4422 if (!tcp_prune_ofo_queue(sk))
4423 return -1;
4424 }
4425 }
4426 return 0;
4427}
4428
4429static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4430{
4431 struct tcp_sock *tp = tcp_sk(sk);
4432 struct rb_node **p, *q, *parent;
4433 struct sk_buff *skb1;
4434 u32 seq, end_seq;
4435 bool fragstolen;
4436
4437 tcp_ecn_check_ce(tp, skb);
4438
4439 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4440 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
4441 tcp_drop(sk, skb);
4442 return;
4443 }
4444
4445 /* Disable header prediction. */
4446 tp->pred_flags = 0;
4447 inet_csk_schedule_ack(sk);
4448
4449 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4450 seq = TCP_SKB_CB(skb)->seq;
4451 end_seq = TCP_SKB_CB(skb)->end_seq;
4452 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4453 tp->rcv_nxt, seq, end_seq);
4454
4455 p = &tp->out_of_order_queue.rb_node;
4456 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4457 /* Initial out of order segment, build 1 SACK. */
4458 if (tcp_is_sack(tp)) {
4459 tp->rx_opt.num_sacks = 1;
4460 tp->selective_acks[0].start_seq = seq;
4461 tp->selective_acks[0].end_seq = end_seq;
4462 }
4463 rb_link_node(&skb->rbnode, NULL, p);
4464 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4465 tp->ooo_last_skb = skb;
4466 goto end;
4467 }
4468
4469 /* In the typical case, we are adding an skb to the end of the list.
4470 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4471 */
4472 if (tcp_try_coalesce(sk, tp->ooo_last_skb, skb, &fragstolen)) {
4473coalesce_done:
4474 tcp_grow_window(sk, skb);
4475 kfree_skb_partial(skb, fragstolen);
4476 skb = NULL;
4477 goto add_sack;
4478 }
4479 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4480 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
4481 parent = &tp->ooo_last_skb->rbnode;
4482 p = &parent->rb_right;
4483 goto insert;
4484 }
4485
4486 /* Find place to insert this segment. Handle overlaps on the way. */
4487 parent = NULL;
4488 while (*p) {
4489 parent = *p;
4490 skb1 = rb_entry(parent, struct sk_buff, rbnode);
4491 if (before(seq, TCP_SKB_CB(skb1)->seq)) {
4492 p = &parent->rb_left;
4493 continue;
4494 }
4495 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4496 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4497 /* All the bits are present. Drop. */
4498 NET_INC_STATS(sock_net(sk),
4499 LINUX_MIB_TCPOFOMERGE);
4500 __kfree_skb(skb);
4501 skb = NULL;
4502 tcp_dsack_set(sk, seq, end_seq);
4503 goto add_sack;
4504 }
4505 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4506 /* Partial overlap. */
4507 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
4508 } else {
4509 /* skb's seq == skb1's seq and skb covers skb1.
4510 * Replace skb1 with skb.
4511 */
4512 rb_replace_node(&skb1->rbnode, &skb->rbnode,
4513 &tp->out_of_order_queue);
4514 tcp_dsack_extend(sk,
4515 TCP_SKB_CB(skb1)->seq,
4516 TCP_SKB_CB(skb1)->end_seq);
4517 NET_INC_STATS(sock_net(sk),
4518 LINUX_MIB_TCPOFOMERGE);
4519 __kfree_skb(skb1);
4520 goto merge_right;
4521 }
4522 } else if (tcp_try_coalesce(sk, skb1, skb, &fragstolen)) {
4523 goto coalesce_done;
4524 }
4525 p = &parent->rb_right;
4526 }
4527insert:
4528 /* Insert segment into RB tree. */
4529 rb_link_node(&skb->rbnode, parent, p);
4530 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4531
4532merge_right:
4533 /* Remove other segments covered by skb. */
4534 while ((q = rb_next(&skb->rbnode)) != NULL) {
4535 skb1 = rb_entry(q, struct sk_buff, rbnode);
4536
4537 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4538 break;
4539 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4540 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4541 end_seq);
4542 break;
4543 }
4544 rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
4545 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4546 TCP_SKB_CB(skb1)->end_seq);
4547 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4548 tcp_drop(sk, skb1);
4549 }
4550 /* If there is no skb after us, we are the last_skb ! */
4551 if (!q)
4552 tp->ooo_last_skb = skb;
4553
4554add_sack:
4555 if (tcp_is_sack(tp))
4556 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4557end:
4558 if (skb) {
4559 tcp_grow_window(sk, skb);
4560 skb_set_owner_r(skb, sk);
4561 }
4562}
4563
4564static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4565 bool *fragstolen)
4566{
4567 int eaten;
4568 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4569
4570 __skb_pull(skb, hdrlen);
4571 eaten = (tail &&
4572 tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0;
4573 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4574 if (!eaten) {
4575 __skb_queue_tail(&sk->sk_receive_queue, skb);
4576 skb_set_owner_r(skb, sk);
4577 }
4578 return eaten;
4579}
4580
4581int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4582{
4583 struct sk_buff *skb;
4584 int err = -ENOMEM;
4585 int data_len = 0;
4586 bool fragstolen;
4587
4588 if (size == 0)
4589 return 0;
4590
4591 if (size > PAGE_SIZE) {
4592 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4593
4594 data_len = npages << PAGE_SHIFT;
4595 size = data_len + (size & ~PAGE_MASK);
4596 }
4597 skb = alloc_skb_with_frags(size - data_len, data_len,
4598 PAGE_ALLOC_COSTLY_ORDER,
4599 &err, sk->sk_allocation);
4600 if (!skb)
4601 goto err;
4602
4603 skb_put(skb, size - data_len);
4604 skb->data_len = data_len;
4605 skb->len = size;
4606
4607 if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4608 goto err_free;
4609
4610 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
4611 if (err)
4612 goto err_free;
4613
4614 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4615 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4616 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4617
4618 if (tcp_queue_rcv(sk, skb, 0, &fragstolen)) {
4619 WARN_ON_ONCE(fragstolen); /* should not happen */
4620 __kfree_skb(skb);
4621 }
4622 return size;
4623
4624err_free:
4625 kfree_skb(skb);
4626err:
4627 return err;
4628
4629}
4630
4631static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4632{
4633 struct tcp_sock *tp = tcp_sk(sk);
4634 bool fragstolen = false;
4635 int eaten = -1;
4636
4637 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
4638 __kfree_skb(skb);
4639 return;
4640 }
4641 skb_dst_drop(skb);
4642 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
4643
4644 tcp_ecn_accept_cwr(tp, skb);
4645
4646 tp->rx_opt.dsack = 0;
4647
4648 /* Queue data for delivery to the user.
4649 * Packets in sequence go to the receive queue.
4650 * Out of sequence packets to the out_of_order_queue.
4651 */
4652 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4653 if (tcp_receive_window(tp) == 0)
4654 goto out_of_window;
4655
4656 /* Ok. In sequence. In window. */
4657 if (tp->ucopy.task == current &&
4658 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4659 sock_owned_by_user(sk) && !tp->urg_data) {
4660 int chunk = min_t(unsigned int, skb->len,
4661 tp->ucopy.len);
4662
4663 __set_current_state(TASK_RUNNING);
4664
4665 if (!skb_copy_datagram_msg(skb, 0, tp->ucopy.msg, chunk)) {
4666 tp->ucopy.len -= chunk;
4667 tp->copied_seq += chunk;
4668 eaten = (chunk == skb->len);
4669 tcp_rcv_space_adjust(sk);
4670 }
4671 }
4672
4673 if (eaten <= 0) {
4674queue_and_out:
4675 if (eaten < 0) {
4676 if (skb_queue_len(&sk->sk_receive_queue) == 0)
4677 sk_forced_mem_schedule(sk, skb->truesize);
4678 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4679 goto drop;
4680 }
4681 eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4682 }
4683 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4684 if (skb->len)
4685 tcp_event_data_recv(sk, skb);
4686 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4687 tcp_fin(sk);
4688
4689 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4690 tcp_ofo_queue(sk);
4691
4692 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4693 * gap in queue is filled.
4694 */
4695 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
4696 inet_csk(sk)->icsk_ack.pingpong = 0;
4697 }
4698
4699 if (tp->rx_opt.num_sacks)
4700 tcp_sack_remove(tp);
4701
4702 tcp_fast_path_check(sk);
4703
4704 if (eaten > 0)
4705 kfree_skb_partial(skb, fragstolen);
4706 if (!sock_flag(sk, SOCK_DEAD))
4707 sk->sk_data_ready(sk);
4708 return;
4709 }
4710
4711 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4712 /* A retransmit, 2nd most common case. Force an immediate ack. */
4713 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4714 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4715
4716out_of_window:
4717 tcp_enter_quickack_mode(sk);
4718 inet_csk_schedule_ack(sk);
4719drop:
4720 tcp_drop(sk, skb);
4721 return;
4722 }
4723
4724 /* Out of window. F.e. zero window probe. */
4725 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4726 goto out_of_window;
4727
4728 tcp_enter_quickack_mode(sk);
4729
4730 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4731 /* Partial packet, seq < rcv_next < end_seq */
4732 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4733 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4734 TCP_SKB_CB(skb)->end_seq);
4735
4736 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4737
4738 /* If window is closed, drop tail of packet. But after
4739 * remembering D-SACK for its head made in previous line.
4740 */
4741 if (!tcp_receive_window(tp))
4742 goto out_of_window;
4743 goto queue_and_out;
4744 }
4745
4746 tcp_data_queue_ofo(sk, skb);
4747}
4748
4749static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
4750{
4751 if (list)
4752 return !skb_queue_is_last(list, skb) ? skb->next : NULL;
4753
4754 return rb_entry_safe(rb_next(&skb->rbnode), struct sk_buff, rbnode);
4755}
4756
4757static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4758 struct sk_buff_head *list,
4759 struct rb_root *root)
4760{
4761 struct sk_buff *next = tcp_skb_next(skb, list);
4762
4763 if (list)
4764 __skb_unlink(skb, list);
4765 else
4766 rb_erase(&skb->rbnode, root);
4767
4768 __kfree_skb(skb);
4769 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4770
4771 return next;
4772}
4773
4774/* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
4775static void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
4776{
4777 struct rb_node **p = &root->rb_node;
4778 struct rb_node *parent = NULL;
4779 struct sk_buff *skb1;
4780
4781 while (*p) {
4782 parent = *p;
4783 skb1 = rb_entry(parent, struct sk_buff, rbnode);
4784 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
4785 p = &parent->rb_left;
4786 else
4787 p = &parent->rb_right;
4788 }
4789 rb_link_node(&skb->rbnode, parent, p);
4790 rb_insert_color(&skb->rbnode, root);
4791}
4792
4793/* Collapse contiguous sequence of skbs head..tail with
4794 * sequence numbers start..end.
4795 *
4796 * If tail is NULL, this means until the end of the queue.
4797 *
4798 * Segments with FIN/SYN are not collapsed (only because this
4799 * simplifies code)
4800 */
4801static void
4802tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
4803 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
4804{
4805 struct sk_buff *skb = head, *n;
4806 struct sk_buff_head tmp;
4807 bool end_of_skbs;
4808
4809 /* First, check that queue is collapsible and find
4810 * the point where collapsing can be useful.
4811 */
4812restart:
4813 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
4814 n = tcp_skb_next(skb, list);
4815
4816 /* No new bits? It is possible on ofo queue. */
4817 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4818 skb = tcp_collapse_one(sk, skb, list, root);
4819 if (!skb)
4820 break;
4821 goto restart;
4822 }
4823
4824 /* The first skb to collapse is:
4825 * - not SYN/FIN and
4826 * - bloated or contains data before "start" or
4827 * overlaps to the next one.
4828 */
4829 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
4830 (tcp_win_from_space(skb->truesize) > skb->len ||
4831 before(TCP_SKB_CB(skb)->seq, start))) {
4832 end_of_skbs = false;
4833 break;
4834 }
4835
4836 if (n && n != tail &&
4837 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
4838 end_of_skbs = false;
4839 break;
4840 }
4841
4842 /* Decided to skip this, advance start seq. */
4843 start = TCP_SKB_CB(skb)->end_seq;
4844 }
4845 if (end_of_skbs ||
4846 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4847 return;
4848
4849 __skb_queue_head_init(&tmp);
4850
4851 while (before(start, end)) {
4852 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
4853 struct sk_buff *nskb;
4854
4855 nskb = alloc_skb(copy, GFP_ATOMIC);
4856 if (!nskb)
4857 break;
4858
4859 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4860 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4861 if (list)
4862 __skb_queue_before(list, skb, nskb);
4863 else
4864 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
4865 skb_set_owner_r(nskb, sk);
4866
4867 /* Copy data, releasing collapsed skbs. */
4868 while (copy > 0) {
4869 int offset = start - TCP_SKB_CB(skb)->seq;
4870 int size = TCP_SKB_CB(skb)->end_seq - start;
4871
4872 BUG_ON(offset < 0);
4873 if (size > 0) {
4874 size = min(copy, size);
4875 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4876 BUG();
4877 TCP_SKB_CB(nskb)->end_seq += size;
4878 copy -= size;
4879 start += size;
4880 }
4881 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4882 skb = tcp_collapse_one(sk, skb, list, root);
4883 if (!skb ||
4884 skb == tail ||
4885 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4886 goto end;
4887 }
4888 }
4889 }
4890end:
4891 skb_queue_walk_safe(&tmp, skb, n)
4892 tcp_rbtree_insert(root, skb);
4893}
4894
4895/* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4896 * and tcp_collapse() them until all the queue is collapsed.
4897 */
4898static void tcp_collapse_ofo_queue(struct sock *sk)
4899{
4900 struct tcp_sock *tp = tcp_sk(sk);
4901 struct sk_buff *skb, *head;
4902 struct rb_node *p;
4903 u32 start, end;
4904
4905 p = rb_first(&tp->out_of_order_queue);
4906 skb = rb_entry_safe(p, struct sk_buff, rbnode);
4907new_range:
4908 if (!skb) {
4909 p = rb_last(&tp->out_of_order_queue);
4910 /* Note: This is possible p is NULL here. We do not
4911 * use rb_entry_safe(), as ooo_last_skb is valid only
4912 * if rbtree is not empty.
4913 */
4914 tp->ooo_last_skb = rb_entry(p, struct sk_buff, rbnode);
4915 return;
4916 }
4917 start = TCP_SKB_CB(skb)->seq;
4918 end = TCP_SKB_CB(skb)->end_seq;
4919
4920 for (head = skb;;) {
4921 skb = tcp_skb_next(skb, NULL);
4922
4923 /* Range is terminated when we see a gap or when
4924 * we are at the queue end.
4925 */
4926 if (!skb ||
4927 after(TCP_SKB_CB(skb)->seq, end) ||
4928 before(TCP_SKB_CB(skb)->end_seq, start)) {
4929 tcp_collapse(sk, NULL, &tp->out_of_order_queue,
4930 head, skb, start, end);
4931 goto new_range;
4932 }
4933
4934 if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
4935 start = TCP_SKB_CB(skb)->seq;
4936 if (after(TCP_SKB_CB(skb)->end_seq, end))
4937 end = TCP_SKB_CB(skb)->end_seq;
4938 }
4939}
4940
4941/*
4942 * Clean the out-of-order queue to make room.
4943 * We drop high sequences packets to :
4944 * 1) Let a chance for holes to be filled.
4945 * 2) not add too big latencies if thousands of packets sit there.
4946 * (But if application shrinks SO_RCVBUF, we could still end up
4947 * freeing whole queue here)
4948 *
4949 * Return true if queue has shrunk.
4950 */
4951static bool tcp_prune_ofo_queue(struct sock *sk)
4952{
4953 struct tcp_sock *tp = tcp_sk(sk);
4954 struct rb_node *node, *prev;
4955
4956 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
4957 return false;
4958
4959 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
4960 node = &tp->ooo_last_skb->rbnode;
4961 do {
4962 prev = rb_prev(node);
4963 rb_erase(node, &tp->out_of_order_queue);
4964 tcp_drop(sk, rb_entry(node, struct sk_buff, rbnode));
4965 sk_mem_reclaim(sk);
4966 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
4967 !tcp_under_memory_pressure(sk))
4968 break;
4969 node = prev;
4970 } while (node);
4971 tp->ooo_last_skb = rb_entry(prev, struct sk_buff, rbnode);
4972
4973 /* Reset SACK state. A conforming SACK implementation will
4974 * do the same at a timeout based retransmit. When a connection
4975 * is in a sad state like this, we care only about integrity
4976 * of the connection not performance.
4977 */
4978 if (tp->rx_opt.sack_ok)
4979 tcp_sack_reset(&tp->rx_opt);
4980 return true;
4981}
4982
4983/* Reduce allocated memory if we can, trying to get
4984 * the socket within its memory limits again.
4985 *
4986 * Return less than zero if we should start dropping frames
4987 * until the socket owning process reads some of the data
4988 * to stabilize the situation.
4989 */
4990static int tcp_prune_queue(struct sock *sk)
4991{
4992 struct tcp_sock *tp = tcp_sk(sk);
4993
4994 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4995
4996 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
4997
4998 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4999 tcp_clamp_window(sk);
5000 else if (tcp_under_memory_pressure(sk))
5001 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
5002
5003 tcp_collapse_ofo_queue(sk);
5004 if (!skb_queue_empty(&sk->sk_receive_queue))
5005 tcp_collapse(sk, &sk->sk_receive_queue, NULL,
5006 skb_peek(&sk->sk_receive_queue),
5007 NULL,
5008 tp->copied_seq, tp->rcv_nxt);
5009 sk_mem_reclaim(sk);
5010
5011 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5012 return 0;
5013
5014 /* Collapsing did not help, destructive actions follow.
5015 * This must not ever occur. */
5016
5017 tcp_prune_ofo_queue(sk);
5018
5019 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5020 return 0;
5021
5022 /* If we are really being abused, tell the caller to silently
5023 * drop receive data on the floor. It will get retransmitted
5024 * and hopefully then we'll have sufficient space.
5025 */
5026 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
5027
5028 /* Massive buffer overcommit. */
5029 tp->pred_flags = 0;
5030 return -1;
5031}
5032
5033static bool tcp_should_expand_sndbuf(const struct sock *sk)
5034{
5035 const struct tcp_sock *tp = tcp_sk(sk);
5036
5037 /* If the user specified a specific send buffer setting, do
5038 * not modify it.
5039 */
5040 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
5041 return false;
5042
5043 /* If we are under global TCP memory pressure, do not expand. */
5044 if (tcp_under_memory_pressure(sk))
5045 return false;
5046
5047 /* If we are under soft global TCP memory pressure, do not expand. */
5048 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
5049 return false;
5050
5051 /* If we filled the congestion window, do not expand. */
5052 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
5053 return false;
5054
5055 return true;
5056}
5057
5058/* When incoming ACK allowed to free some skb from write_queue,
5059 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
5060 * on the exit from tcp input handler.
5061 *
5062 * PROBLEM: sndbuf expansion does not work well with largesend.
5063 */
5064static void tcp_new_space(struct sock *sk)
5065{
5066 struct tcp_sock *tp = tcp_sk(sk);
5067
5068 if (tcp_should_expand_sndbuf(sk)) {
5069 tcp_sndbuf_expand(sk);
5070 tp->snd_cwnd_stamp = tcp_time_stamp;
5071 }
5072
5073 sk->sk_write_space(sk);
5074}
5075
5076static void tcp_check_space(struct sock *sk)
5077{
5078 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
5079 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
5080 /* pairs with tcp_poll() */
5081 smp_mb();
5082 if (sk->sk_socket &&
5083 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
5084 tcp_new_space(sk);
5085 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5086 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
5087 }
5088 }
5089}
5090
5091static inline void tcp_data_snd_check(struct sock *sk)
5092{
5093 tcp_push_pending_frames(sk);
5094 tcp_check_space(sk);
5095}
5096
5097/*
5098 * Check if sending an ack is needed.
5099 */
5100static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5101{
5102 struct tcp_sock *tp = tcp_sk(sk);
5103
5104 /* More than one full frame received... */
5105 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5106 /* ... and right edge of window advances far enough.
5107 * (tcp_recvmsg() will send ACK otherwise). Or...
5108 */
5109 __tcp_select_window(sk) >= tp->rcv_wnd) ||
5110 /* We ACK each frame or... */
5111 tcp_in_quickack_mode(sk) ||
5112 /* We have out of order data. */
5113 (ofo_possible && !RB_EMPTY_ROOT(&tp->out_of_order_queue))) {
5114 /* Then ack it now */
5115 tcp_send_ack(sk);
5116 } else {
5117 /* Else, send delayed ack. */
5118 tcp_send_delayed_ack(sk);
5119 }
5120}
5121
5122static inline void tcp_ack_snd_check(struct sock *sk)
5123{
5124 if (!inet_csk_ack_scheduled(sk)) {
5125 /* We sent a data segment already. */
5126 return;
5127 }
5128 __tcp_ack_snd_check(sk, 1);
5129}
5130
5131/*
5132 * This routine is only called when we have urgent data
5133 * signaled. Its the 'slow' part of tcp_urg. It could be
5134 * moved inline now as tcp_urg is only called from one
5135 * place. We handle URGent data wrong. We have to - as
5136 * BSD still doesn't use the correction from RFC961.
5137 * For 1003.1g we should support a new option TCP_STDURG to permit
5138 * either form (or just set the sysctl tcp_stdurg).
5139 */
5140
5141static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5142{
5143 struct tcp_sock *tp = tcp_sk(sk);
5144 u32 ptr = ntohs(th->urg_ptr);
5145
5146 if (ptr && !sysctl_tcp_stdurg)
5147 ptr--;
5148 ptr += ntohl(th->seq);
5149
5150 /* Ignore urgent data that we've already seen and read. */
5151 if (after(tp->copied_seq, ptr))
5152 return;
5153
5154 /* Do not replay urg ptr.
5155 *
5156 * NOTE: interesting situation not covered by specs.
5157 * Misbehaving sender may send urg ptr, pointing to segment,
5158 * which we already have in ofo queue. We are not able to fetch
5159 * such data and will stay in TCP_URG_NOTYET until will be eaten
5160 * by recvmsg(). Seems, we are not obliged to handle such wicked
5161 * situations. But it is worth to think about possibility of some
5162 * DoSes using some hypothetical application level deadlock.
5163 */
5164 if (before(ptr, tp->rcv_nxt))
5165 return;
5166
5167 /* Do we already have a newer (or duplicate) urgent pointer? */
5168 if (tp->urg_data && !after(ptr, tp->urg_seq))
5169 return;
5170
5171 /* Tell the world about our new urgent pointer. */
5172 sk_send_sigurg(sk);
5173
5174 /* We may be adding urgent data when the last byte read was
5175 * urgent. To do this requires some care. We cannot just ignore
5176 * tp->copied_seq since we would read the last urgent byte again
5177 * as data, nor can we alter copied_seq until this data arrives
5178 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5179 *
5180 * NOTE. Double Dutch. Rendering to plain English: author of comment
5181 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5182 * and expect that both A and B disappear from stream. This is _wrong_.
5183 * Though this happens in BSD with high probability, this is occasional.
5184 * Any application relying on this is buggy. Note also, that fix "works"
5185 * only in this artificial test. Insert some normal data between A and B and we will
5186 * decline of BSD again. Verdict: it is better to remove to trap
5187 * buggy users.
5188 */
5189 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5190 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5191 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5192 tp->copied_seq++;
5193 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5194 __skb_unlink(skb, &sk->sk_receive_queue);
5195 __kfree_skb(skb);
5196 }
5197 }
5198
5199 tp->urg_data = TCP_URG_NOTYET;
5200 tp->urg_seq = ptr;
5201
5202 /* Disable header prediction. */
5203 tp->pred_flags = 0;
5204}
5205
5206/* This is the 'fast' part of urgent handling. */
5207static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5208{
5209 struct tcp_sock *tp = tcp_sk(sk);
5210
5211 /* Check if we get a new urgent pointer - normally not. */
5212 if (th->urg)
5213 tcp_check_urg(sk, th);
5214
5215 /* Do we wait for any urgent data? - normally not... */
5216 if (tp->urg_data == TCP_URG_NOTYET) {
5217 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5218 th->syn;
5219
5220 /* Is the urgent pointer pointing into this packet? */
5221 if (ptr < skb->len) {
5222 u8 tmp;
5223 if (skb_copy_bits(skb, ptr, &tmp, 1))
5224 BUG();
5225 tp->urg_data = TCP_URG_VALID | tmp;
5226 if (!sock_flag(sk, SOCK_DEAD))
5227 sk->sk_data_ready(sk);
5228 }
5229 }
5230}
5231
5232static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
5233{
5234 struct tcp_sock *tp = tcp_sk(sk);
5235 int chunk = skb->len - hlen;
5236 int err;
5237
5238 if (skb_csum_unnecessary(skb))
5239 err = skb_copy_datagram_msg(skb, hlen, tp->ucopy.msg, chunk);
5240 else
5241 err = skb_copy_and_csum_datagram_msg(skb, hlen, tp->ucopy.msg);
5242
5243 if (!err) {
5244 tp->ucopy.len -= chunk;
5245 tp->copied_seq += chunk;
5246 tcp_rcv_space_adjust(sk);
5247 }
5248
5249 return err;
5250}
5251
5252/* Does PAWS and seqno based validation of an incoming segment, flags will
5253 * play significant role here.
5254 */
5255static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5256 const struct tcphdr *th, int syn_inerr)
5257{
5258 struct tcp_sock *tp = tcp_sk(sk);
5259 bool rst_seq_match = false;
5260
5261 /* RFC1323: H1. Apply PAWS check first. */
5262 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
5263 tcp_paws_discard(sk, skb)) {
5264 if (!th->rst) {
5265 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5266 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5267 LINUX_MIB_TCPACKSKIPPEDPAWS,
5268 &tp->last_oow_ack_time))
5269 tcp_send_dupack(sk, skb);
5270 goto discard;
5271 }
5272 /* Reset is accepted even if it did not pass PAWS. */
5273 }
5274
5275 /* Step 1: check sequence number */
5276 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5277 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5278 * (RST) segments are validated by checking their SEQ-fields."
5279 * And page 69: "If an incoming segment is not acceptable,
5280 * an acknowledgment should be sent in reply (unless the RST
5281 * bit is set, if so drop the segment and return)".
5282 */
5283 if (!th->rst) {
5284 if (th->syn)
5285 goto syn_challenge;
5286 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5287 LINUX_MIB_TCPACKSKIPPEDSEQ,
5288 &tp->last_oow_ack_time))
5289 tcp_send_dupack(sk, skb);
5290 }
5291 goto discard;
5292 }
5293
5294 /* Step 2: check RST bit */
5295 if (th->rst) {
5296 /* RFC 5961 3.2 (extend to match against SACK too if available):
5297 * If seq num matches RCV.NXT or the right-most SACK block,
5298 * then
5299 * RESET the connection
5300 * else
5301 * Send a challenge ACK
5302 */
5303 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
5304 rst_seq_match = true;
5305 } else if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
5306 struct tcp_sack_block *sp = &tp->selective_acks[0];
5307 int max_sack = sp[0].end_seq;
5308 int this_sack;
5309
5310 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
5311 ++this_sack) {
5312 max_sack = after(sp[this_sack].end_seq,
5313 max_sack) ?
5314 sp[this_sack].end_seq : max_sack;
5315 }
5316
5317 if (TCP_SKB_CB(skb)->seq == max_sack)
5318 rst_seq_match = true;
5319 }
5320
5321 if (rst_seq_match)
5322 tcp_reset(sk);
5323 else
5324 tcp_send_challenge_ack(sk, skb);
5325 goto discard;
5326 }
5327
5328 /* step 3: check security and precedence [ignored] */
5329
5330 /* step 4: Check for a SYN
5331 * RFC 5961 4.2 : Send a challenge ack
5332 */
5333 if (th->syn) {
5334syn_challenge:
5335 if (syn_inerr)
5336 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5337 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5338 tcp_send_challenge_ack(sk, skb);
5339 goto discard;
5340 }
5341
5342 return true;
5343
5344discard:
5345 tcp_drop(sk, skb);
5346 return false;
5347}
5348
5349/*
5350 * TCP receive function for the ESTABLISHED state.
5351 *
5352 * It is split into a fast path and a slow path. The fast path is
5353 * disabled when:
5354 * - A zero window was announced from us - zero window probing
5355 * is only handled properly in the slow path.
5356 * - Out of order segments arrived.
5357 * - Urgent data is expected.
5358 * - There is no buffer space left
5359 * - Unexpected TCP flags/window values/header lengths are received
5360 * (detected by checking the TCP header against pred_flags)
5361 * - Data is sent in both directions. Fast path only supports pure senders
5362 * or pure receivers (this means either the sequence number or the ack
5363 * value must stay constant)
5364 * - Unexpected TCP option.
5365 *
5366 * When these conditions are not satisfied it drops into a standard
5367 * receive procedure patterned after RFC793 to handle all cases.
5368 * The first three cases are guaranteed by proper pred_flags setting,
5369 * the rest is checked inline. Fast processing is turned on in
5370 * tcp_data_queue when everything is OK.
5371 */
5372void tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5373 const struct tcphdr *th, unsigned int len)
5374{
5375 struct tcp_sock *tp = tcp_sk(sk);
5376
5377 if (unlikely(!sk->sk_rx_dst))
5378 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5379 /*
5380 * Header prediction.
5381 * The code loosely follows the one in the famous
5382 * "30 instruction TCP receive" Van Jacobson mail.
5383 *
5384 * Van's trick is to deposit buffers into socket queue
5385 * on a device interrupt, to call tcp_recv function
5386 * on the receive process context and checksum and copy
5387 * the buffer to user space. smart...
5388 *
5389 * Our current scheme is not silly either but we take the
5390 * extra cost of the net_bh soft interrupt processing...
5391 * We do checksum and copy also but from device to kernel.
5392 */
5393
5394 tp->rx_opt.saw_tstamp = 0;
5395
5396 /* pred_flags is 0xS?10 << 16 + snd_wnd
5397 * if header_prediction is to be made
5398 * 'S' will always be tp->tcp_header_len >> 2
5399 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5400 * turn it off (when there are holes in the receive
5401 * space for instance)
5402 * PSH flag is ignored.
5403 */
5404
5405 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5406 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5407 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5408 int tcp_header_len = tp->tcp_header_len;
5409
5410 /* Timestamp header prediction: tcp_header_len
5411 * is automatically equal to th->doff*4 due to pred_flags
5412 * match.
5413 */
5414
5415 /* Check timestamp */
5416 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5417 /* No? Slow path! */
5418 if (!tcp_parse_aligned_timestamp(tp, th))
5419 goto slow_path;
5420
5421 /* If PAWS failed, check it more carefully in slow path */
5422 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5423 goto slow_path;
5424
5425 /* DO NOT update ts_recent here, if checksum fails
5426 * and timestamp was corrupted part, it will result
5427 * in a hung connection since we will drop all
5428 * future packets due to the PAWS test.
5429 */
5430 }
5431
5432 if (len <= tcp_header_len) {
5433 /* Bulk data transfer: sender */
5434 if (len == tcp_header_len) {
5435 /* Predicted packet is in window by definition.
5436 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5437 * Hence, check seq<=rcv_wup reduces to:
5438 */
5439 if (tcp_header_len ==
5440 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5441 tp->rcv_nxt == tp->rcv_wup)
5442 tcp_store_ts_recent(tp);
5443
5444 /* We know that such packets are checksummed
5445 * on entry.
5446 */
5447 tcp_ack(sk, skb, 0);
5448 __kfree_skb(skb);
5449 tcp_data_snd_check(sk);
5450 return;
5451 } else { /* Header too small */
5452 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5453 goto discard;
5454 }
5455 } else {
5456 int eaten = 0;
5457 bool fragstolen = false;
5458
5459 if (tp->ucopy.task == current &&
5460 tp->copied_seq == tp->rcv_nxt &&
5461 len - tcp_header_len <= tp->ucopy.len &&
5462 sock_owned_by_user(sk)) {
5463 __set_current_state(TASK_RUNNING);
5464
5465 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) {
5466 /* Predicted packet is in window by definition.
5467 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5468 * Hence, check seq<=rcv_wup reduces to:
5469 */
5470 if (tcp_header_len ==
5471 (sizeof(struct tcphdr) +
5472 TCPOLEN_TSTAMP_ALIGNED) &&
5473 tp->rcv_nxt == tp->rcv_wup)
5474 tcp_store_ts_recent(tp);
5475
5476 tcp_rcv_rtt_measure_ts(sk, skb);
5477
5478 __skb_pull(skb, tcp_header_len);
5479 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
5480 NET_INC_STATS(sock_net(sk),
5481 LINUX_MIB_TCPHPHITSTOUSER);
5482 eaten = 1;
5483 }
5484 }
5485 if (!eaten) {
5486 if (tcp_checksum_complete(skb))
5487 goto csum_error;
5488
5489 if ((int)skb->truesize > sk->sk_forward_alloc)
5490 goto step5;
5491
5492 /* Predicted packet is in window by definition.
5493 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5494 * Hence, check seq<=rcv_wup reduces to:
5495 */
5496 if (tcp_header_len ==
5497 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5498 tp->rcv_nxt == tp->rcv_wup)
5499 tcp_store_ts_recent(tp);
5500
5501 tcp_rcv_rtt_measure_ts(sk, skb);
5502
5503 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
5504
5505 /* Bulk data transfer: receiver */
5506 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5507 &fragstolen);
5508 }
5509
5510 tcp_event_data_recv(sk, skb);
5511
5512 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5513 /* Well, only one small jumplet in fast path... */
5514 tcp_ack(sk, skb, FLAG_DATA);
5515 tcp_data_snd_check(sk);
5516 if (!inet_csk_ack_scheduled(sk))
5517 goto no_ack;
5518 }
5519
5520 __tcp_ack_snd_check(sk, 0);
5521no_ack:
5522 if (eaten)
5523 kfree_skb_partial(skb, fragstolen);
5524 sk->sk_data_ready(sk);
5525 return;
5526 }
5527 }
5528
5529slow_path:
5530 if (len < (th->doff << 2) || tcp_checksum_complete(skb))
5531 goto csum_error;
5532
5533 if (!th->ack && !th->rst && !th->syn)
5534 goto discard;
5535
5536 /*
5537 * Standard slow path.
5538 */
5539
5540 if (!tcp_validate_incoming(sk, skb, th, 1))
5541 return;
5542
5543step5:
5544 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5545 goto discard;
5546
5547 tcp_rcv_rtt_measure_ts(sk, skb);
5548
5549 /* Process urgent data. */
5550 tcp_urg(sk, skb, th);
5551
5552 /* step 7: process the segment text */
5553 tcp_data_queue(sk, skb);
5554
5555 tcp_data_snd_check(sk);
5556 tcp_ack_snd_check(sk);
5557 return;
5558
5559csum_error:
5560 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
5561 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5562
5563discard:
5564 tcp_drop(sk, skb);
5565}
5566EXPORT_SYMBOL(tcp_rcv_established);
5567
5568void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5569{
5570 struct tcp_sock *tp = tcp_sk(sk);
5571 struct inet_connection_sock *icsk = inet_csk(sk);
5572
5573 tcp_set_state(sk, TCP_ESTABLISHED);
5574 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5575
5576 if (skb) {
5577 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5578 security_inet_conn_established(sk, skb);
5579 }
5580
5581 /* Make sure socket is routed, for correct metrics. */
5582 icsk->icsk_af_ops->rebuild_header(sk);
5583
5584 tcp_init_metrics(sk);
5585
5586 tcp_init_congestion_control(sk);
5587
5588 /* Prevent spurious tcp_cwnd_restart() on first data
5589 * packet.
5590 */
5591 tp->lsndtime = tcp_time_stamp;
5592
5593 tcp_init_buffer_space(sk);
5594
5595 if (sock_flag(sk, SOCK_KEEPOPEN))
5596 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5597
5598 if (!tp->rx_opt.snd_wscale)
5599 __tcp_fast_path_on(tp, tp->snd_wnd);
5600 else
5601 tp->pred_flags = 0;
5602
5603 if (!sock_flag(sk, SOCK_DEAD)) {
5604 sk->sk_state_change(sk);
5605 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5606 }
5607}
5608
5609static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5610 struct tcp_fastopen_cookie *cookie)
5611{
5612 struct tcp_sock *tp = tcp_sk(sk);
5613 struct sk_buff *data = tp->syn_data ? tcp_write_queue_head(sk) : NULL;
5614 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
5615 bool syn_drop = false;
5616
5617 if (mss == tp->rx_opt.user_mss) {
5618 struct tcp_options_received opt;
5619
5620 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5621 tcp_clear_options(&opt);
5622 opt.user_mss = opt.mss_clamp = 0;
5623 tcp_parse_options(synack, &opt, 0, NULL);
5624 mss = opt.mss_clamp;
5625 }
5626
5627 if (!tp->syn_fastopen) {
5628 /* Ignore an unsolicited cookie */
5629 cookie->len = -1;
5630 } else if (tp->total_retrans) {
5631 /* SYN timed out and the SYN-ACK neither has a cookie nor
5632 * acknowledges data. Presumably the remote received only
5633 * the retransmitted (regular) SYNs: either the original
5634 * SYN-data or the corresponding SYN-ACK was dropped.
5635 */
5636 syn_drop = (cookie->len < 0 && data);
5637 } else if (cookie->len < 0 && !tp->syn_data) {
5638 /* We requested a cookie but didn't get it. If we did not use
5639 * the (old) exp opt format then try so next time (try_exp=1).
5640 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5641 */
5642 try_exp = tp->syn_fastopen_exp ? 2 : 1;
5643 }
5644
5645 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
5646
5647 if (data) { /* Retransmit unacked data in SYN */
5648 tcp_for_write_queue_from(data, sk) {
5649 if (data == tcp_send_head(sk) ||
5650 __tcp_retransmit_skb(sk, data, 1))
5651 break;
5652 }
5653 tcp_rearm_rto(sk);
5654 NET_INC_STATS(sock_net(sk),
5655 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
5656 return true;
5657 }
5658 tp->syn_data_acked = tp->syn_data;
5659 if (tp->syn_data_acked)
5660 NET_INC_STATS(sock_net(sk),
5661 LINUX_MIB_TCPFASTOPENACTIVE);
5662
5663 tcp_fastopen_add_skb(sk, synack);
5664
5665 return false;
5666}
5667
5668static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5669 const struct tcphdr *th)
5670{
5671 struct inet_connection_sock *icsk = inet_csk(sk);
5672 struct tcp_sock *tp = tcp_sk(sk);
5673 struct tcp_fastopen_cookie foc = { .len = -1 };
5674 int saved_clamp = tp->rx_opt.mss_clamp;
5675
5676 tcp_parse_options(skb, &tp->rx_opt, 0, &foc);
5677 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5678 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5679
5680 if (th->ack) {
5681 /* rfc793:
5682 * "If the state is SYN-SENT then
5683 * first check the ACK bit
5684 * If the ACK bit is set
5685 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5686 * a reset (unless the RST bit is set, if so drop
5687 * the segment and return)"
5688 */
5689 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5690 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5691 goto reset_and_undo;
5692
5693 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5694 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5695 tcp_time_stamp)) {
5696 NET_INC_STATS(sock_net(sk),
5697 LINUX_MIB_PAWSACTIVEREJECTED);
5698 goto reset_and_undo;
5699 }
5700
5701 /* Now ACK is acceptable.
5702 *
5703 * "If the RST bit is set
5704 * If the ACK was acceptable then signal the user "error:
5705 * connection reset", drop the segment, enter CLOSED state,
5706 * delete TCB, and return."
5707 */
5708
5709 if (th->rst) {
5710 tcp_reset(sk);
5711 goto discard;
5712 }
5713
5714 /* rfc793:
5715 * "fifth, if neither of the SYN or RST bits is set then
5716 * drop the segment and return."
5717 *
5718 * See note below!
5719 * --ANK(990513)
5720 */
5721 if (!th->syn)
5722 goto discard_and_undo;
5723
5724 /* rfc793:
5725 * "If the SYN bit is on ...
5726 * are acceptable then ...
5727 * (our SYN has been ACKed), change the connection
5728 * state to ESTABLISHED..."
5729 */
5730
5731 tcp_ecn_rcv_synack(tp, th);
5732
5733 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5734 tcp_ack(sk, skb, FLAG_SLOWPATH);
5735
5736 /* Ok.. it's good. Set up sequence numbers and
5737 * move to established.
5738 */
5739 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5740 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5741
5742 /* RFC1323: The window in SYN & SYN/ACK segments is
5743 * never scaled.
5744 */
5745 tp->snd_wnd = ntohs(th->window);
5746
5747 if (!tp->rx_opt.wscale_ok) {
5748 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5749 tp->window_clamp = min(tp->window_clamp, 65535U);
5750 }
5751
5752 if (tp->rx_opt.saw_tstamp) {
5753 tp->rx_opt.tstamp_ok = 1;
5754 tp->tcp_header_len =
5755 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5756 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5757 tcp_store_ts_recent(tp);
5758 } else {
5759 tp->tcp_header_len = sizeof(struct tcphdr);
5760 }
5761
5762 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5763 tcp_enable_fack(tp);
5764
5765 tcp_mtup_init(sk);
5766 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5767 tcp_initialize_rcv_mss(sk);
5768
5769 /* Remember, tcp_poll() does not lock socket!
5770 * Change state from SYN-SENT only after copied_seq
5771 * is initialized. */
5772 tp->copied_seq = tp->rcv_nxt;
5773
5774 smp_mb();
5775
5776 tcp_finish_connect(sk, skb);
5777
5778 if ((tp->syn_fastopen || tp->syn_data) &&
5779 tcp_rcv_fastopen_synack(sk, skb, &foc))
5780 return -1;
5781
5782 if (sk->sk_write_pending ||
5783 icsk->icsk_accept_queue.rskq_defer_accept ||
5784 icsk->icsk_ack.pingpong) {
5785 /* Save one ACK. Data will be ready after
5786 * several ticks, if write_pending is set.
5787 *
5788 * It may be deleted, but with this feature tcpdumps
5789 * look so _wonderfully_ clever, that I was not able
5790 * to stand against the temptation 8) --ANK
5791 */
5792 inet_csk_schedule_ack(sk);
5793 tcp_enter_quickack_mode(sk);
5794 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5795 TCP_DELACK_MAX, TCP_RTO_MAX);
5796
5797discard:
5798 tcp_drop(sk, skb);
5799 return 0;
5800 } else {
5801 tcp_send_ack(sk);
5802 }
5803 return -1;
5804 }
5805
5806 /* No ACK in the segment */
5807
5808 if (th->rst) {
5809 /* rfc793:
5810 * "If the RST bit is set
5811 *
5812 * Otherwise (no ACK) drop the segment and return."
5813 */
5814
5815 goto discard_and_undo;
5816 }
5817
5818 /* PAWS check. */
5819 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5820 tcp_paws_reject(&tp->rx_opt, 0))
5821 goto discard_and_undo;
5822
5823 if (th->syn) {
5824 /* We see SYN without ACK. It is attempt of
5825 * simultaneous connect with crossed SYNs.
5826 * Particularly, it can be connect to self.
5827 */
5828 tcp_set_state(sk, TCP_SYN_RECV);
5829
5830 if (tp->rx_opt.saw_tstamp) {
5831 tp->rx_opt.tstamp_ok = 1;
5832 tcp_store_ts_recent(tp);
5833 tp->tcp_header_len =
5834 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5835 } else {
5836 tp->tcp_header_len = sizeof(struct tcphdr);
5837 }
5838
5839 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5840 tp->copied_seq = tp->rcv_nxt;
5841 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5842
5843 /* RFC1323: The window in SYN & SYN/ACK segments is
5844 * never scaled.
5845 */
5846 tp->snd_wnd = ntohs(th->window);
5847 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5848 tp->max_window = tp->snd_wnd;
5849
5850 tcp_ecn_rcv_syn(tp, th);
5851
5852 tcp_mtup_init(sk);
5853 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5854 tcp_initialize_rcv_mss(sk);
5855
5856 tcp_send_synack(sk);
5857#if 0
5858 /* Note, we could accept data and URG from this segment.
5859 * There are no obstacles to make this (except that we must
5860 * either change tcp_recvmsg() to prevent it from returning data
5861 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5862 *
5863 * However, if we ignore data in ACKless segments sometimes,
5864 * we have no reasons to accept it sometimes.
5865 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5866 * is not flawless. So, discard packet for sanity.
5867 * Uncomment this return to process the data.
5868 */
5869 return -1;
5870#else
5871 goto discard;
5872#endif
5873 }
5874 /* "fifth, if neither of the SYN or RST bits is set then
5875 * drop the segment and return."
5876 */
5877
5878discard_and_undo:
5879 tcp_clear_options(&tp->rx_opt);
5880 tp->rx_opt.mss_clamp = saved_clamp;
5881 goto discard;
5882
5883reset_and_undo:
5884 tcp_clear_options(&tp->rx_opt);
5885 tp->rx_opt.mss_clamp = saved_clamp;
5886 return 1;
5887}
5888
5889/*
5890 * This function implements the receiving procedure of RFC 793 for
5891 * all states except ESTABLISHED and TIME_WAIT.
5892 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5893 * address independent.
5894 */
5895
5896int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
5897{
5898 struct tcp_sock *tp = tcp_sk(sk);
5899 struct inet_connection_sock *icsk = inet_csk(sk);
5900 const struct tcphdr *th = tcp_hdr(skb);
5901 struct request_sock *req;
5902 int queued = 0;
5903 bool acceptable;
5904
5905 switch (sk->sk_state) {
5906 case TCP_CLOSE:
5907 goto discard;
5908
5909 case TCP_LISTEN:
5910 if (th->ack)
5911 return 1;
5912
5913 if (th->rst)
5914 goto discard;
5915
5916 if (th->syn) {
5917 if (th->fin)
5918 goto discard;
5919 /* It is possible that we process SYN packets from backlog,
5920 * so we need to make sure to disable BH right there.
5921 */
5922 local_bh_disable();
5923 acceptable = icsk->icsk_af_ops->conn_request(sk, skb) >= 0;
5924 local_bh_enable();
5925
5926 if (!acceptable)
5927 return 1;
5928 consume_skb(skb);
5929 return 0;
5930 }
5931 goto discard;
5932
5933 case TCP_SYN_SENT:
5934 tp->rx_opt.saw_tstamp = 0;
5935 queued = tcp_rcv_synsent_state_process(sk, skb, th);
5936 if (queued >= 0)
5937 return queued;
5938
5939 /* Do step6 onward by hand. */
5940 tcp_urg(sk, skb, th);
5941 __kfree_skb(skb);
5942 tcp_data_snd_check(sk);
5943 return 0;
5944 }
5945
5946 tp->rx_opt.saw_tstamp = 0;
5947 req = tp->fastopen_rsk;
5948 if (req) {
5949 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
5950 sk->sk_state != TCP_FIN_WAIT1);
5951
5952 if (!tcp_check_req(sk, skb, req, true))
5953 goto discard;
5954 }
5955
5956 if (!th->ack && !th->rst && !th->syn)
5957 goto discard;
5958
5959 if (!tcp_validate_incoming(sk, skb, th, 0))
5960 return 0;
5961
5962 /* step 5: check the ACK field */
5963 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
5964 FLAG_UPDATE_TS_RECENT) > 0;
5965
5966 switch (sk->sk_state) {
5967 case TCP_SYN_RECV:
5968 if (!acceptable)
5969 return 1;
5970
5971 if (!tp->srtt_us)
5972 tcp_synack_rtt_meas(sk, req);
5973
5974 /* Once we leave TCP_SYN_RECV, we no longer need req
5975 * so release it.
5976 */
5977 if (req) {
5978 inet_csk(sk)->icsk_retransmits = 0;
5979 reqsk_fastopen_remove(sk, req, false);
5980 } else {
5981 /* Make sure socket is routed, for correct metrics. */
5982 icsk->icsk_af_ops->rebuild_header(sk);
5983 tcp_init_congestion_control(sk);
5984
5985 tcp_mtup_init(sk);
5986 tp->copied_seq = tp->rcv_nxt;
5987 tcp_init_buffer_space(sk);
5988 }
5989 smp_mb();
5990 tcp_set_state(sk, TCP_ESTABLISHED);
5991 sk->sk_state_change(sk);
5992
5993 /* Note, that this wakeup is only for marginal crossed SYN case.
5994 * Passively open sockets are not waked up, because
5995 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5996 */
5997 if (sk->sk_socket)
5998 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5999
6000 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
6001 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
6002 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6003
6004 if (tp->rx_opt.tstamp_ok)
6005 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6006
6007 if (req) {
6008 /* Re-arm the timer because data may have been sent out.
6009 * This is similar to the regular data transmission case
6010 * when new data has just been ack'ed.
6011 *
6012 * (TFO) - we could try to be more aggressive and
6013 * retransmitting any data sooner based on when they
6014 * are sent out.
6015 */
6016 tcp_rearm_rto(sk);
6017 } else
6018 tcp_init_metrics(sk);
6019
6020 if (!inet_csk(sk)->icsk_ca_ops->cong_control)
6021 tcp_update_pacing_rate(sk);
6022
6023 /* Prevent spurious tcp_cwnd_restart() on first data packet */
6024 tp->lsndtime = tcp_time_stamp;
6025
6026 tcp_initialize_rcv_mss(sk);
6027 tcp_fast_path_on(tp);
6028 break;
6029
6030 case TCP_FIN_WAIT1: {
6031 struct dst_entry *dst;
6032 int tmo;
6033
6034 /* If we enter the TCP_FIN_WAIT1 state and we are a
6035 * Fast Open socket and this is the first acceptable
6036 * ACK we have received, this would have acknowledged
6037 * our SYNACK so stop the SYNACK timer.
6038 */
6039 if (req) {
6040 /* Return RST if ack_seq is invalid.
6041 * Note that RFC793 only says to generate a
6042 * DUPACK for it but for TCP Fast Open it seems
6043 * better to treat this case like TCP_SYN_RECV
6044 * above.
6045 */
6046 if (!acceptable)
6047 return 1;
6048 /* We no longer need the request sock. */
6049 reqsk_fastopen_remove(sk, req, false);
6050 tcp_rearm_rto(sk);
6051 }
6052 if (tp->snd_una != tp->write_seq)
6053 break;
6054
6055 tcp_set_state(sk, TCP_FIN_WAIT2);
6056 sk->sk_shutdown |= SEND_SHUTDOWN;
6057
6058 dst = __sk_dst_get(sk);
6059 if (dst)
6060 dst_confirm(dst);
6061
6062 if (!sock_flag(sk, SOCK_DEAD)) {
6063 /* Wake up lingering close() */
6064 sk->sk_state_change(sk);
6065 break;
6066 }
6067
6068 if (tp->linger2 < 0 ||
6069 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6070 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
6071 tcp_done(sk);
6072 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6073 return 1;
6074 }
6075
6076 tmo = tcp_fin_time(sk);
6077 if (tmo > TCP_TIMEWAIT_LEN) {
6078 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
6079 } else if (th->fin || sock_owned_by_user(sk)) {
6080 /* Bad case. We could lose such FIN otherwise.
6081 * It is not a big problem, but it looks confusing
6082 * and not so rare event. We still can lose it now,
6083 * if it spins in bh_lock_sock(), but it is really
6084 * marginal case.
6085 */
6086 inet_csk_reset_keepalive_timer(sk, tmo);
6087 } else {
6088 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6089 goto discard;
6090 }
6091 break;
6092 }
6093
6094 case TCP_CLOSING:
6095 if (tp->snd_una == tp->write_seq) {
6096 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6097 goto discard;
6098 }
6099 break;
6100
6101 case TCP_LAST_ACK:
6102 if (tp->snd_una == tp->write_seq) {
6103 tcp_update_metrics(sk);
6104 tcp_done(sk);
6105 goto discard;
6106 }
6107 break;
6108 }
6109
6110 /* step 6: check the URG bit */
6111 tcp_urg(sk, skb, th);
6112
6113 /* step 7: process the segment text */
6114 switch (sk->sk_state) {
6115 case TCP_CLOSE_WAIT:
6116 case TCP_CLOSING:
6117 case TCP_LAST_ACK:
6118 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
6119 break;
6120 case TCP_FIN_WAIT1:
6121 case TCP_FIN_WAIT2:
6122 /* RFC 793 says to queue data in these states,
6123 * RFC 1122 says we MUST send a reset.
6124 * BSD 4.4 also does reset.
6125 */
6126 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6127 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6128 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6129 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6130 tcp_reset(sk);
6131 return 1;
6132 }
6133 }
6134 /* Fall through */
6135 case TCP_ESTABLISHED:
6136 tcp_data_queue(sk, skb);
6137 queued = 1;
6138 break;
6139 }
6140
6141 /* tcp_data could move socket to TIME-WAIT */
6142 if (sk->sk_state != TCP_CLOSE) {
6143 tcp_data_snd_check(sk);
6144 tcp_ack_snd_check(sk);
6145 }
6146
6147 if (!queued) {
6148discard:
6149 tcp_drop(sk, skb);
6150 }
6151 return 0;
6152}
6153EXPORT_SYMBOL(tcp_rcv_state_process);
6154
6155static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6156{
6157 struct inet_request_sock *ireq = inet_rsk(req);
6158
6159 if (family == AF_INET)
6160 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6161 &ireq->ir_rmt_addr, port);
6162#if IS_ENABLED(CONFIG_IPV6)
6163 else if (family == AF_INET6)
6164 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6165 &ireq->ir_v6_rmt_addr, port);
6166#endif
6167}
6168
6169/* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6170 *
6171 * If we receive a SYN packet with these bits set, it means a
6172 * network is playing bad games with TOS bits. In order to
6173 * avoid possible false congestion notifications, we disable
6174 * TCP ECN negotiation.
6175 *
6176 * Exception: tcp_ca wants ECN. This is required for DCTCP
6177 * congestion control: Linux DCTCP asserts ECT on all packets,
6178 * including SYN, which is most optimal solution; however,
6179 * others, such as FreeBSD do not.
6180 */
6181static void tcp_ecn_create_request(struct request_sock *req,
6182 const struct sk_buff *skb,
6183 const struct sock *listen_sk,
6184 const struct dst_entry *dst)
6185{
6186 const struct tcphdr *th = tcp_hdr(skb);
6187 const struct net *net = sock_net(listen_sk);
6188 bool th_ecn = th->ece && th->cwr;
6189 bool ect, ecn_ok;
6190 u32 ecn_ok_dst;
6191
6192 if (!th_ecn)
6193 return;
6194
6195 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
6196 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6197 ecn_ok = net->ipv4.sysctl_tcp_ecn || ecn_ok_dst;
6198
6199 if ((!ect && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
6200 (ecn_ok_dst & DST_FEATURE_ECN_CA))
6201 inet_rsk(req)->ecn_ok = 1;
6202}
6203
6204static void tcp_openreq_init(struct request_sock *req,
6205 const struct tcp_options_received *rx_opt,
6206 struct sk_buff *skb, const struct sock *sk)
6207{
6208 struct inet_request_sock *ireq = inet_rsk(req);
6209
6210 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
6211 req->cookie_ts = 0;
6212 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6213 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6214 skb_mstamp_get(&tcp_rsk(req)->snt_synack);
6215 tcp_rsk(req)->last_oow_ack_time = 0;
6216 req->mss = rx_opt->mss_clamp;
6217 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6218 ireq->tstamp_ok = rx_opt->tstamp_ok;
6219 ireq->sack_ok = rx_opt->sack_ok;
6220 ireq->snd_wscale = rx_opt->snd_wscale;
6221 ireq->wscale_ok = rx_opt->wscale_ok;
6222 ireq->acked = 0;
6223 ireq->ecn_ok = 0;
6224 ireq->ir_rmt_port = tcp_hdr(skb)->source;
6225 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6226 ireq->ir_mark = inet_request_mark(sk, skb);
6227}
6228
6229struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6230 struct sock *sk_listener,
6231 bool attach_listener)
6232{
6233 struct request_sock *req = reqsk_alloc(ops, sk_listener,
6234 attach_listener);
6235
6236 if (req) {
6237 struct inet_request_sock *ireq = inet_rsk(req);
6238
6239 kmemcheck_annotate_bitfield(ireq, flags);
6240 ireq->opt = NULL;
6241#if IS_ENABLED(CONFIG_IPV6)
6242 ireq->pktopts = NULL;
6243#endif
6244 atomic64_set(&ireq->ir_cookie, 0);
6245 ireq->ireq_state = TCP_NEW_SYN_RECV;
6246 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
6247 ireq->ireq_family = sk_listener->sk_family;
6248 }
6249
6250 return req;
6251}
6252EXPORT_SYMBOL(inet_reqsk_alloc);
6253
6254/*
6255 * Return true if a syncookie should be sent
6256 */
6257static bool tcp_syn_flood_action(const struct sock *sk,
6258 const struct sk_buff *skb,
6259 const char *proto)
6260{
6261 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
6262 const char *msg = "Dropping request";
6263 bool want_cookie = false;
6264 struct net *net = sock_net(sk);
6265
6266#ifdef CONFIG_SYN_COOKIES
6267 if (net->ipv4.sysctl_tcp_syncookies) {
6268 msg = "Sending cookies";
6269 want_cookie = true;
6270 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
6271 } else
6272#endif
6273 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
6274
6275 if (!queue->synflood_warned &&
6276 net->ipv4.sysctl_tcp_syncookies != 2 &&
6277 xchg(&queue->synflood_warned, 1) == 0)
6278 pr_info("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6279 proto, ntohs(tcp_hdr(skb)->dest), msg);
6280
6281 return want_cookie;
6282}
6283
6284static void tcp_reqsk_record_syn(const struct sock *sk,
6285 struct request_sock *req,
6286 const struct sk_buff *skb)
6287{
6288 if (tcp_sk(sk)->save_syn) {
6289 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
6290 u32 *copy;
6291
6292 copy = kmalloc(len + sizeof(u32), GFP_ATOMIC);
6293 if (copy) {
6294 copy[0] = len;
6295 memcpy(©[1], skb_network_header(skb), len);
6296 req->saved_syn = copy;
6297 }
6298 }
6299}
6300
6301int tcp_conn_request(struct request_sock_ops *rsk_ops,
6302 const struct tcp_request_sock_ops *af_ops,
6303 struct sock *sk, struct sk_buff *skb)
6304{
6305 struct tcp_fastopen_cookie foc = { .len = -1 };
6306 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6307 struct tcp_options_received tmp_opt;
6308 struct tcp_sock *tp = tcp_sk(sk);
6309 struct net *net = sock_net(sk);
6310 struct sock *fastopen_sk = NULL;
6311 struct dst_entry *dst = NULL;
6312 struct request_sock *req;
6313 bool want_cookie = false;
6314 struct flowi fl;
6315
6316 /* TW buckets are converted to open requests without
6317 * limitations, they conserve resources and peer is
6318 * evidently real one.
6319 */
6320 if ((net->ipv4.sysctl_tcp_syncookies == 2 ||
6321 inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6322 want_cookie = tcp_syn_flood_action(sk, skb, rsk_ops->slab_name);
6323 if (!want_cookie)
6324 goto drop;
6325 }
6326
6327 if (sk_acceptq_is_full(sk)) {
6328 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6329 goto drop;
6330 }
6331
6332 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
6333 if (!req)
6334 goto drop;
6335
6336 tcp_rsk(req)->af_specific = af_ops;
6337 tcp_rsk(req)->ts_off = 0;
6338
6339 tcp_clear_options(&tmp_opt);
6340 tmp_opt.mss_clamp = af_ops->mss_clamp;
6341 tmp_opt.user_mss = tp->rx_opt.user_mss;
6342 tcp_parse_options(skb, &tmp_opt, 0, want_cookie ? NULL : &foc);
6343
6344 if (want_cookie && !tmp_opt.saw_tstamp)
6345 tcp_clear_options(&tmp_opt);
6346
6347 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6348 tcp_openreq_init(req, &tmp_opt, skb, sk);
6349 inet_rsk(req)->no_srccheck = inet_sk(sk)->transparent;
6350
6351 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6352 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
6353
6354 af_ops->init_req(req, sk, skb);
6355
6356 if (security_inet_conn_request(sk, skb, req))
6357 goto drop_and_free;
6358
6359 if (isn && tmp_opt.tstamp_ok)
6360 af_ops->init_seq(skb, &tcp_rsk(req)->ts_off);
6361
6362 if (!want_cookie && !isn) {
6363 /* VJ's idea. We save last timestamp seen
6364 * from the destination in peer table, when entering
6365 * state TIME-WAIT, and check against it before
6366 * accepting new connection request.
6367 *
6368 * If "isn" is not zero, this request hit alive
6369 * timewait bucket, so that all the necessary checks
6370 * are made in the function processing timewait state.
6371 */
6372 if (tcp_death_row.sysctl_tw_recycle) {
6373 bool strict;
6374
6375 dst = af_ops->route_req(sk, &fl, req, &strict);
6376
6377 if (dst && strict &&
6378 !tcp_peer_is_proven(req, dst, true,
6379 tmp_opt.saw_tstamp)) {
6380 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSPASSIVEREJECTED);
6381 goto drop_and_release;
6382 }
6383 }
6384 /* Kill the following clause, if you dislike this way. */
6385 else if (!net->ipv4.sysctl_tcp_syncookies &&
6386 (sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6387 (sysctl_max_syn_backlog >> 2)) &&
6388 !tcp_peer_is_proven(req, dst, false,
6389 tmp_opt.saw_tstamp)) {
6390 /* Without syncookies last quarter of
6391 * backlog is filled with destinations,
6392 * proven to be alive.
6393 * It means that we continue to communicate
6394 * to destinations, already remembered
6395 * to the moment of synflood.
6396 */
6397 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6398 rsk_ops->family);
6399 goto drop_and_release;
6400 }
6401
6402 isn = af_ops->init_seq(skb, &tcp_rsk(req)->ts_off);
6403 }
6404 if (!dst) {
6405 dst = af_ops->route_req(sk, &fl, req, NULL);
6406 if (!dst)
6407 goto drop_and_free;
6408 }
6409
6410 tcp_ecn_create_request(req, skb, sk, dst);
6411
6412 if (want_cookie) {
6413 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6414 tcp_rsk(req)->ts_off = 0;
6415 req->cookie_ts = tmp_opt.tstamp_ok;
6416 if (!tmp_opt.tstamp_ok)
6417 inet_rsk(req)->ecn_ok = 0;
6418 }
6419
6420 tcp_rsk(req)->snt_isn = isn;
6421 tcp_rsk(req)->txhash = net_tx_rndhash();
6422 tcp_openreq_init_rwin(req, sk, dst);
6423 if (!want_cookie) {
6424 tcp_reqsk_record_syn(sk, req, skb);
6425 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
6426 }
6427 if (fastopen_sk) {
6428 af_ops->send_synack(fastopen_sk, dst, &fl, req,
6429 &foc, TCP_SYNACK_FASTOPEN);
6430 /* Add the child socket directly into the accept queue */
6431 inet_csk_reqsk_queue_add(sk, req, fastopen_sk);
6432 sk->sk_data_ready(sk);
6433 bh_unlock_sock(fastopen_sk);
6434 sock_put(fastopen_sk);
6435 } else {
6436 tcp_rsk(req)->tfo_listener = false;
6437 if (!want_cookie)
6438 inet_csk_reqsk_queue_hash_add(sk, req, TCP_TIMEOUT_INIT);
6439 af_ops->send_synack(sk, dst, &fl, req, &foc,
6440 !want_cookie ? TCP_SYNACK_NORMAL :
6441 TCP_SYNACK_COOKIE);
6442 if (want_cookie) {
6443 reqsk_free(req);
6444 return 0;
6445 }
6446 }
6447 reqsk_put(req);
6448 return 0;
6449
6450drop_and_release:
6451 dst_release(dst);
6452drop_and_free:
6453 reqsk_free(req);
6454drop:
6455 tcp_listendrop(sk);
6456 return 0;
6457}
6458EXPORT_SYMBOL(tcp_conn_request);