blob: 2d656eef7f8e16458127ba591e0a4154365f6a83 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
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 *
Jesper Juhl02c30a82005-05-05 16:16:16 -07008 * Authors: Ross Biro
Linus Torvalds1da177e2005-04-16 15:20:36 -07009 * 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.
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -080043 * Andrey Savochkin: Fix RTT measurements in the presence of
Linus Torvalds1da177e2005-04-16 15:20:36 -070044 * 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.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +090051 * Andi Kleen: Add tcp_measure_rcv_mss to make
Linus Torvalds1da177e2005-04-16 15:20:36 -070052 * connections with MSS<min(MTU,ann. MSS)
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +090053 * work without delayed acks.
Linus Torvalds1da177e2005-04-16 15:20:36 -070054 * 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
Linus Torvalds1da177e2005-04-16 15:20:36 -070062 */
63
Joe Perchesafd465032012-03-12 07:03:32 +000064#define pr_fmt(fmt) "TCP: " fmt
65
Linus Torvalds1da177e2005-04-16 15:20:36 -070066#include <linux/mm.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090067#include <linux/slab.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070068#include <linux/module.h>
69#include <linux/sysctl.h>
Ilpo Järvinena0bffff2009-03-21 13:36:17 -070070#include <linux/kernel.h>
Eric Dumazetad971f62014-10-11 15:17:29 -070071#include <linux/prefetch.h>
Satoru SATOH5ffc02a2008-05-04 22:14:42 -070072#include <net/dst.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070073#include <net/tcp.h>
74#include <net/inet_common.h>
75#include <linux/ipsec.h>
76#include <asm/unaligned.h>
Willem de Bruijne1c8a602014-08-04 22:11:50 -040077#include <linux/errqueue.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070078
Brian Haleyab32ea52006-09-22 14:15:41 -070079int 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_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
Eric Dumazetdca145f2014-10-27 21:45:24 -070084int sysctl_tcp_max_reordering __read_mostly = 300;
Eric Dumazet4bc2f182010-07-09 21:22:10 +000085EXPORT_SYMBOL(sysctl_tcp_reordering);
Brian Haleyab32ea52006-09-22 14:15:41 -070086int sysctl_tcp_dsack __read_mostly = 1;
87int sysctl_tcp_app_win __read_mostly = 31;
Eric Dumazetb49960a2012-05-02 02:28:41 +000088int sysctl_tcp_adv_win_scale __read_mostly = 1;
Eric Dumazet4bc2f182010-07-09 21:22:10 +000089EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
Linus Torvalds1da177e2005-04-16 15:20:36 -070090
Eric Dumazet282f23c2012-07-17 10:13:05 +020091/* rfc5961 challenge ack rate limiting */
92int sysctl_tcp_challenge_ack_limit = 100;
93
Brian Haleyab32ea52006-09-22 14:15:41 -070094int sysctl_tcp_stdurg __read_mostly;
95int sysctl_tcp_rfc1337 __read_mostly;
96int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
Ilpo Järvinenc96fd3d2007-09-20 11:36:37 -070097int sysctl_tcp_frto __read_mostly = 2;
Yuchung Chengf6722582015-10-16 21:57:42 -070098int sysctl_tcp_min_rtt_wlen __read_mostly = 300;
Linus Torvalds1da177e2005-04-16 15:20:36 -070099
Andreas Petlund7e380172010-02-18 04:48:19 +0000100int sysctl_tcp_thin_dupack __read_mostly;
101
Brian Haleyab32ea52006-09-22 14:15:41 -0700102int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +0000103int sysctl_tcp_early_retrans __read_mostly = 3;
Neal Cardwell032ee422015-02-06 16:04:38 -0500104int sysctl_tcp_invalid_ratelimit __read_mostly = HZ/2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700105
Linus Torvalds1da177e2005-04-16 15:20:36 -0700106#define FLAG_DATA 0x01 /* Incoming frame contained data. */
107#define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
108#define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
109#define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
110#define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
111#define FLAG_DATA_SACKED 0x20 /* New SACK. */
112#define FLAG_ECE 0x40 /* ECE in this ACK */
Yuchung Cheng291a00d2015-07-01 14:11:14 -0700113#define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700114#define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
Yuchung Chenge33099f2013-03-20 13:33:00 +0000115#define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
Ilpo Järvinen2e605292007-08-02 19:46:58 -0700116#define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
Ryousei Takano564262c12007-10-25 23:03:52 -0700117#define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
Ilpo Järvinencadbd032007-12-31 04:49:21 -0800118#define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
Eric Dumazet12fb3dd2013-04-19 07:19:48 +0000119#define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700120
121#define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
122#define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
123#define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
124#define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
125
Linus Torvalds1da177e2005-04-16 15:20:36 -0700126#define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700127#define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700128
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900129/* Adapt the MSS value used to make delayed ack decision to the
Linus Torvalds1da177e2005-04-16 15:20:36 -0700130 * real world.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900131 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800132static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700133{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700134 struct inet_connection_sock *icsk = inet_csk(sk);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900135 const unsigned int lss = icsk->icsk_ack.last_seg_size;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700136 unsigned int len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700137
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900138 icsk->icsk_ack.last_seg_size = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700139
140 /* skb->len may jitter because of SACKs, even if peer
141 * sends good full-sized frames.
142 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800143 len = skb_shinfo(skb)->gso_size ? : skb->len;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700144 if (len >= icsk->icsk_ack.rcv_mss) {
145 icsk->icsk_ack.rcv_mss = len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700146 } else {
147 /* Otherwise, we make more careful check taking into account,
148 * that SACKs block is variable.
149 *
150 * "len" is invariant segment length, including TCP header.
151 */
Arnaldo Carvalho de Melo9c702202007-04-25 18:04:18 -0700152 len += skb->data - skb_transport_header(skb);
William Allen Simpsonbee7ca92009-11-10 09:51:18 +0000153 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -0700154 /* If PSH is not set, packet should be
155 * full sized, provided peer TCP is not badly broken.
156 * This observation (if it is correct 8)) allows
157 * to handle super-low mtu links fairly.
158 */
159 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -0700160 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700161 /* Subtract also invariant (if peer is RFC compliant),
162 * tcp header plus fixed timestamp option length.
163 * Resulting "len" is MSS free of SACK jitter.
164 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700165 len -= tcp_sk(sk)->tcp_header_len;
166 icsk->icsk_ack.last_seg_size = len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700167 if (len == lss) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700168 icsk->icsk_ack.rcv_mss = len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700169 return;
170 }
171 }
Alexey Kuznetsov1ef96962006-09-19 12:52:50 -0700172 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
173 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700174 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700175 }
176}
177
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700178static void tcp_incr_quickack(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700179{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700180 struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazet95c96172012-04-15 05:58:06 +0000181 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700182
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800183 if (quickacks == 0)
184 quickacks = 2;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700185 if (quickacks > icsk->icsk_ack.quick)
186 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700187}
188
stephen hemminger1b9f4092010-09-28 19:30:14 +0000189static void tcp_enter_quickack_mode(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700190{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700191 struct inet_connection_sock *icsk = inet_csk(sk);
192 tcp_incr_quickack(sk);
193 icsk->icsk_ack.pingpong = 0;
194 icsk->icsk_ack.ato = TCP_ATO_MIN;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700195}
196
197/* Send ACKs quickly, if "quick" count is not exhausted
198 * and the session is not interactive.
199 */
200
Jon Maxwell2251ae42015-07-08 10:12:28 +1000201static bool tcp_in_quickack_mode(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700202{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700203 const struct inet_connection_sock *icsk = inet_csk(sk);
Jon Maxwell2251ae42015-07-08 10:12:28 +1000204 const struct dst_entry *dst = __sk_dst_get(sk);
Eric Dumazeta2a385d2012-05-16 23:15:34 +0000205
Jon Maxwell2251ae42015-07-08 10:12:28 +1000206 return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
207 (icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700208}
209
Florian Westphal735d3832014-09-29 13:08:30 +0200210static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700211{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800212 if (tp->ecn_flags & TCP_ECN_OK)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700213 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
214}
215
Florian Westphal735d3832014-09-29 13:08:30 +0200216static void tcp_ecn_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700217{
218 if (tcp_hdr(skb)->cwr)
219 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
220}
221
Florian Westphal735d3832014-09-29 13:08:30 +0200222static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700223{
224 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
225}
226
Florian Westphal735d3832014-09-29 13:08:30 +0200227static void __tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700228{
Eric Dumazetb82d1bb2011-09-27 02:20:08 -0400229 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
Eric Dumazet7a269ff2011-09-22 20:02:19 +0000230 case INET_ECN_NOT_ECT:
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700231 /* Funny extension: if ECT is not set on a segment,
Eric Dumazet7a269ff2011-09-22 20:02:19 +0000232 * and we already seen ECT on a previous segment,
233 * it is probably a retransmit.
234 */
235 if (tp->ecn_flags & TCP_ECN_SEEN)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700236 tcp_enter_quickack_mode((struct sock *)tp);
Eric Dumazet7a269ff2011-09-22 20:02:19 +0000237 break;
238 case INET_ECN_CE:
Florian Westphal98900922014-09-26 22:37:35 +0200239 if (tcp_ca_needs_ecn((struct sock *)tp))
240 tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_IS_CE);
241
Eric Dumazetaae06bf2012-08-06 11:04:43 +0000242 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
243 /* Better not delay acks, sender can have a very low cwnd */
244 tcp_enter_quickack_mode((struct sock *)tp);
245 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
246 }
Eric Dumazet7a269ff2011-09-22 20:02:19 +0000247 tp->ecn_flags |= TCP_ECN_SEEN;
Florian Westphal98900922014-09-26 22:37:35 +0200248 break;
249 default:
250 if (tcp_ca_needs_ecn((struct sock *)tp))
251 tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_NO_CE);
252 tp->ecn_flags |= TCP_ECN_SEEN;
253 break;
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700254 }
255}
256
Florian Westphal735d3832014-09-29 13:08:30 +0200257static void tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
258{
259 if (tp->ecn_flags & TCP_ECN_OK)
260 __tcp_ecn_check_ce(tp, skb);
261}
262
263static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700264{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800265 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700266 tp->ecn_flags &= ~TCP_ECN_OK;
267}
268
Florian Westphal735d3832014-09-29 13:08:30 +0200269static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700270{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800271 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700272 tp->ecn_flags &= ~TCP_ECN_OK;
273}
274
Florian Westphal735d3832014-09-29 13:08:30 +0200275static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700276{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800277 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
Eric Dumazeta2a385d2012-05-16 23:15:34 +0000278 return true;
279 return false;
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700280}
281
Linus Torvalds1da177e2005-04-16 15:20:36 -0700282/* Buffer size and advertised window tuning.
283 *
284 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
285 */
286
Eric Dumazet6ae70532013-10-01 10:23:44 -0700287static void tcp_sndbuf_expand(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700288{
Eric Dumazet6ae70532013-10-01 10:23:44 -0700289 const struct tcp_sock *tp = tcp_sk(sk);
290 int sndmem, per_mss;
291 u32 nr_segs;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700292
Eric Dumazet6ae70532013-10-01 10:23:44 -0700293 /* Worst case is non GSO/TSO : each frame consumes one skb
294 * and skb->head is kmalloced using power of two area of memory
295 */
296 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
297 MAX_TCP_HEADER +
298 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
299
300 per_mss = roundup_pow_of_two(per_mss) +
301 SKB_DATA_ALIGN(sizeof(struct sk_buff));
302
303 nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
304 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
305
306 /* Fast Recovery (RFC 5681 3.2) :
307 * Cubic needs 1.7 factor, rounded to 2 to include
308 * extra cushion (application might react slowly to POLLOUT)
309 */
310 sndmem = 2 * nr_segs * per_mss;
311
Eric Dumazet06a59ec2011-10-13 18:24:42 +0000312 if (sk->sk_sndbuf < sndmem)
313 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700314}
315
316/* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
317 *
318 * All tcp_full_space() is split to two parts: "network" buffer, allocated
319 * forward and advertised in receiver window (tp->rcv_wnd) and
320 * "application buffer", required to isolate scheduling/application
321 * latencies from network.
322 * window_clamp is maximal advertised window. It can be less than
323 * tcp_full_space(), in this case tcp_full_space() - window_clamp
324 * is reserved for "application" buffer. The less window_clamp is
325 * the smoother our behaviour from viewpoint of network, but the lower
326 * throughput and the higher sensitivity of the connection to losses. 8)
327 *
328 * rcv_ssthresh is more strict window_clamp used at "slow start"
329 * phase to predict further behaviour of this connection.
330 * It is used for two goals:
331 * - to enforce header prediction at sender, even when application
332 * requires some significant "application buffer". It is check #1.
333 * - to prevent pruning of receive queue because of misprediction
334 * of receiver window. Check #2.
335 *
336 * The scheme does not work when sender sends good segments opening
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800337 * window and then starts to feed us spaghetti. But it should work
Linus Torvalds1da177e2005-04-16 15:20:36 -0700338 * in common situations. Otherwise, we have to rely on queue collapsing.
339 */
340
341/* Slow part of check#2. */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700342static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700343{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700344 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700345 /* Optimize this! */
Eric Dumazetdfd4f0a2007-12-21 06:07:53 -0800346 int truesize = tcp_win_from_space(skb->truesize) >> 1;
347 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700348
349 while (tp->rcv_ssthresh <= window) {
350 if (truesize <= skb->len)
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700351 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700352
353 truesize >>= 1;
354 window >>= 1;
355 }
356 return 0;
357}
358
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400359static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700360{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700361 struct tcp_sock *tp = tcp_sk(sk);
362
Linus Torvalds1da177e2005-04-16 15:20:36 -0700363 /* Check #1 */
364 if (tp->rcv_ssthresh < tp->window_clamp &&
365 (int)tp->rcv_ssthresh < tcp_space(sk) &&
Eric Dumazetb8da51e2015-05-15 12:39:27 -0700366 !tcp_under_memory_pressure(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700367 int incr;
368
369 /* Check #2. Increase window, if skb with such overhead
370 * will fit to rcvbuf in future.
371 */
372 if (tcp_win_from_space(skb->truesize) <= skb->len)
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800373 incr = 2 * tp->advmss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700374 else
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700375 incr = __tcp_grow_window(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700376
377 if (incr) {
Eric Dumazet4d846f02012-04-16 23:28:07 +0000378 incr = max_t(int, incr, 2 * skb->len);
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800379 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
380 tp->window_clamp);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700381 inet_csk(sk)->icsk_ack.quick |= 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700382 }
383 }
384}
385
386/* 3. Tuning rcvbuf, when connection enters established state. */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700387static void tcp_fixup_rcvbuf(struct sock *sk)
388{
Eric Dumazete9266a02011-10-20 16:53:56 -0400389 u32 mss = tcp_sk(sk)->advmss;
Eric Dumazete9266a02011-10-20 16:53:56 -0400390 int rcvmem;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700391
Yuchung Cheng85f16522013-06-11 15:35:32 -0700392 rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) *
393 tcp_default_init_rwnd(mss);
Eric Dumazete9266a02011-10-20 16:53:56 -0400394
Eric Dumazetb0983d32013-09-20 13:56:58 -0700395 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
396 * Allow enough cushion so that sender is not limited by our window
397 */
398 if (sysctl_tcp_moderate_rcvbuf)
399 rcvmem <<= 2;
400
Eric Dumazete9266a02011-10-20 16:53:56 -0400401 if (sk->sk_rcvbuf < rcvmem)
402 sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700403}
404
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800405/* 4. Try to fixup all. It is made immediately after connection enters
Linus Torvalds1da177e2005-04-16 15:20:36 -0700406 * established state.
407 */
Jerry Chu10467162012-08-31 12:29:11 +0000408void tcp_init_buffer_space(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700409{
410 struct tcp_sock *tp = tcp_sk(sk);
411 int maxwin;
412
413 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
414 tcp_fixup_rcvbuf(sk);
415 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
Eric Dumazet6ae70532013-10-01 10:23:44 -0700416 tcp_sndbuf_expand(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700417
418 tp->rcvq_space.space = tp->rcv_wnd;
Eric Dumazetb0983d32013-09-20 13:56:58 -0700419 tp->rcvq_space.time = tcp_time_stamp;
420 tp->rcvq_space.seq = tp->copied_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700421
422 maxwin = tcp_full_space(sk);
423
424 if (tp->window_clamp >= maxwin) {
425 tp->window_clamp = maxwin;
426
427 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
428 tp->window_clamp = max(maxwin -
429 (maxwin >> sysctl_tcp_app_win),
430 4 * tp->advmss);
431 }
432
433 /* Force reservation of one segment. */
434 if (sysctl_tcp_app_win &&
435 tp->window_clamp > 2 * tp->advmss &&
436 tp->window_clamp + tp->advmss > maxwin)
437 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
438
439 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
440 tp->snd_cwnd_stamp = tcp_time_stamp;
441}
442
Linus Torvalds1da177e2005-04-16 15:20:36 -0700443/* 5. Recalculate window clamp after socket hit its memory bounds. */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700444static void tcp_clamp_window(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700445{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700446 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300447 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700448
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300449 icsk->icsk_ack.quick = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700450
John Heffner326f36e2005-11-10 17:11:48 -0800451 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
452 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
Eric Dumazetb8da51e2015-05-15 12:39:27 -0700453 !tcp_under_memory_pressure(sk) &&
Glauber Costa180d8cd2011-12-11 21:47:02 +0000454 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
John Heffner326f36e2005-11-10 17:11:48 -0800455 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
456 sysctl_tcp_rmem[2]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700457 }
John Heffner326f36e2005-11-10 17:11:48 -0800458 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800459 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700460}
461
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800462/* Initialize RCV_MSS value.
463 * RCV_MSS is an our guess about MSS used by the peer.
464 * We haven't any direct information about the MSS.
465 * It's better to underestimate the RCV_MSS rather than overestimate.
466 * Overestimations make us ACKing less frequently than needed.
467 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
468 */
469void tcp_initialize_rcv_mss(struct sock *sk)
470{
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400471 const struct tcp_sock *tp = tcp_sk(sk);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800472 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
473
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800474 hint = min(hint, tp->rcv_wnd / 2);
William Allen Simpsonbee7ca92009-11-10 09:51:18 +0000475 hint = min(hint, TCP_MSS_DEFAULT);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800476 hint = max(hint, TCP_MIN_MSS);
477
478 inet_csk(sk)->icsk_ack.rcv_mss = hint;
479}
Eric Dumazet4bc2f182010-07-09 21:22:10 +0000480EXPORT_SYMBOL(tcp_initialize_rcv_mss);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800481
Linus Torvalds1da177e2005-04-16 15:20:36 -0700482/* Receiver "autotuning" code.
483 *
484 * The algorithm for RTT estimation w/o timestamps is based on
485 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
Justin P. Mattock631dd1a2010-10-18 11:03:14 +0200486 * <http://public.lanl.gov/radiant/pubs.html#DRS>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700487 *
488 * More detail on this code can be found at
Justin P. Mattock631dd1a2010-10-18 11:03:14 +0200489 * <http://staff.psc.edu/jheffner/>,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700490 * though this reference is out of date. A new paper
491 * is pending.
492 */
493static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
494{
495 u32 new_sample = tp->rcv_rtt_est.rtt;
496 long m = sample;
497
498 if (m == 0)
499 m = 1;
500
501 if (new_sample != 0) {
502 /* If we sample in larger samples in the non-timestamp
503 * case, we could grossly overestimate the RTT especially
504 * with chatty applications or bulk transfer apps which
505 * are stalled on filesystem I/O.
506 *
507 * Also, since we are only going for a minimum in the
Stephen Hemminger31f34262005-11-15 15:17:10 -0800508 * non-timestamp case, we do not smooth things out
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800509 * else with timestamps disabled convergence takes too
Linus Torvalds1da177e2005-04-16 15:20:36 -0700510 * long.
511 */
512 if (!win_dep) {
513 m -= (new_sample >> 3);
514 new_sample += m;
Neal Cardwell18a223e2012-04-10 07:59:20 +0000515 } else {
516 m <<= 3;
517 if (m < new_sample)
518 new_sample = m;
519 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700520 } else {
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800521 /* No previous measure. */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700522 new_sample = m << 3;
523 }
524
525 if (tp->rcv_rtt_est.rtt != new_sample)
526 tp->rcv_rtt_est.rtt = new_sample;
527}
528
529static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
530{
531 if (tp->rcv_rtt_est.time == 0)
532 goto new_measure;
533 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
534 return;
Neal Cardwell651913c2012-04-27 11:29:37 -0400535 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700536
537new_measure:
538 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
539 tp->rcv_rtt_est.time = tcp_time_stamp;
540}
541
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800542static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
543 const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700544{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700545 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700546 if (tp->rx_opt.rcv_tsecr &&
547 (TCP_SKB_CB(skb)->end_seq -
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700548 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700549 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
550}
551
552/*
553 * This function should be called every time data is copied to user space.
554 * It calculates the appropriate TCP receive buffer space.
555 */
556void tcp_rcv_space_adjust(struct sock *sk)
557{
558 struct tcp_sock *tp = tcp_sk(sk);
559 int time;
Eric Dumazetb0983d32013-09-20 13:56:58 -0700560 int copied;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900561
Linus Torvalds1da177e2005-04-16 15:20:36 -0700562 time = tcp_time_stamp - tp->rcvq_space.time;
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800563 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700564 return;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900565
Eric Dumazetb0983d32013-09-20 13:56:58 -0700566 /* Number of bytes copied to user in last RTT */
567 copied = tp->copied_seq - tp->rcvq_space.seq;
568 if (copied <= tp->rcvq_space.space)
569 goto new_measure;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700570
Eric Dumazetb0983d32013-09-20 13:56:58 -0700571 /* A bit of theory :
572 * copied = bytes received in previous RTT, our base window
573 * To cope with packet losses, we need a 2x factor
574 * To cope with slow start, and sender growing its cwin by 100 %
575 * every RTT, we need a 4x factor, because the ACK we are sending
576 * now is for the next RTT, not the current one :
577 * <prev RTT . ><current RTT .. ><next RTT .... >
578 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700579
Eric Dumazetb0983d32013-09-20 13:56:58 -0700580 if (sysctl_tcp_moderate_rcvbuf &&
581 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
582 int rcvwin, rcvmem, rcvbuf;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700583
Eric Dumazetb0983d32013-09-20 13:56:58 -0700584 /* minimal window to cope with packet losses, assuming
585 * steady state. Add some cushion because of small variations.
586 */
587 rcvwin = (copied << 1) + 16 * tp->advmss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700588
Eric Dumazetb0983d32013-09-20 13:56:58 -0700589 /* If rate increased by 25%,
590 * assume slow start, rcvwin = 3 * copied
591 * If rate increased by 50%,
592 * assume sender can use 2x growth, rcvwin = 4 * copied
593 */
594 if (copied >=
595 tp->rcvq_space.space + (tp->rcvq_space.space >> 2)) {
596 if (copied >=
597 tp->rcvq_space.space + (tp->rcvq_space.space >> 1))
598 rcvwin <<= 1;
599 else
600 rcvwin += (rcvwin >> 1);
601 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700602
Eric Dumazetb0983d32013-09-20 13:56:58 -0700603 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
604 while (tcp_win_from_space(rcvmem) < tp->advmss)
605 rcvmem += 128;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700606
Eric Dumazetb0983d32013-09-20 13:56:58 -0700607 rcvbuf = min(rcvwin / tp->advmss * rcvmem, sysctl_tcp_rmem[2]);
608 if (rcvbuf > sk->sk_rcvbuf) {
609 sk->sk_rcvbuf = rcvbuf;
610
611 /* Make the window clamp follow along. */
612 tp->window_clamp = rcvwin;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700613 }
614 }
Eric Dumazetb0983d32013-09-20 13:56:58 -0700615 tp->rcvq_space.space = copied;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900616
Linus Torvalds1da177e2005-04-16 15:20:36 -0700617new_measure:
618 tp->rcvq_space.seq = tp->copied_seq;
619 tp->rcvq_space.time = tcp_time_stamp;
620}
621
622/* There is something which you must keep in mind when you analyze the
623 * behavior of the tp->ato delayed ack timeout interval. When a
624 * connection starts up, we want to ack as quickly as possible. The
625 * problem is that "good" TCP's do slow start at the beginning of data
626 * transmission. The means that until we send the first few ACK's the
627 * sender will sit on his end and only queue most of his data, because
628 * he can only send snd_cwnd unacked packets at any given time. For
629 * each ACK we send, he increments snd_cwnd and transmits more of his
630 * queue. -DaveM
631 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700632static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700633{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700634 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700635 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700636 u32 now;
637
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700638 inet_csk_schedule_ack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700639
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700640 tcp_measure_rcv_mss(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700641
642 tcp_rcv_rtt_measure(tp);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900643
Linus Torvalds1da177e2005-04-16 15:20:36 -0700644 now = tcp_time_stamp;
645
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700646 if (!icsk->icsk_ack.ato) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700647 /* The _first_ data packet received, initialize
648 * delayed ACK engine.
649 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700650 tcp_incr_quickack(sk);
651 icsk->icsk_ack.ato = TCP_ATO_MIN;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700652 } else {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700653 int m = now - icsk->icsk_ack.lrcvtime;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700654
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800655 if (m <= TCP_ATO_MIN / 2) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700656 /* The fastest case is the first. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700657 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
658 } else if (m < icsk->icsk_ack.ato) {
659 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
660 if (icsk->icsk_ack.ato > icsk->icsk_rto)
661 icsk->icsk_ack.ato = icsk->icsk_rto;
662 } else if (m > icsk->icsk_rto) {
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800663 /* Too long gap. Apparently sender failed to
Linus Torvalds1da177e2005-04-16 15:20:36 -0700664 * restart window, so that we send ACKs quickly.
665 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700666 tcp_incr_quickack(sk);
Hideo Aoki3ab224b2007-12-31 00:11:19 -0800667 sk_mem_reclaim(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700668 }
669 }
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700670 icsk->icsk_ack.lrcvtime = now;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700671
Florian Westphal735d3832014-09-29 13:08:30 +0200672 tcp_ecn_check_ce(tp, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700673
674 if (skb->len >= 128)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700675 tcp_grow_window(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700676}
677
Linus Torvalds1da177e2005-04-16 15:20:36 -0700678/* Called to compute a smoothed rtt estimate. The data fed to this
679 * routine either comes from timestamps, or from segments that were
680 * known _not_ to have been retransmitted [see Karn/Partridge
681 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
682 * piece by Van Jacobson.
683 * NOTE: the next three routines used to be one big routine.
684 * To save cycles in the RFC 1323 implementation it was better to break
685 * it up into three procedures. -- erics
686 */
Eric Dumazet740b0f12014-02-26 14:02:48 -0800687static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700688{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300689 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet740b0f12014-02-26 14:02:48 -0800690 long m = mrtt_us; /* RTT */
691 u32 srtt = tp->srtt_us;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700692
Linus Torvalds1da177e2005-04-16 15:20:36 -0700693 /* The following amusing code comes from Jacobson's
694 * article in SIGCOMM '88. Note that rtt and mdev
695 * are scaled versions of rtt and mean deviation.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900696 * This is designed to be as fast as possible
Linus Torvalds1da177e2005-04-16 15:20:36 -0700697 * m stands for "measurement".
698 *
699 * On a 1990 paper the rto value is changed to:
700 * RTO = rtt + 4 * mdev
701 *
702 * Funny. This algorithm seems to be very broken.
703 * These formulae increase RTO, when it should be decreased, increase
Stephen Hemminger31f34262005-11-15 15:17:10 -0800704 * too slowly, when it should be increased quickly, decrease too quickly
Linus Torvalds1da177e2005-04-16 15:20:36 -0700705 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
706 * does not matter how to _calculate_ it. Seems, it was trap
707 * that VJ failed to avoid. 8)
708 */
Eric Dumazet4a5ab4e2014-02-06 15:57:10 -0800709 if (srtt != 0) {
710 m -= (srtt >> 3); /* m is now error in rtt est */
711 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700712 if (m < 0) {
713 m = -m; /* m is now abs(error) */
Eric Dumazet740b0f12014-02-26 14:02:48 -0800714 m -= (tp->mdev_us >> 2); /* similar update on mdev */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700715 /* This is similar to one of Eifel findings.
716 * Eifel blocks mdev updates when rtt decreases.
717 * This solution is a bit different: we use finer gain
718 * for mdev in this case (alpha*beta).
719 * Like Eifel it also prevents growth of rto,
720 * but also it limits too fast rto decreases,
721 * happening in pure Eifel.
722 */
723 if (m > 0)
724 m >>= 3;
725 } else {
Eric Dumazet740b0f12014-02-26 14:02:48 -0800726 m -= (tp->mdev_us >> 2); /* similar update on mdev */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700727 }
Eric Dumazet740b0f12014-02-26 14:02:48 -0800728 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
729 if (tp->mdev_us > tp->mdev_max_us) {
730 tp->mdev_max_us = tp->mdev_us;
731 if (tp->mdev_max_us > tp->rttvar_us)
732 tp->rttvar_us = tp->mdev_max_us;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700733 }
734 if (after(tp->snd_una, tp->rtt_seq)) {
Eric Dumazet740b0f12014-02-26 14:02:48 -0800735 if (tp->mdev_max_us < tp->rttvar_us)
736 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700737 tp->rtt_seq = tp->snd_nxt;
Eric Dumazet740b0f12014-02-26 14:02:48 -0800738 tp->mdev_max_us = tcp_rto_min_us(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700739 }
740 } else {
741 /* no previous measure. */
Eric Dumazet4a5ab4e2014-02-06 15:57:10 -0800742 srtt = m << 3; /* take the measured time to be rtt */
Eric Dumazet740b0f12014-02-26 14:02:48 -0800743 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
744 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
745 tp->mdev_max_us = tp->rttvar_us;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700746 tp->rtt_seq = tp->snd_nxt;
747 }
Eric Dumazet740b0f12014-02-26 14:02:48 -0800748 tp->srtt_us = max(1U, srtt);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700749}
750
Eric Dumazet95bd09e2013-08-27 05:46:32 -0700751/* Set the sk_pacing_rate to allow proper sizing of TSO packets.
752 * Note: TCP stack does not yet implement pacing.
753 * FQ packet scheduler can be used to implement cheap but effective
754 * TCP pacing, to smooth the burst on large writes when packets
755 * in flight is significantly lower than cwnd (or rwin)
756 */
Eric Dumazet43e122b2015-08-21 17:38:02 -0700757int sysctl_tcp_pacing_ss_ratio __read_mostly = 200;
758int sysctl_tcp_pacing_ca_ratio __read_mostly = 120;
759
Eric Dumazet95bd09e2013-08-27 05:46:32 -0700760static void tcp_update_pacing_rate(struct sock *sk)
761{
762 const struct tcp_sock *tp = tcp_sk(sk);
763 u64 rate;
764
765 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
Eric Dumazet43e122b2015-08-21 17:38:02 -0700766 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
767
768 /* current rate is (cwnd * mss) / srtt
769 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
770 * In Congestion Avoidance phase, set it to 120 % the current rate.
771 *
772 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
773 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
774 * end of slow start and should slow down.
775 */
776 if (tp->snd_cwnd < tp->snd_ssthresh / 2)
777 rate *= sysctl_tcp_pacing_ss_ratio;
778 else
779 rate *= sysctl_tcp_pacing_ca_ratio;
Eric Dumazet95bd09e2013-08-27 05:46:32 -0700780
781 rate *= max(tp->snd_cwnd, tp->packets_out);
782
Eric Dumazet740b0f12014-02-26 14:02:48 -0800783 if (likely(tp->srtt_us))
784 do_div(rate, tp->srtt_us);
Eric Dumazet95bd09e2013-08-27 05:46:32 -0700785
Eric Dumazetba537422013-10-09 17:14:52 -0700786 /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate
787 * without any lock. We want to make sure compiler wont store
788 * intermediate values in this location.
789 */
790 ACCESS_ONCE(sk->sk_pacing_rate) = min_t(u64, rate,
791 sk->sk_max_pacing_rate);
Eric Dumazet95bd09e2013-08-27 05:46:32 -0700792}
793
Linus Torvalds1da177e2005-04-16 15:20:36 -0700794/* Calculate rto without backoff. This is the second half of Van Jacobson's
795 * routine referred to above.
796 */
stephen hemmingerf7e56a72013-12-29 11:39:51 -0800797static void tcp_set_rto(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700798{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700799 const struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700800 /* Old crap is replaced with new one. 8)
801 *
802 * More seriously:
803 * 1. If rtt variance happened to be less 50msec, it is hallucination.
804 * It cannot be less due to utterly erratic ACK generation made
805 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
806 * to do with delayed acks, because at cwnd>2 true delack timeout
807 * is invisible. Actually, Linux-2.4 also generates erratic
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800808 * ACKs in some circumstances.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700809 */
Damian Lukowskif1ecd5d2009-08-26 00:16:31 +0000810 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700811
812 /* 2. Fixups made earlier cannot be right.
813 * If we do not estimate RTO correctly without them,
814 * all the algo is pure shit and should be replaced
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800815 * with correct one. It is exactly, which we pretend to do.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700816 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700817
Ilpo Järvinenee6aac52008-12-05 22:43:08 -0800818 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
819 * guarantees that rto is higher.
820 */
Damian Lukowskif1ecd5d2009-08-26 00:16:31 +0000821 tcp_bound_rto(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700822}
823
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400824__u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700825{
826 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
827
Gerrit Renker22b71c82010-08-29 19:23:12 +0000828 if (!cwnd)
David S. Miller442b9632011-02-02 17:05:11 -0800829 cwnd = TCP_INIT_CWND;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700830 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
831}
832
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300833/*
834 * Packet counting of FACK is based on in-order assumptions, therefore TCP
835 * disables it when reordering is detected
836 */
David S. Miller4aabd8e2012-07-09 16:07:30 -0700837void tcp_disable_fack(struct tcp_sock *tp)
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300838{
Ilpo Järvinen85cc3912007-11-15 19:39:31 -0800839 /* RFC3517 uses different metric in lost marker => reset on change */
840 if (tcp_is_fack(tp))
841 tp->lost_skb_hint = NULL;
Vijay Subramanianab562222011-12-20 13:23:24 +0000842 tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300843}
844
Ryousei Takano564262c12007-10-25 23:03:52 -0700845/* Take a notice that peer is sending D-SACKs */
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300846static void tcp_dsack_seen(struct tcp_sock *tp)
847{
Vijay Subramanianab562222011-12-20 13:23:24 +0000848 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300849}
850
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300851static void tcp_update_reordering(struct sock *sk, const int metric,
852 const int ts)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700853{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300854 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700855 if (metric > tp->reordering) {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -0700856 int mib_idx;
857
Eric Dumazetdca145f2014-10-27 21:45:24 -0700858 tp->reordering = min(sysctl_tcp_max_reordering, metric);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700859
860 /* This exciting event is worth to be remembered. 8) */
861 if (ts)
Pavel Emelyanov40b215e2008-07-03 01:05:41 -0700862 mib_idx = LINUX_MIB_TCPTSREORDER;
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300863 else if (tcp_is_reno(tp))
Pavel Emelyanov40b215e2008-07-03 01:05:41 -0700864 mib_idx = LINUX_MIB_TCPRENOREORDER;
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300865 else if (tcp_is_fack(tp))
Pavel Emelyanov40b215e2008-07-03 01:05:41 -0700866 mib_idx = LINUX_MIB_TCPFACKREORDER;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700867 else
Pavel Emelyanov40b215e2008-07-03 01:05:41 -0700868 mib_idx = LINUX_MIB_TCPSACKREORDER;
869
Pavel Emelyanovde0744a2008-07-16 20:31:16 -0700870 NET_INC_STATS_BH(sock_net(sk), mib_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700871#if FASTRETRANS_DEBUG > 1
Joe Perches91df42b2012-05-15 14:11:54 +0000872 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
873 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
874 tp->reordering,
875 tp->fackets_out,
876 tp->sacked_out,
877 tp->undo_marker ? tp->undo_retrans : 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700878#endif
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300879 tcp_disable_fack(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700880 }
Yuchung Chengeed530b2012-05-02 13:30:03 +0000881
882 if (metric > 0)
883 tcp_disable_early_retrans(tp);
Yuchung Cheng4f41b1c2015-10-16 21:57:47 -0700884 tp->rack.reord = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700885}
886
Ilpo Järvinen006f5822008-09-20 21:20:20 -0700887/* This must be called before lost_out is incremented */
Ilpo Järvinenc8c213f2008-09-20 21:18:55 -0700888static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
889{
Ian Morris51456b22015-04-03 09:17:26 +0100890 if (!tp->retransmit_skb_hint ||
Ilpo Järvinenc8c213f2008-09-20 21:18:55 -0700891 before(TCP_SKB_CB(skb)->seq,
892 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
Ilpo Järvinen006f5822008-09-20 21:20:20 -0700893 tp->retransmit_skb_hint = skb;
894
895 if (!tp->lost_out ||
896 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
897 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
Ilpo Järvinenc8c213f2008-09-20 21:18:55 -0700898}
899
Ilpo Järvinen41ea36e2008-09-20 21:19:22 -0700900static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
901{
902 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
903 tcp_verify_retransmit_hint(tp, skb);
904
905 tp->lost_out += tcp_skb_pcount(skb);
906 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
907 }
908}
909
Yuchung Cheng4f41b1c2015-10-16 21:57:47 -0700910void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb)
Ilpo Järvinen006f5822008-09-20 21:20:20 -0700911{
912 tcp_verify_retransmit_hint(tp, skb);
913
914 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
915 tp->lost_out += tcp_skb_pcount(skb);
916 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
917 }
918}
919
Linus Torvalds1da177e2005-04-16 15:20:36 -0700920/* This procedure tags the retransmission queue when SACKs arrive.
921 *
922 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
923 * Packets in queue with these bits set are counted in variables
924 * sacked_out, retrans_out and lost_out, correspondingly.
925 *
926 * Valid combinations are:
927 * Tag InFlight Description
928 * 0 1 - orig segment is in flight.
929 * S 0 - nothing flies, orig reached receiver.
930 * L 0 - nothing flies, orig lost by net.
931 * R 2 - both orig and retransmit are in flight.
932 * L|R 1 - orig is lost, retransmit is in flight.
933 * S|R 1 - orig reached receiver, retrans is still in flight.
934 * (L|S|R is logically valid, it could occur when L|R is sacked,
935 * but it is equivalent to plain S and code short-curcuits it to S.
936 * L|S is logically invalid, it would mean -1 packet in flight 8))
937 *
938 * These 6 states form finite state machine, controlled by the following events:
939 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
940 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
Yuchung Cheng974c1232012-01-19 14:42:21 +0000941 * 3. Loss detection event of two flavors:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700942 * A. Scoreboard estimator decided the packet is lost.
943 * A'. Reno "three dupacks" marks head of queue lost.
Yuchung Cheng974c1232012-01-19 14:42:21 +0000944 * A''. Its FACK modification, head until snd.fack is lost.
945 * B. SACK arrives sacking SND.NXT at the moment, when the
Linus Torvalds1da177e2005-04-16 15:20:36 -0700946 * segment was retransmitted.
947 * 4. D-SACK added new rule: D-SACK changes any tag to S.
948 *
949 * It is pleasant to note, that state diagram turns out to be commutative,
950 * so that we are allowed not to be bothered by order of our actions,
951 * when multiple events arrive simultaneously. (see the function below).
952 *
953 * Reordering detection.
954 * --------------------
955 * Reordering metric is maximal distance, which a packet can be displaced
956 * in packet stream. With SACKs we can estimate it:
957 *
958 * 1. SACK fills old hole and the corresponding segment was not
959 * ever retransmitted -> reordering. Alas, we cannot use it
960 * when segment was retransmitted.
961 * 2. The last flaw is solved with D-SACK. D-SACK arrives
962 * for retransmitted and already SACKed segment -> reordering..
963 * Both of these heuristics are not used in Loss state, when we cannot
964 * account for retransmits accurately.
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -0700965 *
966 * SACK block validation.
967 * ----------------------
968 *
969 * SACK block range validation checks that the received SACK block fits to
970 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
971 * Note that SND.UNA is not included to the range though being valid because
Ilpo Järvinen0e835332007-10-01 15:28:17 -0700972 * it means that the receiver is rather inconsistent with itself reporting
973 * SACK reneging when it should advance SND.UNA. Such SACK block this is
974 * perfectly valid, however, in light of RFC2018 which explicitly states
975 * that "SACK block MUST reflect the newest segment. Even if the newest
976 * segment is going to be discarded ...", not that it looks very clever
977 * in case of head skb. Due to potentional receiver driven attacks, we
978 * choose to avoid immediate execution of a walk in write queue due to
979 * reneging and defer head skb's loss recovery to standard loss recovery
980 * procedure that will eventually trigger (nothing forbids us doing this).
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -0700981 *
982 * Implements also blockage to start_seq wrap-around. Problem lies in the
983 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
984 * there's no guarantee that it will be before snd_nxt (n). The problem
985 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
986 * wrap (s_w):
987 *
988 * <- outs wnd -> <- wrapzone ->
989 * u e n u_w e_w s n_w
990 * | | | | | | |
991 * |<------------+------+----- TCP seqno space --------------+---------->|
992 * ...-- <2^31 ->| |<--------...
993 * ...---- >2^31 ------>| |<--------...
994 *
995 * Current code wouldn't be vulnerable but it's better still to discard such
996 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
997 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
998 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
999 * equal to the ideal case (infinite seqno space without wrap caused issues).
1000 *
1001 * With D-SACK the lower bound is extended to cover sequence space below
1002 * SND.UNA down to undo_marker, which is the last point of interest. Yet
Ryousei Takano564262c12007-10-25 23:03:52 -07001003 * again, D-SACK block must not to go across snd_una (for the same reason as
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001004 * for the normal SACK blocks, explained above). But there all simplicity
1005 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1006 * fully below undo_marker they do not affect behavior in anyway and can
1007 * therefore be safely ignored. In rare cases (which are more or less
1008 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1009 * fragmentation and packet reordering past skb's retransmission. To consider
1010 * them correctly, the acceptable range must be extended even more though
1011 * the exact amount is rather hard to quantify. However, tp->max_window can
1012 * be used as an exaggerated estimate.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001013 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001014static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1015 u32 start_seq, u32 end_seq)
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001016{
1017 /* Too far in future, or reversed (interpretation is ambiguous) */
1018 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001019 return false;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001020
1021 /* Nasty start_seq wrap-around check (see comments above) */
1022 if (!before(start_seq, tp->snd_nxt))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001023 return false;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001024
Ryousei Takano564262c12007-10-25 23:03:52 -07001025 /* In outstanding window? ...This is valid exit for D-SACKs too.
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001026 * start_seq == snd_una is non-sensical (see comments above)
1027 */
1028 if (after(start_seq, tp->snd_una))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001029 return true;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001030
1031 if (!is_dsack || !tp->undo_marker)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001032 return false;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001033
1034 /* ...Then it's D-SACK, and must reside below snd_una completely */
Zheng Yanf779b2d2011-09-18 22:37:34 -04001035 if (after(end_seq, tp->snd_una))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001036 return false;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001037
1038 if (!before(start_seq, tp->undo_marker))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001039 return true;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001040
1041 /* Too old */
1042 if (!after(end_seq, tp->undo_marker))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001043 return false;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001044
1045 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1046 * start_seq < undo_marker and end_seq >= undo_marker.
1047 */
1048 return !before(start_seq, end_seq - tp->max_window);
1049}
1050
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001051static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1052 struct tcp_sack_block_wire *sp, int num_sacks,
1053 u32 prior_snd_una)
David S. Millerd06e0212007-06-18 22:43:06 -07001054{
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07001055 struct tcp_sock *tp = tcp_sk(sk);
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07001056 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1057 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001058 bool dup_sack = false;
David S. Millerd06e0212007-06-18 22:43:06 -07001059
1060 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001061 dup_sack = true;
Ilpo Järvinene60402d2007-08-09 15:14:46 +03001062 tcp_dsack_seen(tp);
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07001063 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
David S. Millerd06e0212007-06-18 22:43:06 -07001064 } else if (num_sacks > 1) {
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07001065 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1066 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
David S. Millerd06e0212007-06-18 22:43:06 -07001067
1068 if (!after(end_seq_0, end_seq_1) &&
1069 !before(start_seq_0, start_seq_1)) {
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001070 dup_sack = true;
Ilpo Järvinene60402d2007-08-09 15:14:46 +03001071 tcp_dsack_seen(tp);
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07001072 NET_INC_STATS_BH(sock_net(sk),
1073 LINUX_MIB_TCPDSACKOFORECV);
David S. Millerd06e0212007-06-18 22:43:06 -07001074 }
1075 }
1076
1077 /* D-SACK for already forgotten data... Do dumb counting. */
Yuchung Cheng6e08d5e2014-07-02 12:07:16 -07001078 if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
David S. Millerd06e0212007-06-18 22:43:06 -07001079 !after(end_seq_0, prior_snd_una) &&
1080 after(end_seq_0, tp->undo_marker))
1081 tp->undo_retrans--;
1082
1083 return dup_sack;
1084}
1085
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001086struct tcp_sacktag_state {
Eric Dumazet740b0f12014-02-26 14:02:48 -08001087 int reord;
1088 int fack_count;
Kenneth Klette Jonassen31231a82015-05-01 01:10:58 +02001089 /* Timestamps for earliest and latest never-retransmitted segment
1090 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1091 * but congestion control should still get an accurate delay signal.
1092 */
1093 struct skb_mstamp first_sackt;
1094 struct skb_mstamp last_sackt;
Eric Dumazet740b0f12014-02-26 14:02:48 -08001095 int flag;
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001096};
1097
Ilpo Järvinend1935942007-10-11 17:34:25 -07001098/* Check if skb is fully within the SACK block. In presence of GSO skbs,
1099 * the incoming SACK may not exactly match but we can find smaller MSS
1100 * aligned portion of it that matches. Therefore we might need to fragment
1101 * which may fail and creates some hassle (caller must handle error case
1102 * returns).
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001103 *
1104 * FIXME: this could be merged to shift decision code
Ilpo Järvinend1935942007-10-11 17:34:25 -07001105 */
Adrian Bunk0f79efd2007-10-26 03:57:36 -07001106static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001107 u32 start_seq, u32 end_seq)
Ilpo Järvinend1935942007-10-11 17:34:25 -07001108{
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001109 int err;
1110 bool in_sack;
Ilpo Järvinend1935942007-10-11 17:34:25 -07001111 unsigned int pkt_len;
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001112 unsigned int mss;
Ilpo Järvinend1935942007-10-11 17:34:25 -07001113
1114 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1115 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1116
1117 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1118 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001119 mss = tcp_skb_mss(skb);
Ilpo Järvinend1935942007-10-11 17:34:25 -07001120 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1121
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001122 if (!in_sack) {
Ilpo Järvinend1935942007-10-11 17:34:25 -07001123 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001124 if (pkt_len < mss)
1125 pkt_len = mss;
1126 } else {
Ilpo Järvinend1935942007-10-11 17:34:25 -07001127 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001128 if (pkt_len < mss)
1129 return -EINVAL;
1130 }
1131
1132 /* Round if necessary so that SACKs cover only full MSSes
1133 * and/or the remaining small portion (if present)
1134 */
1135 if (pkt_len > mss) {
1136 unsigned int new_len = (pkt_len / mss) * mss;
1137 if (!in_sack && new_len < pkt_len) {
1138 new_len += mss;
Neal Cardwell2cd0d742014-06-18 21:15:03 -04001139 if (new_len >= skb->len)
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001140 return 0;
1141 }
1142 pkt_len = new_len;
1143 }
Octavian Purdila6cc55e02014-06-06 17:32:37 +03001144 err = tcp_fragment(sk, skb, pkt_len, mss, GFP_ATOMIC);
Ilpo Järvinend1935942007-10-11 17:34:25 -07001145 if (err < 0)
1146 return err;
1147 }
1148
1149 return in_sack;
1150}
1151
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001152/* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1153static u8 tcp_sacktag_one(struct sock *sk,
1154 struct tcp_sacktag_state *state, u8 sacked,
1155 u32 start_seq, u32 end_seq,
Eric Dumazet740b0f12014-02-26 14:02:48 -08001156 int dup_sack, int pcount,
1157 const struct skb_mstamp *xmit_time)
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001158{
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001159 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001160 int fack_count = state->fack_count;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001161
1162 /* Account D-SACK for retransmitted packet. */
1163 if (dup_sack && (sacked & TCPCB_RETRANS)) {
Yuchung Cheng6e08d5e2014-07-02 12:07:16 -07001164 if (tp->undo_marker && tp->undo_retrans > 0 &&
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001165 after(end_seq, tp->undo_marker))
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001166 tp->undo_retrans--;
Ilpo Järvinenede9f3b2007-12-02 00:47:58 +02001167 if (sacked & TCPCB_SACKED_ACKED)
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001168 state->reord = min(fack_count, state->reord);
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001169 }
1170
1171 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001172 if (!after(end_seq, tp->snd_una))
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001173 return sacked;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001174
1175 if (!(sacked & TCPCB_SACKED_ACKED)) {
Yuchung Cheng659a8ad2015-10-16 21:57:46 -07001176 tcp_rack_advance(tp, xmit_time, sacked);
1177
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001178 if (sacked & TCPCB_SACKED_RETRANS) {
1179 /* If the segment is not tagged as lost,
1180 * we do not clear RETRANS, believing
1181 * that retransmission is still in flight.
1182 */
1183 if (sacked & TCPCB_LOST) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001184 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001185 tp->lost_out -= pcount;
1186 tp->retrans_out -= pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001187 }
1188 } else {
1189 if (!(sacked & TCPCB_RETRANS)) {
1190 /* New sack for not retransmitted frame,
1191 * which was in hole. It is reordering.
1192 */
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001193 if (before(start_seq,
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001194 tcp_highest_sack_seq(tp)))
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001195 state->reord = min(fack_count,
1196 state->reord);
Yuchung Chenge33099f2013-03-20 13:33:00 +00001197 if (!after(end_seq, tp->high_seq))
1198 state->flag |= FLAG_ORIG_SACK_ACKED;
Kenneth Klette Jonassen31231a82015-05-01 01:10:58 +02001199 if (state->first_sackt.v64 == 0)
1200 state->first_sackt = *xmit_time;
1201 state->last_sackt = *xmit_time;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001202 }
1203
1204 if (sacked & TCPCB_LOST) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001205 sacked &= ~TCPCB_LOST;
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001206 tp->lost_out -= pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001207 }
1208 }
1209
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001210 sacked |= TCPCB_SACKED_ACKED;
1211 state->flag |= FLAG_DATA_SACKED;
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001212 tp->sacked_out += pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001213
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001214 fack_count += pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001215
1216 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
Ian Morris00db4122015-04-03 09:17:27 +01001217 if (!tcp_is_fack(tp) && tp->lost_skb_hint &&
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001218 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001219 tp->lost_cnt_hint += pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001220
1221 if (fack_count > tp->fackets_out)
1222 tp->fackets_out = fack_count;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001223 }
1224
1225 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1226 * frames and clear it. undo_retrans is decreased above, L|R frames
1227 * are accounted above as well.
1228 */
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001229 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1230 sacked &= ~TCPCB_SACKED_RETRANS;
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001231 tp->retrans_out -= pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001232 }
1233
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001234 return sacked;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001235}
1236
Neal Cardwelldaef52b2012-02-12 18:37:10 +00001237/* Shift newly-SACKed bytes from this skb to the immediately previous
1238 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1239 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001240static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1241 struct tcp_sacktag_state *state,
1242 unsigned int pcount, int shifted, int mss,
1243 bool dup_sack)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001244{
1245 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen50133162008-12-05 22:42:41 -08001246 struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
Neal Cardwelldaef52b2012-02-12 18:37:10 +00001247 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1248 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001249
1250 BUG_ON(!pcount);
1251
Neal Cardwell4c90d3b2012-02-26 10:06:19 +00001252 /* Adjust counters and hints for the newly sacked sequence
1253 * range but discard the return value since prev is already
1254 * marked. We must tag the range first because the seq
1255 * advancement below implicitly advances
1256 * tcp_highest_sack_seq() when skb is highest_sack.
1257 */
1258 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
Yuchung Cheng59c9af42013-07-22 16:20:47 -07001259 start_seq, end_seq, dup_sack, pcount,
Eric Dumazet740b0f12014-02-26 14:02:48 -08001260 &skb->skb_mstamp);
Neal Cardwell4c90d3b2012-02-26 10:06:19 +00001261
1262 if (skb == tp->lost_skb_hint)
Neal Cardwell0af2a0d2012-02-13 20:22:08 +00001263 tp->lost_cnt_hint += pcount;
1264
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001265 TCP_SKB_CB(prev)->end_seq += shifted;
1266 TCP_SKB_CB(skb)->seq += shifted;
1267
Eric Dumazetcd7d8492014-09-24 04:11:22 -07001268 tcp_skb_pcount_add(prev, pcount);
1269 BUG_ON(tcp_skb_pcount(skb) < pcount);
1270 tcp_skb_pcount_add(skb, -pcount);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001271
1272 /* When we're adding to gso_segs == 1, gso_size will be zero,
1273 * in theory this shouldn't be necessary but as long as DSACK
1274 * code can come after this skb later on it's better to keep
1275 * setting gso_size to something.
1276 */
Eric Dumazetf69ad292015-06-11 09:15:18 -07001277 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1278 TCP_SKB_CB(prev)->tcp_gso_size = mss;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001279
1280 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
Eric Dumazet51466a72015-06-11 09:15:16 -07001281 if (tcp_skb_pcount(skb) <= 1)
Eric Dumazetf69ad292015-06-11 09:15:18 -07001282 TCP_SKB_CB(skb)->tcp_gso_size = 0;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001283
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001284 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1285 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1286
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001287 if (skb->len > 0) {
1288 BUG_ON(!tcp_skb_pcount(skb));
Ilpo Järvinen111cc8b2008-11-24 21:27:22 -08001289 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001290 return false;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001291 }
1292
1293 /* Whole SKB was eaten :-) */
1294
Ilpo Järvinen92ee76b2008-11-24 21:26:56 -08001295 if (skb == tp->retransmit_skb_hint)
1296 tp->retransmit_skb_hint = prev;
Ilpo Järvinen92ee76b2008-11-24 21:26:56 -08001297 if (skb == tp->lost_skb_hint) {
1298 tp->lost_skb_hint = prev;
1299 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1300 }
1301
Eric Dumazet5e8a402f2013-10-04 10:31:41 -07001302 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1303 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1304 TCP_SKB_CB(prev)->end_seq++;
1305
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001306 if (skb == tcp_highest_sack(sk))
1307 tcp_advance_highest_sack(sk, skb);
1308
1309 tcp_unlink_write_queue(skb, sk);
1310 sk_wmem_free_skb(sk, skb);
1311
Ilpo Järvinen111cc8b2008-11-24 21:27:22 -08001312 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1313
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001314 return true;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001315}
1316
1317/* I wish gso_size would have a bit more sane initialization than
1318 * something-or-zero which complicates things
1319 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001320static int tcp_skb_seglen(const struct sk_buff *skb)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001321{
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001322 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001323}
1324
1325/* Shifting pages past head area doesn't work */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001326static int skb_can_shift(const struct sk_buff *skb)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001327{
1328 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1329}
1330
1331/* Try collapsing SACK blocks spanning across multiple skbs to a single
1332 * skb.
1333 */
1334static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001335 struct tcp_sacktag_state *state,
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001336 u32 start_seq, u32 end_seq,
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001337 bool dup_sack)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001338{
1339 struct tcp_sock *tp = tcp_sk(sk);
1340 struct sk_buff *prev;
1341 int mss;
1342 int pcount = 0;
1343 int len;
1344 int in_sack;
1345
1346 if (!sk_can_gso(sk))
1347 goto fallback;
1348
1349 /* Normally R but no L won't result in plain S */
1350 if (!dup_sack &&
Ilpo Järvinen9969ca52008-12-05 22:41:06 -08001351 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001352 goto fallback;
1353 if (!skb_can_shift(skb))
1354 goto fallback;
1355 /* This frame is about to be dropped (was ACKed). */
1356 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1357 goto fallback;
1358
1359 /* Can only happen with delayed DSACK + discard craziness */
1360 if (unlikely(skb == tcp_write_queue_head(sk)))
1361 goto fallback;
1362 prev = tcp_write_queue_prev(sk, skb);
1363
1364 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1365 goto fallback;
1366
1367 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1368 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1369
1370 if (in_sack) {
1371 len = skb->len;
1372 pcount = tcp_skb_pcount(skb);
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001373 mss = tcp_skb_seglen(skb);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001374
1375 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1376 * drop this restriction as unnecessary
1377 */
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001378 if (mss != tcp_skb_seglen(prev))
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001379 goto fallback;
1380 } else {
1381 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1382 goto noop;
1383 /* CHECKME: This is non-MSS split case only?, this will
1384 * cause skipped skbs due to advancing loop btw, original
1385 * has that feature too
1386 */
1387 if (tcp_skb_pcount(skb) <= 1)
1388 goto noop;
1389
1390 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1391 if (!in_sack) {
1392 /* TODO: head merge to next could be attempted here
1393 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1394 * though it might not be worth of the additional hassle
1395 *
1396 * ...we can probably just fallback to what was done
1397 * previously. We could try merging non-SACKed ones
1398 * as well but it probably isn't going to buy off
1399 * because later SACKs might again split them, and
1400 * it would make skb timestamp tracking considerably
1401 * harder problem.
1402 */
1403 goto fallback;
1404 }
1405
1406 len = end_seq - TCP_SKB_CB(skb)->seq;
1407 BUG_ON(len < 0);
1408 BUG_ON(len > skb->len);
1409
1410 /* MSS boundaries should be honoured or else pcount will
1411 * severely break even though it makes things bit trickier.
1412 * Optimize common case to avoid most of the divides
1413 */
1414 mss = tcp_skb_mss(skb);
1415
1416 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1417 * drop this restriction as unnecessary
1418 */
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001419 if (mss != tcp_skb_seglen(prev))
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001420 goto fallback;
1421
1422 if (len == mss) {
1423 pcount = 1;
1424 } else if (len < mss) {
1425 goto noop;
1426 } else {
1427 pcount = len / mss;
1428 len = pcount * mss;
1429 }
1430 }
1431
Neal Cardwell4648dc92012-03-05 19:35:04 +00001432 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1433 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1434 goto fallback;
1435
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001436 if (!skb_shift(prev, skb, len))
1437 goto fallback;
Ilpo Järvinen9ec06ff2009-03-01 00:21:36 -08001438 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001439 goto out;
1440
1441 /* Hole filled allows collapsing with the next as well, this is very
1442 * useful when hole on every nth skb pattern happens
1443 */
1444 if (prev == tcp_write_queue_tail(sk))
1445 goto out;
1446 skb = tcp_write_queue_next(sk, prev);
1447
Ilpo Järvinenf0bc52f2008-12-05 22:40:47 -08001448 if (!skb_can_shift(skb) ||
1449 (skb == tcp_send_head(sk)) ||
1450 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001451 (mss != tcp_skb_seglen(skb)))
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001452 goto out;
1453
1454 len = skb->len;
1455 if (skb_shift(prev, skb, len)) {
1456 pcount += tcp_skb_pcount(skb);
Ilpo Järvinen9ec06ff2009-03-01 00:21:36 -08001457 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001458 }
1459
1460out:
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001461 state->fack_count += pcount;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001462 return prev;
1463
1464noop:
1465 return skb;
1466
1467fallback:
Ilpo Järvinen111cc8b2008-11-24 21:27:22 -08001468 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001469 return NULL;
1470}
1471
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001472static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1473 struct tcp_sack_block *next_dup,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001474 struct tcp_sacktag_state *state,
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001475 u32 start_seq, u32 end_seq,
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001476 bool dup_sack_in)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001477{
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001478 struct tcp_sock *tp = tcp_sk(sk);
1479 struct sk_buff *tmp;
1480
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001481 tcp_for_write_queue_from(skb, sk) {
1482 int in_sack = 0;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001483 bool dup_sack = dup_sack_in;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001484
1485 if (skb == tcp_send_head(sk))
1486 break;
1487
1488 /* queue is in-order => we can short-circuit the walk early */
1489 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1490 break;
1491
Ian Morris00db4122015-04-03 09:17:27 +01001492 if (next_dup &&
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001493 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1494 in_sack = tcp_match_skb_to_sack(sk, skb,
1495 next_dup->start_seq,
1496 next_dup->end_seq);
1497 if (in_sack > 0)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001498 dup_sack = true;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001499 }
1500
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001501 /* skb reference here is a bit tricky to get right, since
1502 * shifting can eat and free both this skb and the next,
1503 * so not even _safe variant of the loop is enough.
1504 */
1505 if (in_sack <= 0) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001506 tmp = tcp_shift_skb_data(sk, skb, state,
1507 start_seq, end_seq, dup_sack);
Ian Morris00db4122015-04-03 09:17:27 +01001508 if (tmp) {
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001509 if (tmp != skb) {
1510 skb = tmp;
1511 continue;
1512 }
1513
1514 in_sack = 0;
1515 } else {
1516 in_sack = tcp_match_skb_to_sack(sk, skb,
1517 start_seq,
1518 end_seq);
1519 }
1520 }
1521
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001522 if (unlikely(in_sack < 0))
1523 break;
1524
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001525 if (in_sack) {
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001526 TCP_SKB_CB(skb)->sacked =
1527 tcp_sacktag_one(sk,
1528 state,
1529 TCP_SKB_CB(skb)->sacked,
1530 TCP_SKB_CB(skb)->seq,
1531 TCP_SKB_CB(skb)->end_seq,
1532 dup_sack,
Yuchung Cheng59c9af42013-07-22 16:20:47 -07001533 tcp_skb_pcount(skb),
Eric Dumazet740b0f12014-02-26 14:02:48 -08001534 &skb->skb_mstamp);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001535
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001536 if (!before(TCP_SKB_CB(skb)->seq,
1537 tcp_highest_sack_seq(tp)))
1538 tcp_advance_highest_sack(sk, skb);
1539 }
1540
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001541 state->fack_count += tcp_skb_pcount(skb);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001542 }
1543 return skb;
1544}
1545
1546/* Avoid all extra work that is being done by sacktag while walking in
1547 * a normal way
1548 */
1549static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001550 struct tcp_sacktag_state *state,
1551 u32 skip_to_seq)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001552{
1553 tcp_for_write_queue_from(skb, sk) {
1554 if (skb == tcp_send_head(sk))
1555 break;
1556
Ilpo Järvinene8bae272008-11-24 21:12:28 -08001557 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001558 break;
Ilpo Järvinend152a7d2008-03-03 12:10:16 -08001559
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001560 state->fack_count += tcp_skb_pcount(skb);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001561 }
1562 return skb;
1563}
1564
1565static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1566 struct sock *sk,
1567 struct tcp_sack_block *next_dup,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001568 struct tcp_sacktag_state *state,
1569 u32 skip_to_seq)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001570{
Ian Morris51456b22015-04-03 09:17:26 +01001571 if (!next_dup)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001572 return skb;
1573
1574 if (before(next_dup->start_seq, skip_to_seq)) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001575 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1576 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1577 next_dup->start_seq, next_dup->end_seq,
1578 1);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001579 }
1580
1581 return skb;
1582}
1583
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001584static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001585{
1586 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1587}
1588
Linus Torvalds1da177e2005-04-16 15:20:36 -07001589static int
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001590tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001591 u32 prior_snd_una, struct tcp_sacktag_state *state)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001592{
1593 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001594 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1595 TCP_SKB_CB(ack_skb)->sacked);
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001596 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
Adam Langley4389dde2008-07-19 00:07:02 -07001597 struct tcp_sack_block sp[TCP_NUM_SACKS];
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001598 struct tcp_sack_block *cache;
1599 struct sk_buff *skb;
Adam Langley4389dde2008-07-19 00:07:02 -07001600 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001601 int used_sacks;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001602 bool found_dup_sack = false;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001603 int i, j;
Baruch Evenfda03fb2007-02-04 23:35:57 -08001604 int first_sack_index;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001605
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001606 state->flag = 0;
1607 state->reord = tp->packets_out;
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001608
Ilpo Järvinend738cd82007-03-24 21:03:23 -07001609 if (!tp->sacked_out) {
Ilpo Järvinende83c052007-10-07 23:37:55 -07001610 if (WARN_ON(tp->fackets_out))
1611 tp->fackets_out = 0;
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001612 tcp_highest_sack_reset(sk);
Ilpo Järvinend738cd82007-03-24 21:03:23 -07001613 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001614
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07001615 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
David S. Millerd06e0212007-06-18 22:43:06 -07001616 num_sacks, prior_snd_una);
1617 if (found_dup_sack)
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001618 state->flag |= FLAG_DSACKING_ACK;
Baruch Even6f746512007-02-04 23:36:42 -08001619
1620 /* Eliminate too old ACKs, but take into
1621 * account more or less fresh ones, they can
1622 * contain valid SACK info.
1623 */
1624 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1625 return 0;
1626
Ilpo Järvinen96a2d412007-11-14 15:47:18 -08001627 if (!tp->packets_out)
1628 goto out;
1629
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001630 used_sacks = 0;
1631 first_sack_index = 0;
1632 for (i = 0; i < num_sacks; i++) {
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001633 bool dup_sack = !i && found_dup_sack;
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001634
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07001635 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1636 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001637
1638 if (!tcp_is_sackblock_valid(tp, dup_sack,
1639 sp[used_sacks].start_seq,
1640 sp[used_sacks].end_seq)) {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001641 int mib_idx;
1642
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001643 if (dup_sack) {
1644 if (!tp->undo_marker)
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001645 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001646 else
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001647 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001648 } else {
1649 /* Don't count olds caused by ACK reordering */
1650 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1651 !after(sp[used_sacks].end_seq, tp->snd_una))
1652 continue;
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001653 mib_idx = LINUX_MIB_TCPSACKDISCARD;
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001654 }
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001655
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07001656 NET_INC_STATS_BH(sock_net(sk), mib_idx);
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001657 if (i == 0)
1658 first_sack_index = -1;
1659 continue;
1660 }
1661
1662 /* Ignore very old stuff early */
1663 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1664 continue;
1665
1666 used_sacks++;
1667 }
1668
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001669 /* order SACK blocks to allow in order walk of the retrans queue */
1670 for (i = used_sacks - 1; i > 0; i--) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001671 for (j = 0; j < i; j++) {
1672 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
Ilpo Järvinena0bffff2009-03-21 13:36:17 -07001673 swap(sp[j], sp[j + 1]);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001674
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001675 /* Track where the first SACK block goes to */
1676 if (j == first_sack_index)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001677 first_sack_index = j + 1;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001678 }
1679 }
1680 }
1681
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001682 skb = tcp_write_queue_head(sk);
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001683 state->fack_count = 0;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001684 i = 0;
1685
1686 if (!tp->sacked_out) {
1687 /* It's already past, so skip checking against it */
1688 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1689 } else {
1690 cache = tp->recv_sack_cache;
1691 /* Skip empty blocks in at head of the cache */
1692 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1693 !cache->end_seq)
1694 cache++;
Baruch Evenfda03fb2007-02-04 23:35:57 -08001695 }
1696
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001697 while (i < used_sacks) {
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001698 u32 start_seq = sp[i].start_seq;
1699 u32 end_seq = sp[i].end_seq;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001700 bool dup_sack = (found_dup_sack && (i == first_sack_index));
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001701 struct tcp_sack_block *next_dup = NULL;
Ilpo Järvinene56d6cd2007-11-01 00:09:37 -07001702
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001703 if (found_dup_sack && ((i + 1) == first_sack_index))
1704 next_dup = &sp[i + 1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001705
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001706 /* Skip too early cached blocks */
1707 while (tcp_sack_cache_ok(tp, cache) &&
1708 !before(start_seq, cache->end_seq))
1709 cache++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001710
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001711 /* Can skip some work by looking recv_sack_cache? */
1712 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1713 after(end_seq, cache->start_seq)) {
David S. Millerfe067e82007-03-07 12:12:44 -08001714
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001715 /* Head todo? */
1716 if (before(start_seq, cache->start_seq)) {
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001717 skb = tcp_sacktag_skip(skb, sk, state,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001718 start_seq);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001719 skb = tcp_sacktag_walk(skb, sk, next_dup,
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001720 state,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001721 start_seq,
1722 cache->start_seq,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001723 dup_sack);
Baruch Evenfda03fb2007-02-04 23:35:57 -08001724 }
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001725
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001726 /* Rest of the block already fully processed? */
Ilpo Järvinen20de20b2007-11-16 16:17:05 -08001727 if (!after(end_seq, cache->end_seq))
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001728 goto advance_sp;
Ilpo Järvinen20de20b2007-11-16 16:17:05 -08001729
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001730 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001731 state,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001732 cache->end_seq);
Ilpo Järvinene56d6cd2007-11-01 00:09:37 -07001733
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001734 /* ...tail remains todo... */
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001735 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
Ilpo Järvinen20de20b2007-11-16 16:17:05 -08001736 /* ...but better entrypoint exists! */
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001737 skb = tcp_highest_sack(sk);
Ian Morris51456b22015-04-03 09:17:26 +01001738 if (!skb)
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001739 break;
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001740 state->fack_count = tp->fackets_out;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001741 cache++;
1742 goto walk;
1743 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001744
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001745 skb = tcp_sacktag_skip(skb, sk, state, cache->end_seq);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001746 /* Check overlap against next cached too (past this one already) */
1747 cache++;
1748 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001749 }
Ilpo Järvinenfbd52eb2007-11-10 21:24:19 -08001750
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001751 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1752 skb = tcp_highest_sack(sk);
Ian Morris51456b22015-04-03 09:17:26 +01001753 if (!skb)
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001754 break;
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001755 state->fack_count = tp->fackets_out;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001756 }
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001757 skb = tcp_sacktag_skip(skb, sk, state, start_seq);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001758
1759walk:
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001760 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001761 start_seq, end_seq, dup_sack);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001762
1763advance_sp:
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001764 i++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001765 }
1766
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001767 /* Clear the head of the cache sack blocks so we can skip it next time */
1768 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1769 tp->recv_sack_cache[i].start_seq = 0;
1770 tp->recv_sack_cache[i].end_seq = 0;
1771 }
1772 for (j = 0; j < used_sacks; j++)
1773 tp->recv_sack_cache[i++] = sp[j];
1774
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001775 if ((state->reord < tp->fackets_out) &&
Yuchung Cheng9b441902013-03-20 13:32:58 +00001776 ((inet_csk(sk)->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker))
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001777 tcp_update_reordering(sk, tp->fackets_out - state->reord, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001778
Yuchung Cheng9dac8832015-04-29 11:28:30 -07001779 tcp_verify_left_out(tp);
Ilpo Järvinen96a2d412007-11-14 15:47:18 -08001780out:
1781
Linus Torvalds1da177e2005-04-16 15:20:36 -07001782#if FASTRETRANS_DEBUG > 0
Ilpo Järvinen547b7922008-07-25 21:43:18 -07001783 WARN_ON((int)tp->sacked_out < 0);
1784 WARN_ON((int)tp->lost_out < 0);
1785 WARN_ON((int)tp->retrans_out < 0);
1786 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001787#endif
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001788 return state->flag;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001789}
1790
Ilpo Järvinen882beba2008-04-07 22:33:07 -07001791/* Limits sacked_out so that sum with lost_out isn't ever larger than
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001792 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
Ilpo Järvinen30935cf2007-02-21 23:01:36 -08001793 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001794static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001795{
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001796 u32 holes;
1797
1798 holes = max(tp->lost_out, 1U);
1799 holes = min(holes, tp->packets_out);
1800
1801 if ((tp->sacked_out + holes) > tp->packets_out) {
1802 tp->sacked_out = tp->packets_out - holes;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001803 return true;
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001804 }
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001805 return false;
Ilpo Järvinen882beba2008-04-07 22:33:07 -07001806}
1807
1808/* If we receive more dupacks than we expected counting segments
1809 * in assumption of absent reordering, interpret this as reordering.
1810 * The only another reason could be bug in receiver TCP.
1811 */
1812static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1813{
1814 struct tcp_sock *tp = tcp_sk(sk);
1815 if (tcp_limit_reno_sacked(tp))
1816 tcp_update_reordering(sk, tp->packets_out + addend, 0);
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001817}
1818
1819/* Emulate SACKs for SACKless connection: account for a new dupack. */
1820
1821static void tcp_add_reno_sack(struct sock *sk)
1822{
1823 struct tcp_sock *tp = tcp_sk(sk);
1824 tp->sacked_out++;
1825 tcp_check_reno_reordering(sk, 0);
Ilpo Järvinen005903b2007-08-09 14:44:16 +03001826 tcp_verify_left_out(tp);
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001827}
1828
1829/* Account for ACK, ACKing some data in Reno Recovery phase. */
1830
1831static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1832{
1833 struct tcp_sock *tp = tcp_sk(sk);
1834
1835 if (acked > 0) {
1836 /* One ACK acked hole. The rest eat duplicate ACKs. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001837 if (acked - 1 >= tp->sacked_out)
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001838 tp->sacked_out = 0;
1839 else
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001840 tp->sacked_out -= acked - 1;
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001841 }
1842 tcp_check_reno_reordering(sk, acked);
Ilpo Järvinen005903b2007-08-09 14:44:16 +03001843 tcp_verify_left_out(tp);
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001844}
1845
1846static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1847{
1848 tp->sacked_out = 0;
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001849}
1850
Yuchung Cheng989e04c2014-08-22 14:15:22 -07001851void tcp_clear_retrans(struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001852{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001853 tp->retrans_out = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001854 tp->lost_out = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001855 tp->undo_marker = 0;
Yuchung Cheng6e08d5e2014-07-02 12:07:16 -07001856 tp->undo_retrans = -1;
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07001857 tp->fackets_out = 0;
1858 tp->sacked_out = 0;
1859}
1860
Yuchung Cheng989e04c2014-08-22 14:15:22 -07001861static inline void tcp_init_undo(struct tcp_sock *tp)
1862{
1863 tp->undo_marker = tp->snd_una;
1864 /* Retransmission still in flight may cause DSACKs later. */
1865 tp->undo_retrans = tp->retrans_out ? : -1;
1866}
1867
Neal Cardwell5ae344c2014-08-04 19:12:29 -04001868/* Enter Loss state. If we detect SACK reneging, forget all SACK information
Linus Torvalds1da177e2005-04-16 15:20:36 -07001869 * and reset tags completely, otherwise preserve SACKs. If receiver
1870 * dropped its ofo queue, we will know this due to reneging detection.
1871 */
Neal Cardwell5ae344c2014-08-04 19:12:29 -04001872void tcp_enter_loss(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001873{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001874 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001875 struct tcp_sock *tp = tcp_sk(sk);
1876 struct sk_buff *skb;
Yuchung Chengf82b6812015-07-13 12:10:20 -07001877 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
Neal Cardwell5ae344c2014-08-04 19:12:29 -04001878 bool is_reneg; /* is receiver reneging on SACKs? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001879
1880 /* Reduce ssthresh if it has not yet been made inside this window. */
Yuchung Chenge33099f2013-03-20 13:33:00 +00001881 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1882 !after(tp->high_seq, tp->snd_una) ||
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001883 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1884 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1885 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1886 tcp_ca_event(sk, CA_EVENT_LOSS);
Yuchung Cheng989e04c2014-08-22 14:15:22 -07001887 tcp_init_undo(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001888 }
1889 tp->snd_cwnd = 1;
1890 tp->snd_cwnd_cnt = 0;
1891 tp->snd_cwnd_stamp = tcp_time_stamp;
1892
Yuchung Cheng989e04c2014-08-22 14:15:22 -07001893 tp->retrans_out = 0;
1894 tp->lost_out = 0;
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07001895
1896 if (tcp_is_reno(tp))
1897 tcp_reset_reno_sack(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001898
Neal Cardwell5ae344c2014-08-04 19:12:29 -04001899 skb = tcp_write_queue_head(sk);
1900 is_reneg = skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED);
1901 if (is_reneg) {
1902 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07001903 tp->sacked_out = 0;
1904 tp->fackets_out = 0;
Ilpo Järvinenb7689202007-09-20 11:37:19 -07001905 }
Ilpo Järvinen64edc272008-09-20 21:18:32 -07001906 tcp_clear_all_retrans_hints(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001907
David S. Millerfe067e82007-03-07 12:12:44 -08001908 tcp_for_write_queue(skb, sk) {
1909 if (skb == tcp_send_head(sk))
1910 break;
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07001911
Linus Torvalds1da177e2005-04-16 15:20:36 -07001912 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
Neal Cardwell5ae344c2014-08-04 19:12:29 -04001913 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || is_reneg) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001914 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1915 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1916 tp->lost_out += tcp_skb_pcount(skb);
Ilpo Järvinen006f5822008-09-20 21:20:20 -07001917 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001918 }
1919 }
Ilpo Järvinen005903b2007-08-09 14:44:16 +03001920 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001921
Yuchung Cheng74c181d2013-08-12 16:41:25 -07001922 /* Timeout in disordered state after receiving substantial DUPACKs
1923 * suggests that the degree of reordering is over-estimated.
1924 */
1925 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
1926 tp->sacked_out >= sysctl_tcp_reordering)
1927 tp->reordering = min_t(unsigned int, tp->reordering,
1928 sysctl_tcp_reordering);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001929 tcp_set_ca_state(sk, TCP_CA_Loss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001930 tp->high_seq = tp->snd_nxt;
Florian Westphal735d3832014-09-29 13:08:30 +02001931 tcp_ecn_queue_cwr(tp);
Yuchung Chenge33099f2013-03-20 13:33:00 +00001932
1933 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1934 * loss recovery is underway except recurring timeout(s) on
1935 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1936 */
1937 tp->frto = sysctl_tcp_frto &&
1938 (new_recovery || icsk->icsk_retransmits) &&
1939 !inet_csk(sk)->icsk_mtup.probe_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001940}
1941
Ilpo Järvinencadbd032007-12-31 04:49:21 -08001942/* If ACK arrived pointing to a remembered SACK, it means that our
1943 * remembered SACKs do not reflect real state of receiver i.e.
1944 * receiver _host_ is heavily congested (or buggy).
1945 *
Neal Cardwell5ae344c2014-08-04 19:12:29 -04001946 * To avoid big spurious retransmission bursts due to transient SACK
1947 * scoreboard oddities that look like reneging, we give the receiver a
1948 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
1949 * restore sanity to the SACK scoreboard. If the apparent reneging
1950 * persists until this RTO then we'll clear the SACK scoreboard.
Ilpo Järvinencadbd032007-12-31 04:49:21 -08001951 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001952static bool tcp_check_sack_reneging(struct sock *sk, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001953{
Ilpo Järvinencadbd032007-12-31 04:49:21 -08001954 if (flag & FLAG_SACK_RENEGING) {
Neal Cardwell5ae344c2014-08-04 19:12:29 -04001955 struct tcp_sock *tp = tcp_sk(sk);
1956 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
1957 msecs_to_jiffies(10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001958
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001959 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
Neal Cardwell5ae344c2014-08-04 19:12:29 -04001960 delay, TCP_RTO_MAX);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001961 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001962 }
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001963 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001964}
1965
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001966static inline int tcp_fackets_out(const struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001967{
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001968 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001969}
1970
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08001971/* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
1972 * counter when SACK is enabled (without SACK, sacked_out is used for
1973 * that purpose).
1974 *
1975 * Instead, with FACK TCP uses fackets_out that includes both SACKed
1976 * segments up to the highest received SACK block so far and holes in
1977 * between them.
1978 *
1979 * With reordering, holes may still be in flight, so RFC3517 recovery
1980 * uses pure sacked_out (total number of SACKed segments) even though
1981 * it violates the RFC that uses duplicate ACKs, often these are equal
1982 * but when e.g. out-of-window ACKs or packet duplication occurs,
1983 * they differ. Since neither occurs due to loss, TCP should really
1984 * ignore them.
1985 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001986static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08001987{
1988 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
1989}
1990
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00001991static bool tcp_pause_early_retransmit(struct sock *sk, int flag)
1992{
1993 struct tcp_sock *tp = tcp_sk(sk);
1994 unsigned long delay;
1995
1996 /* Delay early retransmit and entering fast recovery for
1997 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
1998 * available, or RTO is scheduled to fire first.
1999 */
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002000 if (sysctl_tcp_early_retrans < 2 || sysctl_tcp_early_retrans > 3 ||
Eric Dumazet740b0f12014-02-26 14:02:48 -08002001 (flag & FLAG_ECE) || !tp->srtt_us)
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00002002 return false;
2003
Eric Dumazet740b0f12014-02-26 14:02:48 -08002004 delay = max(usecs_to_jiffies(tp->srtt_us >> 5),
2005 msecs_to_jiffies(2));
2006
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00002007 if (!time_after(inet_csk(sk)->icsk_timeout, (jiffies + delay)))
2008 return false;
2009
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002010 inet_csk_reset_xmit_timer(sk, ICSK_TIME_EARLY_RETRANS, delay,
2011 TCP_RTO_MAX);
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00002012 return true;
2013}
2014
Linus Torvalds1da177e2005-04-16 15:20:36 -07002015/* Linux NewReno/SACK/FACK/ECN state machine.
2016 * --------------------------------------
2017 *
2018 * "Open" Normal state, no dubious events, fast path.
2019 * "Disorder" In all the respects it is "Open",
2020 * but requires a bit more attention. It is entered when
2021 * we see some SACKs or dupacks. It is split of "Open"
2022 * mainly to move some processing from fast path to slow one.
2023 * "CWR" CWND was reduced due to some Congestion Notification event.
2024 * It can be ECN, ICMP source quench, local device congestion.
2025 * "Recovery" CWND was reduced, we are fast-retransmitting.
2026 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2027 *
2028 * tcp_fastretrans_alert() is entered:
2029 * - each incoming ACK, if state is not "Open"
2030 * - when arrived ACK is unusual, namely:
2031 * * SACK
2032 * * Duplicate ACK.
2033 * * ECN ECE.
2034 *
2035 * Counting packets in flight is pretty simple.
2036 *
2037 * in_flight = packets_out - left_out + retrans_out
2038 *
2039 * packets_out is SND.NXT-SND.UNA counted in packets.
2040 *
2041 * retrans_out is number of retransmitted segments.
2042 *
2043 * left_out is number of segments left network, but not ACKed yet.
2044 *
2045 * left_out = sacked_out + lost_out
2046 *
2047 * sacked_out: Packets, which arrived to receiver out of order
2048 * and hence not ACKed. With SACKs this number is simply
2049 * amount of SACKed data. Even without SACKs
2050 * it is easy to give pretty reliable estimate of this number,
2051 * counting duplicate ACKs.
2052 *
2053 * lost_out: Packets lost by network. TCP has no explicit
2054 * "loss notification" feedback from network (for now).
2055 * It means that this number can be only _guessed_.
2056 * Actually, it is the heuristics to predict lossage that
2057 * distinguishes different algorithms.
2058 *
2059 * F.e. after RTO, when all the queue is considered as lost,
2060 * lost_out = packets_out and in_flight = retrans_out.
2061 *
2062 * Essentially, we have now two algorithms counting
2063 * lost packets.
2064 *
2065 * FACK: It is the simplest heuristics. As soon as we decided
2066 * that something is lost, we decide that _all_ not SACKed
2067 * packets until the most forward SACK are lost. I.e.
2068 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2069 * It is absolutely correct estimate, if network does not reorder
2070 * packets. And it loses any connection to reality when reordering
2071 * takes place. We use FACK by default until reordering
2072 * is suspected on the path to this destination.
2073 *
2074 * NewReno: when Recovery is entered, we assume that one segment
2075 * is lost (classic Reno). While we are in Recovery and
2076 * a partial ACK arrives, we assume that one more packet
2077 * is lost (NewReno). This heuristics are the same in NewReno
2078 * and SACK.
2079 *
2080 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2081 * deflation etc. CWND is real congestion window, never inflated, changes
2082 * only according to classic VJ rules.
2083 *
2084 * Really tricky (and requiring careful tuning) part of algorithm
2085 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2086 * The first determines the moment _when_ we should reduce CWND and,
2087 * hence, slow down forward transmission. In fact, it determines the moment
2088 * when we decide that hole is caused by loss, rather than by a reorder.
2089 *
2090 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2091 * holes, caused by lost packets.
2092 *
2093 * And the most logically complicated part of algorithm is undo
2094 * heuristics. We detect false retransmits due to both too early
2095 * fast retransmit (reordering) and underestimated RTO, analyzing
2096 * timestamps and D-SACKs. When we detect that some segments were
2097 * retransmitted by mistake and CWND reduction was wrong, we undo
2098 * window reduction and abort recovery phase. This logic is hidden
2099 * inside several functions named tcp_try_undo_<something>.
2100 */
2101
2102/* This function decides, when we should leave Disordered state
2103 * and enter Recovery phase, reducing congestion window.
2104 *
2105 * Main question: may we further continue forward transmission
2106 * with the same cwnd?
2107 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002108static bool tcp_time_to_recover(struct sock *sk, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002109{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002110 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002111 __u32 packets_out;
2112
2113 /* Trick#1: The loss is proven. */
2114 if (tp->lost_out)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002115 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002116
2117 /* Not-A-Trick#2 : Classic rule... */
Andreas Petlundea84e552009-10-27 03:27:21 +00002118 if (tcp_dupack_heuristics(tp) > tp->reordering)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002119 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002120
Linus Torvalds1da177e2005-04-16 15:20:36 -07002121 /* Trick#4: It is still not OK... But will it be useful to delay
2122 * recovery more?
2123 */
2124 packets_out = tp->packets_out;
2125 if (packets_out <= tp->reordering &&
2126 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002127 !tcp_may_send_now(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002128 /* We have nothing to send. This connection is limited
2129 * either by receiver window or by application.
2130 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002131 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002132 }
2133
Andreas Petlund7e380172010-02-18 04:48:19 +00002134 /* If a thin stream is detected, retransmit after first
2135 * received dupack. Employ only if SACK is supported in order
2136 * to avoid possible corner-case series of spurious retransmissions
2137 * Use only if there are no unsent data.
2138 */
2139 if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2140 tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2141 tcp_is_sack(tp) && !tcp_send_head(sk))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002142 return true;
Andreas Petlund7e380172010-02-18 04:48:19 +00002143
Yuchung Chengeed530b2012-05-02 13:30:03 +00002144 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious
2145 * retransmissions due to small network reorderings, we implement
2146 * Mitigation A.3 in the RFC and delay the retransmission for a short
2147 * interval if appropriate.
2148 */
2149 if (tp->do_early_retrans && !tp->retrans_out && tp->sacked_out &&
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002150 (tp->packets_out >= (tp->sacked_out + 1) && tp->packets_out < 4) &&
Yuchung Chengeed530b2012-05-02 13:30:03 +00002151 !tcp_may_send_now(sk))
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00002152 return !tcp_pause_early_retransmit(sk, flag);
Yuchung Chengeed530b2012-05-02 13:30:03 +00002153
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002154 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002155}
2156
Yuchung Cheng974c1232012-01-19 14:42:21 +00002157/* Detect loss in event "A" above by marking head of queue up as lost.
2158 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2159 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2160 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2161 * the maximum SACKed segments to pass before reaching this limit.
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002162 */
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002163static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002164{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002165 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002166 struct sk_buff *skb;
Ilpo Järvinenc137f3d2008-04-07 22:32:38 -07002167 int cnt, oldcnt;
2168 int err;
2169 unsigned int mss;
Yuchung Cheng974c1232012-01-19 14:42:21 +00002170 /* Use SACK to deduce losses of new sequences sent during recovery */
2171 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002172
Ilpo Järvinen547b7922008-07-25 21:43:18 -07002173 WARN_ON(packets > tp->packets_out);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002174 if (tp->lost_skb_hint) {
2175 skb = tp->lost_skb_hint;
2176 cnt = tp->lost_cnt_hint;
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002177 /* Head already handled? */
2178 if (mark_head && skb != tcp_write_queue_head(sk))
2179 return;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002180 } else {
David S. Millerfe067e82007-03-07 12:12:44 -08002181 skb = tcp_write_queue_head(sk);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002182 cnt = 0;
2183 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002184
David S. Millerfe067e82007-03-07 12:12:44 -08002185 tcp_for_write_queue_from(skb, sk) {
2186 if (skb == tcp_send_head(sk))
2187 break;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002188 /* TODO: do this better */
2189 /* this is not the most efficient way to do this... */
2190 tp->lost_skb_hint = skb;
2191 tp->lost_cnt_hint = cnt;
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002192
Yuchung Cheng974c1232012-01-19 14:42:21 +00002193 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
Ilpo Järvinenc137f3d2008-04-07 22:32:38 -07002194 break;
2195
2196 oldcnt = cnt;
Ilpo Järvinenad1984e2008-01-30 20:06:02 -08002197 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002198 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2199 cnt += tcp_skb_pcount(skb);
2200
Ilpo Järvinenc137f3d2008-04-07 22:32:38 -07002201 if (cnt > packets) {
Yuchung Chengb3de7552010-09-24 13:22:06 +00002202 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
Neal Cardwellc0638c22012-03-02 21:36:51 +00002203 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
Yuchung Chengb3de7552010-09-24 13:22:06 +00002204 (oldcnt >= packets))
Ilpo Järvinenc137f3d2008-04-07 22:32:38 -07002205 break;
2206
Eric Dumazetf69ad292015-06-11 09:15:18 -07002207 mss = tcp_skb_mss(skb);
Octavian Purdila6cc55e02014-06-06 17:32:37 +03002208 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss,
2209 mss, GFP_ATOMIC);
Ilpo Järvinenc137f3d2008-04-07 22:32:38 -07002210 if (err < 0)
2211 break;
2212 cnt = packets;
2213 }
2214
Ilpo Järvinen41ea36e2008-09-20 21:19:22 -07002215 tcp_skb_mark_lost(tp, skb);
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002216
2217 if (mark_head)
2218 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002219 }
Ilpo Järvinen005903b2007-08-09 14:44:16 +03002220 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002221}
2222
2223/* Account newly detected lost packet(s) */
2224
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002225static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002226{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002227 struct tcp_sock *tp = tcp_sk(sk);
2228
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002229 if (tcp_is_reno(tp)) {
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002230 tcp_mark_head_lost(sk, 1, 1);
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002231 } else if (tcp_is_fack(tp)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002232 int lost = tp->fackets_out - tp->reordering;
2233 if (lost <= 0)
2234 lost = 1;
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002235 tcp_mark_head_lost(sk, lost, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002236 } else {
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002237 int sacked_upto = tp->sacked_out - tp->reordering;
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002238 if (sacked_upto >= 0)
2239 tcp_mark_head_lost(sk, sacked_upto, 0);
2240 else if (fast_rexmit)
2241 tcp_mark_head_lost(sk, 1, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002242 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002243}
2244
2245/* CWND moderation, preventing bursts due to too big ACKs
2246 * in dubious situations.
2247 */
2248static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2249{
2250 tp->snd_cwnd = min(tp->snd_cwnd,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002251 tcp_packets_in_flight(tp) + tcp_max_burst(tp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002252 tp->snd_cwnd_stamp = tcp_time_stamp;
2253}
2254
Yuchung Cheng77c63122015-10-16 21:57:44 -07002255static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2256{
2257 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2258 before(tp->rx_opt.rcv_tsecr, when);
2259}
2260
Yuchung Cheng659a8ad2015-10-16 21:57:46 -07002261/* skb is spurious retransmitted if the returned timestamp echo
2262 * reply is prior to the skb transmission time
2263 */
2264static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2265 const struct sk_buff *skb)
2266{
2267 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2268 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb));
2269}
2270
Linus Torvalds1da177e2005-04-16 15:20:36 -07002271/* Nothing was retransmitted or returned timestamp is less
2272 * than timestamp of the first retransmission.
2273 */
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00002274static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002275{
2276 return !tp->retrans_stamp ||
Yuchung Cheng77c63122015-10-16 21:57:44 -07002277 tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002278}
2279
2280/* Undo procedures. */
2281
Marcelo Leitner1f37bf82014-11-04 17:15:08 -02002282/* We can clear retrans_stamp when there are no retransmissions in the
2283 * window. It would seem that it is trivially available for us in
2284 * tp->retrans_out, however, that kind of assumptions doesn't consider
2285 * what will happen if errors occur when sending retransmission for the
2286 * second time. ...It could the that such segment has only
2287 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2288 * the head skb is enough except for some reneging corner cases that
2289 * are not worth the effort.
2290 *
2291 * Main reason for all this complexity is the fact that connection dying
2292 * time now depends on the validity of the retrans_stamp, in particular,
2293 * that successive retransmissions of a segment must not advance
2294 * retrans_stamp under any conditions.
2295 */
2296static bool tcp_any_retrans_done(const struct sock *sk)
2297{
2298 const struct tcp_sock *tp = tcp_sk(sk);
2299 struct sk_buff *skb;
2300
2301 if (tp->retrans_out)
2302 return true;
2303
2304 skb = tcp_write_queue_head(sk);
2305 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2306 return true;
2307
2308 return false;
2309}
2310
Linus Torvalds1da177e2005-04-16 15:20:36 -07002311#if FASTRETRANS_DEBUG > 1
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002312static void DBGUNDO(struct sock *sk, const char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002313{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002314 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002315 struct inet_sock *inet = inet_sk(sk);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002316
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002317 if (sk->sk_family == AF_INET) {
Joe Perches91df42b2012-05-15 14:11:54 +00002318 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2319 msg,
2320 &inet->inet_daddr, ntohs(inet->inet_dport),
2321 tp->snd_cwnd, tcp_left_out(tp),
2322 tp->snd_ssthresh, tp->prior_ssthresh,
2323 tp->packets_out);
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002324 }
Eric Dumazetdfd56b82011-12-10 09:48:31 +00002325#if IS_ENABLED(CONFIG_IPV6)
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002326 else if (sk->sk_family == AF_INET6) {
2327 struct ipv6_pinfo *np = inet6_sk(sk);
Joe Perches91df42b2012-05-15 14:11:54 +00002328 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2329 msg,
2330 &np->daddr, ntohs(inet->inet_dport),
2331 tp->snd_cwnd, tcp_left_out(tp),
2332 tp->snd_ssthresh, tp->prior_ssthresh,
2333 tp->packets_out);
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002334 }
2335#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002336}
2337#else
2338#define DBGUNDO(x...) do { } while (0)
2339#endif
2340
Yuchung Cheng7026b912013-05-29 14:20:13 +00002341static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002342{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002343 struct tcp_sock *tp = tcp_sk(sk);
2344
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002345 if (unmark_loss) {
2346 struct sk_buff *skb;
2347
2348 tcp_for_write_queue(skb, sk) {
2349 if (skb == tcp_send_head(sk))
2350 break;
2351 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2352 }
2353 tp->lost_out = 0;
2354 tcp_clear_all_retrans_hints(tp);
2355 }
2356
Linus Torvalds1da177e2005-04-16 15:20:36 -07002357 if (tp->prior_ssthresh) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002358 const struct inet_connection_sock *icsk = inet_csk(sk);
2359
2360 if (icsk->icsk_ca_ops->undo_cwnd)
2361 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002362 else
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002363 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002364
Yuchung Cheng7026b912013-05-29 14:20:13 +00002365 if (tp->prior_ssthresh > tp->snd_ssthresh) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002366 tp->snd_ssthresh = tp->prior_ssthresh;
Florian Westphal735d3832014-09-29 13:08:30 +02002367 tcp_ecn_withdraw_cwr(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002368 }
2369 } else {
2370 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2371 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002372 tp->snd_cwnd_stamp = tcp_time_stamp;
Yuchung Cheng7026b912013-05-29 14:20:13 +00002373 tp->undo_marker = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002374}
2375
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00002376static inline bool tcp_may_undo(const struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002377{
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002378 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002379}
2380
2381/* People celebrate: "We love our President!" */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002382static bool tcp_try_undo_recovery(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002383{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002384 struct tcp_sock *tp = tcp_sk(sk);
2385
Linus Torvalds1da177e2005-04-16 15:20:36 -07002386 if (tcp_may_undo(tp)) {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07002387 int mib_idx;
2388
Linus Torvalds1da177e2005-04-16 15:20:36 -07002389 /* Happy end! We did not retransmit anything
2390 * or our original transmission succeeded.
2391 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002392 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
Yuchung Cheng7026b912013-05-29 14:20:13 +00002393 tcp_undo_cwnd_reduction(sk, false);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002394 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07002395 mib_idx = LINUX_MIB_TCPLOSSUNDO;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002396 else
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07002397 mib_idx = LINUX_MIB_TCPFULLUNDO;
2398
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07002399 NET_INC_STATS_BH(sock_net(sk), mib_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002400 }
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002401 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002402 /* Hold old state until something *above* high_seq
2403 * is ACKed. For Reno it is MUST to prevent false
2404 * fast retransmits (RFC2582). SACK TCP is safe. */
2405 tcp_moderate_cwnd(tp);
Marcelo Leitner1f37bf82014-11-04 17:15:08 -02002406 if (!tcp_any_retrans_done(sk))
2407 tp->retrans_stamp = 0;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002408 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002409 }
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002410 tcp_set_ca_state(sk, TCP_CA_Open);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002411 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002412}
2413
2414/* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
Yuchung Chengc7d9d6a2013-05-29 14:20:14 +00002415static bool tcp_try_undo_dsack(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002416{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002417 struct tcp_sock *tp = tcp_sk(sk);
2418
Linus Torvalds1da177e2005-04-16 15:20:36 -07002419 if (tp->undo_marker && !tp->undo_retrans) {
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002420 DBGUNDO(sk, "D-SACK");
Yuchung Cheng7026b912013-05-29 14:20:13 +00002421 tcp_undo_cwnd_reduction(sk, false);
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07002422 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
Yuchung Chengc7d9d6a2013-05-29 14:20:14 +00002423 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002424 }
Yuchung Chengc7d9d6a2013-05-29 14:20:14 +00002425 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002426}
2427
Yuchung Chenge33099f2013-03-20 13:33:00 +00002428/* Undo during loss recovery after partial ACK or using F-RTO. */
2429static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002430{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002431 struct tcp_sock *tp = tcp_sk(sk);
2432
Yuchung Chenge33099f2013-03-20 13:33:00 +00002433 if (frto_undo || tcp_may_undo(tp)) {
Yuchung Cheng7026b912013-05-29 14:20:13 +00002434 tcp_undo_cwnd_reduction(sk, true);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002435
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002436 DBGUNDO(sk, "partial loss");
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07002437 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
Yuchung Chenge33099f2013-03-20 13:33:00 +00002438 if (frto_undo)
2439 NET_INC_STATS_BH(sock_net(sk),
2440 LINUX_MIB_TCPSPURIOUSRTOS);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002441 inet_csk(sk)->icsk_retransmits = 0;
Yuchung Chenge33099f2013-03-20 13:33:00 +00002442 if (frto_undo || tcp_is_sack(tp))
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002443 tcp_set_ca_state(sk, TCP_CA_Open);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002444 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002445 }
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002446 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002447}
2448
Yuchung Cheng37598242015-07-01 14:11:15 -07002449/* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002450 * It computes the number of packets to send (sndcnt) based on packets newly
2451 * delivered:
2452 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2453 * cwnd reductions across a full RTT.
Yuchung Cheng37598242015-07-01 14:11:15 -07002454 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2455 * But when the retransmits are acked without further losses, PRR
2456 * slow starts cwnd up to ssthresh to speed up the recovery.
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002457 */
Christoph Paasch5ee2c942014-07-14 16:58:32 +02002458static void tcp_init_cwnd_reduction(struct sock *sk)
Yuchung Cheng684bad12012-09-02 17:38:04 +00002459{
2460 struct tcp_sock *tp = tcp_sk(sk);
2461
2462 tp->high_seq = tp->snd_nxt;
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00002463 tp->tlp_high_seq = 0;
Yuchung Cheng684bad12012-09-02 17:38:04 +00002464 tp->snd_cwnd_cnt = 0;
2465 tp->prior_cwnd = tp->snd_cwnd;
2466 tp->prr_delivered = 0;
2467 tp->prr_out = 0;
Christoph Paasch5ee2c942014-07-14 16:58:32 +02002468 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
Florian Westphal735d3832014-09-29 13:08:30 +02002469 tcp_ecn_queue_cwr(tp);
Yuchung Cheng684bad12012-09-02 17:38:04 +00002470}
2471
Yuchung Cheng68049732013-05-29 14:20:11 +00002472static void tcp_cwnd_reduction(struct sock *sk, const int prior_unsacked,
Yuchung Cheng37598242015-07-01 14:11:15 -07002473 int fast_rexmit, int flag)
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002474{
2475 struct tcp_sock *tp = tcp_sk(sk);
2476 int sndcnt = 0;
2477 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
Yuchung Cheng68049732013-05-29 14:20:11 +00002478 int newly_acked_sacked = prior_unsacked -
2479 (tp->packets_out - tp->sacked_out);
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002480
Yuchung Cheng684bad12012-09-02 17:38:04 +00002481 tp->prr_delivered += newly_acked_sacked;
Yuchung Cheng37598242015-07-01 14:11:15 -07002482 if (delta < 0) {
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002483 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2484 tp->prior_cwnd - 1;
2485 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
Yuchung Cheng37598242015-07-01 14:11:15 -07002486 } else if ((flag & FLAG_RETRANS_DATA_ACKED) &&
2487 !(flag & FLAG_LOST_RETRANS)) {
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002488 sndcnt = min_t(int, delta,
2489 max_t(int, tp->prr_delivered - tp->prr_out,
2490 newly_acked_sacked) + 1);
Yuchung Cheng37598242015-07-01 14:11:15 -07002491 } else {
2492 sndcnt = min(delta, newly_acked_sacked);
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002493 }
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002494 sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0));
2495 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2496}
2497
Yuchung Cheng684bad12012-09-02 17:38:04 +00002498static inline void tcp_end_cwnd_reduction(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002499{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002500 struct tcp_sock *tp = tcp_sk(sk);
Nandita Dukkipatia262f0c2011-08-21 20:21:57 +00002501
Yuchung Cheng684bad12012-09-02 17:38:04 +00002502 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2503 if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR ||
2504 (tp->undo_marker && tp->snd_ssthresh < TCP_INFINITE_SSTHRESH)) {
2505 tp->snd_cwnd = tp->snd_ssthresh;
2506 tp->snd_cwnd_stamp = tcp_time_stamp;
Yuchung Cheng67d41202011-03-14 10:57:03 +00002507 }
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002508 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002509}
2510
Yuchung Cheng684bad12012-09-02 17:38:04 +00002511/* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
Christoph Paasch5ee2c942014-07-14 16:58:32 +02002512void tcp_enter_cwr(struct sock *sk)
Yuchung Cheng09484d12012-09-02 17:38:02 +00002513{
2514 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng09484d12012-09-02 17:38:02 +00002515
2516 tp->prior_ssthresh = 0;
Yuchung Cheng684bad12012-09-02 17:38:04 +00002517 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
Yuchung Cheng09484d12012-09-02 17:38:02 +00002518 tp->undo_marker = 0;
Christoph Paasch5ee2c942014-07-14 16:58:32 +02002519 tcp_init_cwnd_reduction(sk);
Yuchung Cheng09484d12012-09-02 17:38:02 +00002520 tcp_set_ca_state(sk, TCP_CA_CWR);
2521 }
2522}
Kenneth Klette Jonassen7782ad82015-06-10 19:08:16 +02002523EXPORT_SYMBOL(tcp_enter_cwr);
Yuchung Cheng09484d12012-09-02 17:38:02 +00002524
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07002525static void tcp_try_keep_open(struct sock *sk)
2526{
2527 struct tcp_sock *tp = tcp_sk(sk);
2528 int state = TCP_CA_Open;
2529
Neal Cardwellf6982042011-11-16 08:58:04 +00002530 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07002531 state = TCP_CA_Disorder;
2532
2533 if (inet_csk(sk)->icsk_ca_state != state) {
2534 tcp_set_ca_state(sk, state);
2535 tp->high_seq = tp->snd_nxt;
2536 }
2537}
2538
Yuchung Cheng68049732013-05-29 14:20:11 +00002539static void tcp_try_to_open(struct sock *sk, int flag, const int prior_unsacked)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002540{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002541 struct tcp_sock *tp = tcp_sk(sk);
2542
Ilpo Järvinen86426c22007-08-09 14:45:17 +03002543 tcp_verify_left_out(tp);
2544
Yuchung Cheng9b441902013-03-20 13:32:58 +00002545 if (!tcp_any_retrans_done(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002546 tp->retrans_stamp = 0;
2547
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002548 if (flag & FLAG_ECE)
Christoph Paasch5ee2c942014-07-14 16:58:32 +02002549 tcp_enter_cwr(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002550
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002551 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07002552 tcp_try_keep_open(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002553 } else {
Yuchung Cheng37598242015-07-01 14:11:15 -07002554 tcp_cwnd_reduction(sk, prior_unsacked, 0, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002555 }
2556}
2557
John Heffner5d424d52006-03-20 17:53:41 -08002558static void tcp_mtup_probe_failed(struct sock *sk)
2559{
2560 struct inet_connection_sock *icsk = inet_csk(sk);
2561
2562 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2563 icsk->icsk_mtup.probe_size = 0;
Rick Jonesb56ea292015-07-21 16:14:13 -07002564 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
John Heffner5d424d52006-03-20 17:53:41 -08002565}
2566
Ilpo Järvinen72211e92009-03-14 14:23:04 +00002567static void tcp_mtup_probe_success(struct sock *sk)
John Heffner5d424d52006-03-20 17:53:41 -08002568{
2569 struct tcp_sock *tp = tcp_sk(sk);
2570 struct inet_connection_sock *icsk = inet_csk(sk);
2571
2572 /* FIXME: breaks with very large cwnd */
2573 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2574 tp->snd_cwnd = tp->snd_cwnd *
2575 tcp_mss_to_mtu(sk, tp->mss_cache) /
2576 icsk->icsk_mtup.probe_size;
2577 tp->snd_cwnd_cnt = 0;
2578 tp->snd_cwnd_stamp = tcp_time_stamp;
John Heffner9c6d5e52010-10-06 21:18:02 -07002579 tp->snd_ssthresh = tcp_current_ssthresh(sk);
John Heffner5d424d52006-03-20 17:53:41 -08002580
2581 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2582 icsk->icsk_mtup.probe_size = 0;
2583 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
Rick Jonesb56ea292015-07-21 16:14:13 -07002584 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
John Heffner5d424d52006-03-20 17:53:41 -08002585}
2586
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08002587/* Do a simple retransmit without using the backoff mechanisms in
2588 * tcp_timer. This is used for path mtu discovery.
2589 * The socket is already locked here.
2590 */
2591void tcp_simple_retransmit(struct sock *sk)
2592{
2593 const struct inet_connection_sock *icsk = inet_csk(sk);
2594 struct tcp_sock *tp = tcp_sk(sk);
2595 struct sk_buff *skb;
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00002596 unsigned int mss = tcp_current_mss(sk);
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08002597 u32 prior_lost = tp->lost_out;
2598
2599 tcp_for_write_queue(skb, sk) {
2600 if (skb == tcp_send_head(sk))
2601 break;
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08002602 if (tcp_skb_seglen(skb) > mss &&
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08002603 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2604 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2605 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2606 tp->retrans_out -= tcp_skb_pcount(skb);
2607 }
2608 tcp_skb_mark_lost_uncond_verify(tp, skb);
2609 }
2610 }
2611
2612 tcp_clear_retrans_hints_partial(tp);
2613
2614 if (prior_lost == tp->lost_out)
2615 return;
2616
2617 if (tcp_is_reno(tp))
2618 tcp_limit_reno_sacked(tp);
2619
2620 tcp_verify_left_out(tp);
2621
2622 /* Don't muck with the congestion window here.
2623 * Reason is that we do not increase amount of _data_
2624 * in network, but units changed and effective
2625 * cwnd/ssthresh really reduced now.
2626 */
2627 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2628 tp->high_seq = tp->snd_nxt;
2629 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2630 tp->prior_ssthresh = 0;
2631 tp->undo_marker = 0;
2632 tcp_set_ca_state(sk, TCP_CA_Loss);
2633 }
2634 tcp_xmit_retransmit_queue(sk);
2635}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00002636EXPORT_SYMBOL(tcp_simple_retransmit);
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08002637
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002638static void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2639{
2640 struct tcp_sock *tp = tcp_sk(sk);
2641 int mib_idx;
2642
2643 if (tcp_is_reno(tp))
2644 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2645 else
2646 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2647
2648 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2649
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002650 tp->prior_ssthresh = 0;
Yuchung Cheng989e04c2014-08-22 14:15:22 -07002651 tcp_init_undo(tp);
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002652
Yuchung Cheng291a00d2015-07-01 14:11:14 -07002653 if (!tcp_in_cwnd_reduction(sk)) {
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002654 if (!ece_ack)
2655 tp->prior_ssthresh = tcp_current_ssthresh(sk);
Christoph Paasch5ee2c942014-07-14 16:58:32 +02002656 tcp_init_cwnd_reduction(sk);
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002657 }
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002658 tcp_set_ca_state(sk, TCP_CA_Recovery);
2659}
2660
Yuchung Chengab42d9e2013-03-20 13:32:59 +00002661/* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2662 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2663 */
Yuchung Chenge33099f2013-03-20 13:33:00 +00002664static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack)
Yuchung Chengab42d9e2013-03-20 13:32:59 +00002665{
Yuchung Chengab42d9e2013-03-20 13:32:59 +00002666 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Chenge33099f2013-03-20 13:33:00 +00002667 bool recovered = !before(tp->snd_una, tp->high_seq);
Yuchung Chengab42d9e2013-03-20 13:32:59 +00002668
Yuchung Chengda34ac72015-05-18 12:31:44 -07002669 if ((flag & FLAG_SND_UNA_ADVANCED) &&
2670 tcp_try_undo_loss(sk, false))
2671 return;
2672
Yuchung Chenge33099f2013-03-20 13:33:00 +00002673 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
Yuchung Cheng0cfa5c02014-05-30 15:25:59 -07002674 /* Step 3.b. A timeout is spurious if not all data are
2675 * lost, i.e., never-retransmitted data are (s)acked.
2676 */
Yuchung Chengda34ac72015-05-18 12:31:44 -07002677 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2678 tcp_try_undo_loss(sk, true))
Yuchung Chenge33099f2013-03-20 13:33:00 +00002679 return;
Yuchung Cheng0cfa5c02014-05-30 15:25:59 -07002680
Yuchung Chengb7b0ed92015-05-18 12:31:45 -07002681 if (after(tp->snd_nxt, tp->high_seq)) {
2682 if (flag & FLAG_DATA_SACKED || is_dupack)
2683 tp->frto = 0; /* Step 3.a. loss was real */
Yuchung Chenge33099f2013-03-20 13:33:00 +00002684 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2685 tp->high_seq = tp->snd_nxt;
2686 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
2687 TCP_NAGLE_OFF);
2688 if (after(tp->snd_nxt, tp->high_seq))
2689 return; /* Step 2.b */
2690 tp->frto = 0;
2691 }
2692 }
2693
2694 if (recovered) {
2695 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
Yuchung Chengab42d9e2013-03-20 13:32:59 +00002696 tcp_try_undo_recovery(sk);
2697 return;
2698 }
Yuchung Chenge33099f2013-03-20 13:33:00 +00002699 if (tcp_is_reno(tp)) {
2700 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2701 * delivered. Lower inflight to clock out (re)tranmissions.
2702 */
2703 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2704 tcp_add_reno_sack(sk);
2705 else if (flag & FLAG_SND_UNA_ADVANCED)
2706 tcp_reset_reno_sack(tp);
2707 }
Yuchung Chengab42d9e2013-03-20 13:32:59 +00002708 tcp_xmit_retransmit_queue(sk);
2709}
2710
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002711/* Undo during fast recovery after partial ACK. */
Yuchung Cheng7026b912013-05-29 14:20:13 +00002712static bool tcp_try_undo_partial(struct sock *sk, const int acked,
Yuchung Cheng37598242015-07-01 14:11:15 -07002713 const int prior_unsacked, int flag)
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002714{
2715 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002716
Yuchung Cheng7026b912013-05-29 14:20:13 +00002717 if (tp->undo_marker && tcp_packet_delayed(tp)) {
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002718 /* Plain luck! Hole if filled with delayed
2719 * packet, rather than with a retransmit.
2720 */
Yuchung Cheng7026b912013-05-29 14:20:13 +00002721 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2722
2723 /* We are getting evidence that the reordering degree is higher
2724 * than we realized. If there are no retransmits out then we
2725 * can undo. Otherwise we clock out new packets but do not
2726 * mark more packets lost or retransmit more.
2727 */
2728 if (tp->retrans_out) {
Yuchung Cheng37598242015-07-01 14:11:15 -07002729 tcp_cwnd_reduction(sk, prior_unsacked, 0, flag);
Yuchung Cheng7026b912013-05-29 14:20:13 +00002730 return true;
2731 }
2732
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002733 if (!tcp_any_retrans_done(sk))
2734 tp->retrans_stamp = 0;
2735
Yuchung Cheng7026b912013-05-29 14:20:13 +00002736 DBGUNDO(sk, "partial recovery");
2737 tcp_undo_cwnd_reduction(sk, true);
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002738 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
Yuchung Cheng7026b912013-05-29 14:20:13 +00002739 tcp_try_keep_open(sk);
2740 return true;
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002741 }
Yuchung Cheng7026b912013-05-29 14:20:13 +00002742 return false;
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002743}
2744
Linus Torvalds1da177e2005-04-16 15:20:36 -07002745/* Process an event, which can update packets-in-flight not trivially.
2746 * Main goal of this function is to calculate new estimate for left_out,
2747 * taking into account both packets sitting in receiver's buffer and
2748 * packets lost by network.
2749 *
2750 * Besides that it does CWND reduction, when packet loss is detected
2751 * and changes state of machine.
2752 *
2753 * It does _not_ decide what to send, it is made in function
2754 * tcp_xmit_retransmit_queue().
2755 */
Yuchung Cheng68049732013-05-29 14:20:11 +00002756static void tcp_fastretrans_alert(struct sock *sk, const int acked,
2757 const int prior_unsacked,
Nandita Dukkipati35f079e2013-05-21 15:12:07 +00002758 bool is_dupack, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002759{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002760 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002761 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002762 bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002763 (tcp_fackets_out(tp) > tp->reordering));
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002764 int fast_rexmit = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002765
Ilpo Järvinen3ccd3132007-12-31 04:43:32 -08002766 if (WARN_ON(!tp->packets_out && tp->sacked_out))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002767 tp->sacked_out = 0;
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07002768 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002769 tp->fackets_out = 0;
2770
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002771 /* Now state machine starts.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002772 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002773 if (flag & FLAG_ECE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002774 tp->prior_ssthresh = 0;
2775
2776 /* B. In all the states check for reneging SACKs. */
Ilpo Järvinencadbd032007-12-31 04:49:21 -08002777 if (tcp_check_sack_reneging(sk, flag))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002778 return;
2779
Yuchung Cheng974c1232012-01-19 14:42:21 +00002780 /* C. Check consistency of the current state. */
Ilpo Järvinen005903b2007-08-09 14:44:16 +03002781 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002782
Yuchung Cheng974c1232012-01-19 14:42:21 +00002783 /* D. Check state exit conditions. State can be terminated
Linus Torvalds1da177e2005-04-16 15:20:36 -07002784 * when high_seq is ACKed. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002785 if (icsk->icsk_ca_state == TCP_CA_Open) {
Ilpo Järvinen547b7922008-07-25 21:43:18 -07002786 WARN_ON(tp->retrans_out != 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002787 tp->retrans_stamp = 0;
2788 } else if (!before(tp->snd_una, tp->high_seq)) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002789 switch (icsk->icsk_ca_state) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002790 case TCP_CA_CWR:
2791 /* CWR is to be held something *above* high_seq
2792 * is ACKed for CWR bit to reach receiver. */
2793 if (tp->snd_una != tp->high_seq) {
Yuchung Cheng684bad12012-09-02 17:38:04 +00002794 tcp_end_cwnd_reduction(sk);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002795 tcp_set_ca_state(sk, TCP_CA_Open);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002796 }
2797 break;
2798
Linus Torvalds1da177e2005-04-16 15:20:36 -07002799 case TCP_CA_Recovery:
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002800 if (tcp_is_reno(tp))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002801 tcp_reset_reno_sack(tp);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002802 if (tcp_try_undo_recovery(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002803 return;
Yuchung Cheng684bad12012-09-02 17:38:04 +00002804 tcp_end_cwnd_reduction(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002805 break;
2806 }
2807 }
2808
Yuchung Cheng4f41b1c2015-10-16 21:57:47 -07002809 /* Use RACK to detect loss */
2810 if (sysctl_tcp_recovery & TCP_RACK_LOST_RETRANS &&
2811 tcp_rack_mark_lost(sk))
2812 flag |= FLAG_LOST_RETRANS;
2813
Yuchung Cheng974c1232012-01-19 14:42:21 +00002814 /* E. Process state. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002815 switch (icsk->icsk_ca_state) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002816 case TCP_CA_Recovery:
Ilpo Järvinen2e605292007-08-02 19:46:58 -07002817 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002818 if (tcp_is_reno(tp) && is_dupack)
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002819 tcp_add_reno_sack(sk);
Yuchung Cheng7026b912013-05-29 14:20:13 +00002820 } else {
Yuchung Cheng37598242015-07-01 14:11:15 -07002821 if (tcp_try_undo_partial(sk, acked, prior_unsacked, flag))
Yuchung Cheng7026b912013-05-29 14:20:13 +00002822 return;
2823 /* Partial ACK arrived. Force fast retransmit. */
2824 do_lost = tcp_is_reno(tp) ||
2825 tcp_fackets_out(tp) > tp->reordering;
2826 }
Yuchung Chengc7d9d6a2013-05-29 14:20:14 +00002827 if (tcp_try_undo_dsack(sk)) {
2828 tcp_try_keep_open(sk);
2829 return;
2830 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002831 break;
2832 case TCP_CA_Loss:
Yuchung Chenge33099f2013-03-20 13:33:00 +00002833 tcp_process_loss(sk, flag, is_dupack);
Yuchung Cheng291a00d2015-07-01 14:11:14 -07002834 if (icsk->icsk_ca_state != TCP_CA_Open &&
2835 !(flag & FLAG_LOST_RETRANS))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002836 return;
Yuchung Cheng291a00d2015-07-01 14:11:14 -07002837 /* Change state if cwnd is undone or retransmits are lost */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002838 default:
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002839 if (tcp_is_reno(tp)) {
Ilpo Järvinen2e605292007-08-02 19:46:58 -07002840 if (flag & FLAG_SND_UNA_ADVANCED)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002841 tcp_reset_reno_sack(tp);
2842 if (is_dupack)
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002843 tcp_add_reno_sack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002844 }
2845
Neal Cardwellf6982042011-11-16 08:58:04 +00002846 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002847 tcp_try_undo_dsack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002848
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00002849 if (!tcp_time_to_recover(sk, flag)) {
Yuchung Cheng68049732013-05-29 14:20:11 +00002850 tcp_try_to_open(sk, flag, prior_unsacked);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002851 return;
2852 }
2853
John Heffner5d424d52006-03-20 17:53:41 -08002854 /* MTU probe failure: don't reduce cwnd */
2855 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2856 icsk->icsk_mtup.probe_size &&
John Heffner0e7b1362006-03-20 21:32:58 -08002857 tp->snd_una == tp->mtu_probe.probe_seq_start) {
John Heffner5d424d52006-03-20 17:53:41 -08002858 tcp_mtup_probe_failed(sk);
2859 /* Restores the reduction we did in tcp_mtup_probe() */
2860 tp->snd_cwnd++;
2861 tcp_simple_retransmit(sk);
2862 return;
2863 }
2864
Linus Torvalds1da177e2005-04-16 15:20:36 -07002865 /* Otherwise enter Recovery state */
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002866 tcp_enter_recovery(sk, (flag & FLAG_ECE));
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002867 fast_rexmit = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002868 }
2869
Yuchung Cheng3e59cb02013-05-17 13:45:05 +00002870 if (do_lost)
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002871 tcp_update_scoreboard(sk, fast_rexmit);
Yuchung Cheng37598242015-07-01 14:11:15 -07002872 tcp_cwnd_reduction(sk, prior_unsacked, fast_rexmit, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002873 tcp_xmit_retransmit_queue(sk);
2874}
2875
Yuchung Chengf6722582015-10-16 21:57:42 -07002876/* Kathleen Nichols' algorithm for tracking the minimum value of
2877 * a data stream over some fixed time interval. (E.g., the minimum
2878 * RTT over the past five minutes.) It uses constant space and constant
2879 * time per update yet almost always delivers the same minimum as an
2880 * implementation that has to keep all the data in the window.
2881 *
2882 * The algorithm keeps track of the best, 2nd best & 3rd best min
2883 * values, maintaining an invariant that the measurement time of the
2884 * n'th best >= n-1'th best. It also makes sure that the three values
2885 * are widely separated in the time window since that bounds the worse
2886 * case error when that data is monotonically increasing over the window.
2887 *
2888 * Upon getting a new min, we can forget everything earlier because it
2889 * has no value - the new min is <= everything else in the window by
2890 * definition and it's the most recent. So we restart fresh on every new min
2891 * and overwrites 2nd & 3rd choices. The same property holds for 2nd & 3rd
2892 * best.
2893 */
2894static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us)
2895{
2896 const u32 now = tcp_time_stamp, wlen = sysctl_tcp_min_rtt_wlen * HZ;
2897 struct rtt_meas *m = tcp_sk(sk)->rtt_min;
2898 struct rtt_meas rttm = { .rtt = (rtt_us ? : 1), .ts = now };
2899 u32 elapsed;
2900
2901 /* Check if the new measurement updates the 1st, 2nd, or 3rd choices */
2902 if (unlikely(rttm.rtt <= m[0].rtt))
2903 m[0] = m[1] = m[2] = rttm;
2904 else if (rttm.rtt <= m[1].rtt)
2905 m[1] = m[2] = rttm;
2906 else if (rttm.rtt <= m[2].rtt)
2907 m[2] = rttm;
2908
2909 elapsed = now - m[0].ts;
2910 if (unlikely(elapsed > wlen)) {
2911 /* Passed entire window without a new min so make 2nd choice
2912 * the new min & 3rd choice the new 2nd. So forth and so on.
2913 */
2914 m[0] = m[1];
2915 m[1] = m[2];
2916 m[2] = rttm;
2917 if (now - m[0].ts > wlen) {
2918 m[0] = m[1];
2919 m[1] = rttm;
2920 if (now - m[0].ts > wlen)
2921 m[0] = rttm;
2922 }
2923 } else if (m[1].ts == m[0].ts && elapsed > wlen / 4) {
2924 /* Passed a quarter of the window without a new min so
2925 * take 2nd choice from the 2nd quarter of the window.
2926 */
2927 m[2] = m[1] = rttm;
2928 } else if (m[2].ts == m[1].ts && elapsed > wlen / 2) {
2929 /* Passed half the window without a new min so take the 3rd
2930 * choice from the last half of the window.
2931 */
2932 m[2] = rttm;
2933 }
2934}
2935
Yuchung Chenged084952013-07-22 16:20:48 -07002936static inline bool tcp_ack_update_rtt(struct sock *sk, const int flag,
Yuchung Chengf6722582015-10-16 21:57:42 -07002937 long seq_rtt_us, long sack_rtt_us,
2938 long ca_rtt_us)
Ilpo Järvinen41834b72008-12-05 22:43:26 -08002939{
Yuchung Cheng5b08e472013-07-22 16:20:46 -07002940 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen41834b72008-12-05 22:43:26 -08002941
Yuchung Cheng5b08e472013-07-22 16:20:46 -07002942 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2943 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2944 * Karn's algorithm forbids taking RTT if some retransmitted data
2945 * is acked (RFC6298).
2946 */
Eric Dumazet740b0f12014-02-26 14:02:48 -08002947 if (seq_rtt_us < 0)
2948 seq_rtt_us = sack_rtt_us;
Yuchung Chenged084952013-07-22 16:20:48 -07002949
Linus Torvalds1da177e2005-04-16 15:20:36 -07002950 /* RTTM Rule: A TSecr value received in a segment is used to
2951 * update the averaged RTT measurement only if the segment
2952 * acknowledges some new data, i.e., only if it advances the
2953 * left edge of the send window.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002954 * See draft-ietf-tcplw-high-performance-00, section 3.3.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002955 */
Eric Dumazet740b0f12014-02-26 14:02:48 -08002956 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
Yuchung Cheng2909d872013-10-24 08:55:25 -07002957 flag & FLAG_ACKED)
Yuchung Chengf6722582015-10-16 21:57:42 -07002958 seq_rtt_us = ca_rtt_us = jiffies_to_usecs(tcp_time_stamp -
2959 tp->rx_opt.rcv_tsecr);
Eric Dumazet740b0f12014-02-26 14:02:48 -08002960 if (seq_rtt_us < 0)
Yuchung Chenged084952013-07-22 16:20:48 -07002961 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002962
Yuchung Chengf6722582015-10-16 21:57:42 -07002963 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2964 * always taken together with ACK, SACK, or TS-opts. Any negative
2965 * values will be skipped with the seq_rtt_us < 0 check above.
2966 */
2967 tcp_update_rtt_min(sk, ca_rtt_us);
Eric Dumazet740b0f12014-02-26 14:02:48 -08002968 tcp_rtt_estimator(sk, seq_rtt_us);
Yuchung Cheng5b08e472013-07-22 16:20:46 -07002969 tcp_set_rto(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002970
Yuchung Cheng5b08e472013-07-22 16:20:46 -07002971 /* RFC6298: only reset backoff on valid RTT measurement. */
2972 inet_csk(sk)->icsk_backoff = 0;
Yuchung Chenged084952013-07-22 16:20:48 -07002973 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002974}
2975
Yuchung Cheng375fe022013-07-22 16:20:45 -07002976/* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
Yuchung Cheng0f1c28a2015-09-18 11:36:14 -07002977void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
Yuchung Cheng375fe022013-07-22 16:20:45 -07002978{
Yuchung Cheng0f1c28a2015-09-18 11:36:14 -07002979 long rtt_us = -1L;
Yuchung Cheng375fe022013-07-22 16:20:45 -07002980
Yuchung Cheng0f1c28a2015-09-18 11:36:14 -07002981 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack.v64) {
2982 struct skb_mstamp now;
Yuchung Chengbc15afa2013-10-24 08:44:25 -07002983
Yuchung Cheng0f1c28a2015-09-18 11:36:14 -07002984 skb_mstamp_get(&now);
2985 rtt_us = skb_mstamp_us_delta(&now, &tcp_rsk(req)->snt_synack);
2986 }
2987
Yuchung Chengf6722582015-10-16 21:57:42 -07002988 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us);
Yuchung Cheng375fe022013-07-22 16:20:45 -07002989}
2990
Yuchung Cheng0f1c28a2015-09-18 11:36:14 -07002991
Eric Dumazet24901552014-05-02 21:18:05 -07002992static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002993{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002994 const struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazet24901552014-05-02 21:18:05 -07002995
2996 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002997 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002998}
2999
Linus Torvalds1da177e2005-04-16 15:20:36 -07003000/* Restart timer after forward progress on connection.
3001 * RFC2988 recommends to restart timer to now+rto.
3002 */
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00003003void tcp_rearm_rto(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003004{
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00003005 const struct inet_connection_sock *icsk = inet_csk(sk);
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00003006 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003007
Jerry Chu168a8f52012-08-31 12:29:13 +00003008 /* If the retrans timer is currently being used by Fast Open
3009 * for SYN-ACK retrans purpose, stay put.
3010 */
3011 if (tp->fastopen_rsk)
3012 return;
3013
Linus Torvalds1da177e2005-04-16 15:20:36 -07003014 if (!tp->packets_out) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003015 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003016 } else {
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00003017 u32 rto = inet_csk(sk)->icsk_rto;
3018 /* Offset the time elapsed after installing regular RTO */
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00003019 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
3020 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00003021 struct sk_buff *skb = tcp_write_queue_head(sk);
Eric Dumazet7faee5c2014-09-05 15:33:33 -07003022 const u32 rto_time_stamp =
3023 tcp_skb_timestamp(skb) + rto;
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00003024 s32 delta = (s32)(rto_time_stamp - tcp_time_stamp);
3025 /* delta may not be positive if the socket is locked
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00003026 * when the retrans timer fires and is rescheduled.
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00003027 */
3028 if (delta > 0)
3029 rto = delta;
3030 }
3031 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3032 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003033 }
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00003034}
3035
3036/* This function is called when the delayed ER timer fires. TCP enters
3037 * fast recovery and performs fast-retransmit.
3038 */
3039void tcp_resume_early_retransmit(struct sock *sk)
3040{
3041 struct tcp_sock *tp = tcp_sk(sk);
3042
3043 tcp_rearm_rto(sk);
3044
3045 /* Stop if ER is disabled after the delayed ER timer is scheduled */
3046 if (!tp->do_early_retrans)
3047 return;
3048
3049 tcp_enter_recovery(sk, false);
3050 tcp_update_scoreboard(sk, 1);
3051 tcp_xmit_retransmit_queue(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003052}
3053
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003054/* If we get here, the whole TSO packet has not been acked. */
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003055static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003056{
3057 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003058 u32 packets_acked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003059
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003060 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003061
3062 packets_acked = tcp_skb_pcount(skb);
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003063 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003064 return 0;
3065 packets_acked -= tcp_skb_pcount(skb);
3066
3067 if (packets_acked) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003068 BUG_ON(tcp_skb_pcount(skb) == 0);
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003069 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003070 }
3071
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003072 return packets_acked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003073}
3074
Eric Dumazetad971f62014-10-11 15:17:29 -07003075static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3076 u32 prior_snd_una)
3077{
3078 const struct skb_shared_info *shinfo;
3079
3080 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3081 if (likely(!(sk->sk_tsflags & SOF_TIMESTAMPING_TX_ACK)))
3082 return;
3083
3084 shinfo = skb_shinfo(skb);
3085 if ((shinfo->tx_flags & SKBTX_ACK_TSTAMP) &&
3086 between(shinfo->tskey, prior_snd_una, tcp_sk(sk)->snd_una - 1))
3087 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3088}
3089
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003090/* Remove acknowledged frames from the retransmission queue. If our packet
3091 * is before the ack sequence we can discard it as it's confirmed to have
3092 * arrived at the other end.
3093 */
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07003094static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
Kenneth Klette Jonassen31231a82015-05-01 01:10:58 +02003095 u32 prior_snd_una,
3096 struct tcp_sacktag_state *sack)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003097{
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08003098 const struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazet740b0f12014-02-26 14:02:48 -08003099 struct skb_mstamp first_ackt, last_ackt, now;
3100 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen90638a02008-09-20 21:25:52 -07003101 u32 prior_sacked = tp->sacked_out;
Eric Dumazet740b0f12014-02-26 14:02:48 -08003102 u32 reord = tp->packets_out;
3103 bool fully_acked = true;
Kenneth Klette Jonassen31231a82015-05-01 01:10:58 +02003104 long sack_rtt_us = -1L;
Eric Dumazet740b0f12014-02-26 14:02:48 -08003105 long seq_rtt_us = -1L;
Kenneth Klette Jonassen31231a82015-05-01 01:10:58 +02003106 long ca_rtt_us = -1L;
Eric Dumazet740b0f12014-02-26 14:02:48 -08003107 struct sk_buff *skb;
3108 u32 pkts_acked = 0;
Yuchung Cheng2f715c12013-10-24 08:59:27 -07003109 bool rtt_update;
Eric Dumazet740b0f12014-02-26 14:02:48 -08003110 int flag = 0;
3111
3112 first_ackt.v64 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003113
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003114 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09003115 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003116 u8 sacked = scb->sacked;
Eric Dumazet740b0f12014-02-26 14:02:48 -08003117 u32 acked_pcount;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003118
Eric Dumazetad971f62014-10-11 15:17:29 -07003119 tcp_ack_tstamp(sk, skb, prior_snd_una);
Willem de Bruijn712a7222014-08-12 14:53:16 -04003120
Gavin McCullagh2072c222007-12-29 19:11:21 -08003121 /* Determine how many packets and what bytes were acked, tso and else */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003122 if (after(scb->end_seq, tp->snd_una)) {
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003123 if (tcp_skb_pcount(skb) == 1 ||
3124 !after(tp->snd_una, scb->seq))
3125 break;
3126
Ilpo Järvinen72018832007-12-30 04:37:55 -08003127 acked_pcount = tcp_tso_acked(sk, skb);
3128 if (!acked_pcount)
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003129 break;
3130
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003131 fully_acked = false;
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003132 } else {
Eric Dumazetad971f62014-10-11 15:17:29 -07003133 /* Speedup tcp_unlink_write_queue() and next loop */
3134 prefetchw(skb->next);
Ilpo Järvinen72018832007-12-30 04:37:55 -08003135 acked_pcount = tcp_skb_pcount(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003136 }
3137
Eric Dumazetad971f62014-10-11 15:17:29 -07003138 if (unlikely(sacked & TCPCB_RETRANS)) {
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003139 if (sacked & TCPCB_SACKED_RETRANS)
Ilpo Järvinen72018832007-12-30 04:37:55 -08003140 tp->retrans_out -= acked_pcount;
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003141 flag |= FLAG_RETRANS_DATA_ACKED;
Kenneth Klette Jonassen3d0d26c2015-04-11 02:17:49 +02003142 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
Eric Dumazet740b0f12014-02-26 14:02:48 -08003143 last_ackt = skb->skb_mstamp;
Eric Dumazet431a9122014-03-09 17:36:02 -07003144 WARN_ON_ONCE(last_ackt.v64 == 0);
Eric Dumazet740b0f12014-02-26 14:02:48 -08003145 if (!first_ackt.v64)
3146 first_ackt = last_ackt;
3147
Kenneth Klette Jonassen3d0d26c2015-04-11 02:17:49 +02003148 reord = min(pkts_acked, reord);
3149 if (!after(scb->end_seq, tp->high_seq))
3150 flag |= FLAG_ORIG_SACK_ACKED;
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08003151 }
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003152
3153 if (sacked & TCPCB_SACKED_ACKED)
Ilpo Järvinen72018832007-12-30 04:37:55 -08003154 tp->sacked_out -= acked_pcount;
Yuchung Cheng659a8ad2015-10-16 21:57:46 -07003155 else if (tcp_is_sack(tp) && !tcp_skb_spurious_retrans(tp, skb))
3156 tcp_rack_advance(tp, &skb->skb_mstamp, sacked);
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003157 if (sacked & TCPCB_LOST)
Ilpo Järvinen72018832007-12-30 04:37:55 -08003158 tp->lost_out -= acked_pcount;
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003159
Ilpo Järvinen72018832007-12-30 04:37:55 -08003160 tp->packets_out -= acked_pcount;
3161 pkts_acked += acked_pcount;
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003162
Ilpo Järvinen009a2e32007-09-20 11:34:38 -07003163 /* Initial outgoing SYN's get put onto the write_queue
3164 * just like anything else we transmit. It is not
3165 * true data, and if we misinform our callers that
3166 * this ACK acks real data, we will erroneously exit
3167 * connection startup slow start one packet too
3168 * quickly. This is severely frowned upon behavior.
3169 */
Eric Dumazetad971f62014-10-11 15:17:29 -07003170 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
Ilpo Järvinen009a2e32007-09-20 11:34:38 -07003171 flag |= FLAG_DATA_ACKED;
3172 } else {
3173 flag |= FLAG_SYN_ACKED;
3174 tp->retrans_stamp = 0;
3175 }
3176
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003177 if (!fully_acked)
3178 break;
3179
David S. Millerfe067e82007-03-07 12:12:44 -08003180 tcp_unlink_write_queue(skb, sk);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08003181 sk_wmem_free_skb(sk, skb);
Eric Dumazetad971f62014-10-11 15:17:29 -07003182 if (unlikely(skb == tp->retransmit_skb_hint))
Ilpo Järvinenef9da472008-09-20 21:25:15 -07003183 tp->retransmit_skb_hint = NULL;
Eric Dumazetad971f62014-10-11 15:17:29 -07003184 if (unlikely(skb == tp->lost_skb_hint))
Ilpo Järvinen90638a02008-09-20 21:25:52 -07003185 tp->lost_skb_hint = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003186 }
3187
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07003188 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3189 tp->snd_up = tp->snd_una;
3190
Ilpo Järvinencadbd032007-12-31 04:49:21 -08003191 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3192 flag |= FLAG_SACK_RENEGING;
3193
Eric Dumazet740b0f12014-02-26 14:02:48 -08003194 skb_mstamp_get(&now);
Yuchung Cheng9e45a3e2015-10-16 21:57:41 -07003195 if (likely(first_ackt.v64) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
Eric Dumazet740b0f12014-02-26 14:02:48 -08003196 seq_rtt_us = skb_mstamp_us_delta(&now, &first_ackt);
Kenneth Klette Jonassen31231a82015-05-01 01:10:58 +02003197 ca_rtt_us = skb_mstamp_us_delta(&now, &last_ackt);
3198 }
3199 if (sack->first_sackt.v64) {
3200 sack_rtt_us = skb_mstamp_us_delta(&now, &sack->first_sackt);
3201 ca_rtt_us = skb_mstamp_us_delta(&now, &sack->last_sackt);
Eric Dumazet740b0f12014-02-26 14:02:48 -08003202 }
3203
Yuchung Chengf6722582015-10-16 21:57:42 -07003204 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3205 ca_rtt_us);
Yuchung Chenged084952013-07-22 16:20:48 -07003206
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003207 if (flag & FLAG_ACKED) {
Yuchung Cheng2f715c12013-10-24 08:59:27 -07003208 tcp_rearm_rto(sk);
Ilpo Järvinen72211e92009-03-14 14:23:04 +00003209 if (unlikely(icsk->icsk_mtup.probe_size &&
3210 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3211 tcp_mtup_probe_success(sk);
3212 }
3213
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003214 if (tcp_is_reno(tp)) {
3215 tcp_remove_reno_sacks(sk, pkts_acked);
3216 } else {
Ilpo Järvinen59a08cb2009-02-28 04:44:28 +00003217 int delta;
3218
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003219 /* Non-retransmitted hole got filled? That's reordering */
3220 if (reord < prior_fackets)
3221 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
Ilpo Järvinen90638a02008-09-20 21:25:52 -07003222
Ilpo Järvinen59a08cb2009-02-28 04:44:28 +00003223 delta = tcp_is_fack(tp) ? pkts_acked :
3224 prior_sacked - tp->sacked_out;
3225 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003226 }
3227
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07003228 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08003229
Eric Dumazet740b0f12014-02-26 14:02:48 -08003230 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3231 sack_rtt_us > skb_mstamp_us_delta(&now, &skb->skb_mstamp)) {
Yuchung Cheng2f715c12013-10-24 08:59:27 -07003232 /* Do not re-arm RTO if the sack RTT is measured from data sent
3233 * after when the head was last (re)transmitted. Otherwise the
3234 * timeout may continue to extend in loss recovery.
3235 */
3236 tcp_rearm_rto(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003237 }
3238
Kenneth Klette Jonassen138998f2015-05-01 01:10:59 +02003239 if (icsk->icsk_ca_ops->pkts_acked)
3240 icsk->icsk_ca_ops->pkts_acked(sk, pkts_acked, ca_rtt_us);
3241
Linus Torvalds1da177e2005-04-16 15:20:36 -07003242#if FASTRETRANS_DEBUG > 0
Ilpo Järvinen547b7922008-07-25 21:43:18 -07003243 WARN_ON((int)tp->sacked_out < 0);
3244 WARN_ON((int)tp->lost_out < 0);
3245 WARN_ON((int)tp->retrans_out < 0);
Ilpo Järvinene60402d2007-08-09 15:14:46 +03003246 if (!tp->packets_out && tcp_is_sack(tp)) {
Stephen Hemmingercfcabdc2007-10-09 01:59:42 -07003247 icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003248 if (tp->lost_out) {
Joe Perches91df42b2012-05-15 14:11:54 +00003249 pr_debug("Leak l=%u %d\n",
3250 tp->lost_out, icsk->icsk_ca_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003251 tp->lost_out = 0;
3252 }
3253 if (tp->sacked_out) {
Joe Perches91df42b2012-05-15 14:11:54 +00003254 pr_debug("Leak s=%u %d\n",
3255 tp->sacked_out, icsk->icsk_ca_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003256 tp->sacked_out = 0;
3257 }
3258 if (tp->retrans_out) {
Joe Perches91df42b2012-05-15 14:11:54 +00003259 pr_debug("Leak r=%u %d\n",
3260 tp->retrans_out, icsk->icsk_ca_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003261 tp->retrans_out = 0;
3262 }
3263 }
3264#endif
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003265 return flag;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003266}
3267
3268static void tcp_ack_probe(struct sock *sk)
3269{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003270 const struct tcp_sock *tp = tcp_sk(sk);
3271 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003272
3273 /* Was it a usable window open? */
3274
Ilpo Järvinen90840de2007-12-31 04:48:41 -08003275 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003276 icsk->icsk_backoff = 0;
3277 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003278 /* Socket must be waked up by subsequent tcp_data_snd_check().
3279 * This function is not for random using!
3280 */
3281 } else {
Eric Dumazet21c8fe92015-05-06 14:26:24 -07003282 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
Eric Dumazetfcdd1cf2014-09-22 13:19:44 -07003283
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003284 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
Eric Dumazetfcdd1cf2014-09-22 13:19:44 -07003285 when, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003286 }
3287}
3288
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00003289static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003290{
Eric Dumazeta02cec22010-09-22 20:43:57 +00003291 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3292 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003293}
3294
Yuchung Cheng0f7cc9a2013-08-21 17:29:23 -07003295/* Decide wheather to run the increase function of congestion control. */
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00003296static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003297{
Yuchung Cheng0f7cc9a2013-08-21 17:29:23 -07003298 if (tcp_in_cwnd_reduction(sk))
3299 return false;
3300
3301 /* If reordering is high then always grow cwnd whenever data is
3302 * delivered regardless of its ordering. Otherwise stay conservative
Yuchung Cheng16edfe72013-09-05 15:36:09 -07003303 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
Yuchung Cheng0f7cc9a2013-08-21 17:29:23 -07003304 * new SACK or ECE mark may first advance cwnd here and later reduce
3305 * cwnd in tcp_fastretrans_alert() based on more states.
3306 */
3307 if (tcp_sk(sk)->reordering > sysctl_tcp_reordering)
3308 return flag & FLAG_FORWARD_PROGRESS;
3309
Yuchung Cheng16edfe72013-09-05 15:36:09 -07003310 return flag & FLAG_DATA_ACKED;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003311}
3312
3313/* Check that window update is acceptable.
3314 * The function assumes that snd_una<=ack<=snd_next.
3315 */
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00003316static inline bool tcp_may_update_window(const struct tcp_sock *tp,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003317 const u32 ack, const u32 ack_seq,
3318 const u32 nwin)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003319{
Eric Dumazeta02cec22010-09-22 20:43:57 +00003320 return after(ack, tp->snd_una) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07003321 after(ack_seq, tp->snd_wl1) ||
Eric Dumazeta02cec22010-09-22 20:43:57 +00003322 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003323}
3324
Eric Dumazet0df48c22015-04-28 15:28:17 -07003325/* If we update tp->snd_una, also update tp->bytes_acked */
3326static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3327{
3328 u32 delta = ack - tp->snd_una;
3329
Eric Dumazetd6549762015-05-21 21:51:19 -07003330 u64_stats_update_begin(&tp->syncp);
Eric Dumazet0df48c22015-04-28 15:28:17 -07003331 tp->bytes_acked += delta;
Eric Dumazetd6549762015-05-21 21:51:19 -07003332 u64_stats_update_end(&tp->syncp);
Eric Dumazet0df48c22015-04-28 15:28:17 -07003333 tp->snd_una = ack;
3334}
3335
Eric Dumazetbdd1f9e2015-04-28 15:28:18 -07003336/* If we update tp->rcv_nxt, also update tp->bytes_received */
3337static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3338{
3339 u32 delta = seq - tp->rcv_nxt;
3340
Eric Dumazetd6549762015-05-21 21:51:19 -07003341 u64_stats_update_begin(&tp->syncp);
Eric Dumazetbdd1f9e2015-04-28 15:28:18 -07003342 tp->bytes_received += delta;
Eric Dumazetd6549762015-05-21 21:51:19 -07003343 u64_stats_update_end(&tp->syncp);
Eric Dumazetbdd1f9e2015-04-28 15:28:18 -07003344 tp->rcv_nxt = seq;
3345}
3346
Linus Torvalds1da177e2005-04-16 15:20:36 -07003347/* Update our send window.
3348 *
3349 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3350 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3351 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003352static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003353 u32 ack_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003354{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003355 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003356 int flag = 0;
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07003357 u32 nwin = ntohs(tcp_hdr(skb)->window);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003358
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07003359 if (likely(!tcp_hdr(skb)->syn))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003360 nwin <<= tp->rx_opt.snd_wscale;
3361
3362 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3363 flag |= FLAG_WIN_UPDATE;
Hantzis Fotisee7537b2009-03-02 22:42:02 -08003364 tcp_update_wl(tp, ack_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003365
3366 if (tp->snd_wnd != nwin) {
3367 tp->snd_wnd = nwin;
3368
3369 /* Note, it is the only place, where
3370 * fast path is recovered for sending TCP.
3371 */
Herbert Xu2ad41062005-10-27 18:47:46 +10003372 tp->pred_flags = 0;
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003373 tcp_fast_path_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003374
Eric Dumazet6f021c62015-08-21 12:30:00 -07003375 if (tcp_send_head(sk))
3376 tcp_slow_start_after_idle_check(sk);
3377
Linus Torvalds1da177e2005-04-16 15:20:36 -07003378 if (nwin > tp->max_window) {
3379 tp->max_window = nwin;
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08003380 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003381 }
3382 }
3383 }
3384
Eric Dumazet0df48c22015-04-28 15:28:17 -07003385 tcp_snd_una_update(tp, ack);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003386
3387 return flag;
3388}
3389
Eric Dumazet7970ddc2015-03-16 21:06:20 -07003390/* Return true if we're currently rate-limiting out-of-window ACKs and
3391 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3392 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3393 * attacks that send repeated SYNs or ACKs for the same connection. To
3394 * do this, we do not send a duplicate SYNACK or ACK if the remote
3395 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3396 */
3397bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3398 int mib_idx, u32 *last_oow_ack_time)
3399{
3400 /* Data packets without SYNs are not likely part of an ACK loop. */
3401 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3402 !tcp_hdr(skb)->syn)
3403 goto not_rate_limited;
3404
3405 if (*last_oow_ack_time) {
3406 s32 elapsed = (s32)(tcp_time_stamp - *last_oow_ack_time);
3407
3408 if (0 <= elapsed && elapsed < sysctl_tcp_invalid_ratelimit) {
3409 NET_INC_STATS_BH(net, mib_idx);
3410 return true; /* rate-limited: don't send yet! */
3411 }
3412 }
3413
3414 *last_oow_ack_time = tcp_time_stamp;
3415
3416not_rate_limited:
3417 return false; /* not rate-limited: go ahead, send dupack now! */
3418}
3419
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003420/* RFC 5961 7 [ACK Throttling] */
Neal Cardwellf2b2c582015-02-06 16:04:40 -05003421static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003422{
3423 /* unprotected vars, we dont care of overwrites */
3424 static u32 challenge_timestamp;
3425 static unsigned int challenge_count;
Neal Cardwellf2b2c582015-02-06 16:04:40 -05003426 struct tcp_sock *tp = tcp_sk(sk);
3427 u32 now;
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003428
Neal Cardwellf2b2c582015-02-06 16:04:40 -05003429 /* First check our per-socket dupack rate limit. */
3430 if (tcp_oow_rate_limited(sock_net(sk), skb,
3431 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3432 &tp->last_oow_ack_time))
3433 return;
3434
3435 /* Then check the check host-wide RFC 5961 rate limit. */
3436 now = jiffies / HZ;
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003437 if (now != challenge_timestamp) {
3438 challenge_timestamp = now;
3439 challenge_count = 0;
3440 }
3441 if (++challenge_count <= sysctl_tcp_challenge_ack_limit) {
3442 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK);
3443 tcp_send_ack(sk);
3444 }
3445}
3446
Eric Dumazet12fb3dd2013-04-19 07:19:48 +00003447static void tcp_store_ts_recent(struct tcp_sock *tp)
3448{
3449 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3450 tp->rx_opt.ts_recent_stamp = get_seconds();
3451}
3452
3453static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3454{
3455 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3456 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3457 * extra check below makes sure this can only happen
3458 * for pure ACK frames. -DaveM
3459 *
3460 * Not only, also it occurs for expired timestamps.
3461 */
3462
3463 if (tcp_paws_check(&tp->rx_opt, 0))
3464 tcp_store_ts_recent(tp);
3465 }
3466}
3467
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003468/* This routine deals with acks during a TLP episode.
Sébastien Barré08abdff2015-01-12 10:30:40 +01003469 * We mark the end of a TLP episode on receiving TLP dupack or when
3470 * ack is after tlp_high_seq.
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003471 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3472 */
3473static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3474{
3475 struct tcp_sock *tp = tcp_sk(sk);
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003476
Sébastien Barré08abdff2015-01-12 10:30:40 +01003477 if (before(ack, tp->tlp_high_seq))
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003478 return;
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003479
Sébastien Barré08abdff2015-01-12 10:30:40 +01003480 if (flag & FLAG_DSACKING_ACK) {
3481 /* This DSACK means original and TLP probe arrived; no loss */
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003482 tp->tlp_high_seq = 0;
Sébastien Barré08abdff2015-01-12 10:30:40 +01003483 } else if (after(ack, tp->tlp_high_seq)) {
3484 /* ACK advances: there was a loss, so reduce cwnd. Reset
3485 * tlp_high_seq in tcp_init_cwnd_reduction()
3486 */
3487 tcp_init_cwnd_reduction(sk);
3488 tcp_set_ca_state(sk, TCP_CA_CWR);
3489 tcp_end_cwnd_reduction(sk);
3490 tcp_try_keep_open(sk);
3491 NET_INC_STATS_BH(sock_net(sk),
3492 LINUX_MIB_TCPLOSSPROBERECOVERY);
3493 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3494 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3495 /* Pure dupack: original and TLP probe arrived; no loss */
3496 tp->tlp_high_seq = 0;
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003497 }
3498}
3499
Florian Westphal7354c8c2014-09-26 22:37:34 +02003500static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3501{
3502 const struct inet_connection_sock *icsk = inet_csk(sk);
3503
3504 if (icsk->icsk_ca_ops->in_ack_event)
3505 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3506}
3507
Linus Torvalds1da177e2005-04-16 15:20:36 -07003508/* This routine deals with incoming acks, but not outgoing ones. */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003509static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003510{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003511 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003512 struct tcp_sock *tp = tcp_sk(sk);
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02003513 struct tcp_sacktag_state sack_state;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003514 u32 prior_snd_una = tp->snd_una;
3515 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3516 u32 ack = TCP_SKB_CB(skb)->ack_seq;
Neal Cardwell7d2b55f2011-11-16 08:58:01 +00003517 bool is_dupack = false;
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003518 u32 prior_fackets;
Nandita Dukkipati35f079e2013-05-21 15:12:07 +00003519 int prior_packets = tp->packets_out;
Yuchung Cheng68049732013-05-29 14:20:11 +00003520 const int prior_unsacked = tp->packets_out - tp->sacked_out;
3521 int acked = 0; /* Number of packets newly acked */
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02003522
Kenneth Klette Jonassen31231a82015-05-01 01:10:58 +02003523 sack_state.first_sackt.v64 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003524
Eric Dumazetad971f62014-10-11 15:17:29 -07003525 /* We very likely will need to access write queue head. */
3526 prefetchw(sk->sk_write_queue.next);
3527
John Dykstra96e0bf42009-03-22 21:49:57 -07003528 /* If the ack is older than previous acks
Linus Torvalds1da177e2005-04-16 15:20:36 -07003529 * then we can probably ignore it.
3530 */
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003531 if (before(ack, prior_snd_una)) {
3532 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3533 if (before(ack, prior_snd_una - tp->max_window)) {
Neal Cardwellf2b2c582015-02-06 16:04:40 -05003534 tcp_send_challenge_ack(sk, skb);
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003535 return -1;
3536 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003537 goto old_ack;
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003538 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003539
John Dykstra96e0bf42009-03-22 21:49:57 -07003540 /* If the ack includes data we haven't sent yet, discard
3541 * this segment (RFC793 Section 3.9).
3542 */
3543 if (after(ack, tp->snd_nxt))
3544 goto invalid_ack;
3545
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00003546 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
3547 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00003548 tcp_rearm_rto(sk);
3549
Neal Cardwell0c9ab092014-08-14 16:13:07 -04003550 if (after(ack, prior_snd_una)) {
Ilpo Järvinen2e605292007-08-02 19:46:58 -07003551 flag |= FLAG_SND_UNA_ADVANCED;
Neal Cardwell0c9ab092014-08-14 16:13:07 -04003552 icsk->icsk_retransmits = 0;
3553 }
Ilpo Järvinen2e605292007-08-02 19:46:58 -07003554
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003555 prior_fackets = tp->fackets_out;
3556
Eric Dumazet12fb3dd2013-04-19 07:19:48 +00003557 /* ts_recent update must be made after we are sure that the packet
3558 * is in window.
3559 */
3560 if (flag & FLAG_UPDATE_TS_RECENT)
3561 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3562
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003563 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003564 /* Window is constant, pure forward advance.
3565 * No more checks are required.
3566 * Note, we use the fact that SND.UNA>=SND.WL2.
3567 */
Hantzis Fotisee7537b2009-03-02 22:42:02 -08003568 tcp_update_wl(tp, ack_seq);
Eric Dumazet0df48c22015-04-28 15:28:17 -07003569 tcp_snd_una_update(tp, ack);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003570 flag |= FLAG_WIN_UPDATE;
3571
Florian Westphal98900922014-09-26 22:37:35 +02003572 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
Stephen Hemminger317a76f2005-06-23 12:19:55 -07003573
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07003574 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003575 } else {
Florian Westphal98900922014-09-26 22:37:35 +02003576 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3577
Linus Torvalds1da177e2005-04-16 15:20:36 -07003578 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3579 flag |= FLAG_DATA;
3580 else
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07003581 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003582
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003583 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003584
3585 if (TCP_SKB_CB(skb)->sacked)
Yuchung Cheng59c9af42013-07-22 16:20:47 -07003586 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02003587 &sack_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003588
Florian Westphal735d3832014-09-29 13:08:30 +02003589 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003590 flag |= FLAG_ECE;
Florian Westphal98900922014-09-26 22:37:35 +02003591 ack_ev_flags |= CA_ACK_ECE;
3592 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003593
Florian Westphal98900922014-09-26 22:37:35 +02003594 if (flag & FLAG_WIN_UPDATE)
3595 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3596
3597 tcp_in_ack_event(sk, ack_ev_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003598 }
3599
3600 /* We passed data and got it acked, remove any soft error
3601 * log. Something worked...
3602 */
3603 sk->sk_err_soft = 0;
David S. Miller4b53fb62008-07-23 16:38:45 -07003604 icsk->icsk_probes_out = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003605 tp->rcv_tstamp = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003606 if (!prior_packets)
3607 goto no_queue;
3608
Linus Torvalds1da177e2005-04-16 15:20:36 -07003609 /* See if we can take anything off of the retransmit queue. */
Yuchung Cheng68049732013-05-29 14:20:11 +00003610 acked = tp->packets_out;
Eric Dumazet740b0f12014-02-26 14:02:48 -08003611 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una,
Kenneth Klette Jonassen31231a82015-05-01 01:10:58 +02003612 &sack_state);
Yuchung Cheng68049732013-05-29 14:20:11 +00003613 acked -= tp->packets_out;
Nandita Dukkipatia262f0c2011-08-21 20:21:57 +00003614
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003615 if (tcp_ack_is_dubious(sk, flag)) {
Neal Cardwell7d2b55f2011-11-16 08:58:01 +00003616 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
Yuchung Cheng68049732013-05-29 14:20:11 +00003617 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3618 is_dupack, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003619 }
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003620 if (tp->tlp_high_seq)
3621 tcp_process_tlp_ack(sk, ack, flag);
3622
Yuchung Chengb20a3fa2015-07-09 13:16:31 -07003623 /* Advance cwnd if state allows */
3624 if (tcp_may_raise_cwnd(sk, flag))
3625 tcp_cong_avoid(sk, ack, acked);
3626
David S. Miller5110effe2012-07-02 02:21:03 -07003627 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) {
3628 struct dst_entry *dst = __sk_dst_get(sk);
3629 if (dst)
3630 dst_confirm(dst);
3631 }
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00003632
3633 if (icsk->icsk_pending == ICSK_TIME_RETRANS)
3634 tcp_schedule_loss_probe(sk);
Eric Dumazet740b0f12014-02-26 14:02:48 -08003635 tcp_update_pacing_rate(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003636 return 1;
3637
3638no_queue:
Neal Cardwell5628adf2011-11-16 08:58:02 +00003639 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3640 if (flag & FLAG_DSACKING_ACK)
Yuchung Cheng68049732013-05-29 14:20:11 +00003641 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3642 is_dupack, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003643 /* If this ack opens up a zero window, clear backoff. It was
3644 * being used to time the probes, and is probably far higher than
3645 * it needs to be for normal retransmission.
3646 */
David S. Millerfe067e82007-03-07 12:12:44 -08003647 if (tcp_send_head(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003648 tcp_ack_probe(sk);
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003649
3650 if (tp->tlp_high_seq)
3651 tcp_process_tlp_ack(sk, ack, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003652 return 1;
3653
John Dykstra96e0bf42009-03-22 21:49:57 -07003654invalid_ack:
3655 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3656 return -1;
3657
Linus Torvalds1da177e2005-04-16 15:20:36 -07003658old_ack:
Neal Cardwelle95ae2f2011-11-16 08:58:03 +00003659 /* If data was SACKed, tag it and see if we should send more data.
3660 * If data was DSACKed, see if we can undo a cwnd reduction.
3661 */
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07003662 if (TCP_SKB_CB(skb)->sacked) {
Yuchung Cheng59c9af42013-07-22 16:20:47 -07003663 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02003664 &sack_state);
Yuchung Cheng68049732013-05-29 14:20:11 +00003665 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3666 is_dupack, flag);
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07003667 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003668
John Dykstra96e0bf42009-03-22 21:49:57 -07003669 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003670 return 0;
3671}
3672
Daniel Lee7f9b8382015-04-06 14:37:26 -07003673static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3674 bool syn, struct tcp_fastopen_cookie *foc,
3675 bool exp_opt)
3676{
3677 /* Valid only in SYN or SYN-ACK with an even length. */
3678 if (!foc || !syn || len < 0 || (len & 1))
3679 return;
3680
3681 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3682 len <= TCP_FASTOPEN_COOKIE_MAX)
3683 memcpy(foc->val, cookie, len);
3684 else if (len != 0)
3685 len = -1;
3686 foc->len = len;
3687 foc->exp = exp_opt;
3688}
3689
Linus Torvalds1da177e2005-04-16 15:20:36 -07003690/* Look for tcp options. Normally only called on SYN and SYNACK packets.
3691 * But, this can also be called on packets in the established flow when
3692 * the fast version below fails.
3693 */
Christoph Paasch1a2c6182013-03-17 08:23:34 +00003694void tcp_parse_options(const struct sk_buff *skb,
3695 struct tcp_options_received *opt_rx, int estab,
Yuchung Cheng2100c8d2012-07-19 06:43:05 +00003696 struct tcp_fastopen_cookie *foc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003697{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003698 const unsigned char *ptr;
3699 const struct tcphdr *th = tcp_hdr(skb);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003700 int length = (th->doff * 4) - sizeof(struct tcphdr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003701
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003702 ptr = (const unsigned char *)(th + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003703 opt_rx->saw_tstamp = 0;
3704
Stephen Hemminger2de979b2007-03-08 20:45:19 -08003705 while (length > 0) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003706 int opcode = *ptr++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003707 int opsize;
3708
3709 switch (opcode) {
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003710 case TCPOPT_EOL:
3711 return;
3712 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3713 length--;
3714 continue;
3715 default:
3716 opsize = *ptr++;
3717 if (opsize < 2) /* "silly options" */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003718 return;
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003719 if (opsize > length)
3720 return; /* don't parse partial options */
3721 switch (opcode) {
3722 case TCPOPT_MSS:
3723 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07003724 u16 in_mss = get_unaligned_be16(ptr);
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003725 if (in_mss) {
3726 if (opt_rx->user_mss &&
3727 opt_rx->user_mss < in_mss)
3728 in_mss = opt_rx->user_mss;
3729 opt_rx->mss_clamp = in_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003730 }
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07003731 }
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003732 break;
3733 case TCPOPT_WINDOW:
3734 if (opsize == TCPOLEN_WINDOW && th->syn &&
David S. Millerbb5b7c12009-12-15 20:56:42 -08003735 !estab && sysctl_tcp_window_scaling) {
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003736 __u8 snd_wscale = *(__u8 *)ptr;
3737 opt_rx->wscale_ok = 1;
3738 if (snd_wscale > 14) {
Joe Perchese87cc472012-05-13 21:56:26 +00003739 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3740 __func__,
3741 snd_wscale);
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003742 snd_wscale = 14;
3743 }
3744 opt_rx->snd_wscale = snd_wscale;
3745 }
3746 break;
3747 case TCPOPT_TIMESTAMP:
3748 if ((opsize == TCPOLEN_TIMESTAMP) &&
3749 ((estab && opt_rx->tstamp_ok) ||
David S. Millerbb5b7c12009-12-15 20:56:42 -08003750 (!estab && sysctl_tcp_timestamps))) {
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003751 opt_rx->saw_tstamp = 1;
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07003752 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3753 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003754 }
3755 break;
3756 case TCPOPT_SACK_PERM:
3757 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
David S. Millerbb5b7c12009-12-15 20:56:42 -08003758 !estab && sysctl_tcp_sack) {
Vijay Subramanianab562222011-12-20 13:23:24 +00003759 opt_rx->sack_ok = TCP_SACK_SEEN;
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003760 tcp_sack_reset(opt_rx);
3761 }
3762 break;
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07003763
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003764 case TCPOPT_SACK:
3765 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3766 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3767 opt_rx->sack_ok) {
3768 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3769 }
3770 break;
3771#ifdef CONFIG_TCP_MD5SIG
3772 case TCPOPT_MD5SIG:
3773 /*
3774 * The MD5 Hash has already been
3775 * checked (see tcp_v{4,6}_do_rcv()).
3776 */
3777 break;
3778#endif
Daniel Lee7f9b8382015-04-06 14:37:26 -07003779 case TCPOPT_FASTOPEN:
3780 tcp_parse_fastopen_option(
3781 opsize - TCPOLEN_FASTOPEN_BASE,
3782 ptr, th->syn, foc, false);
3783 break;
3784
Yuchung Cheng2100c8d2012-07-19 06:43:05 +00003785 case TCPOPT_EXP:
3786 /* Fast Open option shares code 254 using a
Daniel Lee7f9b8382015-04-06 14:37:26 -07003787 * 16 bits magic number.
Yuchung Cheng2100c8d2012-07-19 06:43:05 +00003788 */
Daniel Lee7f9b8382015-04-06 14:37:26 -07003789 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
3790 get_unaligned_be16(ptr) ==
3791 TCPOPT_FASTOPEN_MAGIC)
3792 tcp_parse_fastopen_option(opsize -
3793 TCPOLEN_EXP_FASTOPEN_BASE,
3794 ptr + 2, th->syn, foc, true);
Yuchung Cheng2100c8d2012-07-19 06:43:05 +00003795 break;
3796
3797 }
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003798 ptr += opsize-2;
3799 length -= opsize;
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07003800 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003801 }
3802}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00003803EXPORT_SYMBOL(tcp_parse_options);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003804
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003805static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
Ilpo Järvinena4356b22008-08-23 05:12:29 -07003806{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003807 const __be32 *ptr = (const __be32 *)(th + 1);
Ilpo Järvinena4356b22008-08-23 05:12:29 -07003808
3809 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3810 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3811 tp->rx_opt.saw_tstamp = 1;
3812 ++ptr;
3813 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3814 ++ptr;
Andrew Vagine3e12022013-08-27 12:21:55 +04003815 if (*ptr)
3816 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3817 else
3818 tp->rx_opt.rcv_tsecr = 0;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003819 return true;
Ilpo Järvinena4356b22008-08-23 05:12:29 -07003820 }
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003821 return false;
Ilpo Järvinena4356b22008-08-23 05:12:29 -07003822}
3823
Linus Torvalds1da177e2005-04-16 15:20:36 -07003824/* Fast parse options. This hopes to only see timestamps.
3825 * If it is wrong it falls back on tcp_parse_options().
3826 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003827static bool tcp_fast_parse_options(const struct sk_buff *skb,
Christoph Paasch1a2c6182013-03-17 08:23:34 +00003828 const struct tcphdr *th, struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003829{
William Allen Simpson4957faade2009-12-02 18:25:27 +00003830 /* In the spirit of fast parsing, compare doff directly to constant
3831 * values. Because equality is used, short doff can be ignored here.
3832 */
3833 if (th->doff == (sizeof(*th) / 4)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003834 tp->rx_opt.saw_tstamp = 0;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003835 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003836 } else if (tp->rx_opt.tstamp_ok &&
William Allen Simpson4957faade2009-12-02 18:25:27 +00003837 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
Ilpo Järvinena4356b22008-08-23 05:12:29 -07003838 if (tcp_parse_aligned_timestamp(tp, th))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003839 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003840 }
Andrey Vaginee684b62013-02-11 05:50:19 +00003841
Christoph Paasch1a2c6182013-03-17 08:23:34 +00003842 tcp_parse_options(skb, &tp->rx_opt, 1, NULL);
Andrew Vagine3e12022013-08-27 12:21:55 +04003843 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
Andrey Vaginee684b62013-02-11 05:50:19 +00003844 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3845
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003846 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003847}
3848
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003849#ifdef CONFIG_TCP_MD5SIG
3850/*
3851 * Parse MD5 Signature option
3852 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003853const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003854{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003855 int length = (th->doff << 2) - sizeof(*th);
3856 const u8 *ptr = (const u8 *)(th + 1);
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003857
3858 /* If the TCP option is too short, we can short cut */
3859 if (length < TCPOLEN_MD5SIG)
3860 return NULL;
3861
3862 while (length > 0) {
3863 int opcode = *ptr++;
3864 int opsize;
3865
Weilong Chena22318e2013-12-23 14:37:26 +08003866 switch (opcode) {
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003867 case TCPOPT_EOL:
3868 return NULL;
3869 case TCPOPT_NOP:
3870 length--;
3871 continue;
3872 default:
3873 opsize = *ptr++;
3874 if (opsize < 2 || opsize > length)
3875 return NULL;
3876 if (opcode == TCPOPT_MD5SIG)
Dmitry Popovba78e2d2010-08-07 20:24:28 -07003877 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003878 }
3879 ptr += opsize - 2;
3880 length -= opsize;
3881 }
3882 return NULL;
3883}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00003884EXPORT_SYMBOL(tcp_parse_md5sig_option);
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003885#endif
3886
Linus Torvalds1da177e2005-04-16 15:20:36 -07003887/* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3888 *
3889 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3890 * it can pass through stack. So, the following predicate verifies that
3891 * this segment is not used for anything but congestion avoidance or
3892 * fast retransmit. Moreover, we even are able to eliminate most of such
3893 * second order effects, if we apply some small "replay" window (~RTO)
3894 * to timestamp space.
3895 *
3896 * All these measures still do not guarantee that we reject wrapped ACKs
3897 * on networks with high bandwidth, when sequence space is recycled fastly,
3898 * but it guarantees that such events will be very rare and do not affect
3899 * connection seriously. This doesn't look nice, but alas, PAWS is really
3900 * buggy extension.
3901 *
3902 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3903 * states that events when retransmit arrives after original data are rare.
3904 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3905 * the biggest problem on large power networks even with minor reordering.
3906 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3907 * up to bandwidth of 18Gigabit/sec. 8) ]
3908 */
3909
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003910static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003911{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003912 const struct tcp_sock *tp = tcp_sk(sk);
3913 const struct tcphdr *th = tcp_hdr(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003914 u32 seq = TCP_SKB_CB(skb)->seq;
3915 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3916
3917 return (/* 1. Pure ACK with correct sequence number. */
3918 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3919
3920 /* 2. ... and duplicate ACK. */
3921 ack == tp->snd_una &&
3922
3923 /* 3. ... and does not update window. */
3924 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3925
3926 /* 4. ... and sits in replay window. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003927 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003928}
3929
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00003930static inline bool tcp_paws_discard(const struct sock *sk,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003931 const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003932{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003933 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinenc887e6d2009-03-14 14:23:03 +00003934
3935 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
3936 !tcp_disordered_ack(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003937}
3938
3939/* Check segment sequence number for validity.
3940 *
3941 * Segment controls are considered valid, if the segment
3942 * fits to the window after truncation to the window. Acceptability
3943 * of data (and SYN, FIN, of course) is checked separately.
3944 * See tcp_data_queue(), for example.
3945 *
3946 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3947 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3948 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3949 * (borrowed from freebsd)
3950 */
3951
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00003952static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003953{
3954 return !before(end_seq, tp->rcv_wup) &&
3955 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3956}
3957
3958/* When we get a reset we do this. */
Jerry Chu10467162012-08-31 12:29:11 +00003959void tcp_reset(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003960{
3961 /* We want the right error as BSD sees it (and indeed as we do). */
3962 switch (sk->sk_state) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003963 case TCP_SYN_SENT:
3964 sk->sk_err = ECONNREFUSED;
3965 break;
3966 case TCP_CLOSE_WAIT:
3967 sk->sk_err = EPIPE;
3968 break;
3969 case TCP_CLOSE:
3970 return;
3971 default:
3972 sk->sk_err = ECONNRESET;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003973 }
Tom Marshalla4d25802010-09-20 15:42:05 -07003974 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3975 smp_wmb();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003976
3977 if (!sock_flag(sk, SOCK_DEAD))
3978 sk->sk_error_report(sk);
3979
3980 tcp_done(sk);
3981}
3982
3983/*
3984 * Process the FIN bit. This now behaves as it is supposed to work
3985 * and the FIN takes effect when it is validly part of sequence
3986 * space. Not before when we get holes.
3987 *
3988 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3989 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3990 * TIME-WAIT)
3991 *
3992 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3993 * close and we go into CLOSING (and later onto TIME-WAIT)
3994 *
3995 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3996 */
Eric Dumazet20c4cb72011-10-20 17:44:03 -04003997static void tcp_fin(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003998{
3999 struct tcp_sock *tp = tcp_sk(sk);
4000
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004001 inet_csk_schedule_ack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004002
4003 sk->sk_shutdown |= RCV_SHUTDOWN;
4004 sock_set_flag(sk, SOCK_DONE);
4005
4006 switch (sk->sk_state) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004007 case TCP_SYN_RECV:
4008 case TCP_ESTABLISHED:
4009 /* Move to CLOSE_WAIT */
4010 tcp_set_state(sk, TCP_CLOSE_WAIT);
Jon Maxwell2251ae42015-07-08 10:12:28 +10004011 inet_csk(sk)->icsk_ack.pingpong = 1;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004012 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004013
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004014 case TCP_CLOSE_WAIT:
4015 case TCP_CLOSING:
4016 /* Received a retransmission of the FIN, do
4017 * nothing.
4018 */
4019 break;
4020 case TCP_LAST_ACK:
4021 /* RFC793: Remain in the LAST-ACK state. */
4022 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004023
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004024 case TCP_FIN_WAIT1:
4025 /* This case occurs when a simultaneous close
4026 * happens, we must ack the received FIN and
4027 * enter the CLOSING state.
4028 */
4029 tcp_send_ack(sk);
4030 tcp_set_state(sk, TCP_CLOSING);
4031 break;
4032 case TCP_FIN_WAIT2:
4033 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4034 tcp_send_ack(sk);
4035 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4036 break;
4037 default:
4038 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4039 * cases we should never reach this piece of code.
4040 */
Joe Perches058bd4d2012-03-11 18:36:11 +00004041 pr_err("%s: Impossible, sk->sk_state=%d\n",
Harvey Harrison0dc47872008-03-05 20:47:47 -08004042 __func__, sk->sk_state);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004043 break;
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07004044 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004045
4046 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4047 * Probably, we should reset in this case. For now drop them.
4048 */
4049 __skb_queue_purge(&tp->out_of_order_queue);
Ilpo Järvinene60402d2007-08-09 15:14:46 +03004050 if (tcp_is_sack(tp))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004051 tcp_sack_reset(&tp->rx_opt);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08004052 sk_mem_reclaim(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004053
4054 if (!sock_flag(sk, SOCK_DEAD)) {
4055 sk->sk_state_change(sk);
4056
4057 /* Do not send POLL_HUP for half duplex close. */
4058 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4059 sk->sk_state == TCP_CLOSE)
Pavel Emelyanov8d8ad9d2007-11-26 20:10:50 +08004060 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004061 else
Pavel Emelyanov8d8ad9d2007-11-26 20:10:50 +08004062 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004063 }
4064}
4065
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004066static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004067 u32 end_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004068{
4069 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4070 if (before(seq, sp->start_seq))
4071 sp->start_seq = seq;
4072 if (after(end_seq, sp->end_seq))
4073 sp->end_seq = end_seq;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004074 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004075 }
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004076 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004077}
4078
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004079static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004080{
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004081 struct tcp_sock *tp = tcp_sk(sk);
4082
David S. Millerbb5b7c12009-12-15 20:56:42 -08004083 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07004084 int mib_idx;
4085
Linus Torvalds1da177e2005-04-16 15:20:36 -07004086 if (before(seq, tp->rcv_nxt))
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07004087 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004088 else
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07004089 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4090
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07004091 NET_INC_STATS_BH(sock_net(sk), mib_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004092
4093 tp->rx_opt.dsack = 1;
4094 tp->duplicate_sack[0].start_seq = seq;
4095 tp->duplicate_sack[0].end_seq = end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004096 }
4097}
4098
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004099static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004100{
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004101 struct tcp_sock *tp = tcp_sk(sk);
4102
Linus Torvalds1da177e2005-04-16 15:20:36 -07004103 if (!tp->rx_opt.dsack)
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004104 tcp_dsack_set(sk, seq, end_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004105 else
4106 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4107}
4108
Eric Dumazetcf533ea2011-10-21 05:22:42 -04004109static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004110{
4111 struct tcp_sock *tp = tcp_sk(sk);
4112
4113 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4114 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07004115 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004116 tcp_enter_quickack_mode(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004117
David S. Millerbb5b7c12009-12-15 20:56:42 -08004118 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004119 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4120
4121 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4122 end_seq = tp->rcv_nxt;
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004123 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004124 }
4125 }
4126
4127 tcp_send_ack(sk);
4128}
4129
4130/* These routines update the SACK block as out-of-order packets arrive or
4131 * in-order packets close up the sequence space.
4132 */
4133static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4134{
4135 int this_sack;
4136 struct tcp_sack_block *sp = &tp->selective_acks[0];
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004137 struct tcp_sack_block *swalk = sp + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004138
4139 /* See if the recent change to the first SACK eats into
4140 * or hits the sequence space of other SACK blocks, if so coalesce.
4141 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004142 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004143 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4144 int i;
4145
4146 /* Zap SWALK, by moving every further SACK up by one slot.
4147 * Decrease num_sacks.
4148 */
4149 tp->rx_opt.num_sacks--;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004150 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4151 sp[i] = sp[i + 1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07004152 continue;
4153 }
4154 this_sack++, swalk++;
4155 }
4156}
4157
Linus Torvalds1da177e2005-04-16 15:20:36 -07004158static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4159{
4160 struct tcp_sock *tp = tcp_sk(sk);
4161 struct tcp_sack_block *sp = &tp->selective_acks[0];
4162 int cur_sacks = tp->rx_opt.num_sacks;
4163 int this_sack;
4164
4165 if (!cur_sacks)
4166 goto new_sack;
4167
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004168 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004169 if (tcp_sack_extend(sp, seq, end_seq)) {
4170 /* Rotate this_sack to the first one. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004171 for (; this_sack > 0; this_sack--, sp--)
Ilpo Järvinena0bffff2009-03-21 13:36:17 -07004172 swap(*sp, *(sp - 1));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004173 if (cur_sacks > 1)
4174 tcp_sack_maybe_coalesce(tp);
4175 return;
4176 }
4177 }
4178
4179 /* Could not find an adjacent existing SACK, build a new one,
4180 * put it at the front, and shift everyone else down. We
4181 * always know there is at least one SACK present already here.
4182 *
4183 * If the sack array is full, forget about the last one.
4184 */
Adam Langley4389dde2008-07-19 00:07:02 -07004185 if (this_sack >= TCP_NUM_SACKS) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004186 this_sack--;
4187 tp->rx_opt.num_sacks--;
4188 sp--;
4189 }
Stephen Hemminger2de979b2007-03-08 20:45:19 -08004190 for (; this_sack > 0; this_sack--, sp--)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004191 *sp = *(sp - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004192
4193new_sack:
4194 /* Build the new head SACK, and we're done. */
4195 sp->start_seq = seq;
4196 sp->end_seq = end_seq;
4197 tp->rx_opt.num_sacks++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004198}
4199
4200/* RCV.NXT advances, some SACKs should be eaten. */
4201
4202static void tcp_sack_remove(struct tcp_sock *tp)
4203{
4204 struct tcp_sack_block *sp = &tp->selective_acks[0];
4205 int num_sacks = tp->rx_opt.num_sacks;
4206 int this_sack;
4207
4208 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
David S. Millerb03efcf2005-07-08 14:57:23 -07004209 if (skb_queue_empty(&tp->out_of_order_queue)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004210 tp->rx_opt.num_sacks = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004211 return;
4212 }
4213
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004214 for (this_sack = 0; this_sack < num_sacks;) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004215 /* Check if the start of the sack is covered by RCV.NXT. */
4216 if (!before(tp->rcv_nxt, sp->start_seq)) {
4217 int i;
4218
4219 /* RCV.NXT must cover all the block! */
Ilpo Järvinen547b7922008-07-25 21:43:18 -07004220 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004221
4222 /* Zap this SACK, by moving forward any other SACKS. */
Weilong Chena22318e2013-12-23 14:37:26 +08004223 for (i = this_sack+1; i < num_sacks; i++)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004224 tp->selective_acks[i-1] = tp->selective_acks[i];
4225 num_sacks--;
4226 continue;
4227 }
4228 this_sack++;
4229 sp++;
4230 }
Ilpo Järvinen5861f8e2009-03-14 14:23:01 +00004231 tp->rx_opt.num_sacks = num_sacks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004232}
4233
Eric Dumazet1402d362012-04-23 07:11:42 +00004234/**
4235 * tcp_try_coalesce - try to merge skb to prior one
4236 * @sk: socket
4237 * @to: prior buffer
4238 * @from: buffer to add in queue
Eric Dumazet923dd342012-05-02 07:55:58 +00004239 * @fragstolen: pointer to boolean
Eric Dumazet1402d362012-04-23 07:11:42 +00004240 *
4241 * Before queueing skb @from after @to, try to merge them
4242 * to reduce overall memory use and queue lengths, if cost is small.
4243 * Packets in ofo or receive queues can stay a long time.
4244 * Better try to coalesce them right now to avoid future collapses.
Eric Dumazet783c1752012-04-23 17:34:36 +00004245 * Returns true if caller should free @from instead of queueing it
Eric Dumazet1402d362012-04-23 07:11:42 +00004246 */
Eric Dumazet783c1752012-04-23 17:34:36 +00004247static bool tcp_try_coalesce(struct sock *sk,
4248 struct sk_buff *to,
Eric Dumazet329033f2012-04-27 00:38:33 +00004249 struct sk_buff *from,
4250 bool *fragstolen)
Eric Dumazet1402d362012-04-23 07:11:42 +00004251{
Eric Dumazetbad43ca2012-05-19 03:02:02 +00004252 int delta;
Eric Dumazet1402d362012-04-23 07:11:42 +00004253
Eric Dumazet329033f2012-04-27 00:38:33 +00004254 *fragstolen = false;
Alexander Duyck34a802a2012-05-02 21:19:09 +00004255
Eric Dumazet1ca7ee32012-05-23 17:51:37 +00004256 /* Its possible this segment overlaps with prior segment in queue */
4257 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4258 return false;
4259
Eric Dumazetbad43ca2012-05-19 03:02:02 +00004260 if (!skb_try_coalesce(to, from, fragstolen, &delta))
Eric Dumazet783c1752012-04-23 17:34:36 +00004261 return false;
Alexander Duyck34a802a2012-05-02 21:19:09 +00004262
Alexander Duyck34a802a2012-05-02 21:19:09 +00004263 atomic_add(delta, &sk->sk_rmem_alloc);
4264 sk_mem_charge(sk, delta);
Alexander Duyck34a802a2012-05-02 21:19:09 +00004265 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4266 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4267 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
Eric Dumazete93a0432014-09-15 04:19:52 -07004268 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
Alexander Duyck34a802a2012-05-02 21:19:09 +00004269 return true;
Eric Dumazet1402d362012-04-23 07:11:42 +00004270}
4271
Linus Torvalds1da177e2005-04-16 15:20:36 -07004272/* This one checks to see if we can put data from the
4273 * out_of_order queue into the receive_queue.
4274 */
4275static void tcp_ofo_queue(struct sock *sk)
4276{
4277 struct tcp_sock *tp = tcp_sk(sk);
4278 __u32 dsack_high = tp->rcv_nxt;
Eric Dumazetbd1e75a2014-09-19 08:26:20 -07004279 struct sk_buff *skb, *tail;
4280 bool fragstolen, eaten;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004281
4282 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4283 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4284 break;
4285
4286 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4287 __u32 dsack = dsack_high;
4288 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4289 dsack_high = TCP_SKB_CB(skb)->end_seq;
4290 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4291 }
4292
Eric Dumazetbd1e75a2014-09-19 08:26:20 -07004293 __skb_unlink(skb, &tp->out_of_order_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004294 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4295 SOCK_DEBUG(sk, "ofo packet was already received\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07004296 __kfree_skb(skb);
4297 continue;
4298 }
4299 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4300 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4301 TCP_SKB_CB(skb)->end_seq);
4302
Eric Dumazetbd1e75a2014-09-19 08:26:20 -07004303 tail = skb_peek_tail(&sk->sk_receive_queue);
4304 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
Eric Dumazetbdd1f9e2015-04-28 15:28:18 -07004305 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
Eric Dumazetbd1e75a2014-09-19 08:26:20 -07004306 if (!eaten)
4307 __skb_queue_tail(&sk->sk_receive_queue, skb);
4308 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004309 tcp_fin(sk);
Eric Dumazetbd1e75a2014-09-19 08:26:20 -07004310 if (eaten)
4311 kfree_skb_partial(skb, fragstolen);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004312 }
4313}
4314
4315static bool tcp_prune_ofo_queue(struct sock *sk);
4316static int tcp_prune_queue(struct sock *sk);
4317
4318static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4319 unsigned int size)
4320{
4321 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4322 !sk_rmem_schedule(sk, skb, size)) {
4323
4324 if (tcp_prune_queue(sk) < 0)
4325 return -1;
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004326
Linus Torvalds1da177e2005-04-16 15:20:36 -07004327 if (!sk_rmem_schedule(sk, skb, size)) {
4328 if (!tcp_prune_ofo_queue(sk))
4329 return -1;
Frans Popb1383382010-03-24 07:57:28 +00004330
David S. Miller8728b832005-08-09 19:25:21 -07004331 if (!sk_rmem_schedule(sk, skb, size))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004332 return -1;
4333 }
4334 }
4335 return 0;
4336}
4337
Eric Dumazete86b29192012-03-18 11:06:44 +00004338static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4339{
4340 struct tcp_sock *tp = tcp_sk(sk);
4341 struct sk_buff *skb1;
4342 u32 seq, end_seq;
4343
Florian Westphal735d3832014-09-29 13:08:30 +02004344 tcp_ecn_check_ce(tp, skb);
Eric Dumazete86b29192012-03-18 11:06:44 +00004345
Mel Gormanc76562b2012-07-31 16:44:41 -07004346 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
Eric Dumazeta6df1ae2012-07-16 01:41:36 +00004347 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFODROP);
Eric Dumazete86b29192012-03-18 11:06:44 +00004348 __kfree_skb(skb);
4349 return;
4350 }
4351
4352 /* Disable header prediction. */
4353 tp->pred_flags = 0;
4354 inet_csk_schedule_ack(sk);
4355
Eric Dumazeta6df1ae2012-07-16 01:41:36 +00004356 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
Eric Dumazete86b29192012-03-18 11:06:44 +00004357 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4358 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4359
4360 skb1 = skb_peek_tail(&tp->out_of_order_queue);
4361 if (!skb1) {
4362 /* Initial out of order segment, build 1 SACK. */
4363 if (tcp_is_sack(tp)) {
4364 tp->rx_opt.num_sacks = 1;
4365 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4366 tp->selective_acks[0].end_seq =
4367 TCP_SKB_CB(skb)->end_seq;
4368 }
4369 __skb_queue_head(&tp->out_of_order_queue, skb);
4370 goto end;
4371 }
4372
4373 seq = TCP_SKB_CB(skb)->seq;
4374 end_seq = TCP_SKB_CB(skb)->end_seq;
4375
4376 if (seq == TCP_SKB_CB(skb1)->end_seq) {
Eric Dumazet329033f2012-04-27 00:38:33 +00004377 bool fragstolen;
4378
4379 if (!tcp_try_coalesce(sk, skb1, skb, &fragstolen)) {
Eric Dumazet1402d362012-04-23 07:11:42 +00004380 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4381 } else {
Eric Dumazet4e4f1fc2013-09-06 10:35:58 -07004382 tcp_grow_window(sk, skb);
Eric Dumazet923dd342012-05-02 07:55:58 +00004383 kfree_skb_partial(skb, fragstolen);
Eric Dumazetc8628152012-03-18 11:07:47 +00004384 skb = NULL;
Eric Dumazetc8628152012-03-18 11:07:47 +00004385 }
Eric Dumazete86b29192012-03-18 11:06:44 +00004386
4387 if (!tp->rx_opt.num_sacks ||
4388 tp->selective_acks[0].end_seq != seq)
4389 goto add_sack;
4390
4391 /* Common case: data arrive in order after hole. */
4392 tp->selective_acks[0].end_seq = end_seq;
4393 goto end;
4394 }
4395
4396 /* Find place to insert this segment. */
4397 while (1) {
4398 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4399 break;
4400 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4401 skb1 = NULL;
4402 break;
4403 }
4404 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4405 }
4406
4407 /* Do skb overlap to previous one? */
4408 if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4409 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4410 /* All the bits are present. Drop. */
Eric Dumazeta6df1ae2012-07-16 01:41:36 +00004411 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
Eric Dumazete86b29192012-03-18 11:06:44 +00004412 __kfree_skb(skb);
4413 skb = NULL;
4414 tcp_dsack_set(sk, seq, end_seq);
4415 goto add_sack;
4416 }
4417 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4418 /* Partial overlap. */
4419 tcp_dsack_set(sk, seq,
4420 TCP_SKB_CB(skb1)->end_seq);
4421 } else {
4422 if (skb_queue_is_first(&tp->out_of_order_queue,
4423 skb1))
4424 skb1 = NULL;
4425 else
4426 skb1 = skb_queue_prev(
4427 &tp->out_of_order_queue,
4428 skb1);
4429 }
4430 }
4431 if (!skb1)
4432 __skb_queue_head(&tp->out_of_order_queue, skb);
4433 else
4434 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4435
4436 /* And clean segments covered by new one as whole. */
4437 while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4438 skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4439
4440 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4441 break;
4442 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4443 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4444 end_seq);
4445 break;
4446 }
4447 __skb_unlink(skb1, &tp->out_of_order_queue);
4448 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4449 TCP_SKB_CB(skb1)->end_seq);
Eric Dumazeta6df1ae2012-07-16 01:41:36 +00004450 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
Eric Dumazete86b29192012-03-18 11:06:44 +00004451 __kfree_skb(skb1);
4452 }
4453
4454add_sack:
4455 if (tcp_is_sack(tp))
4456 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4457end:
Eric Dumazet4e4f1fc2013-09-06 10:35:58 -07004458 if (skb) {
4459 tcp_grow_window(sk, skb);
Eric Dumazete86b29192012-03-18 11:06:44 +00004460 skb_set_owner_r(skb, sk);
Eric Dumazet4e4f1fc2013-09-06 10:35:58 -07004461 }
Eric Dumazete86b29192012-03-18 11:06:44 +00004462}
4463
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004464static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
Eric Dumazetb081f852012-05-02 09:58:29 +00004465 bool *fragstolen)
4466{
4467 int eaten;
4468 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4469
4470 __skb_pull(skb, hdrlen);
4471 eaten = (tail &&
4472 tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0;
Eric Dumazetbdd1f9e2015-04-28 15:28:18 -07004473 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
Eric Dumazetb081f852012-05-02 09:58:29 +00004474 if (!eaten) {
4475 __skb_queue_tail(&sk->sk_receive_queue, skb);
4476 skb_set_owner_r(skb, sk);
4477 }
4478 return eaten;
4479}
Eric Dumazete86b29192012-03-18 11:06:44 +00004480
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004481int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4482{
Eric Dumazetcb934712014-09-17 03:14:42 -07004483 struct sk_buff *skb;
Eric Dumazet5d4c9bf2015-11-18 21:03:33 -08004484 int err = -ENOMEM;
4485 int data_len = 0;
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004486 bool fragstolen;
4487
Pavel Emelyanovc454e612012-10-29 05:05:33 +00004488 if (size == 0)
4489 return 0;
4490
Eric Dumazet5d4c9bf2015-11-18 21:03:33 -08004491 if (size > PAGE_SIZE) {
4492 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4493
4494 data_len = npages << PAGE_SHIFT;
4495 size = data_len + (size & ~PAGE_MASK);
4496 }
4497 skb = alloc_skb_with_frags(size - data_len, data_len,
4498 PAGE_ALLOC_COSTLY_ORDER,
4499 &err, sk->sk_allocation);
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004500 if (!skb)
4501 goto err;
4502
Eric Dumazet5d4c9bf2015-11-18 21:03:33 -08004503 skb_put(skb, size - data_len);
4504 skb->data_len = data_len;
4505 skb->len = size;
4506
Eric Dumazetcb934712014-09-17 03:14:42 -07004507 if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
Mel Gormanc76562b2012-07-31 16:44:41 -07004508 goto err_free;
4509
Eric Dumazet5d4c9bf2015-11-18 21:03:33 -08004510 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
4511 if (err)
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004512 goto err_free;
4513
4514 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4515 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4516 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4517
Eric Dumazetcb934712014-09-17 03:14:42 -07004518 if (tcp_queue_rcv(sk, skb, 0, &fragstolen)) {
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004519 WARN_ON_ONCE(fragstolen); /* should not happen */
4520 __kfree_skb(skb);
4521 }
4522 return size;
4523
4524err_free:
4525 kfree_skb(skb);
4526err:
Eric Dumazet5d4c9bf2015-11-18 21:03:33 -08004527 return err;
4528
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004529}
4530
Linus Torvalds1da177e2005-04-16 15:20:36 -07004531static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4532{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004533 struct tcp_sock *tp = tcp_sk(sk);
4534 int eaten = -1;
Eric Dumazet329033f2012-04-27 00:38:33 +00004535 bool fragstolen = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004536
4537 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4538 goto drop;
4539
Eric Dumazetf84af322010-04-28 15:31:51 -07004540 skb_dst_drop(skb);
Peter Pan(潘卫平)155c6e12014-09-24 22:17:02 +08004541 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004542
Florian Westphal735d3832014-09-29 13:08:30 +02004543 tcp_ecn_accept_cwr(tp, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004544
Ilpo Järvinen5861f8e2009-03-14 14:23:01 +00004545 tp->rx_opt.dsack = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004546
4547 /* Queue data for delivery to the user.
4548 * Packets in sequence go to the receive queue.
4549 * Out of sequence packets to the out_of_order_queue.
4550 */
4551 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4552 if (tcp_receive_window(tp) == 0)
4553 goto out_of_window;
4554
4555 /* Ok. In sequence. In window. */
4556 if (tp->ucopy.task == current &&
4557 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4558 sock_owned_by_user(sk) && !tp->urg_data) {
4559 int chunk = min_t(unsigned int, skb->len,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004560 tp->ucopy.len);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004561
4562 __set_current_state(TASK_RUNNING);
4563
4564 local_bh_enable();
Al Virof4362a22014-11-24 13:26:06 -05004565 if (!skb_copy_datagram_msg(skb, 0, tp->ucopy.msg, chunk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004566 tp->ucopy.len -= chunk;
4567 tp->copied_seq += chunk;
Jerry Chu44f53242011-01-25 13:46:30 -08004568 eaten = (chunk == skb->len);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004569 tcp_rcv_space_adjust(sk);
4570 }
4571 local_bh_disable();
4572 }
4573
4574 if (eaten <= 0) {
4575queue_and_out:
Eric Dumazet76dfa602015-05-15 12:39:29 -07004576 if (eaten < 0) {
4577 if (skb_queue_len(&sk->sk_receive_queue) == 0)
4578 sk_forced_mem_schedule(sk, skb->truesize);
4579 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4580 goto drop;
4581 }
Eric Dumazetb081f852012-05-02 09:58:29 +00004582 eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004583 }
Eric Dumazetbdd1f9e2015-04-28 15:28:18 -07004584 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
Stephen Hemminger2de979b2007-03-08 20:45:19 -08004585 if (skb->len)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004586 tcp_event_data_recv(sk, skb);
Peter Pan(潘卫平)155c6e12014-09-24 22:17:02 +08004587 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
Eric Dumazet20c4cb72011-10-20 17:44:03 -04004588 tcp_fin(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004589
David S. Millerb03efcf2005-07-08 14:57:23 -07004590 if (!skb_queue_empty(&tp->out_of_order_queue)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004591 tcp_ofo_queue(sk);
4592
4593 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4594 * gap in queue is filled.
4595 */
David S. Millerb03efcf2005-07-08 14:57:23 -07004596 if (skb_queue_empty(&tp->out_of_order_queue))
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004597 inet_csk(sk)->icsk_ack.pingpong = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004598 }
4599
4600 if (tp->rx_opt.num_sacks)
4601 tcp_sack_remove(tp);
4602
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004603 tcp_fast_path_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004604
Eric Dumazet923dd342012-05-02 07:55:58 +00004605 if (eaten > 0)
4606 kfree_skb_partial(skb, fragstolen);
Eric Dumazet1d57f192012-09-17 12:51:39 +00004607 if (!sock_flag(sk, SOCK_DEAD))
David S. Miller676d2362014-04-11 16:15:36 -04004608 sk->sk_data_ready(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004609 return;
4610 }
4611
4612 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4613 /* A retransmit, 2nd most common case. Force an immediate ack. */
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07004614 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004615 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004616
4617out_of_window:
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004618 tcp_enter_quickack_mode(sk);
4619 inet_csk_schedule_ack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004620drop:
4621 __kfree_skb(skb);
4622 return;
4623 }
4624
4625 /* Out of window. F.e. zero window probe. */
4626 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4627 goto out_of_window;
4628
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004629 tcp_enter_quickack_mode(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004630
4631 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4632 /* Partial packet, seq < rcv_next < end_seq */
4633 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4634 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4635 TCP_SKB_CB(skb)->end_seq);
4636
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004637 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004638
Linus Torvalds1da177e2005-04-16 15:20:36 -07004639 /* If window is closed, drop tail of packet. But after
4640 * remembering D-SACK for its head made in previous line.
4641 */
4642 if (!tcp_receive_window(tp))
4643 goto out_of_window;
4644 goto queue_and_out;
4645 }
4646
Eric Dumazete86b29192012-03-18 11:06:44 +00004647 tcp_data_queue_ofo(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004648}
4649
Ilpo Järvinen2cf46632008-08-23 05:11:41 -07004650static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4651 struct sk_buff_head *list)
4652{
David S. Miller91521942009-05-28 21:35:47 -07004653 struct sk_buff *next = NULL;
4654
4655 if (!skb_queue_is_last(list, skb))
4656 next = skb_queue_next(list, skb);
Ilpo Järvinen2cf46632008-08-23 05:11:41 -07004657
4658 __skb_unlink(skb, list);
4659 __kfree_skb(skb);
4660 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4661
4662 return next;
4663}
4664
Linus Torvalds1da177e2005-04-16 15:20:36 -07004665/* Collapse contiguous sequence of skbs head..tail with
4666 * sequence numbers start..end.
David S. Miller91521942009-05-28 21:35:47 -07004667 *
4668 * If tail is NULL, this means until the end of the list.
4669 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07004670 * Segments with FIN/SYN are not collapsed (only because this
4671 * simplifies code)
4672 */
4673static void
David S. Miller8728b832005-08-09 19:25:21 -07004674tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4675 struct sk_buff *head, struct sk_buff *tail,
4676 u32 start, u32 end)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004677{
David S. Miller91521942009-05-28 21:35:47 -07004678 struct sk_buff *skb, *n;
4679 bool end_of_skbs;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004680
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08004681 /* First, check that queue is collapsible and find
Linus Torvalds1da177e2005-04-16 15:20:36 -07004682 * the point where collapsing can be useful. */
David S. Miller91521942009-05-28 21:35:47 -07004683 skb = head;
4684restart:
4685 end_of_skbs = true;
4686 skb_queue_walk_from_safe(list, skb, n) {
4687 if (skb == tail)
4688 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004689 /* No new bits? It is possible on ofo queue. */
4690 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
Ilpo Järvinen2cf46632008-08-23 05:11:41 -07004691 skb = tcp_collapse_one(sk, skb, list);
David S. Miller91521942009-05-28 21:35:47 -07004692 if (!skb)
4693 break;
4694 goto restart;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004695 }
4696
4697 /* The first skb to collapse is:
4698 * - not SYN/FIN and
4699 * - bloated or contains data before "start" or
4700 * overlaps to the next one.
4701 */
Eric Dumazete11ecdd2014-09-15 04:19:51 -07004702 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07004703 (tcp_win_from_space(skb->truesize) > skb->len ||
David S. Miller91521942009-05-28 21:35:47 -07004704 before(TCP_SKB_CB(skb)->seq, start))) {
4705 end_of_skbs = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004706 break;
David S. Miller91521942009-05-28 21:35:47 -07004707 }
4708
4709 if (!skb_queue_is_last(list, skb)) {
4710 struct sk_buff *next = skb_queue_next(list, skb);
4711 if (next != tail &&
4712 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4713 end_of_skbs = false;
4714 break;
4715 }
4716 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004717
4718 /* Decided to skip this, advance start seq. */
4719 start = TCP_SKB_CB(skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004720 }
Eric Dumazete11ecdd2014-09-15 04:19:51 -07004721 if (end_of_skbs ||
4722 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004723 return;
4724
4725 while (before(start, end)) {
Eric Dumazetb3d6cb92014-09-15 04:19:53 -07004726 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004727 struct sk_buff *nskb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004728
Eric Dumazetb3d6cb92014-09-15 04:19:53 -07004729 nskb = alloc_skb(copy, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004730 if (!nskb)
4731 return;
Arnaldo Carvalho de Meloc51957d2007-03-10 12:47:22 -03004732
Linus Torvalds1da177e2005-04-16 15:20:36 -07004733 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4734 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
David S. Miller43f59c82008-09-21 21:28:51 -07004735 __skb_queue_before(list, skb, nskb);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08004736 skb_set_owner_r(nskb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004737
4738 /* Copy data, releasing collapsed skbs. */
4739 while (copy > 0) {
4740 int offset = start - TCP_SKB_CB(skb)->seq;
4741 int size = TCP_SKB_CB(skb)->end_seq - start;
4742
Kris Katterjohn09a62662006-01-08 22:24:28 -08004743 BUG_ON(offset < 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004744 if (size > 0) {
4745 size = min(copy, size);
4746 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4747 BUG();
4748 TCP_SKB_CB(nskb)->end_seq += size;
4749 copy -= size;
4750 start += size;
4751 }
4752 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
Ilpo Järvinen2cf46632008-08-23 05:11:41 -07004753 skb = tcp_collapse_one(sk, skb, list);
David S. Miller91521942009-05-28 21:35:47 -07004754 if (!skb ||
4755 skb == tail ||
Eric Dumazete11ecdd2014-09-15 04:19:51 -07004756 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004757 return;
4758 }
4759 }
4760 }
4761}
4762
4763/* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4764 * and tcp_collapse() them until all the queue is collapsed.
4765 */
4766static void tcp_collapse_ofo_queue(struct sock *sk)
4767{
4768 struct tcp_sock *tp = tcp_sk(sk);
4769 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4770 struct sk_buff *head;
4771 u32 start, end;
4772
Ian Morris51456b22015-04-03 09:17:26 +01004773 if (!skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004774 return;
4775
4776 start = TCP_SKB_CB(skb)->seq;
4777 end = TCP_SKB_CB(skb)->end_seq;
4778 head = skb;
4779
4780 for (;;) {
David S. Miller91521942009-05-28 21:35:47 -07004781 struct sk_buff *next = NULL;
4782
4783 if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4784 next = skb_queue_next(&tp->out_of_order_queue, skb);
4785 skb = next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004786
4787 /* Segment is terminated when we see gap or when
4788 * we are at the end of all the queue. */
David S. Miller91521942009-05-28 21:35:47 -07004789 if (!skb ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07004790 after(TCP_SKB_CB(skb)->seq, end) ||
4791 before(TCP_SKB_CB(skb)->end_seq, start)) {
David S. Miller8728b832005-08-09 19:25:21 -07004792 tcp_collapse(sk, &tp->out_of_order_queue,
4793 head, skb, start, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004794 head = skb;
David S. Miller91521942009-05-28 21:35:47 -07004795 if (!skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004796 break;
4797 /* Start new segment */
4798 start = TCP_SKB_CB(skb)->seq;
4799 end = TCP_SKB_CB(skb)->end_seq;
4800 } else {
4801 if (before(TCP_SKB_CB(skb)->seq, start))
4802 start = TCP_SKB_CB(skb)->seq;
4803 if (after(TCP_SKB_CB(skb)->end_seq, end))
4804 end = TCP_SKB_CB(skb)->end_seq;
4805 }
4806 }
4807}
4808
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004809/*
4810 * Purge the out-of-order queue.
Vitaliy Gusev56f367b2008-04-15 20:26:34 -07004811 * Return true if queue was pruned.
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004812 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004813static bool tcp_prune_ofo_queue(struct sock *sk)
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004814{
4815 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004816 bool res = false;
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004817
4818 if (!skb_queue_empty(&tp->out_of_order_queue)) {
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07004819 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004820 __skb_queue_purge(&tp->out_of_order_queue);
4821
4822 /* Reset SACK state. A conforming SACK implementation will
4823 * do the same at a timeout based retransmit. When a connection
4824 * is in a sad state like this, we care only about integrity
4825 * of the connection not performance.
4826 */
4827 if (tp->rx_opt.sack_ok)
4828 tcp_sack_reset(&tp->rx_opt);
4829 sk_mem_reclaim(sk);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004830 res = true;
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004831 }
Vitaliy Gusev56f367b2008-04-15 20:26:34 -07004832 return res;
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004833}
4834
Linus Torvalds1da177e2005-04-16 15:20:36 -07004835/* Reduce allocated memory if we can, trying to get
4836 * the socket within its memory limits again.
4837 *
4838 * Return less than zero if we should start dropping frames
4839 * until the socket owning process reads some of the data
4840 * to stabilize the situation.
4841 */
4842static int tcp_prune_queue(struct sock *sk)
4843{
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004844 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004845
4846 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4847
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07004848 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004849
4850 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004851 tcp_clamp_window(sk);
Eric Dumazetb8da51e2015-05-15 12:39:27 -07004852 else if (tcp_under_memory_pressure(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004853 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4854
4855 tcp_collapse_ofo_queue(sk);
David S. Miller91521942009-05-28 21:35:47 -07004856 if (!skb_queue_empty(&sk->sk_receive_queue))
4857 tcp_collapse(sk, &sk->sk_receive_queue,
4858 skb_peek(&sk->sk_receive_queue),
4859 NULL,
4860 tp->copied_seq, tp->rcv_nxt);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08004861 sk_mem_reclaim(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004862
4863 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4864 return 0;
4865
4866 /* Collapsing did not help, destructive actions follow.
4867 * This must not ever occur. */
4868
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004869 tcp_prune_ofo_queue(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004870
4871 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4872 return 0;
4873
4874 /* If we are really being abused, tell the caller to silently
4875 * drop receive data on the floor. It will get retransmitted
4876 * and hopefully then we'll have sufficient space.
4877 */
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07004878 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004879
4880 /* Massive buffer overcommit. */
4881 tp->pred_flags = 0;
4882 return -1;
4883}
4884
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004885static bool tcp_should_expand_sndbuf(const struct sock *sk)
David S. Miller0d9901d2005-07-05 15:21:10 -07004886{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04004887 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004888
David S. Miller0d9901d2005-07-05 15:21:10 -07004889 /* If the user specified a specific send buffer setting, do
4890 * not modify it.
4891 */
4892 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004893 return false;
David S. Miller0d9901d2005-07-05 15:21:10 -07004894
4895 /* If we are under global TCP memory pressure, do not expand. */
Eric Dumazetb8da51e2015-05-15 12:39:27 -07004896 if (tcp_under_memory_pressure(sk))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004897 return false;
David S. Miller0d9901d2005-07-05 15:21:10 -07004898
4899 /* If we are under soft global TCP memory pressure, do not expand. */
Glauber Costa180d8cd2011-12-11 21:47:02 +00004900 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004901 return false;
David S. Miller0d9901d2005-07-05 15:21:10 -07004902
4903 /* If we filled the congestion window, do not expand. */
Neal Cardwell65148902015-02-20 13:33:16 -05004904 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004905 return false;
David S. Miller0d9901d2005-07-05 15:21:10 -07004906
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004907 return true;
David S. Miller0d9901d2005-07-05 15:21:10 -07004908}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004909
4910/* When incoming ACK allowed to free some skb from write_queue,
4911 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4912 * on the exit from tcp input handler.
4913 *
4914 * PROBLEM: sndbuf expansion does not work well with largesend.
4915 */
4916static void tcp_new_space(struct sock *sk)
4917{
4918 struct tcp_sock *tp = tcp_sk(sk);
4919
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004920 if (tcp_should_expand_sndbuf(sk)) {
Eric Dumazet6ae70532013-10-01 10:23:44 -07004921 tcp_sndbuf_expand(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004922 tp->snd_cwnd_stamp = tcp_time_stamp;
4923 }
4924
4925 sk->sk_write_space(sk);
4926}
4927
Stephen Hemminger40efc6f2006-01-03 16:03:49 -08004928static void tcp_check_space(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004929{
4930 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4931 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
jbaron@akamai.com3c715122015-04-20 20:05:07 +00004932 /* pairs with tcp_poll() */
4933 smp_mb__after_atomic();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004934 if (sk->sk_socket &&
4935 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4936 tcp_new_space(sk);
4937 }
4938}
4939
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004940static inline void tcp_data_snd_check(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004941{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004942 tcp_push_pending_frames(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004943 tcp_check_space(sk);
4944}
4945
4946/*
4947 * Check if sending an ack is needed.
4948 */
4949static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4950{
4951 struct tcp_sock *tp = tcp_sk(sk);
4952
4953 /* More than one full frame received... */
Joe Perches9d4fb272009-11-23 10:41:23 -08004954 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07004955 /* ... and right edge of window advances far enough.
4956 * (tcp_recvmsg() will send ACK otherwise). Or...
4957 */
Joe Perches9d4fb272009-11-23 10:41:23 -08004958 __tcp_select_window(sk) >= tp->rcv_wnd) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07004959 /* We ACK each frame or... */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004960 tcp_in_quickack_mode(sk) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07004961 /* We have out of order data. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004962 (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004963 /* Then ack it now */
4964 tcp_send_ack(sk);
4965 } else {
4966 /* Else, send delayed ack. */
4967 tcp_send_delayed_ack(sk);
4968 }
4969}
4970
Stephen Hemminger40efc6f2006-01-03 16:03:49 -08004971static inline void tcp_ack_snd_check(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004972{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004973 if (!inet_csk_ack_scheduled(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004974 /* We sent a data segment already. */
4975 return;
4976 }
4977 __tcp_ack_snd_check(sk, 1);
4978}
4979
4980/*
4981 * This routine is only called when we have urgent data
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08004982 * signaled. Its the 'slow' part of tcp_urg. It could be
Linus Torvalds1da177e2005-04-16 15:20:36 -07004983 * moved inline now as tcp_urg is only called from one
4984 * place. We handle URGent data wrong. We have to - as
4985 * BSD still doesn't use the correction from RFC961.
4986 * For 1003.1g we should support a new option TCP_STDURG to permit
4987 * either form (or just set the sysctl tcp_stdurg).
4988 */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004989
Eric Dumazetcf533ea2011-10-21 05:22:42 -04004990static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004991{
4992 struct tcp_sock *tp = tcp_sk(sk);
4993 u32 ptr = ntohs(th->urg_ptr);
4994
4995 if (ptr && !sysctl_tcp_stdurg)
4996 ptr--;
4997 ptr += ntohl(th->seq);
4998
4999 /* Ignore urgent data that we've already seen and read. */
5000 if (after(tp->copied_seq, ptr))
5001 return;
5002
5003 /* Do not replay urg ptr.
5004 *
5005 * NOTE: interesting situation not covered by specs.
5006 * Misbehaving sender may send urg ptr, pointing to segment,
5007 * which we already have in ofo queue. We are not able to fetch
5008 * such data and will stay in TCP_URG_NOTYET until will be eaten
5009 * by recvmsg(). Seems, we are not obliged to handle such wicked
5010 * situations. But it is worth to think about possibility of some
5011 * DoSes using some hypothetical application level deadlock.
5012 */
5013 if (before(ptr, tp->rcv_nxt))
5014 return;
5015
5016 /* Do we already have a newer (or duplicate) urgent pointer? */
5017 if (tp->urg_data && !after(ptr, tp->urg_seq))
5018 return;
5019
5020 /* Tell the world about our new urgent pointer. */
5021 sk_send_sigurg(sk);
5022
5023 /* We may be adding urgent data when the last byte read was
5024 * urgent. To do this requires some care. We cannot just ignore
5025 * tp->copied_seq since we would read the last urgent byte again
5026 * as data, nor can we alter copied_seq until this data arrives
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08005027 * or we break the semantics of SIOCATMARK (and thus sockatmark())
Linus Torvalds1da177e2005-04-16 15:20:36 -07005028 *
5029 * NOTE. Double Dutch. Rendering to plain English: author of comment
5030 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5031 * and expect that both A and B disappear from stream. This is _wrong_.
5032 * Though this happens in BSD with high probability, this is occasional.
5033 * Any application relying on this is buggy. Note also, that fix "works"
5034 * only in this artificial test. Insert some normal data between A and B and we will
5035 * decline of BSD again. Verdict: it is better to remove to trap
5036 * buggy users.
5037 */
5038 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005039 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005040 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5041 tp->copied_seq++;
5042 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
David S. Miller8728b832005-08-09 19:25:21 -07005043 __skb_unlink(skb, &sk->sk_receive_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005044 __kfree_skb(skb);
5045 }
5046 }
5047
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005048 tp->urg_data = TCP_URG_NOTYET;
5049 tp->urg_seq = ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005050
5051 /* Disable header prediction. */
5052 tp->pred_flags = 0;
5053}
5054
5055/* This is the 'fast' part of urgent handling. */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04005056static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005057{
5058 struct tcp_sock *tp = tcp_sk(sk);
5059
5060 /* Check if we get a new urgent pointer - normally not. */
5061 if (th->urg)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005062 tcp_check_urg(sk, th);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005063
5064 /* Do we wait for any urgent data? - normally not... */
5065 if (tp->urg_data == TCP_URG_NOTYET) {
5066 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5067 th->syn;
5068
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005069 /* Is the urgent pointer pointing into this packet? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005070 if (ptr < skb->len) {
5071 u8 tmp;
5072 if (skb_copy_bits(skb, ptr, &tmp, 1))
5073 BUG();
5074 tp->urg_data = TCP_URG_VALID | tmp;
5075 if (!sock_flag(sk, SOCK_DEAD))
David S. Miller676d2362014-04-11 16:15:36 -04005076 sk->sk_data_ready(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005077 }
5078 }
5079}
5080
5081static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
5082{
5083 struct tcp_sock *tp = tcp_sk(sk);
5084 int chunk = skb->len - hlen;
5085 int err;
5086
5087 local_bh_enable();
Herbert Xu60476372007-04-09 11:59:39 -07005088 if (skb_csum_unnecessary(skb))
Al Virof4362a22014-11-24 13:26:06 -05005089 err = skb_copy_datagram_msg(skb, hlen, tp->ucopy.msg, chunk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005090 else
Al Virof4362a22014-11-24 13:26:06 -05005091 err = skb_copy_and_csum_datagram_msg(skb, hlen, tp->ucopy.msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005092
5093 if (!err) {
5094 tp->ucopy.len -= chunk;
5095 tp->copied_seq += chunk;
5096 tcp_rcv_space_adjust(sk);
5097 }
5098
5099 local_bh_disable();
5100 return err;
5101}
5102
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005103static __sum16 __tcp_checksum_complete_user(struct sock *sk,
5104 struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005105{
Al Virob51655b2006-11-14 21:40:42 -08005106 __sum16 result;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005107
5108 if (sock_owned_by_user(sk)) {
5109 local_bh_enable();
5110 result = __tcp_checksum_complete(skb);
5111 local_bh_disable();
5112 } else {
5113 result = __tcp_checksum_complete(skb);
5114 }
5115 return result;
5116}
5117
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00005118static inline bool tcp_checksum_complete_user(struct sock *sk,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005119 struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005120{
Herbert Xu60476372007-04-09 11:59:39 -07005121 return !skb_csum_unnecessary(skb) &&
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005122 __tcp_checksum_complete_user(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005123}
5124
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005125/* Does PAWS and seqno based validation of an incoming segment, flags will
5126 * play significant role here.
5127 */
Eric Dumazet0c246042012-07-17 01:41:30 +00005128static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5129 const struct tcphdr *th, int syn_inerr)
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005130{
5131 struct tcp_sock *tp = tcp_sk(sk);
5132
5133 /* RFC1323: H1. Apply PAWS check first. */
Christoph Paasch1a2c6182013-03-17 08:23:34 +00005134 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005135 tcp_paws_discard(sk, skb)) {
5136 if (!th->rst) {
5137 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
Neal Cardwellf2b2c582015-02-06 16:04:40 -05005138 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5139 LINUX_MIB_TCPACKSKIPPEDPAWS,
5140 &tp->last_oow_ack_time))
5141 tcp_send_dupack(sk, skb);
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005142 goto discard;
5143 }
5144 /* Reset is accepted even if it did not pass PAWS. */
5145 }
5146
5147 /* Step 1: check sequence number */
5148 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5149 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5150 * (RST) segments are validated by checking their SEQ-fields."
5151 * And page 69: "If an incoming segment is not acceptable,
5152 * an acknowledgment should be sent in reply (unless the RST
5153 * bit is set, if so drop the segment and return)".
5154 */
Eric Dumazete3715892012-07-17 12:29:30 +00005155 if (!th->rst) {
5156 if (th->syn)
5157 goto syn_challenge;
Neal Cardwellf2b2c582015-02-06 16:04:40 -05005158 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5159 LINUX_MIB_TCPACKSKIPPEDSEQ,
5160 &tp->last_oow_ack_time))
5161 tcp_send_dupack(sk, skb);
Eric Dumazete3715892012-07-17 12:29:30 +00005162 }
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005163 goto discard;
5164 }
5165
5166 /* Step 2: check RST bit */
5167 if (th->rst) {
Eric Dumazet282f23c2012-07-17 10:13:05 +02005168 /* RFC 5961 3.2 :
5169 * If sequence number exactly matches RCV.NXT, then
5170 * RESET the connection
5171 * else
5172 * Send a challenge ACK
5173 */
5174 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt)
5175 tcp_reset(sk);
5176 else
Neal Cardwellf2b2c582015-02-06 16:04:40 -05005177 tcp_send_challenge_ack(sk, skb);
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005178 goto discard;
5179 }
5180
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005181 /* step 3: check security and precedence [ignored] */
5182
Eric Dumazet0c246042012-07-17 01:41:30 +00005183 /* step 4: Check for a SYN
Sowmini Varadhancd214532014-10-30 12:48:08 -04005184 * RFC 5961 4.2 : Send a challenge ack
Eric Dumazet0c246042012-07-17 01:41:30 +00005185 */
5186 if (th->syn) {
Eric Dumazete3715892012-07-17 12:29:30 +00005187syn_challenge:
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005188 if (syn_inerr)
5189 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
Eric Dumazet0c246042012-07-17 01:41:30 +00005190 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
Neal Cardwellf2b2c582015-02-06 16:04:40 -05005191 tcp_send_challenge_ack(sk, skb);
Eric Dumazet0c246042012-07-17 01:41:30 +00005192 goto discard;
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005193 }
5194
Eric Dumazet0c246042012-07-17 01:41:30 +00005195 return true;
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005196
5197discard:
5198 __kfree_skb(skb);
Eric Dumazet0c246042012-07-17 01:41:30 +00005199 return false;
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005200}
5201
Linus Torvalds1da177e2005-04-16 15:20:36 -07005202/*
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005203 * TCP receive function for the ESTABLISHED state.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005204 *
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005205 * It is split into a fast path and a slow path. The fast path is
Linus Torvalds1da177e2005-04-16 15:20:36 -07005206 * disabled when:
5207 * - A zero window was announced from us - zero window probing
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005208 * is only handled properly in the slow path.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005209 * - Out of order segments arrived.
5210 * - Urgent data is expected.
5211 * - There is no buffer space left
5212 * - Unexpected TCP flags/window values/header lengths are received
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005213 * (detected by checking the TCP header against pred_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005214 * - Data is sent in both directions. Fast path only supports pure senders
5215 * or pure receivers (this means either the sequence number or the ack
5216 * value must stay constant)
5217 * - Unexpected TCP option.
5218 *
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005219 * When these conditions are not satisfied it drops into a standard
Linus Torvalds1da177e2005-04-16 15:20:36 -07005220 * receive procedure patterned after RFC793 to handle all cases.
5221 * The first three cases are guaranteed by proper pred_flags setting,
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005222 * the rest is checked inline. Fast processing is turned on in
Linus Torvalds1da177e2005-04-16 15:20:36 -07005223 * tcp_data_queue when everything is OK.
5224 */
Vijay Subramanianc995ae22013-09-03 12:23:22 -07005225void tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5226 const struct tcphdr *th, unsigned int len)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005227{
5228 struct tcp_sock *tp = tcp_sk(sk);
5229
Ian Morris51456b22015-04-03 09:17:26 +01005230 if (unlikely(!sk->sk_rx_dst))
Eric Dumazet5d299f32012-08-06 05:09:33 +00005231 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005232 /*
5233 * Header prediction.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005234 * The code loosely follows the one in the famous
Linus Torvalds1da177e2005-04-16 15:20:36 -07005235 * "30 instruction TCP receive" Van Jacobson mail.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005236 *
5237 * Van's trick is to deposit buffers into socket queue
Linus Torvalds1da177e2005-04-16 15:20:36 -07005238 * on a device interrupt, to call tcp_recv function
5239 * on the receive process context and checksum and copy
5240 * the buffer to user space. smart...
5241 *
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005242 * Our current scheme is not silly either but we take the
Linus Torvalds1da177e2005-04-16 15:20:36 -07005243 * extra cost of the net_bh soft interrupt processing...
5244 * We do checksum and copy also but from device to kernel.
5245 */
5246
5247 tp->rx_opt.saw_tstamp = 0;
5248
5249 /* pred_flags is 0xS?10 << 16 + snd_wnd
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08005250 * if header_prediction is to be made
Linus Torvalds1da177e2005-04-16 15:20:36 -07005251 * 'S' will always be tp->tcp_header_len >> 2
5252 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005253 * turn it off (when there are holes in the receive
Linus Torvalds1da177e2005-04-16 15:20:36 -07005254 * space for instance)
5255 * PSH flag is ignored.
5256 */
5257
5258 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
John Dykstra96e0bf42009-03-22 21:49:57 -07005259 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5260 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005261 int tcp_header_len = tp->tcp_header_len;
5262
5263 /* Timestamp header prediction: tcp_header_len
5264 * is automatically equal to th->doff*4 due to pred_flags
5265 * match.
5266 */
5267
5268 /* Check timestamp */
5269 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005270 /* No? Slow path! */
Ilpo Järvinena4356b22008-08-23 05:12:29 -07005271 if (!tcp_parse_aligned_timestamp(tp, th))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005272 goto slow_path;
5273
Linus Torvalds1da177e2005-04-16 15:20:36 -07005274 /* If PAWS failed, check it more carefully in slow path */
5275 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5276 goto slow_path;
5277
5278 /* DO NOT update ts_recent here, if checksum fails
5279 * and timestamp was corrupted part, it will result
5280 * in a hung connection since we will drop all
5281 * future packets due to the PAWS test.
5282 */
5283 }
5284
5285 if (len <= tcp_header_len) {
5286 /* Bulk data transfer: sender */
5287 if (len == tcp_header_len) {
5288 /* Predicted packet is in window by definition.
5289 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5290 * Hence, check seq<=rcv_wup reduces to:
5291 */
5292 if (tcp_header_len ==
5293 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5294 tp->rcv_nxt == tp->rcv_wup)
5295 tcp_store_ts_recent(tp);
5296
Linus Torvalds1da177e2005-04-16 15:20:36 -07005297 /* We know that such packets are checksummed
5298 * on entry.
5299 */
5300 tcp_ack(sk, skb, 0);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005301 __kfree_skb(skb);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005302 tcp_data_snd_check(sk);
Vijay Subramanianc995ae22013-09-03 12:23:22 -07005303 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005304 } else { /* Header too small */
Pavel Emelyanov63231bd2008-07-16 20:22:25 -07005305 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005306 goto discard;
5307 }
5308 } else {
5309 int eaten = 0;
Eric Dumazetb081f852012-05-02 09:58:29 +00005310 bool fragstolen = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005311
Dan Williamsd27f9bc2013-12-30 11:37:15 -08005312 if (tp->ucopy.task == current &&
5313 tp->copied_seq == tp->rcv_nxt &&
5314 len - tcp_header_len <= tp->ucopy.len &&
5315 sock_owned_by_user(sk)) {
5316 __set_current_state(TASK_RUNNING);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005317
Dan Williamsd27f9bc2013-12-30 11:37:15 -08005318 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005319 /* Predicted packet is in window by definition.
5320 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5321 * Hence, check seq<=rcv_wup reduces to:
5322 */
5323 if (tcp_header_len ==
5324 (sizeof(struct tcphdr) +
5325 TCPOLEN_TSTAMP_ALIGNED) &&
5326 tp->rcv_nxt == tp->rcv_wup)
5327 tcp_store_ts_recent(tp);
5328
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005329 tcp_rcv_rtt_measure_ts(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005330
5331 __skb_pull(skb, tcp_header_len);
Eric Dumazetbdd1f9e2015-04-28 15:28:18 -07005332 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07005333 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
Dan Williamsd27f9bc2013-12-30 11:37:15 -08005334 eaten = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005335 }
5336 }
5337 if (!eaten) {
5338 if (tcp_checksum_complete_user(sk, skb))
5339 goto csum_error;
5340
Neal Cardwellaab2b4b2013-03-04 06:23:05 +00005341 if ((int)skb->truesize > sk->sk_forward_alloc)
5342 goto step5;
5343
Linus Torvalds1da177e2005-04-16 15:20:36 -07005344 /* Predicted packet is in window by definition.
5345 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5346 * Hence, check seq<=rcv_wup reduces to:
5347 */
5348 if (tcp_header_len ==
5349 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5350 tp->rcv_nxt == tp->rcv_wup)
5351 tcp_store_ts_recent(tp);
5352
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005353 tcp_rcv_rtt_measure_ts(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005354
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07005355 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005356
5357 /* Bulk data transfer: receiver */
Eric Dumazetb081f852012-05-02 09:58:29 +00005358 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5359 &fragstolen);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005360 }
5361
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005362 tcp_event_data_recv(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005363
5364 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5365 /* Well, only one small jumplet in fast path... */
5366 tcp_ack(sk, skb, FLAG_DATA);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005367 tcp_data_snd_check(sk);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005368 if (!inet_csk_ack_scheduled(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005369 goto no_ack;
5370 }
5371
Dan Williamsd27f9bc2013-12-30 11:37:15 -08005372 __tcp_ack_snd_check(sk, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005373no_ack:
5374 if (eaten)
Eric Dumazetb081f852012-05-02 09:58:29 +00005375 kfree_skb_partial(skb, fragstolen);
David S. Miller676d2362014-04-11 16:15:36 -04005376 sk->sk_data_ready(sk);
Vijay Subramanianc995ae22013-09-03 12:23:22 -07005377 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005378 }
5379 }
5380
5381slow_path:
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005382 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005383 goto csum_error;
5384
Calvin Owens0c228e82014-11-20 15:09:53 -08005385 if (!th->ack && !th->rst && !th->syn)
Eric Dumazetc3ae62af2012-12-26 12:44:34 +00005386 goto discard;
5387
Linus Torvalds1da177e2005-04-16 15:20:36 -07005388 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07005389 * Standard slow path.
5390 */
5391
Eric Dumazet0c246042012-07-17 01:41:30 +00005392 if (!tcp_validate_incoming(sk, skb, th, 1))
Vijay Subramanianc995ae22013-09-03 12:23:22 -07005393 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005394
5395step5:
Eric Dumazet12fb3dd2013-04-19 07:19:48 +00005396 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
John Dykstra96e0bf42009-03-22 21:49:57 -07005397 goto discard;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005398
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005399 tcp_rcv_rtt_measure_ts(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005400
5401 /* Process urgent data. */
5402 tcp_urg(sk, skb, th);
5403
5404 /* step 7: process the segment text */
5405 tcp_data_queue(sk, skb);
5406
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005407 tcp_data_snd_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005408 tcp_ack_snd_check(sk);
Vijay Subramanianc995ae22013-09-03 12:23:22 -07005409 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005410
5411csum_error:
Eric Dumazet6a5dc9e2013-04-29 08:39:56 +00005412 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_CSUMERRORS);
Pavel Emelyanov63231bd2008-07-16 20:22:25 -07005413 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005414
5415discard:
5416 __kfree_skb(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005417}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00005418EXPORT_SYMBOL(tcp_rcv_established);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005419
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005420void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5421{
5422 struct tcp_sock *tp = tcp_sk(sk);
5423 struct inet_connection_sock *icsk = inet_csk(sk);
5424
5425 tcp_set_state(sk, TCP_ESTABLISHED);
5426
Ian Morris00db4122015-04-03 09:17:27 +01005427 if (skb) {
Eric Dumazet5d299f32012-08-06 05:09:33 +00005428 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005429 security_inet_conn_established(sk, skb);
David S. Miller41063e92012-06-19 21:22:05 -07005430 }
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005431
5432 /* Make sure socket is routed, for correct metrics. */
5433 icsk->icsk_af_ops->rebuild_header(sk);
5434
5435 tcp_init_metrics(sk);
5436
5437 tcp_init_congestion_control(sk);
5438
5439 /* Prevent spurious tcp_cwnd_restart() on first data
5440 * packet.
5441 */
5442 tp->lsndtime = tcp_time_stamp;
5443
5444 tcp_init_buffer_space(sk);
5445
5446 if (sock_flag(sk, SOCK_KEEPOPEN))
5447 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5448
5449 if (!tp->rx_opt.snd_wscale)
5450 __tcp_fast_path_on(tp, tp->snd_wnd);
5451 else
5452 tp->pred_flags = 0;
5453
5454 if (!sock_flag(sk, SOCK_DEAD)) {
5455 sk->sk_state_change(sk);
5456 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5457 }
5458}
5459
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005460static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5461 struct tcp_fastopen_cookie *cookie)
5462{
5463 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng67da22d2012-07-19 06:43:11 +00005464 struct sk_buff *data = tp->syn_data ? tcp_write_queue_head(sk) : NULL;
Daniel Lee2646c832015-04-06 14:37:27 -07005465 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
5466 bool syn_drop = false;
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005467
5468 if (mss == tp->rx_opt.user_mss) {
5469 struct tcp_options_received opt;
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005470
5471 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5472 tcp_clear_options(&opt);
5473 opt.user_mss = opt.mss_clamp = 0;
Christoph Paasch1a2c6182013-03-17 08:23:34 +00005474 tcp_parse_options(synack, &opt, 0, NULL);
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005475 mss = opt.mss_clamp;
5476 }
5477
Daniel Lee2646c832015-04-06 14:37:27 -07005478 if (!tp->syn_fastopen) {
5479 /* Ignore an unsolicited cookie */
Yuchung Cheng67da22d2012-07-19 06:43:11 +00005480 cookie->len = -1;
Daniel Lee2646c832015-04-06 14:37:27 -07005481 } else if (tp->total_retrans) {
5482 /* SYN timed out and the SYN-ACK neither has a cookie nor
5483 * acknowledges data. Presumably the remote received only
5484 * the retransmitted (regular) SYNs: either the original
5485 * SYN-data or the corresponding SYN-ACK was dropped.
5486 */
5487 syn_drop = (cookie->len < 0 && data);
5488 } else if (cookie->len < 0 && !tp->syn_data) {
5489 /* We requested a cookie but didn't get it. If we did not use
5490 * the (old) exp opt format then try so next time (try_exp=1).
5491 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5492 */
5493 try_exp = tp->syn_fastopen_exp ? 2 : 1;
5494 }
Yuchung Cheng67da22d2012-07-19 06:43:11 +00005495
Daniel Lee2646c832015-04-06 14:37:27 -07005496 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005497
5498 if (data) { /* Retransmit unacked data in SYN */
Yuchung Cheng93b174a2012-12-06 08:45:32 +00005499 tcp_for_write_queue_from(data, sk) {
5500 if (data == tcp_send_head(sk) ||
5501 __tcp_retransmit_skb(sk, data))
5502 break;
5503 }
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005504 tcp_rearm_rto(sk);
Yuchung Chengf19c29e2014-03-03 12:31:36 -08005505 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVEFAIL);
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005506 return true;
5507 }
Yuchung Cheng6f736012012-10-19 15:14:44 +00005508 tp->syn_data_acked = tp->syn_data;
Yuchung Chengf19c29e2014-03-03 12:31:36 -08005509 if (tp->syn_data_acked)
5510 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005511 return false;
5512}
5513
Linus Torvalds1da177e2005-04-16 15:20:36 -07005514static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
Eric Dumazetbda07a62015-09-29 07:42:40 -07005515 const struct tcphdr *th)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005516{
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08005517 struct inet_connection_sock *icsk = inet_csk(sk);
William Allen Simpson4957faade2009-12-02 18:25:27 +00005518 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005519 struct tcp_fastopen_cookie foc = { .len = -1 };
William Allen Simpson4957faade2009-12-02 18:25:27 +00005520 int saved_clamp = tp->rx_opt.mss_clamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005521
Christoph Paasch1a2c6182013-03-17 08:23:34 +00005522 tcp_parse_options(skb, &tp->rx_opt, 0, &foc);
Andrew Vagine3e12022013-08-27 12:21:55 +04005523 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
Andrey Vaginee684b62013-02-11 05:50:19 +00005524 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005525
5526 if (th->ack) {
5527 /* rfc793:
5528 * "If the state is SYN-SENT then
5529 * first check the ACK bit
5530 * If the ACK bit is set
5531 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5532 * a reset (unless the RST bit is set, if so drop
5533 * the segment and return)"
Linus Torvalds1da177e2005-04-16 15:20:36 -07005534 */
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005535 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5536 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005537 goto reset_and_undo;
5538
5539 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5540 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5541 tcp_time_stamp)) {
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07005542 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005543 goto reset_and_undo;
5544 }
5545
5546 /* Now ACK is acceptable.
5547 *
5548 * "If the RST bit is set
5549 * If the ACK was acceptable then signal the user "error:
5550 * connection reset", drop the segment, enter CLOSED state,
5551 * delete TCB, and return."
5552 */
5553
5554 if (th->rst) {
5555 tcp_reset(sk);
5556 goto discard;
5557 }
5558
5559 /* rfc793:
5560 * "fifth, if neither of the SYN or RST bits is set then
5561 * drop the segment and return."
5562 *
5563 * See note below!
5564 * --ANK(990513)
5565 */
5566 if (!th->syn)
5567 goto discard_and_undo;
5568
5569 /* rfc793:
5570 * "If the SYN bit is on ...
5571 * are acceptable then ...
5572 * (our SYN has been ACKed), change the connection
5573 * state to ESTABLISHED..."
5574 */
5575
Florian Westphal735d3832014-09-29 13:08:30 +02005576 tcp_ecn_rcv_synack(tp, th);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005577
Razvan Ghitulete1b3a6922012-08-14 16:30:20 +03005578 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005579 tcp_ack(sk, skb, FLAG_SLOWPATH);
5580
5581 /* Ok.. it's good. Set up sequence numbers and
5582 * move to established.
5583 */
5584 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5585 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5586
5587 /* RFC1323: The window in SYN & SYN/ACK segments is
5588 * never scaled.
5589 */
5590 tp->snd_wnd = ntohs(th->window);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005591
5592 if (!tp->rx_opt.wscale_ok) {
5593 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5594 tp->window_clamp = min(tp->window_clamp, 65535U);
5595 }
5596
5597 if (tp->rx_opt.saw_tstamp) {
5598 tp->rx_opt.tstamp_ok = 1;
5599 tp->tcp_header_len =
5600 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5601 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5602 tcp_store_ts_recent(tp);
5603 } else {
5604 tp->tcp_header_len = sizeof(struct tcphdr);
5605 }
5606
Ilpo Järvinene60402d2007-08-09 15:14:46 +03005607 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5608 tcp_enable_fack(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005609
John Heffner5d424d52006-03-20 17:53:41 -08005610 tcp_mtup_init(sk);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08005611 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005612 tcp_initialize_rcv_mss(sk);
5613
5614 /* Remember, tcp_poll() does not lock socket!
5615 * Change state from SYN-SENT only after copied_seq
5616 * is initialized. */
5617 tp->copied_seq = tp->rcv_nxt;
William Allen Simpson4957faade2009-12-02 18:25:27 +00005618
Ralf Baechlee16aa202006-12-07 00:11:33 -08005619 smp_mb();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005620
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005621 tcp_finish_connect(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005622
Yuchung Cheng67da22d2012-07-19 06:43:11 +00005623 if ((tp->syn_fastopen || tp->syn_data) &&
5624 tcp_rcv_fastopen_synack(sk, skb, &foc))
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005625 return -1;
5626
Arnaldo Carvalho de Melo295f7322005-08-09 20:11:56 -07005627 if (sk->sk_write_pending ||
5628 icsk->icsk_accept_queue.rskq_defer_accept ||
5629 icsk->icsk_ack.pingpong) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005630 /* Save one ACK. Data will be ready after
5631 * several ticks, if write_pending is set.
5632 *
5633 * It may be deleted, but with this feature tcpdumps
5634 * look so _wonderfully_ clever, that I was not able
5635 * to stand against the temptation 8) --ANK
5636 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005637 inet_csk_schedule_ack(sk);
Arnaldo Carvalho de Melo295f7322005-08-09 20:11:56 -07005638 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005639 tcp_enter_quickack_mode(sk);
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07005640 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5641 TCP_DELACK_MAX, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005642
5643discard:
5644 __kfree_skb(skb);
5645 return 0;
5646 } else {
5647 tcp_send_ack(sk);
5648 }
5649 return -1;
5650 }
5651
5652 /* No ACK in the segment */
5653
5654 if (th->rst) {
5655 /* rfc793:
5656 * "If the RST bit is set
5657 *
5658 * Otherwise (no ACK) drop the segment and return."
5659 */
5660
5661 goto discard_and_undo;
5662 }
5663
5664 /* PAWS check. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005665 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
Ilpo Järvinenc887e6d2009-03-14 14:23:03 +00005666 tcp_paws_reject(&tp->rx_opt, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005667 goto discard_and_undo;
5668
5669 if (th->syn) {
5670 /* We see SYN without ACK. It is attempt of
5671 * simultaneous connect with crossed SYNs.
5672 * Particularly, it can be connect to self.
5673 */
5674 tcp_set_state(sk, TCP_SYN_RECV);
5675
5676 if (tp->rx_opt.saw_tstamp) {
5677 tp->rx_opt.tstamp_ok = 1;
5678 tcp_store_ts_recent(tp);
5679 tp->tcp_header_len =
5680 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5681 } else {
5682 tp->tcp_header_len = sizeof(struct tcphdr);
5683 }
5684
5685 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
Eric Dumazet142a2e72015-11-26 08:18:14 -08005686 tp->copied_seq = tp->rcv_nxt;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005687 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5688
5689 /* RFC1323: The window in SYN & SYN/ACK segments is
5690 * never scaled.
5691 */
5692 tp->snd_wnd = ntohs(th->window);
5693 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5694 tp->max_window = tp->snd_wnd;
5695
Florian Westphal735d3832014-09-29 13:08:30 +02005696 tcp_ecn_rcv_syn(tp, th);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005697
John Heffner5d424d52006-03-20 17:53:41 -08005698 tcp_mtup_init(sk);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08005699 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005700 tcp_initialize_rcv_mss(sk);
5701
Linus Torvalds1da177e2005-04-16 15:20:36 -07005702 tcp_send_synack(sk);
5703#if 0
5704 /* Note, we could accept data and URG from this segment.
Jerry Chu168a8f52012-08-31 12:29:13 +00005705 * There are no obstacles to make this (except that we must
5706 * either change tcp_recvmsg() to prevent it from returning data
5707 * before 3WHS completes per RFC793, or employ TCP Fast Open).
Linus Torvalds1da177e2005-04-16 15:20:36 -07005708 *
5709 * However, if we ignore data in ACKless segments sometimes,
5710 * we have no reasons to accept it sometimes.
5711 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5712 * is not flawless. So, discard packet for sanity.
5713 * Uncomment this return to process the data.
5714 */
5715 return -1;
5716#else
5717 goto discard;
5718#endif
5719 }
5720 /* "fifth, if neither of the SYN or RST bits is set then
5721 * drop the segment and return."
5722 */
5723
5724discard_and_undo:
5725 tcp_clear_options(&tp->rx_opt);
5726 tp->rx_opt.mss_clamp = saved_clamp;
5727 goto discard;
5728
5729reset_and_undo:
5730 tcp_clear_options(&tp->rx_opt);
5731 tp->rx_opt.mss_clamp = saved_clamp;
5732 return 1;
5733}
5734
Linus Torvalds1da177e2005-04-16 15:20:36 -07005735/*
5736 * This function implements the receiving procedure of RFC 793 for
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005737 * all states except ESTABLISHED and TIME_WAIT.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005738 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5739 * address independent.
5740 */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005741
Eric Dumazet72ab4a82015-09-29 07:42:41 -07005742int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005743{
5744 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -08005745 struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazet72ab4a82015-09-29 07:42:41 -07005746 const struct tcphdr *th = tcp_hdr(skb);
Jerry Chu168a8f52012-08-31 12:29:13 +00005747 struct request_sock *req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005748 int queued = 0;
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005749 bool acceptable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005750
5751 tp->rx_opt.saw_tstamp = 0;
5752
5753 switch (sk->sk_state) {
5754 case TCP_CLOSE:
5755 goto discard;
5756
5757 case TCP_LISTEN:
Stephen Hemminger2de979b2007-03-08 20:45:19 -08005758 if (th->ack)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005759 return 1;
5760
Stephen Hemminger2de979b2007-03-08 20:45:19 -08005761 if (th->rst)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005762 goto discard;
5763
Stephen Hemminger2de979b2007-03-08 20:45:19 -08005764 if (th->syn) {
Eric Dumazetfdf5af02011-12-02 23:41:42 +00005765 if (th->fin)
5766 goto discard;
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -08005767 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005768 return 1;
5769
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005770 /* Now we have several options: In theory there is
5771 * nothing else in the frame. KA9Q has an option to
Linus Torvalds1da177e2005-04-16 15:20:36 -07005772 * send data with the syn, BSD accepts data with the
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005773 * syn up to the [to be] advertised window and
5774 * Solaris 2.1 gives you a protocol error. For now
5775 * we just ignore it, that fits the spec precisely
Linus Torvalds1da177e2005-04-16 15:20:36 -07005776 * and avoids incompatibilities. It would be nice in
5777 * future to drop through and process the data.
5778 *
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005779 * Now that TTCP is starting to be used we ought to
Linus Torvalds1da177e2005-04-16 15:20:36 -07005780 * queue this data.
5781 * But, this leaves one open to an easy denial of
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005782 * service attack, and SYN cookies can't defend
Linus Torvalds1da177e2005-04-16 15:20:36 -07005783 * against this problem. So, we drop the data
Masayuki Nakagawafb7e2392007-01-23 20:15:06 -08005784 * in the interest of security over speed unless
5785 * it's still in use.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005786 */
Masayuki Nakagawafb7e2392007-01-23 20:15:06 -08005787 kfree_skb(skb);
5788 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005789 }
5790 goto discard;
5791
5792 case TCP_SYN_SENT:
Eric Dumazetbda07a62015-09-29 07:42:40 -07005793 queued = tcp_rcv_synsent_state_process(sk, skb, th);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005794 if (queued >= 0)
5795 return queued;
5796
5797 /* Do step6 onward by hand. */
5798 tcp_urg(sk, skb, th);
5799 __kfree_skb(skb);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005800 tcp_data_snd_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005801 return 0;
5802 }
5803
Jerry Chu168a8f52012-08-31 12:29:13 +00005804 req = tp->fastopen_rsk;
Ian Morris00db4122015-04-03 09:17:27 +01005805 if (req) {
Jerry Chu37561f62012-10-22 11:26:36 +00005806 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
Jerry Chu168a8f52012-08-31 12:29:13 +00005807 sk->sk_state != TCP_FIN_WAIT1);
5808
Ian Morris51456b22015-04-03 09:17:26 +01005809 if (!tcp_check_req(sk, skb, req, true))
Jerry Chu168a8f52012-08-31 12:29:13 +00005810 goto discard;
Neal Cardwelle69bebd2012-09-22 04:18:57 +00005811 }
Eric Dumazetc3ae62af2012-12-26 12:44:34 +00005812
Calvin Owens0c228e82014-11-20 15:09:53 -08005813 if (!th->ack && !th->rst && !th->syn)
Eric Dumazetc3ae62af2012-12-26 12:44:34 +00005814 goto discard;
5815
Neal Cardwelle69bebd2012-09-22 04:18:57 +00005816 if (!tcp_validate_incoming(sk, skb, th, 0))
Eric Dumazet0c246042012-07-17 01:41:30 +00005817 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005818
5819 /* step 5: check the ACK field */
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005820 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
5821 FLAG_UPDATE_TS_RECENT) > 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005822
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005823 switch (sk->sk_state) {
5824 case TCP_SYN_RECV:
Joe Perches61eb90032013-05-24 18:36:13 +00005825 if (!acceptable)
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005826 return 1;
Joe Perches61eb90032013-05-24 18:36:13 +00005827
Yuchung Cheng0f1c28a2015-09-18 11:36:14 -07005828 if (!tp->srtt_us)
5829 tcp_synack_rtt_meas(sk, req);
5830
Joe Perches61eb90032013-05-24 18:36:13 +00005831 /* Once we leave TCP_SYN_RECV, we no longer need req
5832 * so release it.
5833 */
5834 if (req) {
Joe Perches61eb90032013-05-24 18:36:13 +00005835 tp->total_retrans = req->num_retrans;
Joe Perches61eb90032013-05-24 18:36:13 +00005836 reqsk_fastopen_remove(sk, req, false);
5837 } else {
5838 /* Make sure socket is routed, for correct metrics. */
5839 icsk->icsk_af_ops->rebuild_header(sk);
5840 tcp_init_congestion_control(sk);
5841
5842 tcp_mtup_init(sk);
Joe Perches61eb90032013-05-24 18:36:13 +00005843 tp->copied_seq = tp->rcv_nxt;
Eric Dumazetb0983d32013-09-20 13:56:58 -07005844 tcp_init_buffer_space(sk);
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005845 }
Joe Perches61eb90032013-05-24 18:36:13 +00005846 smp_mb();
5847 tcp_set_state(sk, TCP_ESTABLISHED);
5848 sk->sk_state_change(sk);
5849
5850 /* Note, that this wakeup is only for marginal crossed SYN case.
5851 * Passively open sockets are not waked up, because
5852 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5853 */
5854 if (sk->sk_socket)
5855 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5856
5857 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5858 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
5859 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5860
5861 if (tp->rx_opt.tstamp_ok)
5862 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5863
5864 if (req) {
5865 /* Re-arm the timer because data may have been sent out.
5866 * This is similar to the regular data transmission case
5867 * when new data has just been ack'ed.
5868 *
5869 * (TFO) - we could try to be more aggressive and
5870 * retransmitting any data sooner based on when they
5871 * are sent out.
5872 */
5873 tcp_rearm_rto(sk);
5874 } else
5875 tcp_init_metrics(sk);
5876
Neal Cardwell02cf4eb2013-10-21 15:40:19 -04005877 tcp_update_pacing_rate(sk);
5878
Joe Perches61eb90032013-05-24 18:36:13 +00005879 /* Prevent spurious tcp_cwnd_restart() on first data packet */
5880 tp->lsndtime = tcp_time_stamp;
5881
5882 tcp_initialize_rcv_mss(sk);
5883 tcp_fast_path_on(tp);
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005884 break;
5885
Joe Perchesc48b22d2013-05-24 18:06:58 +00005886 case TCP_FIN_WAIT1: {
5887 struct dst_entry *dst;
5888 int tmo;
5889
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005890 /* If we enter the TCP_FIN_WAIT1 state and we are a
5891 * Fast Open socket and this is the first acceptable
5892 * ACK we have received, this would have acknowledged
5893 * our SYNACK so stop the SYNACK timer.
5894 */
Ian Morris00db4122015-04-03 09:17:27 +01005895 if (req) {
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005896 /* Return RST if ack_seq is invalid.
5897 * Note that RFC793 only says to generate a
5898 * DUPACK for it but for TCP Fast Open it seems
5899 * better to treat this case like TCP_SYN_RECV
5900 * above.
5901 */
5902 if (!acceptable)
5903 return 1;
5904 /* We no longer need the request sock. */
5905 reqsk_fastopen_remove(sk, req, false);
5906 tcp_rearm_rto(sk);
5907 }
Joe Perchesc48b22d2013-05-24 18:06:58 +00005908 if (tp->snd_una != tp->write_seq)
5909 break;
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005910
Joe Perchesc48b22d2013-05-24 18:06:58 +00005911 tcp_set_state(sk, TCP_FIN_WAIT2);
5912 sk->sk_shutdown |= SEND_SHUTDOWN;
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005913
Joe Perchesc48b22d2013-05-24 18:06:58 +00005914 dst = __sk_dst_get(sk);
5915 if (dst)
5916 dst_confirm(dst);
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005917
Joe Perchesc48b22d2013-05-24 18:06:58 +00005918 if (!sock_flag(sk, SOCK_DEAD)) {
5919 /* Wake up lingering close() */
5920 sk->sk_state_change(sk);
5921 break;
5922 }
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005923
Joe Perchesc48b22d2013-05-24 18:06:58 +00005924 if (tp->linger2 < 0 ||
5925 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5926 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5927 tcp_done(sk);
5928 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5929 return 1;
5930 }
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005931
Joe Perchesc48b22d2013-05-24 18:06:58 +00005932 tmo = tcp_fin_time(sk);
5933 if (tmo > TCP_TIMEWAIT_LEN) {
5934 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5935 } else if (th->fin || sock_owned_by_user(sk)) {
5936 /* Bad case. We could lose such FIN otherwise.
5937 * It is not a big problem, but it looks confusing
5938 * and not so rare event. We still can lose it now,
5939 * if it spins in bh_lock_sock(), but it is really
5940 * marginal case.
5941 */
5942 inet_csk_reset_keepalive_timer(sk, tmo);
5943 } else {
5944 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5945 goto discard;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005946 }
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005947 break;
Joe Perchesc48b22d2013-05-24 18:06:58 +00005948 }
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005949
5950 case TCP_CLOSING:
5951 if (tp->snd_una == tp->write_seq) {
5952 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5953 goto discard;
5954 }
5955 break;
5956
5957 case TCP_LAST_ACK:
5958 if (tp->snd_una == tp->write_seq) {
5959 tcp_update_metrics(sk);
5960 tcp_done(sk);
5961 goto discard;
5962 }
5963 break;
Eric Dumazetc3ae62af2012-12-26 12:44:34 +00005964 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07005965
5966 /* step 6: check the URG bit */
5967 tcp_urg(sk, skb, th);
5968
5969 /* step 7: process the segment text */
5970 switch (sk->sk_state) {
5971 case TCP_CLOSE_WAIT:
5972 case TCP_CLOSING:
5973 case TCP_LAST_ACK:
5974 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5975 break;
5976 case TCP_FIN_WAIT1:
5977 case TCP_FIN_WAIT2:
5978 /* RFC 793 says to queue data in these states,
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005979 * RFC 1122 says we MUST send a reset.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005980 * BSD 4.4 also does reset.
5981 */
5982 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5983 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5984 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07005985 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005986 tcp_reset(sk);
5987 return 1;
5988 }
5989 }
5990 /* Fall through */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005991 case TCP_ESTABLISHED:
Linus Torvalds1da177e2005-04-16 15:20:36 -07005992 tcp_data_queue(sk, skb);
5993 queued = 1;
5994 break;
5995 }
5996
5997 /* tcp_data could move socket to TIME-WAIT */
5998 if (sk->sk_state != TCP_CLOSE) {
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005999 tcp_data_snd_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006000 tcp_ack_snd_check(sk);
6001 }
6002
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09006003 if (!queued) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07006004discard:
6005 __kfree_skb(skb);
6006 }
6007 return 0;
6008}
Linus Torvalds1da177e2005-04-16 15:20:36 -07006009EXPORT_SYMBOL(tcp_rcv_state_process);
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006010
6011static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6012{
6013 struct inet_request_sock *ireq = inet_rsk(req);
6014
6015 if (family == AF_INET)
Joe Perchesba7a46f2014-11-11 10:59:17 -08006016 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6017 &ireq->ir_rmt_addr, port);
Octavian Purdila4135ab82014-06-28 21:20:54 +03006018#if IS_ENABLED(CONFIG_IPV6)
6019 else if (family == AF_INET6)
Joe Perchesba7a46f2014-11-11 10:59:17 -08006020 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6021 &ireq->ir_v6_rmt_addr, port);
Octavian Purdila4135ab82014-06-28 21:20:54 +03006022#endif
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006023}
6024
Florian Westphald82bd122014-09-29 13:08:29 +02006025/* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6026 *
6027 * If we receive a SYN packet with these bits set, it means a
6028 * network is playing bad games with TOS bits. In order to
6029 * avoid possible false congestion notifications, we disable
stephen hemmingerf4e715c2014-10-29 16:05:06 -07006030 * TCP ECN negotiation.
Florian Westphald82bd122014-09-29 13:08:29 +02006031 *
6032 * Exception: tcp_ca wants ECN. This is required for DCTCP
Florian Westphal843c2fd2015-01-30 20:45:20 +01006033 * congestion control: Linux DCTCP asserts ECT on all packets,
6034 * including SYN, which is most optimal solution; however,
6035 * others, such as FreeBSD do not.
Florian Westphald82bd122014-09-29 13:08:29 +02006036 */
6037static void tcp_ecn_create_request(struct request_sock *req,
6038 const struct sk_buff *skb,
Florian Westphalf7b3bec2014-11-03 17:35:03 +01006039 const struct sock *listen_sk,
6040 const struct dst_entry *dst)
Florian Westphald82bd122014-09-29 13:08:29 +02006041{
6042 const struct tcphdr *th = tcp_hdr(skb);
6043 const struct net *net = sock_net(listen_sk);
6044 bool th_ecn = th->ece && th->cwr;
Florian Westphal843c2fd2015-01-30 20:45:20 +01006045 bool ect, ecn_ok;
Daniel Borkmannc3a8d942015-08-31 15:58:47 +02006046 u32 ecn_ok_dst;
Florian Westphald82bd122014-09-29 13:08:29 +02006047
6048 if (!th_ecn)
6049 return;
6050
6051 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
Daniel Borkmannc3a8d942015-08-31 15:58:47 +02006052 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6053 ecn_ok = net->ipv4.sysctl_tcp_ecn || ecn_ok_dst;
Florian Westphald82bd122014-09-29 13:08:29 +02006054
Daniel Borkmannc3a8d942015-08-31 15:58:47 +02006055 if ((!ect && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
6056 (ecn_ok_dst & DST_FEATURE_ECN_CA))
Florian Westphald82bd122014-09-29 13:08:29 +02006057 inet_rsk(req)->ecn_ok = 1;
6058}
6059
Eric Dumazet1bfc4432015-03-16 21:06:19 -07006060static void tcp_openreq_init(struct request_sock *req,
6061 const struct tcp_options_received *rx_opt,
6062 struct sk_buff *skb, const struct sock *sk)
6063{
6064 struct inet_request_sock *ireq = inet_rsk(req);
6065
Eric Dumazeted53d0a2015-10-08 19:33:23 -07006066 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
Eric Dumazet1bfc4432015-03-16 21:06:19 -07006067 req->cookie_ts = 0;
6068 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6069 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
Yuchung Cheng0f1c28a2015-09-18 11:36:14 -07006070 skb_mstamp_get(&tcp_rsk(req)->snt_synack);
Eric Dumazet1bfc4432015-03-16 21:06:19 -07006071 tcp_rsk(req)->last_oow_ack_time = 0;
6072 req->mss = rx_opt->mss_clamp;
6073 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6074 ireq->tstamp_ok = rx_opt->tstamp_ok;
6075 ireq->sack_ok = rx_opt->sack_ok;
6076 ireq->snd_wscale = rx_opt->snd_wscale;
6077 ireq->wscale_ok = rx_opt->wscale_ok;
6078 ireq->acked = 0;
6079 ireq->ecn_ok = 0;
6080 ireq->ir_rmt_port = tcp_hdr(skb)->source;
6081 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6082 ireq->ir_mark = inet_request_mark(sk, skb);
6083}
6084
Eric Dumazete49bb332015-03-17 18:32:27 -07006085struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
Eric Dumazeta1a53442015-10-04 21:08:11 -07006086 struct sock *sk_listener,
6087 bool attach_listener)
Eric Dumazete49bb332015-03-17 18:32:27 -07006088{
Eric Dumazeta1a53442015-10-04 21:08:11 -07006089 struct request_sock *req = reqsk_alloc(ops, sk_listener,
6090 attach_listener);
Eric Dumazete49bb332015-03-17 18:32:27 -07006091
6092 if (req) {
6093 struct inet_request_sock *ireq = inet_rsk(req);
6094
6095 kmemcheck_annotate_bitfield(ireq, flags);
6096 ireq->opt = NULL;
6097 atomic64_set(&ireq->ir_cookie, 0);
6098 ireq->ireq_state = TCP_NEW_SYN_RECV;
6099 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
Eric Dumazet0144a812015-03-24 21:45:56 -07006100 ireq->ireq_family = sk_listener->sk_family;
Eric Dumazete49bb332015-03-17 18:32:27 -07006101 }
6102
6103 return req;
6104}
6105EXPORT_SYMBOL(inet_reqsk_alloc);
6106
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006107/*
6108 * Return true if a syncookie should be sent
6109 */
Eric Dumazet2985aaa2015-09-29 07:42:51 -07006110static bool tcp_syn_flood_action(const struct sock *sk,
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006111 const struct sk_buff *skb,
6112 const char *proto)
6113{
Eric Dumazet8d2675f2015-10-02 11:43:25 -07006114 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006115 const char *msg = "Dropping request";
6116 bool want_cookie = false;
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006117
6118#ifdef CONFIG_SYN_COOKIES
6119 if (sysctl_tcp_syncookies) {
6120 msg = "Sending cookies";
6121 want_cookie = true;
6122 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
6123 } else
6124#endif
6125 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
6126
Eric Dumazet8d2675f2015-10-02 11:43:25 -07006127 if (!queue->synflood_warned &&
Eric Dumazet2985aaa2015-09-29 07:42:51 -07006128 sysctl_tcp_syncookies != 2 &&
Eric Dumazet8d2675f2015-10-02 11:43:25 -07006129 xchg(&queue->synflood_warned, 1) == 0)
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006130 pr_info("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6131 proto, ntohs(tcp_hdr(skb)->dest), msg);
Eric Dumazet2985aaa2015-09-29 07:42:51 -07006132
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006133 return want_cookie;
6134}
6135
Eric Dumazetcd8ae852015-05-03 21:34:46 -07006136static void tcp_reqsk_record_syn(const struct sock *sk,
6137 struct request_sock *req,
6138 const struct sk_buff *skb)
6139{
6140 if (tcp_sk(sk)->save_syn) {
6141 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
6142 u32 *copy;
6143
6144 copy = kmalloc(len + sizeof(u32), GFP_ATOMIC);
6145 if (copy) {
6146 copy[0] = len;
6147 memcpy(&copy[1], skb_network_header(skb), len);
6148 req->saved_syn = copy;
6149 }
6150 }
6151}
6152
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006153int tcp_conn_request(struct request_sock_ops *rsk_ops,
6154 const struct tcp_request_sock_ops *af_ops,
6155 struct sock *sk, struct sk_buff *skb)
6156{
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006157 struct tcp_fastopen_cookie foc = { .len = -1 };
Eric Dumazet7c85af82015-09-24 17:16:05 -07006158 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6159 struct tcp_options_received tmp_opt;
6160 struct tcp_sock *tp = tcp_sk(sk);
6161 struct sock *fastopen_sk = NULL;
6162 struct dst_entry *dst = NULL;
6163 struct request_sock *req;
6164 bool want_cookie = false;
6165 struct flowi fl;
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006166
6167 /* TW buckets are converted to open requests without
6168 * limitations, they conserve resources and peer is
6169 * evidently real one.
6170 */
6171 if ((sysctl_tcp_syncookies == 2 ||
6172 inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6173 want_cookie = tcp_syn_flood_action(sk, skb, rsk_ops->slab_name);
6174 if (!want_cookie)
6175 goto drop;
6176 }
6177
6178
6179 /* Accept backlog is full. If we have already queued enough
6180 * of warm entries in syn queue, drop request. It is better than
6181 * clogging syn queue with openreqs with exponentially increasing
6182 * timeout.
6183 */
6184 if (sk_acceptq_is_full(sk) && inet_csk_reqsk_queue_young(sk) > 1) {
6185 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6186 goto drop;
6187 }
6188
Eric Dumazeta1a53442015-10-04 21:08:11 -07006189 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006190 if (!req)
6191 goto drop;
6192
6193 tcp_rsk(req)->af_specific = af_ops;
6194
6195 tcp_clear_options(&tmp_opt);
6196 tmp_opt.mss_clamp = af_ops->mss_clamp;
6197 tmp_opt.user_mss = tp->rx_opt.user_mss;
6198 tcp_parse_options(skb, &tmp_opt, 0, want_cookie ? NULL : &foc);
6199
6200 if (want_cookie && !tmp_opt.saw_tstamp)
6201 tcp_clear_options(&tmp_opt);
6202
6203 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6204 tcp_openreq_init(req, &tmp_opt, skb, sk);
6205
Eric Dumazet16f86162015-03-13 15:51:10 -07006206 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6207 inet_rsk(req)->ir_iif = sk->sk_bound_dev_if;
6208
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006209 af_ops->init_req(req, sk, skb);
6210
6211 if (security_inet_conn_request(sk, skb, req))
6212 goto drop_and_free;
6213
Florian Westphalf7b3bec2014-11-03 17:35:03 +01006214 if (!want_cookie && !isn) {
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006215 /* VJ's idea. We save last timestamp seen
6216 * from the destination in peer table, when entering
6217 * state TIME-WAIT, and check against it before
6218 * accepting new connection request.
6219 *
6220 * If "isn" is not zero, this request hit alive
6221 * timewait bucket, so that all the necessary checks
6222 * are made in the function processing timewait state.
6223 */
Hannes Frederic Sowaa26552a2014-08-14 22:06:12 +02006224 if (tcp_death_row.sysctl_tw_recycle) {
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006225 bool strict;
6226
6227 dst = af_ops->route_req(sk, &fl, req, &strict);
Hannes Frederic Sowaa26552a2014-08-14 22:06:12 +02006228
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006229 if (dst && strict &&
Hannes Frederic Sowaa26552a2014-08-14 22:06:12 +02006230 !tcp_peer_is_proven(req, dst, true,
6231 tmp_opt.saw_tstamp)) {
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006232 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSPASSIVEREJECTED);
6233 goto drop_and_release;
6234 }
6235 }
6236 /* Kill the following clause, if you dislike this way. */
6237 else if (!sysctl_tcp_syncookies &&
6238 (sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6239 (sysctl_max_syn_backlog >> 2)) &&
Hannes Frederic Sowaa26552a2014-08-14 22:06:12 +02006240 !tcp_peer_is_proven(req, dst, false,
6241 tmp_opt.saw_tstamp)) {
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006242 /* Without syncookies last quarter of
6243 * backlog is filled with destinations,
6244 * proven to be alive.
6245 * It means that we continue to communicate
6246 * to destinations, already remembered
6247 * to the moment of synflood.
6248 */
6249 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6250 rsk_ops->family);
6251 goto drop_and_release;
6252 }
6253
6254 isn = af_ops->init_seq(skb);
6255 }
6256 if (!dst) {
6257 dst = af_ops->route_req(sk, &fl, req, NULL);
6258 if (!dst)
6259 goto drop_and_free;
6260 }
6261
Florian Westphalf7b3bec2014-11-03 17:35:03 +01006262 tcp_ecn_create_request(req, skb, sk, dst);
6263
6264 if (want_cookie) {
6265 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6266 req->cookie_ts = tmp_opt.tstamp_ok;
6267 if (!tmp_opt.tstamp_ok)
6268 inet_rsk(req)->ecn_ok = 0;
6269 }
6270
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006271 tcp_rsk(req)->snt_isn = isn;
Eric Dumazet58d607d2015-09-15 15:24:20 -07006272 tcp_rsk(req)->txhash = net_tx_rndhash();
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006273 tcp_openreq_init_rwin(req, sk, dst);
Eric Dumazetca6fb062015-10-02 11:43:35 -07006274 if (!want_cookie) {
Eric Dumazetca6fb062015-10-02 11:43:35 -07006275 tcp_reqsk_record_syn(sk, req, skb);
Eric Dumazet7656d842015-10-04 21:08:07 -07006276 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
Eric Dumazetca6fb062015-10-02 11:43:35 -07006277 }
Eric Dumazet7c85af82015-09-24 17:16:05 -07006278 if (fastopen_sk) {
Eric Dumazetca6fb062015-10-02 11:43:35 -07006279 af_ops->send_synack(fastopen_sk, dst, &fl, req,
Eric Dumazetdc6ef6b2015-10-16 13:00:01 -07006280 &foc, false);
Eric Dumazet7656d842015-10-04 21:08:07 -07006281 /* Add the child socket directly into the accept queue */
6282 inet_csk_reqsk_queue_add(sk, req, fastopen_sk);
6283 sk->sk_data_ready(sk);
6284 bh_unlock_sock(fastopen_sk);
Eric Dumazet7c85af82015-09-24 17:16:05 -07006285 sock_put(fastopen_sk);
6286 } else {
Eric Dumazet9439ce02015-03-17 18:32:29 -07006287 tcp_rsk(req)->tfo_listener = false;
Eric Dumazetca6fb062015-10-02 11:43:35 -07006288 if (!want_cookie)
6289 inet_csk_reqsk_queue_hash_add(sk, req, TCP_TIMEOUT_INIT);
6290 af_ops->send_synack(sk, dst, &fl, req,
Eric Dumazetdc6ef6b2015-10-16 13:00:01 -07006291 &foc, !want_cookie);
Eric Dumazetca6fb062015-10-02 11:43:35 -07006292 if (want_cookie)
6293 goto drop_and_free;
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006294 }
Eric Dumazetca6fb062015-10-02 11:43:35 -07006295 reqsk_put(req);
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006296 return 0;
6297
6298drop_and_release:
6299 dst_release(dst);
6300drop_and_free:
6301 reqsk_free(req);
6302drop:
6303 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_LISTENDROPS);
6304 return 0;
6305}
6306EXPORT_SYMBOL(tcp_conn_request);