blob: 3ebf45b38bc309f448dbc4f27fe8722cefabaf19 [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;
Eric Dumazetdca145f2014-10-27 21:45:24 -070083int sysctl_tcp_max_reordering __read_mostly = 300;
Brian Haleyab32ea52006-09-22 14:15:41 -070084int sysctl_tcp_dsack __read_mostly = 1;
85int sysctl_tcp_app_win __read_mostly = 31;
Eric Dumazetb49960a2012-05-02 02:28:41 +000086int sysctl_tcp_adv_win_scale __read_mostly = 1;
Eric Dumazet4bc2f182010-07-09 21:22:10 +000087EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
Linus Torvalds1da177e2005-04-16 15:20:36 -070088
Eric Dumazet282f23c2012-07-17 10:13:05 +020089/* rfc5961 challenge ack rate limiting */
Eric Dumazet75ff39c2016-07-10 10:04:02 +020090int sysctl_tcp_challenge_ack_limit = 1000;
Eric Dumazet282f23c2012-07-17 10:13:05 +020091
Brian Haleyab32ea52006-09-22 14:15:41 -070092int sysctl_tcp_stdurg __read_mostly;
93int sysctl_tcp_rfc1337 __read_mostly;
94int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
Ilpo Järvinenc96fd3d2007-09-20 11:36:37 -070095int sysctl_tcp_frto __read_mostly = 2;
Yuchung Chengf6722582015-10-16 21:57:42 -070096int sysctl_tcp_min_rtt_wlen __read_mostly = 300;
Linus Torvalds1da177e2005-04-16 15:20:36 -070097
Andreas Petlund7e380172010-02-18 04:48:19 +000098int sysctl_tcp_thin_dupack __read_mostly;
99
Brian Haleyab32ea52006-09-22 14:15:41 -0700100int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +0000101int sysctl_tcp_early_retrans __read_mostly = 3;
Neal Cardwell032ee422015-02-06 16:04:38 -0500102int sysctl_tcp_invalid_ratelimit __read_mostly = HZ/2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700103
Linus Torvalds1da177e2005-04-16 15:20:36 -0700104#define FLAG_DATA 0x01 /* Incoming frame contained data. */
105#define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
106#define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
107#define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
108#define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
109#define FLAG_DATA_SACKED 0x20 /* New SACK. */
110#define FLAG_ECE 0x40 /* ECE in this ACK */
Yuchung Cheng291a00d2015-07-01 14:11:14 -0700111#define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700112#define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
Yuchung Chenge33099f2013-03-20 13:33:00 +0000113#define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
Ilpo Järvinen2e605292007-08-02 19:46:58 -0700114#define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
Ryousei Takano564262c12007-10-25 23:03:52 -0700115#define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
Ilpo Järvinencadbd032007-12-31 04:49:21 -0800116#define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
Eric Dumazet12fb3dd2013-04-19 07:19:48 +0000117#define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700118
119#define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
120#define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
121#define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
122#define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
123
Linus Torvalds1da177e2005-04-16 15:20:36 -0700124#define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700125#define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700126
Yuchung Chenge662ca42016-02-02 10:33:04 -0800127#define REXMIT_NONE 0 /* no loss recovery to do */
128#define REXMIT_LOST 1 /* retransmit packets marked lost */
129#define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */
130
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900131/* Adapt the MSS value used to make delayed ack decision to the
Linus Torvalds1da177e2005-04-16 15:20:36 -0700132 * real world.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900133 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800134static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700135{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700136 struct inet_connection_sock *icsk = inet_csk(sk);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900137 const unsigned int lss = icsk->icsk_ack.last_seg_size;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700138 unsigned int len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700139
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900140 icsk->icsk_ack.last_seg_size = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700141
142 /* skb->len may jitter because of SACKs, even if peer
143 * sends good full-sized frames.
144 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800145 len = skb_shinfo(skb)->gso_size ? : skb->len;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700146 if (len >= icsk->icsk_ack.rcv_mss) {
147 icsk->icsk_ack.rcv_mss = len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700148 } else {
149 /* Otherwise, we make more careful check taking into account,
150 * that SACKs block is variable.
151 *
152 * "len" is invariant segment length, including TCP header.
153 */
Arnaldo Carvalho de Melo9c702202007-04-25 18:04:18 -0700154 len += skb->data - skb_transport_header(skb);
William Allen Simpsonbee7ca92009-11-10 09:51:18 +0000155 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -0700156 /* If PSH is not set, packet should be
157 * full sized, provided peer TCP is not badly broken.
158 * This observation (if it is correct 8)) allows
159 * to handle super-low mtu links fairly.
160 */
161 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -0700162 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700163 /* Subtract also invariant (if peer is RFC compliant),
164 * tcp header plus fixed timestamp option length.
165 * Resulting "len" is MSS free of SACK jitter.
166 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700167 len -= tcp_sk(sk)->tcp_header_len;
168 icsk->icsk_ack.last_seg_size = len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700169 if (len == lss) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700170 icsk->icsk_ack.rcv_mss = len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700171 return;
172 }
173 }
Alexey Kuznetsov1ef96962006-09-19 12:52:50 -0700174 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
175 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700176 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700177 }
178}
179
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700180static void tcp_incr_quickack(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700181{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700182 struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazet95c96172012-04-15 05:58:06 +0000183 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700184
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800185 if (quickacks == 0)
186 quickacks = 2;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700187 if (quickacks > icsk->icsk_ack.quick)
188 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700189}
190
stephen hemminger1b9f4092010-09-28 19:30:14 +0000191static void tcp_enter_quickack_mode(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700192{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700193 struct inet_connection_sock *icsk = inet_csk(sk);
194 tcp_incr_quickack(sk);
195 icsk->icsk_ack.pingpong = 0;
196 icsk->icsk_ack.ato = TCP_ATO_MIN;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700197}
198
199/* Send ACKs quickly, if "quick" count is not exhausted
200 * and the session is not interactive.
201 */
202
Jon Maxwell2251ae42015-07-08 10:12:28 +1000203static bool tcp_in_quickack_mode(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700204{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700205 const struct inet_connection_sock *icsk = inet_csk(sk);
Jon Maxwell2251ae42015-07-08 10:12:28 +1000206 const struct dst_entry *dst = __sk_dst_get(sk);
Eric Dumazeta2a385d2012-05-16 23:15:34 +0000207
Jon Maxwell2251ae42015-07-08 10:12:28 +1000208 return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
209 (icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700210}
211
Florian Westphal735d3832014-09-29 13:08:30 +0200212static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700213{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800214 if (tp->ecn_flags & TCP_ECN_OK)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700215 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
216}
217
Florian Westphal735d3832014-09-29 13:08:30 +0200218static void tcp_ecn_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700219{
220 if (tcp_hdr(skb)->cwr)
221 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
222}
223
Florian Westphal735d3832014-09-29 13:08:30 +0200224static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700225{
226 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
227}
228
Florian Westphal735d3832014-09-29 13:08:30 +0200229static void __tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700230{
Eric Dumazetb82d1bb2011-09-27 02:20:08 -0400231 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
Eric Dumazet7a269ff2011-09-22 20:02:19 +0000232 case INET_ECN_NOT_ECT:
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700233 /* Funny extension: if ECT is not set on a segment,
Eric Dumazet7a269ff2011-09-22 20:02:19 +0000234 * and we already seen ECT on a previous segment,
235 * it is probably a retransmit.
236 */
237 if (tp->ecn_flags & TCP_ECN_SEEN)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700238 tcp_enter_quickack_mode((struct sock *)tp);
Eric Dumazet7a269ff2011-09-22 20:02:19 +0000239 break;
240 case INET_ECN_CE:
Florian Westphal98900922014-09-26 22:37:35 +0200241 if (tcp_ca_needs_ecn((struct sock *)tp))
242 tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_IS_CE);
243
Eric Dumazetaae06bf2012-08-06 11:04:43 +0000244 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
245 /* Better not delay acks, sender can have a very low cwnd */
246 tcp_enter_quickack_mode((struct sock *)tp);
247 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
248 }
Eric Dumazet7a269ff2011-09-22 20:02:19 +0000249 tp->ecn_flags |= TCP_ECN_SEEN;
Florian Westphal98900922014-09-26 22:37:35 +0200250 break;
251 default:
252 if (tcp_ca_needs_ecn((struct sock *)tp))
253 tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_NO_CE);
254 tp->ecn_flags |= TCP_ECN_SEEN;
255 break;
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700256 }
257}
258
Florian Westphal735d3832014-09-29 13:08:30 +0200259static void tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
260{
261 if (tp->ecn_flags & TCP_ECN_OK)
262 __tcp_ecn_check_ce(tp, skb);
263}
264
265static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700266{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800267 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700268 tp->ecn_flags &= ~TCP_ECN_OK;
269}
270
Florian Westphal735d3832014-09-29 13:08:30 +0200271static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700272{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800273 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700274 tp->ecn_flags &= ~TCP_ECN_OK;
275}
276
Florian Westphal735d3832014-09-29 13:08:30 +0200277static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700278{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800279 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
Eric Dumazeta2a385d2012-05-16 23:15:34 +0000280 return true;
281 return false;
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700282}
283
Linus Torvalds1da177e2005-04-16 15:20:36 -0700284/* Buffer size and advertised window tuning.
285 *
286 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
287 */
288
Eric Dumazet6ae70532013-10-01 10:23:44 -0700289static void tcp_sndbuf_expand(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700290{
Eric Dumazet6ae70532013-10-01 10:23:44 -0700291 const struct tcp_sock *tp = tcp_sk(sk);
292 int sndmem, per_mss;
293 u32 nr_segs;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700294
Eric Dumazet6ae70532013-10-01 10:23:44 -0700295 /* Worst case is non GSO/TSO : each frame consumes one skb
296 * and skb->head is kmalloced using power of two area of memory
297 */
298 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
299 MAX_TCP_HEADER +
300 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
301
302 per_mss = roundup_pow_of_two(per_mss) +
303 SKB_DATA_ALIGN(sizeof(struct sk_buff));
304
305 nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
306 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
307
308 /* Fast Recovery (RFC 5681 3.2) :
309 * Cubic needs 1.7 factor, rounded to 2 to include
310 * extra cushion (application might react slowly to POLLOUT)
311 */
312 sndmem = 2 * nr_segs * per_mss;
313
Eric Dumazet06a59ec2011-10-13 18:24:42 +0000314 if (sk->sk_sndbuf < sndmem)
315 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700316}
317
318/* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
319 *
320 * All tcp_full_space() is split to two parts: "network" buffer, allocated
321 * forward and advertised in receiver window (tp->rcv_wnd) and
322 * "application buffer", required to isolate scheduling/application
323 * latencies from network.
324 * window_clamp is maximal advertised window. It can be less than
325 * tcp_full_space(), in this case tcp_full_space() - window_clamp
326 * is reserved for "application" buffer. The less window_clamp is
327 * the smoother our behaviour from viewpoint of network, but the lower
328 * throughput and the higher sensitivity of the connection to losses. 8)
329 *
330 * rcv_ssthresh is more strict window_clamp used at "slow start"
331 * phase to predict further behaviour of this connection.
332 * It is used for two goals:
333 * - to enforce header prediction at sender, even when application
334 * requires some significant "application buffer". It is check #1.
335 * - to prevent pruning of receive queue because of misprediction
336 * of receiver window. Check #2.
337 *
338 * The scheme does not work when sender sends good segments opening
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800339 * window and then starts to feed us spaghetti. But it should work
Linus Torvalds1da177e2005-04-16 15:20:36 -0700340 * in common situations. Otherwise, we have to rely on queue collapsing.
341 */
342
343/* Slow part of check#2. */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700344static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700345{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700346 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700347 /* Optimize this! */
Eric Dumazetdfd4f0a2007-12-21 06:07:53 -0800348 int truesize = tcp_win_from_space(skb->truesize) >> 1;
349 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700350
351 while (tp->rcv_ssthresh <= window) {
352 if (truesize <= skb->len)
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700353 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700354
355 truesize >>= 1;
356 window >>= 1;
357 }
358 return 0;
359}
360
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400361static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700362{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700363 struct tcp_sock *tp = tcp_sk(sk);
364
Linus Torvalds1da177e2005-04-16 15:20:36 -0700365 /* Check #1 */
366 if (tp->rcv_ssthresh < tp->window_clamp &&
367 (int)tp->rcv_ssthresh < tcp_space(sk) &&
Eric Dumazetb8da51e2015-05-15 12:39:27 -0700368 !tcp_under_memory_pressure(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700369 int incr;
370
371 /* Check #2. Increase window, if skb with such overhead
372 * will fit to rcvbuf in future.
373 */
374 if (tcp_win_from_space(skb->truesize) <= skb->len)
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800375 incr = 2 * tp->advmss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700376 else
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700377 incr = __tcp_grow_window(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700378
379 if (incr) {
Eric Dumazet4d846f02012-04-16 23:28:07 +0000380 incr = max_t(int, incr, 2 * skb->len);
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800381 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
382 tp->window_clamp);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700383 inet_csk(sk)->icsk_ack.quick |= 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700384 }
385 }
386}
387
388/* 3. Tuning rcvbuf, when connection enters established state. */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700389static void tcp_fixup_rcvbuf(struct sock *sk)
390{
Eric Dumazete9266a02011-10-20 16:53:56 -0400391 u32 mss = tcp_sk(sk)->advmss;
Eric Dumazete9266a02011-10-20 16:53:56 -0400392 int rcvmem;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700393
Yuchung Cheng85f16522013-06-11 15:35:32 -0700394 rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) *
395 tcp_default_init_rwnd(mss);
Eric Dumazete9266a02011-10-20 16:53:56 -0400396
Eric Dumazetb0983d32013-09-20 13:56:58 -0700397 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
398 * Allow enough cushion so that sender is not limited by our window
399 */
400 if (sysctl_tcp_moderate_rcvbuf)
401 rcvmem <<= 2;
402
Eric Dumazete9266a02011-10-20 16:53:56 -0400403 if (sk->sk_rcvbuf < rcvmem)
404 sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700405}
406
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800407/* 4. Try to fixup all. It is made immediately after connection enters
Linus Torvalds1da177e2005-04-16 15:20:36 -0700408 * established state.
409 */
Jerry Chu10467162012-08-31 12:29:11 +0000410void tcp_init_buffer_space(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700411{
412 struct tcp_sock *tp = tcp_sk(sk);
413 int maxwin;
414
415 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
416 tcp_fixup_rcvbuf(sk);
417 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
Eric Dumazet6ae70532013-10-01 10:23:44 -0700418 tcp_sndbuf_expand(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700419
420 tp->rcvq_space.space = tp->rcv_wnd;
Eric Dumazetb0983d32013-09-20 13:56:58 -0700421 tp->rcvq_space.time = tcp_time_stamp;
422 tp->rcvq_space.seq = tp->copied_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700423
424 maxwin = tcp_full_space(sk);
425
426 if (tp->window_clamp >= maxwin) {
427 tp->window_clamp = maxwin;
428
429 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
430 tp->window_clamp = max(maxwin -
431 (maxwin >> sysctl_tcp_app_win),
432 4 * tp->advmss);
433 }
434
435 /* Force reservation of one segment. */
436 if (sysctl_tcp_app_win &&
437 tp->window_clamp > 2 * tp->advmss &&
438 tp->window_clamp + tp->advmss > maxwin)
439 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
440
441 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
442 tp->snd_cwnd_stamp = tcp_time_stamp;
443}
444
Linus Torvalds1da177e2005-04-16 15:20:36 -0700445/* 5. Recalculate window clamp after socket hit its memory bounds. */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700446static void tcp_clamp_window(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700447{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700448 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300449 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700450
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300451 icsk->icsk_ack.quick = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700452
John Heffner326f36e2005-11-10 17:11:48 -0800453 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
454 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
Eric Dumazetb8da51e2015-05-15 12:39:27 -0700455 !tcp_under_memory_pressure(sk) &&
Glauber Costa180d8cd2011-12-11 21:47:02 +0000456 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
John Heffner326f36e2005-11-10 17:11:48 -0800457 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
458 sysctl_tcp_rmem[2]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700459 }
John Heffner326f36e2005-11-10 17:11:48 -0800460 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800461 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462}
463
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800464/* Initialize RCV_MSS value.
465 * RCV_MSS is an our guess about MSS used by the peer.
466 * We haven't any direct information about the MSS.
467 * It's better to underestimate the RCV_MSS rather than overestimate.
468 * Overestimations make us ACKing less frequently than needed.
469 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
470 */
471void tcp_initialize_rcv_mss(struct sock *sk)
472{
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400473 const struct tcp_sock *tp = tcp_sk(sk);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800474 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
475
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800476 hint = min(hint, tp->rcv_wnd / 2);
William Allen Simpsonbee7ca92009-11-10 09:51:18 +0000477 hint = min(hint, TCP_MSS_DEFAULT);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800478 hint = max(hint, TCP_MIN_MSS);
479
480 inet_csk(sk)->icsk_ack.rcv_mss = hint;
481}
Eric Dumazet4bc2f182010-07-09 21:22:10 +0000482EXPORT_SYMBOL(tcp_initialize_rcv_mss);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800483
Linus Torvalds1da177e2005-04-16 15:20:36 -0700484/* Receiver "autotuning" code.
485 *
486 * The algorithm for RTT estimation w/o timestamps is based on
487 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
Justin P. Mattock631dd1a2010-10-18 11:03:14 +0200488 * <http://public.lanl.gov/radiant/pubs.html#DRS>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700489 *
490 * More detail on this code can be found at
Justin P. Mattock631dd1a2010-10-18 11:03:14 +0200491 * <http://staff.psc.edu/jheffner/>,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700492 * though this reference is out of date. A new paper
493 * is pending.
494 */
495static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
496{
497 u32 new_sample = tp->rcv_rtt_est.rtt;
498 long m = sample;
499
500 if (m == 0)
501 m = 1;
502
503 if (new_sample != 0) {
504 /* If we sample in larger samples in the non-timestamp
505 * case, we could grossly overestimate the RTT especially
506 * with chatty applications or bulk transfer apps which
507 * are stalled on filesystem I/O.
508 *
509 * Also, since we are only going for a minimum in the
Stephen Hemminger31f34262005-11-15 15:17:10 -0800510 * non-timestamp case, we do not smooth things out
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800511 * else with timestamps disabled convergence takes too
Linus Torvalds1da177e2005-04-16 15:20:36 -0700512 * long.
513 */
514 if (!win_dep) {
515 m -= (new_sample >> 3);
516 new_sample += m;
Neal Cardwell18a223e2012-04-10 07:59:20 +0000517 } else {
518 m <<= 3;
519 if (m < new_sample)
520 new_sample = m;
521 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700522 } else {
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800523 /* No previous measure. */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700524 new_sample = m << 3;
525 }
526
527 if (tp->rcv_rtt_est.rtt != new_sample)
528 tp->rcv_rtt_est.rtt = new_sample;
529}
530
531static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
532{
533 if (tp->rcv_rtt_est.time == 0)
534 goto new_measure;
535 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
536 return;
Neal Cardwell651913c2012-04-27 11:29:37 -0400537 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700538
539new_measure:
540 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
541 tp->rcv_rtt_est.time = tcp_time_stamp;
542}
543
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800544static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
545 const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700546{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700547 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700548 if (tp->rx_opt.rcv_tsecr &&
549 (TCP_SKB_CB(skb)->end_seq -
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700550 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700551 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
552}
553
554/*
555 * This function should be called every time data is copied to user space.
556 * It calculates the appropriate TCP receive buffer space.
557 */
558void tcp_rcv_space_adjust(struct sock *sk)
559{
560 struct tcp_sock *tp = tcp_sk(sk);
561 int time;
Eric Dumazetb0983d32013-09-20 13:56:58 -0700562 int copied;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900563
Linus Torvalds1da177e2005-04-16 15:20:36 -0700564 time = tcp_time_stamp - tp->rcvq_space.time;
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800565 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700566 return;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900567
Eric Dumazetb0983d32013-09-20 13:56:58 -0700568 /* Number of bytes copied to user in last RTT */
569 copied = tp->copied_seq - tp->rcvq_space.seq;
570 if (copied <= tp->rcvq_space.space)
571 goto new_measure;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700572
Eric Dumazetb0983d32013-09-20 13:56:58 -0700573 /* A bit of theory :
574 * copied = bytes received in previous RTT, our base window
575 * To cope with packet losses, we need a 2x factor
576 * To cope with slow start, and sender growing its cwin by 100 %
577 * every RTT, we need a 4x factor, because the ACK we are sending
578 * now is for the next RTT, not the current one :
579 * <prev RTT . ><current RTT .. ><next RTT .... >
580 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700581
Eric Dumazetb0983d32013-09-20 13:56:58 -0700582 if (sysctl_tcp_moderate_rcvbuf &&
583 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
584 int rcvwin, rcvmem, rcvbuf;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700585
Eric Dumazetb0983d32013-09-20 13:56:58 -0700586 /* minimal window to cope with packet losses, assuming
587 * steady state. Add some cushion because of small variations.
588 */
589 rcvwin = (copied << 1) + 16 * tp->advmss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700590
Eric Dumazetb0983d32013-09-20 13:56:58 -0700591 /* If rate increased by 25%,
592 * assume slow start, rcvwin = 3 * copied
593 * If rate increased by 50%,
594 * assume sender can use 2x growth, rcvwin = 4 * copied
595 */
596 if (copied >=
597 tp->rcvq_space.space + (tp->rcvq_space.space >> 2)) {
598 if (copied >=
599 tp->rcvq_space.space + (tp->rcvq_space.space >> 1))
600 rcvwin <<= 1;
601 else
602 rcvwin += (rcvwin >> 1);
603 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700604
Eric Dumazetb0983d32013-09-20 13:56:58 -0700605 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
606 while (tcp_win_from_space(rcvmem) < tp->advmss)
607 rcvmem += 128;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700608
Eric Dumazetb0983d32013-09-20 13:56:58 -0700609 rcvbuf = min(rcvwin / tp->advmss * rcvmem, sysctl_tcp_rmem[2]);
610 if (rcvbuf > sk->sk_rcvbuf) {
611 sk->sk_rcvbuf = rcvbuf;
612
613 /* Make the window clamp follow along. */
614 tp->window_clamp = rcvwin;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700615 }
616 }
Eric Dumazetb0983d32013-09-20 13:56:58 -0700617 tp->rcvq_space.space = copied;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900618
Linus Torvalds1da177e2005-04-16 15:20:36 -0700619new_measure:
620 tp->rcvq_space.seq = tp->copied_seq;
621 tp->rcvq_space.time = tcp_time_stamp;
622}
623
624/* There is something which you must keep in mind when you analyze the
625 * behavior of the tp->ato delayed ack timeout interval. When a
626 * connection starts up, we want to ack as quickly as possible. The
627 * problem is that "good" TCP's do slow start at the beginning of data
628 * transmission. The means that until we send the first few ACK's the
629 * sender will sit on his end and only queue most of his data, because
630 * he can only send snd_cwnd unacked packets at any given time. For
631 * each ACK we send, he increments snd_cwnd and transmits more of his
632 * queue. -DaveM
633 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700634static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700635{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700636 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700637 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700638 u32 now;
639
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700640 inet_csk_schedule_ack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700641
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700642 tcp_measure_rcv_mss(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700643
644 tcp_rcv_rtt_measure(tp);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900645
Linus Torvalds1da177e2005-04-16 15:20:36 -0700646 now = tcp_time_stamp;
647
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700648 if (!icsk->icsk_ack.ato) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700649 /* The _first_ data packet received, initialize
650 * delayed ACK engine.
651 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700652 tcp_incr_quickack(sk);
653 icsk->icsk_ack.ato = TCP_ATO_MIN;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700654 } else {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700655 int m = now - icsk->icsk_ack.lrcvtime;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700656
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800657 if (m <= TCP_ATO_MIN / 2) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700658 /* The fastest case is the first. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700659 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
660 } else if (m < icsk->icsk_ack.ato) {
661 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
662 if (icsk->icsk_ack.ato > icsk->icsk_rto)
663 icsk->icsk_ack.ato = icsk->icsk_rto;
664 } else if (m > icsk->icsk_rto) {
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800665 /* Too long gap. Apparently sender failed to
Linus Torvalds1da177e2005-04-16 15:20:36 -0700666 * restart window, so that we send ACKs quickly.
667 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700668 tcp_incr_quickack(sk);
Hideo Aoki3ab224b2007-12-31 00:11:19 -0800669 sk_mem_reclaim(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700670 }
671 }
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700672 icsk->icsk_ack.lrcvtime = now;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700673
Florian Westphal735d3832014-09-29 13:08:30 +0200674 tcp_ecn_check_ce(tp, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700675
676 if (skb->len >= 128)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700677 tcp_grow_window(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700678}
679
Linus Torvalds1da177e2005-04-16 15:20:36 -0700680/* Called to compute a smoothed rtt estimate. The data fed to this
681 * routine either comes from timestamps, or from segments that were
682 * known _not_ to have been retransmitted [see Karn/Partridge
683 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
684 * piece by Van Jacobson.
685 * NOTE: the next three routines used to be one big routine.
686 * To save cycles in the RFC 1323 implementation it was better to break
687 * it up into three procedures. -- erics
688 */
Eric Dumazet740b0f12014-02-26 14:02:48 -0800689static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700690{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300691 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet740b0f12014-02-26 14:02:48 -0800692 long m = mrtt_us; /* RTT */
693 u32 srtt = tp->srtt_us;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700694
Linus Torvalds1da177e2005-04-16 15:20:36 -0700695 /* The following amusing code comes from Jacobson's
696 * article in SIGCOMM '88. Note that rtt and mdev
697 * are scaled versions of rtt and mean deviation.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900698 * This is designed to be as fast as possible
Linus Torvalds1da177e2005-04-16 15:20:36 -0700699 * m stands for "measurement".
700 *
701 * On a 1990 paper the rto value is changed to:
702 * RTO = rtt + 4 * mdev
703 *
704 * Funny. This algorithm seems to be very broken.
705 * These formulae increase RTO, when it should be decreased, increase
Stephen Hemminger31f34262005-11-15 15:17:10 -0800706 * too slowly, when it should be increased quickly, decrease too quickly
Linus Torvalds1da177e2005-04-16 15:20:36 -0700707 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
708 * does not matter how to _calculate_ it. Seems, it was trap
709 * that VJ failed to avoid. 8)
710 */
Eric Dumazet4a5ab4e2014-02-06 15:57:10 -0800711 if (srtt != 0) {
712 m -= (srtt >> 3); /* m is now error in rtt est */
713 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700714 if (m < 0) {
715 m = -m; /* m is now abs(error) */
Eric Dumazet740b0f12014-02-26 14:02:48 -0800716 m -= (tp->mdev_us >> 2); /* similar update on mdev */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700717 /* This is similar to one of Eifel findings.
718 * Eifel blocks mdev updates when rtt decreases.
719 * This solution is a bit different: we use finer gain
720 * for mdev in this case (alpha*beta).
721 * Like Eifel it also prevents growth of rto,
722 * but also it limits too fast rto decreases,
723 * happening in pure Eifel.
724 */
725 if (m > 0)
726 m >>= 3;
727 } else {
Eric Dumazet740b0f12014-02-26 14:02:48 -0800728 m -= (tp->mdev_us >> 2); /* similar update on mdev */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700729 }
Eric Dumazet740b0f12014-02-26 14:02:48 -0800730 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
731 if (tp->mdev_us > tp->mdev_max_us) {
732 tp->mdev_max_us = tp->mdev_us;
733 if (tp->mdev_max_us > tp->rttvar_us)
734 tp->rttvar_us = tp->mdev_max_us;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700735 }
736 if (after(tp->snd_una, tp->rtt_seq)) {
Eric Dumazet740b0f12014-02-26 14:02:48 -0800737 if (tp->mdev_max_us < tp->rttvar_us)
738 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700739 tp->rtt_seq = tp->snd_nxt;
Eric Dumazet740b0f12014-02-26 14:02:48 -0800740 tp->mdev_max_us = tcp_rto_min_us(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700741 }
742 } else {
743 /* no previous measure. */
Eric Dumazet4a5ab4e2014-02-06 15:57:10 -0800744 srtt = m << 3; /* take the measured time to be rtt */
Eric Dumazet740b0f12014-02-26 14:02:48 -0800745 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
746 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
747 tp->mdev_max_us = tp->rttvar_us;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700748 tp->rtt_seq = tp->snd_nxt;
749 }
Eric Dumazet740b0f12014-02-26 14:02:48 -0800750 tp->srtt_us = max(1U, srtt);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700751}
752
Eric Dumazet95bd09e2013-08-27 05:46:32 -0700753/* Set the sk_pacing_rate to allow proper sizing of TSO packets.
754 * Note: TCP stack does not yet implement pacing.
755 * FQ packet scheduler can be used to implement cheap but effective
756 * TCP pacing, to smooth the burst on large writes when packets
757 * in flight is significantly lower than cwnd (or rwin)
758 */
Eric Dumazet43e122b2015-08-21 17:38:02 -0700759int sysctl_tcp_pacing_ss_ratio __read_mostly = 200;
760int sysctl_tcp_pacing_ca_ratio __read_mostly = 120;
761
Eric Dumazet95bd09e2013-08-27 05:46:32 -0700762static void tcp_update_pacing_rate(struct sock *sk)
763{
764 const struct tcp_sock *tp = tcp_sk(sk);
765 u64 rate;
766
767 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
Eric Dumazet43e122b2015-08-21 17:38:02 -0700768 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
769
770 /* current rate is (cwnd * mss) / srtt
771 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
772 * In Congestion Avoidance phase, set it to 120 % the current rate.
773 *
774 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
775 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
776 * end of slow start and should slow down.
777 */
778 if (tp->snd_cwnd < tp->snd_ssthresh / 2)
779 rate *= sysctl_tcp_pacing_ss_ratio;
780 else
781 rate *= sysctl_tcp_pacing_ca_ratio;
Eric Dumazet95bd09e2013-08-27 05:46:32 -0700782
783 rate *= max(tp->snd_cwnd, tp->packets_out);
784
Eric Dumazet740b0f12014-02-26 14:02:48 -0800785 if (likely(tp->srtt_us))
786 do_div(rate, tp->srtt_us);
Eric Dumazet95bd09e2013-08-27 05:46:32 -0700787
Eric Dumazetba537422013-10-09 17:14:52 -0700788 /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate
789 * without any lock. We want to make sure compiler wont store
790 * intermediate values in this location.
791 */
792 ACCESS_ONCE(sk->sk_pacing_rate) = min_t(u64, rate,
793 sk->sk_max_pacing_rate);
Eric Dumazet95bd09e2013-08-27 05:46:32 -0700794}
795
Linus Torvalds1da177e2005-04-16 15:20:36 -0700796/* Calculate rto without backoff. This is the second half of Van Jacobson's
797 * routine referred to above.
798 */
stephen hemmingerf7e56a72013-12-29 11:39:51 -0800799static void tcp_set_rto(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700800{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700801 const struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700802 /* Old crap is replaced with new one. 8)
803 *
804 * More seriously:
805 * 1. If rtt variance happened to be less 50msec, it is hallucination.
806 * It cannot be less due to utterly erratic ACK generation made
807 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
808 * to do with delayed acks, because at cwnd>2 true delack timeout
809 * is invisible. Actually, Linux-2.4 also generates erratic
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800810 * ACKs in some circumstances.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700811 */
Damian Lukowskif1ecd5d2009-08-26 00:16:31 +0000812 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700813
814 /* 2. Fixups made earlier cannot be right.
815 * If we do not estimate RTO correctly without them,
816 * all the algo is pure shit and should be replaced
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800817 * with correct one. It is exactly, which we pretend to do.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700818 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700819
Ilpo Järvinenee6aac52008-12-05 22:43:08 -0800820 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
821 * guarantees that rto is higher.
822 */
Damian Lukowskif1ecd5d2009-08-26 00:16:31 +0000823 tcp_bound_rto(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700824}
825
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400826__u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700827{
828 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
829
Gerrit Renker22b71c82010-08-29 19:23:12 +0000830 if (!cwnd)
David S. Miller442b9632011-02-02 17:05:11 -0800831 cwnd = TCP_INIT_CWND;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700832 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
833}
834
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300835/*
836 * Packet counting of FACK is based on in-order assumptions, therefore TCP
837 * disables it when reordering is detected
838 */
David S. Miller4aabd8e2012-07-09 16:07:30 -0700839void tcp_disable_fack(struct tcp_sock *tp)
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300840{
Ilpo Järvinen85cc3912007-11-15 19:39:31 -0800841 /* RFC3517 uses different metric in lost marker => reset on change */
842 if (tcp_is_fack(tp))
843 tp->lost_skb_hint = NULL;
Vijay Subramanianab562222011-12-20 13:23:24 +0000844 tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300845}
846
Ryousei Takano564262c12007-10-25 23:03:52 -0700847/* Take a notice that peer is sending D-SACKs */
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300848static void tcp_dsack_seen(struct tcp_sock *tp)
849{
Vijay Subramanianab562222011-12-20 13:23:24 +0000850 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300851}
852
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300853static void tcp_update_reordering(struct sock *sk, const int metric,
854 const int ts)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700855{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300856 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700857 if (metric > tp->reordering) {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -0700858 int mib_idx;
859
Eric Dumazetdca145f2014-10-27 21:45:24 -0700860 tp->reordering = min(sysctl_tcp_max_reordering, metric);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700861
862 /* This exciting event is worth to be remembered. 8) */
863 if (ts)
Pavel Emelyanov40b215e2008-07-03 01:05:41 -0700864 mib_idx = LINUX_MIB_TCPTSREORDER;
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300865 else if (tcp_is_reno(tp))
Pavel Emelyanov40b215e2008-07-03 01:05:41 -0700866 mib_idx = LINUX_MIB_TCPRENOREORDER;
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300867 else if (tcp_is_fack(tp))
Pavel Emelyanov40b215e2008-07-03 01:05:41 -0700868 mib_idx = LINUX_MIB_TCPFACKREORDER;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700869 else
Pavel Emelyanov40b215e2008-07-03 01:05:41 -0700870 mib_idx = LINUX_MIB_TCPSACKREORDER;
871
Eric Dumazetc10d9312016-04-29 14:16:47 -0700872 NET_INC_STATS(sock_net(sk), mib_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700873#if FASTRETRANS_DEBUG > 1
Joe Perches91df42b2012-05-15 14:11:54 +0000874 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
875 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
876 tp->reordering,
877 tp->fackets_out,
878 tp->sacked_out,
879 tp->undo_marker ? tp->undo_retrans : 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700880#endif
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300881 tcp_disable_fack(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700882 }
Yuchung Chengeed530b2012-05-02 13:30:03 +0000883
884 if (metric > 0)
885 tcp_disable_early_retrans(tp);
Yuchung Cheng4f41b1c2015-10-16 21:57:47 -0700886 tp->rack.reord = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700887}
888
Ilpo Järvinen006f5822008-09-20 21:20:20 -0700889/* This must be called before lost_out is incremented */
Ilpo Järvinenc8c213f2008-09-20 21:18:55 -0700890static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
891{
Ian Morris51456b22015-04-03 09:17:26 +0100892 if (!tp->retransmit_skb_hint ||
Ilpo Järvinenc8c213f2008-09-20 21:18:55 -0700893 before(TCP_SKB_CB(skb)->seq,
894 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
Ilpo Järvinen006f5822008-09-20 21:20:20 -0700895 tp->retransmit_skb_hint = skb;
896
897 if (!tp->lost_out ||
898 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
899 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
Ilpo Järvinenc8c213f2008-09-20 21:18:55 -0700900}
901
Ilpo Järvinen41ea36e2008-09-20 21:19:22 -0700902static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
903{
904 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
905 tcp_verify_retransmit_hint(tp, skb);
906
907 tp->lost_out += tcp_skb_pcount(skb);
908 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
909 }
910}
911
Yuchung Cheng4f41b1c2015-10-16 21:57:47 -0700912void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb)
Ilpo Järvinen006f5822008-09-20 21:20:20 -0700913{
914 tcp_verify_retransmit_hint(tp, skb);
915
916 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
917 tp->lost_out += tcp_skb_pcount(skb);
918 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
919 }
920}
921
Linus Torvalds1da177e2005-04-16 15:20:36 -0700922/* This procedure tags the retransmission queue when SACKs arrive.
923 *
924 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
925 * Packets in queue with these bits set are counted in variables
926 * sacked_out, retrans_out and lost_out, correspondingly.
927 *
928 * Valid combinations are:
929 * Tag InFlight Description
930 * 0 1 - orig segment is in flight.
931 * S 0 - nothing flies, orig reached receiver.
932 * L 0 - nothing flies, orig lost by net.
933 * R 2 - both orig and retransmit are in flight.
934 * L|R 1 - orig is lost, retransmit is in flight.
935 * S|R 1 - orig reached receiver, retrans is still in flight.
936 * (L|S|R is logically valid, it could occur when L|R is sacked,
937 * but it is equivalent to plain S and code short-curcuits it to S.
938 * L|S is logically invalid, it would mean -1 packet in flight 8))
939 *
940 * These 6 states form finite state machine, controlled by the following events:
941 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
942 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
Yuchung Cheng974c1232012-01-19 14:42:21 +0000943 * 3. Loss detection event of two flavors:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700944 * A. Scoreboard estimator decided the packet is lost.
945 * A'. Reno "three dupacks" marks head of queue lost.
Yuchung Cheng974c1232012-01-19 14:42:21 +0000946 * A''. Its FACK modification, head until snd.fack is lost.
947 * B. SACK arrives sacking SND.NXT at the moment, when the
Linus Torvalds1da177e2005-04-16 15:20:36 -0700948 * segment was retransmitted.
949 * 4. D-SACK added new rule: D-SACK changes any tag to S.
950 *
951 * It is pleasant to note, that state diagram turns out to be commutative,
952 * so that we are allowed not to be bothered by order of our actions,
953 * when multiple events arrive simultaneously. (see the function below).
954 *
955 * Reordering detection.
956 * --------------------
957 * Reordering metric is maximal distance, which a packet can be displaced
958 * in packet stream. With SACKs we can estimate it:
959 *
960 * 1. SACK fills old hole and the corresponding segment was not
961 * ever retransmitted -> reordering. Alas, we cannot use it
962 * when segment was retransmitted.
963 * 2. The last flaw is solved with D-SACK. D-SACK arrives
964 * for retransmitted and already SACKed segment -> reordering..
965 * Both of these heuristics are not used in Loss state, when we cannot
966 * account for retransmits accurately.
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -0700967 *
968 * SACK block validation.
969 * ----------------------
970 *
971 * SACK block range validation checks that the received SACK block fits to
972 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
973 * Note that SND.UNA is not included to the range though being valid because
Ilpo Järvinen0e835332007-10-01 15:28:17 -0700974 * it means that the receiver is rather inconsistent with itself reporting
975 * SACK reneging when it should advance SND.UNA. Such SACK block this is
976 * perfectly valid, however, in light of RFC2018 which explicitly states
977 * that "SACK block MUST reflect the newest segment. Even if the newest
978 * segment is going to be discarded ...", not that it looks very clever
979 * in case of head skb. Due to potentional receiver driven attacks, we
980 * choose to avoid immediate execution of a walk in write queue due to
981 * reneging and defer head skb's loss recovery to standard loss recovery
982 * procedure that will eventually trigger (nothing forbids us doing this).
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -0700983 *
984 * Implements also blockage to start_seq wrap-around. Problem lies in the
985 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
986 * there's no guarantee that it will be before snd_nxt (n). The problem
987 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
988 * wrap (s_w):
989 *
990 * <- outs wnd -> <- wrapzone ->
991 * u e n u_w e_w s n_w
992 * | | | | | | |
993 * |<------------+------+----- TCP seqno space --------------+---------->|
994 * ...-- <2^31 ->| |<--------...
995 * ...---- >2^31 ------>| |<--------...
996 *
997 * Current code wouldn't be vulnerable but it's better still to discard such
998 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
999 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1000 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1001 * equal to the ideal case (infinite seqno space without wrap caused issues).
1002 *
1003 * With D-SACK the lower bound is extended to cover sequence space below
1004 * SND.UNA down to undo_marker, which is the last point of interest. Yet
Ryousei Takano564262c12007-10-25 23:03:52 -07001005 * again, D-SACK block must not to go across snd_una (for the same reason as
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001006 * for the normal SACK blocks, explained above). But there all simplicity
1007 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1008 * fully below undo_marker they do not affect behavior in anyway and can
1009 * therefore be safely ignored. In rare cases (which are more or less
1010 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1011 * fragmentation and packet reordering past skb's retransmission. To consider
1012 * them correctly, the acceptable range must be extended even more though
1013 * the exact amount is rather hard to quantify. However, tp->max_window can
1014 * be used as an exaggerated estimate.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001015 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001016static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1017 u32 start_seq, u32 end_seq)
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001018{
1019 /* Too far in future, or reversed (interpretation is ambiguous) */
1020 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001021 return false;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001022
1023 /* Nasty start_seq wrap-around check (see comments above) */
1024 if (!before(start_seq, tp->snd_nxt))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001025 return false;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001026
Ryousei Takano564262c12007-10-25 23:03:52 -07001027 /* In outstanding window? ...This is valid exit for D-SACKs too.
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001028 * start_seq == snd_una is non-sensical (see comments above)
1029 */
1030 if (after(start_seq, tp->snd_una))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001031 return true;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001032
1033 if (!is_dsack || !tp->undo_marker)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001034 return false;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001035
1036 /* ...Then it's D-SACK, and must reside below snd_una completely */
Zheng Yanf779b2d2011-09-18 22:37:34 -04001037 if (after(end_seq, tp->snd_una))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001038 return false;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001039
1040 if (!before(start_seq, tp->undo_marker))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001041 return true;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001042
1043 /* Too old */
1044 if (!after(end_seq, tp->undo_marker))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001045 return false;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001046
1047 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1048 * start_seq < undo_marker and end_seq >= undo_marker.
1049 */
1050 return !before(start_seq, end_seq - tp->max_window);
1051}
1052
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001053static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1054 struct tcp_sack_block_wire *sp, int num_sacks,
1055 u32 prior_snd_una)
David S. Millerd06e0212007-06-18 22:43:06 -07001056{
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07001057 struct tcp_sock *tp = tcp_sk(sk);
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07001058 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1059 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001060 bool dup_sack = false;
David S. Millerd06e0212007-06-18 22:43:06 -07001061
1062 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001063 dup_sack = true;
Ilpo Järvinene60402d2007-08-09 15:14:46 +03001064 tcp_dsack_seen(tp);
Eric Dumazetc10d9312016-04-29 14:16:47 -07001065 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
David S. Millerd06e0212007-06-18 22:43:06 -07001066 } else if (num_sacks > 1) {
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07001067 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1068 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
David S. Millerd06e0212007-06-18 22:43:06 -07001069
1070 if (!after(end_seq_0, end_seq_1) &&
1071 !before(start_seq_0, start_seq_1)) {
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001072 dup_sack = true;
Ilpo Järvinene60402d2007-08-09 15:14:46 +03001073 tcp_dsack_seen(tp);
Eric Dumazetc10d9312016-04-29 14:16:47 -07001074 NET_INC_STATS(sock_net(sk),
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07001075 LINUX_MIB_TCPDSACKOFORECV);
David S. Millerd06e0212007-06-18 22:43:06 -07001076 }
1077 }
1078
1079 /* D-SACK for already forgotten data... Do dumb counting. */
Yuchung Cheng6e08d5e2014-07-02 12:07:16 -07001080 if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
David S. Millerd06e0212007-06-18 22:43:06 -07001081 !after(end_seq_0, prior_snd_una) &&
1082 after(end_seq_0, tp->undo_marker))
1083 tp->undo_retrans--;
1084
1085 return dup_sack;
1086}
1087
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001088struct tcp_sacktag_state {
Eric Dumazet740b0f12014-02-26 14:02:48 -08001089 int reord;
1090 int fack_count;
Kenneth Klette Jonassen31231a82015-05-01 01:10:58 +02001091 /* Timestamps for earliest and latest never-retransmitted segment
1092 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1093 * but congestion control should still get an accurate delay signal.
1094 */
1095 struct skb_mstamp first_sackt;
1096 struct skb_mstamp last_sackt;
Eric Dumazet740b0f12014-02-26 14:02:48 -08001097 int flag;
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001098};
1099
Ilpo Järvinend1935942007-10-11 17:34:25 -07001100/* Check if skb is fully within the SACK block. In presence of GSO skbs,
1101 * the incoming SACK may not exactly match but we can find smaller MSS
1102 * aligned portion of it that matches. Therefore we might need to fragment
1103 * which may fail and creates some hassle (caller must handle error case
1104 * returns).
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001105 *
1106 * FIXME: this could be merged to shift decision code
Ilpo Järvinend1935942007-10-11 17:34:25 -07001107 */
Adrian Bunk0f79efd2007-10-26 03:57:36 -07001108static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001109 u32 start_seq, u32 end_seq)
Ilpo Järvinend1935942007-10-11 17:34:25 -07001110{
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001111 int err;
1112 bool in_sack;
Ilpo Järvinend1935942007-10-11 17:34:25 -07001113 unsigned int pkt_len;
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001114 unsigned int mss;
Ilpo Järvinend1935942007-10-11 17:34:25 -07001115
1116 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1117 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1118
1119 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1120 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001121 mss = tcp_skb_mss(skb);
Ilpo Järvinend1935942007-10-11 17:34:25 -07001122 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1123
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001124 if (!in_sack) {
Ilpo Järvinend1935942007-10-11 17:34:25 -07001125 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001126 if (pkt_len < mss)
1127 pkt_len = mss;
1128 } else {
Ilpo Järvinend1935942007-10-11 17:34:25 -07001129 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001130 if (pkt_len < mss)
1131 return -EINVAL;
1132 }
1133
1134 /* Round if necessary so that SACKs cover only full MSSes
1135 * and/or the remaining small portion (if present)
1136 */
1137 if (pkt_len > mss) {
1138 unsigned int new_len = (pkt_len / mss) * mss;
1139 if (!in_sack && new_len < pkt_len) {
1140 new_len += mss;
Neal Cardwell2cd0d742014-06-18 21:15:03 -04001141 if (new_len >= skb->len)
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001142 return 0;
1143 }
1144 pkt_len = new_len;
1145 }
Octavian Purdila6cc55e02014-06-06 17:32:37 +03001146 err = tcp_fragment(sk, skb, pkt_len, mss, GFP_ATOMIC);
Ilpo Järvinend1935942007-10-11 17:34:25 -07001147 if (err < 0)
1148 return err;
1149 }
1150
1151 return in_sack;
1152}
1153
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001154/* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1155static u8 tcp_sacktag_one(struct sock *sk,
1156 struct tcp_sacktag_state *state, u8 sacked,
1157 u32 start_seq, u32 end_seq,
Eric Dumazet740b0f12014-02-26 14:02:48 -08001158 int dup_sack, int pcount,
1159 const struct skb_mstamp *xmit_time)
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001160{
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001161 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001162 int fack_count = state->fack_count;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001163
1164 /* Account D-SACK for retransmitted packet. */
1165 if (dup_sack && (sacked & TCPCB_RETRANS)) {
Yuchung Cheng6e08d5e2014-07-02 12:07:16 -07001166 if (tp->undo_marker && tp->undo_retrans > 0 &&
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001167 after(end_seq, tp->undo_marker))
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001168 tp->undo_retrans--;
Ilpo Järvinenede9f3b2007-12-02 00:47:58 +02001169 if (sacked & TCPCB_SACKED_ACKED)
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001170 state->reord = min(fack_count, state->reord);
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001171 }
1172
1173 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001174 if (!after(end_seq, tp->snd_una))
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001175 return sacked;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001176
1177 if (!(sacked & TCPCB_SACKED_ACKED)) {
Yuchung Cheng659a8ad2015-10-16 21:57:46 -07001178 tcp_rack_advance(tp, xmit_time, sacked);
1179
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001180 if (sacked & TCPCB_SACKED_RETRANS) {
1181 /* If the segment is not tagged as lost,
1182 * we do not clear RETRANS, believing
1183 * that retransmission is still in flight.
1184 */
1185 if (sacked & TCPCB_LOST) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001186 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001187 tp->lost_out -= pcount;
1188 tp->retrans_out -= pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001189 }
1190 } else {
1191 if (!(sacked & TCPCB_RETRANS)) {
1192 /* New sack for not retransmitted frame,
1193 * which was in hole. It is reordering.
1194 */
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001195 if (before(start_seq,
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001196 tcp_highest_sack_seq(tp)))
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001197 state->reord = min(fack_count,
1198 state->reord);
Yuchung Chenge33099f2013-03-20 13:33:00 +00001199 if (!after(end_seq, tp->high_seq))
1200 state->flag |= FLAG_ORIG_SACK_ACKED;
Kenneth Klette Jonassen31231a82015-05-01 01:10:58 +02001201 if (state->first_sackt.v64 == 0)
1202 state->first_sackt = *xmit_time;
1203 state->last_sackt = *xmit_time;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001204 }
1205
1206 if (sacked & TCPCB_LOST) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001207 sacked &= ~TCPCB_LOST;
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001208 tp->lost_out -= pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001209 }
1210 }
1211
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001212 sacked |= TCPCB_SACKED_ACKED;
1213 state->flag |= FLAG_DATA_SACKED;
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001214 tp->sacked_out += pcount;
Yuchung Chengddf1af62016-02-02 10:33:06 -08001215 tp->delivered += pcount; /* Out-of-order packets delivered */
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001216
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001217 fack_count += pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001218
1219 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
Ian Morris00db4122015-04-03 09:17:27 +01001220 if (!tcp_is_fack(tp) && tp->lost_skb_hint &&
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001221 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001222 tp->lost_cnt_hint += pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001223
1224 if (fack_count > tp->fackets_out)
1225 tp->fackets_out = fack_count;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001226 }
1227
1228 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1229 * frames and clear it. undo_retrans is decreased above, L|R frames
1230 * are accounted above as well.
1231 */
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001232 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1233 sacked &= ~TCPCB_SACKED_RETRANS;
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001234 tp->retrans_out -= pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001235 }
1236
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001237 return sacked;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001238}
1239
Neal Cardwelldaef52b2012-02-12 18:37:10 +00001240/* Shift newly-SACKed bytes from this skb to the immediately previous
1241 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1242 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001243static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1244 struct tcp_sacktag_state *state,
1245 unsigned int pcount, int shifted, int mss,
1246 bool dup_sack)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001247{
1248 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen50133162008-12-05 22:42:41 -08001249 struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
Neal Cardwelldaef52b2012-02-12 18:37:10 +00001250 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1251 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001252
1253 BUG_ON(!pcount);
1254
Neal Cardwell4c90d3b2012-02-26 10:06:19 +00001255 /* Adjust counters and hints for the newly sacked sequence
1256 * range but discard the return value since prev is already
1257 * marked. We must tag the range first because the seq
1258 * advancement below implicitly advances
1259 * tcp_highest_sack_seq() when skb is highest_sack.
1260 */
1261 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
Yuchung Cheng59c9af42013-07-22 16:20:47 -07001262 start_seq, end_seq, dup_sack, pcount,
Eric Dumazet740b0f12014-02-26 14:02:48 -08001263 &skb->skb_mstamp);
Neal Cardwell4c90d3b2012-02-26 10:06:19 +00001264
1265 if (skb == tp->lost_skb_hint)
Neal Cardwell0af2a0d2012-02-13 20:22:08 +00001266 tp->lost_cnt_hint += pcount;
1267
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001268 TCP_SKB_CB(prev)->end_seq += shifted;
1269 TCP_SKB_CB(skb)->seq += shifted;
1270
Eric Dumazetcd7d8492014-09-24 04:11:22 -07001271 tcp_skb_pcount_add(prev, pcount);
1272 BUG_ON(tcp_skb_pcount(skb) < pcount);
1273 tcp_skb_pcount_add(skb, -pcount);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001274
1275 /* When we're adding to gso_segs == 1, gso_size will be zero,
1276 * in theory this shouldn't be necessary but as long as DSACK
1277 * code can come after this skb later on it's better to keep
1278 * setting gso_size to something.
1279 */
Eric Dumazetf69ad292015-06-11 09:15:18 -07001280 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1281 TCP_SKB_CB(prev)->tcp_gso_size = mss;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001282
1283 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
Eric Dumazet51466a72015-06-11 09:15:16 -07001284 if (tcp_skb_pcount(skb) <= 1)
Eric Dumazetf69ad292015-06-11 09:15:18 -07001285 TCP_SKB_CB(skb)->tcp_gso_size = 0;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001286
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001287 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1288 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1289
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001290 if (skb->len > 0) {
1291 BUG_ON(!tcp_skb_pcount(skb));
Eric Dumazetc10d9312016-04-29 14:16:47 -07001292 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001293 return false;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001294 }
1295
1296 /* Whole SKB was eaten :-) */
1297
Ilpo Järvinen92ee76b2008-11-24 21:26:56 -08001298 if (skb == tp->retransmit_skb_hint)
1299 tp->retransmit_skb_hint = prev;
Ilpo Järvinen92ee76b2008-11-24 21:26:56 -08001300 if (skb == tp->lost_skb_hint) {
1301 tp->lost_skb_hint = prev;
1302 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1303 }
1304
Eric Dumazet5e8a402f2013-10-04 10:31:41 -07001305 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
Martin KaFai Laua643b5d2016-04-25 14:44:49 -07001306 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
Eric Dumazet5e8a402f2013-10-04 10:31:41 -07001307 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1308 TCP_SKB_CB(prev)->end_seq++;
1309
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001310 if (skb == tcp_highest_sack(sk))
1311 tcp_advance_highest_sack(sk, skb);
1312
Martin KaFai Laucfea5a62016-04-19 22:39:29 -07001313 tcp_skb_collapse_tstamp(prev, skb);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001314 tcp_unlink_write_queue(skb, sk);
1315 sk_wmem_free_skb(sk, skb);
1316
Eric Dumazetc10d9312016-04-29 14:16:47 -07001317 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
Ilpo Järvinen111cc8b2008-11-24 21:27:22 -08001318
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001319 return true;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001320}
1321
1322/* I wish gso_size would have a bit more sane initialization than
1323 * something-or-zero which complicates things
1324 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001325static int tcp_skb_seglen(const struct sk_buff *skb)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001326{
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001327 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001328}
1329
1330/* Shifting pages past head area doesn't work */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001331static int skb_can_shift(const struct sk_buff *skb)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001332{
1333 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1334}
1335
1336/* Try collapsing SACK blocks spanning across multiple skbs to a single
1337 * skb.
1338 */
1339static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001340 struct tcp_sacktag_state *state,
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001341 u32 start_seq, u32 end_seq,
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001342 bool dup_sack)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001343{
1344 struct tcp_sock *tp = tcp_sk(sk);
1345 struct sk_buff *prev;
1346 int mss;
1347 int pcount = 0;
1348 int len;
1349 int in_sack;
1350
1351 if (!sk_can_gso(sk))
1352 goto fallback;
1353
1354 /* Normally R but no L won't result in plain S */
1355 if (!dup_sack &&
Ilpo Järvinen9969ca52008-12-05 22:41:06 -08001356 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001357 goto fallback;
1358 if (!skb_can_shift(skb))
1359 goto fallback;
1360 /* This frame is about to be dropped (was ACKed). */
1361 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1362 goto fallback;
1363
1364 /* Can only happen with delayed DSACK + discard craziness */
1365 if (unlikely(skb == tcp_write_queue_head(sk)))
1366 goto fallback;
1367 prev = tcp_write_queue_prev(sk, skb);
1368
1369 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1370 goto fallback;
1371
Martin KaFai Laua643b5d2016-04-25 14:44:49 -07001372 if (!tcp_skb_can_collapse_to(prev))
1373 goto fallback;
1374
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001375 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1376 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1377
1378 if (in_sack) {
1379 len = skb->len;
1380 pcount = tcp_skb_pcount(skb);
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001381 mss = tcp_skb_seglen(skb);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001382
1383 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1384 * drop this restriction as unnecessary
1385 */
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001386 if (mss != tcp_skb_seglen(prev))
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001387 goto fallback;
1388 } else {
1389 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1390 goto noop;
1391 /* CHECKME: This is non-MSS split case only?, this will
1392 * cause skipped skbs due to advancing loop btw, original
1393 * has that feature too
1394 */
1395 if (tcp_skb_pcount(skb) <= 1)
1396 goto noop;
1397
1398 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1399 if (!in_sack) {
1400 /* TODO: head merge to next could be attempted here
1401 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1402 * though it might not be worth of the additional hassle
1403 *
1404 * ...we can probably just fallback to what was done
1405 * previously. We could try merging non-SACKed ones
1406 * as well but it probably isn't going to buy off
1407 * because later SACKs might again split them, and
1408 * it would make skb timestamp tracking considerably
1409 * harder problem.
1410 */
1411 goto fallback;
1412 }
1413
1414 len = end_seq - TCP_SKB_CB(skb)->seq;
1415 BUG_ON(len < 0);
1416 BUG_ON(len > skb->len);
1417
1418 /* MSS boundaries should be honoured or else pcount will
1419 * severely break even though it makes things bit trickier.
1420 * Optimize common case to avoid most of the divides
1421 */
1422 mss = tcp_skb_mss(skb);
1423
1424 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1425 * drop this restriction as unnecessary
1426 */
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001427 if (mss != tcp_skb_seglen(prev))
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001428 goto fallback;
1429
1430 if (len == mss) {
1431 pcount = 1;
1432 } else if (len < mss) {
1433 goto noop;
1434 } else {
1435 pcount = len / mss;
1436 len = pcount * mss;
1437 }
1438 }
1439
Neal Cardwell4648dc92012-03-05 19:35:04 +00001440 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1441 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1442 goto fallback;
1443
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001444 if (!skb_shift(prev, skb, len))
1445 goto fallback;
Ilpo Järvinen9ec06ff2009-03-01 00:21:36 -08001446 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001447 goto out;
1448
1449 /* Hole filled allows collapsing with the next as well, this is very
1450 * useful when hole on every nth skb pattern happens
1451 */
1452 if (prev == tcp_write_queue_tail(sk))
1453 goto out;
1454 skb = tcp_write_queue_next(sk, prev);
1455
Ilpo Järvinenf0bc52f2008-12-05 22:40:47 -08001456 if (!skb_can_shift(skb) ||
1457 (skb == tcp_send_head(sk)) ||
1458 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001459 (mss != tcp_skb_seglen(skb)))
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001460 goto out;
1461
1462 len = skb->len;
1463 if (skb_shift(prev, skb, len)) {
1464 pcount += tcp_skb_pcount(skb);
Ilpo Järvinen9ec06ff2009-03-01 00:21:36 -08001465 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001466 }
1467
1468out:
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001469 state->fack_count += pcount;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001470 return prev;
1471
1472noop:
1473 return skb;
1474
1475fallback:
Eric Dumazetc10d9312016-04-29 14:16:47 -07001476 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001477 return NULL;
1478}
1479
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001480static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1481 struct tcp_sack_block *next_dup,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001482 struct tcp_sacktag_state *state,
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001483 u32 start_seq, u32 end_seq,
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001484 bool dup_sack_in)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001485{
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001486 struct tcp_sock *tp = tcp_sk(sk);
1487 struct sk_buff *tmp;
1488
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001489 tcp_for_write_queue_from(skb, sk) {
1490 int in_sack = 0;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001491 bool dup_sack = dup_sack_in;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001492
1493 if (skb == tcp_send_head(sk))
1494 break;
1495
1496 /* queue is in-order => we can short-circuit the walk early */
1497 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1498 break;
1499
Ian Morris00db4122015-04-03 09:17:27 +01001500 if (next_dup &&
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001501 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1502 in_sack = tcp_match_skb_to_sack(sk, skb,
1503 next_dup->start_seq,
1504 next_dup->end_seq);
1505 if (in_sack > 0)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001506 dup_sack = true;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001507 }
1508
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001509 /* skb reference here is a bit tricky to get right, since
1510 * shifting can eat and free both this skb and the next,
1511 * so not even _safe variant of the loop is enough.
1512 */
1513 if (in_sack <= 0) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001514 tmp = tcp_shift_skb_data(sk, skb, state,
1515 start_seq, end_seq, dup_sack);
Ian Morris00db4122015-04-03 09:17:27 +01001516 if (tmp) {
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001517 if (tmp != skb) {
1518 skb = tmp;
1519 continue;
1520 }
1521
1522 in_sack = 0;
1523 } else {
1524 in_sack = tcp_match_skb_to_sack(sk, skb,
1525 start_seq,
1526 end_seq);
1527 }
1528 }
1529
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001530 if (unlikely(in_sack < 0))
1531 break;
1532
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001533 if (in_sack) {
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001534 TCP_SKB_CB(skb)->sacked =
1535 tcp_sacktag_one(sk,
1536 state,
1537 TCP_SKB_CB(skb)->sacked,
1538 TCP_SKB_CB(skb)->seq,
1539 TCP_SKB_CB(skb)->end_seq,
1540 dup_sack,
Yuchung Cheng59c9af42013-07-22 16:20:47 -07001541 tcp_skb_pcount(skb),
Eric Dumazet740b0f12014-02-26 14:02:48 -08001542 &skb->skb_mstamp);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001543
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001544 if (!before(TCP_SKB_CB(skb)->seq,
1545 tcp_highest_sack_seq(tp)))
1546 tcp_advance_highest_sack(sk, skb);
1547 }
1548
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001549 state->fack_count += tcp_skb_pcount(skb);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001550 }
1551 return skb;
1552}
1553
1554/* Avoid all extra work that is being done by sacktag while walking in
1555 * a normal way
1556 */
1557static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001558 struct tcp_sacktag_state *state,
1559 u32 skip_to_seq)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001560{
1561 tcp_for_write_queue_from(skb, sk) {
1562 if (skb == tcp_send_head(sk))
1563 break;
1564
Ilpo Järvinene8bae272008-11-24 21:12:28 -08001565 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001566 break;
Ilpo Järvinend152a7d2008-03-03 12:10:16 -08001567
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001568 state->fack_count += tcp_skb_pcount(skb);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001569 }
1570 return skb;
1571}
1572
1573static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1574 struct sock *sk,
1575 struct tcp_sack_block *next_dup,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001576 struct tcp_sacktag_state *state,
1577 u32 skip_to_seq)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001578{
Ian Morris51456b22015-04-03 09:17:26 +01001579 if (!next_dup)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001580 return skb;
1581
1582 if (before(next_dup->start_seq, skip_to_seq)) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001583 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1584 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1585 next_dup->start_seq, next_dup->end_seq,
1586 1);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001587 }
1588
1589 return skb;
1590}
1591
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001592static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001593{
1594 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1595}
1596
Linus Torvalds1da177e2005-04-16 15:20:36 -07001597static int
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001598tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001599 u32 prior_snd_una, struct tcp_sacktag_state *state)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001600{
1601 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001602 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1603 TCP_SKB_CB(ack_skb)->sacked);
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001604 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
Adam Langley4389dde2008-07-19 00:07:02 -07001605 struct tcp_sack_block sp[TCP_NUM_SACKS];
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001606 struct tcp_sack_block *cache;
1607 struct sk_buff *skb;
Adam Langley4389dde2008-07-19 00:07:02 -07001608 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001609 int used_sacks;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001610 bool found_dup_sack = false;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001611 int i, j;
Baruch Evenfda03fb2007-02-04 23:35:57 -08001612 int first_sack_index;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001613
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001614 state->flag = 0;
1615 state->reord = tp->packets_out;
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001616
Ilpo Järvinend738cd82007-03-24 21:03:23 -07001617 if (!tp->sacked_out) {
Ilpo Järvinende83c052007-10-07 23:37:55 -07001618 if (WARN_ON(tp->fackets_out))
1619 tp->fackets_out = 0;
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001620 tcp_highest_sack_reset(sk);
Ilpo Järvinend738cd82007-03-24 21:03:23 -07001621 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001622
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07001623 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
David S. Millerd06e0212007-06-18 22:43:06 -07001624 num_sacks, prior_snd_una);
1625 if (found_dup_sack)
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001626 state->flag |= FLAG_DSACKING_ACK;
Baruch Even6f746512007-02-04 23:36:42 -08001627
1628 /* Eliminate too old ACKs, but take into
1629 * account more or less fresh ones, they can
1630 * contain valid SACK info.
1631 */
1632 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1633 return 0;
1634
Ilpo Järvinen96a2d412007-11-14 15:47:18 -08001635 if (!tp->packets_out)
1636 goto out;
1637
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001638 used_sacks = 0;
1639 first_sack_index = 0;
1640 for (i = 0; i < num_sacks; i++) {
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001641 bool dup_sack = !i && found_dup_sack;
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001642
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07001643 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1644 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001645
1646 if (!tcp_is_sackblock_valid(tp, dup_sack,
1647 sp[used_sacks].start_seq,
1648 sp[used_sacks].end_seq)) {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001649 int mib_idx;
1650
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001651 if (dup_sack) {
1652 if (!tp->undo_marker)
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001653 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001654 else
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001655 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001656 } else {
1657 /* Don't count olds caused by ACK reordering */
1658 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1659 !after(sp[used_sacks].end_seq, tp->snd_una))
1660 continue;
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001661 mib_idx = LINUX_MIB_TCPSACKDISCARD;
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001662 }
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001663
Eric Dumazetc10d9312016-04-29 14:16:47 -07001664 NET_INC_STATS(sock_net(sk), mib_idx);
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001665 if (i == 0)
1666 first_sack_index = -1;
1667 continue;
1668 }
1669
1670 /* Ignore very old stuff early */
1671 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1672 continue;
1673
1674 used_sacks++;
1675 }
1676
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001677 /* order SACK blocks to allow in order walk of the retrans queue */
1678 for (i = used_sacks - 1; i > 0; i--) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001679 for (j = 0; j < i; j++) {
1680 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
Ilpo Järvinena0bffff2009-03-21 13:36:17 -07001681 swap(sp[j], sp[j + 1]);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001682
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001683 /* Track where the first SACK block goes to */
1684 if (j == first_sack_index)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001685 first_sack_index = j + 1;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001686 }
1687 }
1688 }
1689
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001690 skb = tcp_write_queue_head(sk);
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001691 state->fack_count = 0;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001692 i = 0;
1693
1694 if (!tp->sacked_out) {
1695 /* It's already past, so skip checking against it */
1696 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1697 } else {
1698 cache = tp->recv_sack_cache;
1699 /* Skip empty blocks in at head of the cache */
1700 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1701 !cache->end_seq)
1702 cache++;
Baruch Evenfda03fb2007-02-04 23:35:57 -08001703 }
1704
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001705 while (i < used_sacks) {
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001706 u32 start_seq = sp[i].start_seq;
1707 u32 end_seq = sp[i].end_seq;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001708 bool dup_sack = (found_dup_sack && (i == first_sack_index));
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001709 struct tcp_sack_block *next_dup = NULL;
Ilpo Järvinene56d6cd2007-11-01 00:09:37 -07001710
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001711 if (found_dup_sack && ((i + 1) == first_sack_index))
1712 next_dup = &sp[i + 1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001713
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001714 /* Skip too early cached blocks */
1715 while (tcp_sack_cache_ok(tp, cache) &&
1716 !before(start_seq, cache->end_seq))
1717 cache++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001718
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001719 /* Can skip some work by looking recv_sack_cache? */
1720 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1721 after(end_seq, cache->start_seq)) {
David S. Millerfe067e82007-03-07 12:12:44 -08001722
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001723 /* Head todo? */
1724 if (before(start_seq, cache->start_seq)) {
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001725 skb = tcp_sacktag_skip(skb, sk, state,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001726 start_seq);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001727 skb = tcp_sacktag_walk(skb, sk, next_dup,
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001728 state,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001729 start_seq,
1730 cache->start_seq,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001731 dup_sack);
Baruch Evenfda03fb2007-02-04 23:35:57 -08001732 }
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001733
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001734 /* Rest of the block already fully processed? */
Ilpo Järvinen20de20b2007-11-16 16:17:05 -08001735 if (!after(end_seq, cache->end_seq))
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001736 goto advance_sp;
Ilpo Järvinen20de20b2007-11-16 16:17:05 -08001737
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001738 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001739 state,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001740 cache->end_seq);
Ilpo Järvinene56d6cd2007-11-01 00:09:37 -07001741
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001742 /* ...tail remains todo... */
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001743 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
Ilpo Järvinen20de20b2007-11-16 16:17:05 -08001744 /* ...but better entrypoint exists! */
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001745 skb = tcp_highest_sack(sk);
Ian Morris51456b22015-04-03 09:17:26 +01001746 if (!skb)
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001747 break;
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001748 state->fack_count = tp->fackets_out;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001749 cache++;
1750 goto walk;
1751 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001752
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001753 skb = tcp_sacktag_skip(skb, sk, state, cache->end_seq);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001754 /* Check overlap against next cached too (past this one already) */
1755 cache++;
1756 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001757 }
Ilpo Järvinenfbd52eb2007-11-10 21:24:19 -08001758
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001759 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1760 skb = tcp_highest_sack(sk);
Ian Morris51456b22015-04-03 09:17:26 +01001761 if (!skb)
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001762 break;
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001763 state->fack_count = tp->fackets_out;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001764 }
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001765 skb = tcp_sacktag_skip(skb, sk, state, start_seq);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001766
1767walk:
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001768 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001769 start_seq, end_seq, dup_sack);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001770
1771advance_sp:
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001772 i++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001773 }
1774
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001775 /* Clear the head of the cache sack blocks so we can skip it next time */
1776 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1777 tp->recv_sack_cache[i].start_seq = 0;
1778 tp->recv_sack_cache[i].end_seq = 0;
1779 }
1780 for (j = 0; j < used_sacks; j++)
1781 tp->recv_sack_cache[i++] = sp[j];
1782
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001783 if ((state->reord < tp->fackets_out) &&
Yuchung Cheng9b441902013-03-20 13:32:58 +00001784 ((inet_csk(sk)->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker))
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001785 tcp_update_reordering(sk, tp->fackets_out - state->reord, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001786
Yuchung Cheng9dac8832015-04-29 11:28:30 -07001787 tcp_verify_left_out(tp);
Ilpo Järvinen96a2d412007-11-14 15:47:18 -08001788out:
1789
Linus Torvalds1da177e2005-04-16 15:20:36 -07001790#if FASTRETRANS_DEBUG > 0
Ilpo Järvinen547b7922008-07-25 21:43:18 -07001791 WARN_ON((int)tp->sacked_out < 0);
1792 WARN_ON((int)tp->lost_out < 0);
1793 WARN_ON((int)tp->retrans_out < 0);
1794 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001795#endif
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001796 return state->flag;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001797}
1798
Ilpo Järvinen882beba2008-04-07 22:33:07 -07001799/* Limits sacked_out so that sum with lost_out isn't ever larger than
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001800 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
Ilpo Järvinen30935cf2007-02-21 23:01:36 -08001801 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001802static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001803{
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001804 u32 holes;
1805
1806 holes = max(tp->lost_out, 1U);
1807 holes = min(holes, tp->packets_out);
1808
1809 if ((tp->sacked_out + holes) > tp->packets_out) {
1810 tp->sacked_out = tp->packets_out - holes;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001811 return true;
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001812 }
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001813 return false;
Ilpo Järvinen882beba2008-04-07 22:33:07 -07001814}
1815
1816/* If we receive more dupacks than we expected counting segments
1817 * in assumption of absent reordering, interpret this as reordering.
1818 * The only another reason could be bug in receiver TCP.
1819 */
1820static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1821{
1822 struct tcp_sock *tp = tcp_sk(sk);
1823 if (tcp_limit_reno_sacked(tp))
1824 tcp_update_reordering(sk, tp->packets_out + addend, 0);
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001825}
1826
1827/* Emulate SACKs for SACKless connection: account for a new dupack. */
1828
1829static void tcp_add_reno_sack(struct sock *sk)
1830{
1831 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Chengddf1af62016-02-02 10:33:06 -08001832 u32 prior_sacked = tp->sacked_out;
1833
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001834 tp->sacked_out++;
1835 tcp_check_reno_reordering(sk, 0);
Yuchung Chengddf1af62016-02-02 10:33:06 -08001836 if (tp->sacked_out > prior_sacked)
1837 tp->delivered++; /* Some out-of-order packet is delivered */
Ilpo Järvinen005903b2007-08-09 14:44:16 +03001838 tcp_verify_left_out(tp);
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001839}
1840
1841/* Account for ACK, ACKing some data in Reno Recovery phase. */
1842
1843static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1844{
1845 struct tcp_sock *tp = tcp_sk(sk);
1846
1847 if (acked > 0) {
1848 /* One ACK acked hole. The rest eat duplicate ACKs. */
Yuchung Chengddf1af62016-02-02 10:33:06 -08001849 tp->delivered += max_t(int, acked - tp->sacked_out, 1);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001850 if (acked - 1 >= tp->sacked_out)
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001851 tp->sacked_out = 0;
1852 else
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001853 tp->sacked_out -= acked - 1;
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001854 }
1855 tcp_check_reno_reordering(sk, acked);
Ilpo Järvinen005903b2007-08-09 14:44:16 +03001856 tcp_verify_left_out(tp);
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001857}
1858
1859static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1860{
1861 tp->sacked_out = 0;
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001862}
1863
Yuchung Cheng989e04c2014-08-22 14:15:22 -07001864void tcp_clear_retrans(struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001865{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001866 tp->retrans_out = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001867 tp->lost_out = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001868 tp->undo_marker = 0;
Yuchung Cheng6e08d5e2014-07-02 12:07:16 -07001869 tp->undo_retrans = -1;
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07001870 tp->fackets_out = 0;
1871 tp->sacked_out = 0;
1872}
1873
Yuchung Cheng989e04c2014-08-22 14:15:22 -07001874static inline void tcp_init_undo(struct tcp_sock *tp)
1875{
1876 tp->undo_marker = tp->snd_una;
1877 /* Retransmission still in flight may cause DSACKs later. */
1878 tp->undo_retrans = tp->retrans_out ? : -1;
1879}
1880
Neal Cardwell5ae344c2014-08-04 19:12:29 -04001881/* Enter Loss state. If we detect SACK reneging, forget all SACK information
Linus Torvalds1da177e2005-04-16 15:20:36 -07001882 * and reset tags completely, otherwise preserve SACKs. If receiver
1883 * dropped its ofo queue, we will know this due to reneging detection.
1884 */
Neal Cardwell5ae344c2014-08-04 19:12:29 -04001885void tcp_enter_loss(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001886{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001887 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001888 struct tcp_sock *tp = tcp_sk(sk);
Nikolay Borisov1043e252016-02-03 09:46:52 +02001889 struct net *net = sock_net(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001890 struct sk_buff *skb;
Yuchung Chengf82b6812015-07-13 12:10:20 -07001891 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
Neal Cardwell5ae344c2014-08-04 19:12:29 -04001892 bool is_reneg; /* is receiver reneging on SACKs? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001893
1894 /* Reduce ssthresh if it has not yet been made inside this window. */
Yuchung Chenge33099f2013-03-20 13:33:00 +00001895 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1896 !after(tp->high_seq, tp->snd_una) ||
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001897 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1898 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1899 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1900 tcp_ca_event(sk, CA_EVENT_LOSS);
Yuchung Cheng989e04c2014-08-22 14:15:22 -07001901 tcp_init_undo(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001902 }
1903 tp->snd_cwnd = 1;
1904 tp->snd_cwnd_cnt = 0;
1905 tp->snd_cwnd_stamp = tcp_time_stamp;
1906
Yuchung Cheng989e04c2014-08-22 14:15:22 -07001907 tp->retrans_out = 0;
1908 tp->lost_out = 0;
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07001909
1910 if (tcp_is_reno(tp))
1911 tcp_reset_reno_sack(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001912
Neal Cardwell5ae344c2014-08-04 19:12:29 -04001913 skb = tcp_write_queue_head(sk);
1914 is_reneg = skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED);
1915 if (is_reneg) {
Eric Dumazetc10d9312016-04-29 14:16:47 -07001916 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07001917 tp->sacked_out = 0;
1918 tp->fackets_out = 0;
Ilpo Järvinenb7689202007-09-20 11:37:19 -07001919 }
Ilpo Järvinen64edc272008-09-20 21:18:32 -07001920 tcp_clear_all_retrans_hints(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001921
David S. Millerfe067e82007-03-07 12:12:44 -08001922 tcp_for_write_queue(skb, sk) {
1923 if (skb == tcp_send_head(sk))
1924 break;
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07001925
Linus Torvalds1da177e2005-04-16 15:20:36 -07001926 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
Neal Cardwell5ae344c2014-08-04 19:12:29 -04001927 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || is_reneg) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001928 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1929 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1930 tp->lost_out += tcp_skb_pcount(skb);
Ilpo Järvinen006f5822008-09-20 21:20:20 -07001931 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001932 }
1933 }
Ilpo Järvinen005903b2007-08-09 14:44:16 +03001934 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001935
Yuchung Cheng74c181d2013-08-12 16:41:25 -07001936 /* Timeout in disordered state after receiving substantial DUPACKs
1937 * suggests that the degree of reordering is over-estimated.
1938 */
1939 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
Nikolay Borisov1043e252016-02-03 09:46:52 +02001940 tp->sacked_out >= net->ipv4.sysctl_tcp_reordering)
Yuchung Cheng74c181d2013-08-12 16:41:25 -07001941 tp->reordering = min_t(unsigned int, tp->reordering,
Nikolay Borisov1043e252016-02-03 09:46:52 +02001942 net->ipv4.sysctl_tcp_reordering);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001943 tcp_set_ca_state(sk, TCP_CA_Loss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001944 tp->high_seq = tp->snd_nxt;
Florian Westphal735d3832014-09-29 13:08:30 +02001945 tcp_ecn_queue_cwr(tp);
Yuchung Chenge33099f2013-03-20 13:33:00 +00001946
1947 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1948 * loss recovery is underway except recurring timeout(s) on
1949 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1950 */
1951 tp->frto = sysctl_tcp_frto &&
1952 (new_recovery || icsk->icsk_retransmits) &&
1953 !inet_csk(sk)->icsk_mtup.probe_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001954}
1955
Ilpo Järvinencadbd032007-12-31 04:49:21 -08001956/* If ACK arrived pointing to a remembered SACK, it means that our
1957 * remembered SACKs do not reflect real state of receiver i.e.
1958 * receiver _host_ is heavily congested (or buggy).
1959 *
Neal Cardwell5ae344c2014-08-04 19:12:29 -04001960 * To avoid big spurious retransmission bursts due to transient SACK
1961 * scoreboard oddities that look like reneging, we give the receiver a
1962 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
1963 * restore sanity to the SACK scoreboard. If the apparent reneging
1964 * persists until this RTO then we'll clear the SACK scoreboard.
Ilpo Järvinencadbd032007-12-31 04:49:21 -08001965 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001966static bool tcp_check_sack_reneging(struct sock *sk, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001967{
Ilpo Järvinencadbd032007-12-31 04:49:21 -08001968 if (flag & FLAG_SACK_RENEGING) {
Neal Cardwell5ae344c2014-08-04 19:12:29 -04001969 struct tcp_sock *tp = tcp_sk(sk);
1970 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
1971 msecs_to_jiffies(10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001972
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001973 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
Neal Cardwell5ae344c2014-08-04 19:12:29 -04001974 delay, TCP_RTO_MAX);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001975 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001976 }
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001977 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001978}
1979
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001980static inline int tcp_fackets_out(const struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001981{
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001982 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001983}
1984
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08001985/* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
1986 * counter when SACK is enabled (without SACK, sacked_out is used for
1987 * that purpose).
1988 *
1989 * Instead, with FACK TCP uses fackets_out that includes both SACKed
1990 * segments up to the highest received SACK block so far and holes in
1991 * between them.
1992 *
1993 * With reordering, holes may still be in flight, so RFC3517 recovery
1994 * uses pure sacked_out (total number of SACKed segments) even though
1995 * it violates the RFC that uses duplicate ACKs, often these are equal
1996 * but when e.g. out-of-window ACKs or packet duplication occurs,
1997 * they differ. Since neither occurs due to loss, TCP should really
1998 * ignore them.
1999 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04002000static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002001{
2002 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2003}
2004
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00002005static bool tcp_pause_early_retransmit(struct sock *sk, int flag)
2006{
2007 struct tcp_sock *tp = tcp_sk(sk);
2008 unsigned long delay;
2009
2010 /* Delay early retransmit and entering fast recovery for
2011 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
2012 * available, or RTO is scheduled to fire first.
2013 */
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002014 if (sysctl_tcp_early_retrans < 2 || sysctl_tcp_early_retrans > 3 ||
Eric Dumazet740b0f12014-02-26 14:02:48 -08002015 (flag & FLAG_ECE) || !tp->srtt_us)
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00002016 return false;
2017
Eric Dumazet740b0f12014-02-26 14:02:48 -08002018 delay = max(usecs_to_jiffies(tp->srtt_us >> 5),
2019 msecs_to_jiffies(2));
2020
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00002021 if (!time_after(inet_csk(sk)->icsk_timeout, (jiffies + delay)))
2022 return false;
2023
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002024 inet_csk_reset_xmit_timer(sk, ICSK_TIME_EARLY_RETRANS, delay,
2025 TCP_RTO_MAX);
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00002026 return true;
2027}
2028
Linus Torvalds1da177e2005-04-16 15:20:36 -07002029/* Linux NewReno/SACK/FACK/ECN state machine.
2030 * --------------------------------------
2031 *
2032 * "Open" Normal state, no dubious events, fast path.
2033 * "Disorder" In all the respects it is "Open",
2034 * but requires a bit more attention. It is entered when
2035 * we see some SACKs or dupacks. It is split of "Open"
2036 * mainly to move some processing from fast path to slow one.
2037 * "CWR" CWND was reduced due to some Congestion Notification event.
2038 * It can be ECN, ICMP source quench, local device congestion.
2039 * "Recovery" CWND was reduced, we are fast-retransmitting.
2040 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2041 *
2042 * tcp_fastretrans_alert() is entered:
2043 * - each incoming ACK, if state is not "Open"
2044 * - when arrived ACK is unusual, namely:
2045 * * SACK
2046 * * Duplicate ACK.
2047 * * ECN ECE.
2048 *
2049 * Counting packets in flight is pretty simple.
2050 *
2051 * in_flight = packets_out - left_out + retrans_out
2052 *
2053 * packets_out is SND.NXT-SND.UNA counted in packets.
2054 *
2055 * retrans_out is number of retransmitted segments.
2056 *
2057 * left_out is number of segments left network, but not ACKed yet.
2058 *
2059 * left_out = sacked_out + lost_out
2060 *
2061 * sacked_out: Packets, which arrived to receiver out of order
2062 * and hence not ACKed. With SACKs this number is simply
2063 * amount of SACKed data. Even without SACKs
2064 * it is easy to give pretty reliable estimate of this number,
2065 * counting duplicate ACKs.
2066 *
2067 * lost_out: Packets lost by network. TCP has no explicit
2068 * "loss notification" feedback from network (for now).
2069 * It means that this number can be only _guessed_.
2070 * Actually, it is the heuristics to predict lossage that
2071 * distinguishes different algorithms.
2072 *
2073 * F.e. after RTO, when all the queue is considered as lost,
2074 * lost_out = packets_out and in_flight = retrans_out.
2075 *
2076 * Essentially, we have now two algorithms counting
2077 * lost packets.
2078 *
2079 * FACK: It is the simplest heuristics. As soon as we decided
2080 * that something is lost, we decide that _all_ not SACKed
2081 * packets until the most forward SACK are lost. I.e.
2082 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2083 * It is absolutely correct estimate, if network does not reorder
2084 * packets. And it loses any connection to reality when reordering
2085 * takes place. We use FACK by default until reordering
2086 * is suspected on the path to this destination.
2087 *
2088 * NewReno: when Recovery is entered, we assume that one segment
2089 * is lost (classic Reno). While we are in Recovery and
2090 * a partial ACK arrives, we assume that one more packet
2091 * is lost (NewReno). This heuristics are the same in NewReno
2092 * and SACK.
2093 *
2094 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2095 * deflation etc. CWND is real congestion window, never inflated, changes
2096 * only according to classic VJ rules.
2097 *
2098 * Really tricky (and requiring careful tuning) part of algorithm
2099 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2100 * The first determines the moment _when_ we should reduce CWND and,
2101 * hence, slow down forward transmission. In fact, it determines the moment
2102 * when we decide that hole is caused by loss, rather than by a reorder.
2103 *
2104 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2105 * holes, caused by lost packets.
2106 *
2107 * And the most logically complicated part of algorithm is undo
2108 * heuristics. We detect false retransmits due to both too early
2109 * fast retransmit (reordering) and underestimated RTO, analyzing
2110 * timestamps and D-SACKs. When we detect that some segments were
2111 * retransmitted by mistake and CWND reduction was wrong, we undo
2112 * window reduction and abort recovery phase. This logic is hidden
2113 * inside several functions named tcp_try_undo_<something>.
2114 */
2115
2116/* This function decides, when we should leave Disordered state
2117 * and enter Recovery phase, reducing congestion window.
2118 *
2119 * Main question: may we further continue forward transmission
2120 * with the same cwnd?
2121 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002122static bool tcp_time_to_recover(struct sock *sk, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002123{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002124 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002125 __u32 packets_out;
Nikolay Borisov1043e252016-02-03 09:46:52 +02002126 int tcp_reordering = sock_net(sk)->ipv4.sysctl_tcp_reordering;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002127
2128 /* Trick#1: The loss is proven. */
2129 if (tp->lost_out)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002130 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002131
2132 /* Not-A-Trick#2 : Classic rule... */
Andreas Petlundea84e552009-10-27 03:27:21 +00002133 if (tcp_dupack_heuristics(tp) > tp->reordering)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002134 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002135
Linus Torvalds1da177e2005-04-16 15:20:36 -07002136 /* Trick#4: It is still not OK... But will it be useful to delay
2137 * recovery more?
2138 */
2139 packets_out = tp->packets_out;
2140 if (packets_out <= tp->reordering &&
Nikolay Borisov1043e252016-02-03 09:46:52 +02002141 tp->sacked_out >= max_t(__u32, packets_out/2, tcp_reordering) &&
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002142 !tcp_may_send_now(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002143 /* We have nothing to send. This connection is limited
2144 * either by receiver window or by application.
2145 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002146 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002147 }
2148
Andreas Petlund7e380172010-02-18 04:48:19 +00002149 /* If a thin stream is detected, retransmit after first
2150 * received dupack. Employ only if SACK is supported in order
2151 * to avoid possible corner-case series of spurious retransmissions
2152 * Use only if there are no unsent data.
2153 */
2154 if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2155 tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2156 tcp_is_sack(tp) && !tcp_send_head(sk))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002157 return true;
Andreas Petlund7e380172010-02-18 04:48:19 +00002158
Yuchung Chengeed530b2012-05-02 13:30:03 +00002159 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious
2160 * retransmissions due to small network reorderings, we implement
2161 * Mitigation A.3 in the RFC and delay the retransmission for a short
2162 * interval if appropriate.
2163 */
2164 if (tp->do_early_retrans && !tp->retrans_out && tp->sacked_out &&
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002165 (tp->packets_out >= (tp->sacked_out + 1) && tp->packets_out < 4) &&
Yuchung Chengeed530b2012-05-02 13:30:03 +00002166 !tcp_may_send_now(sk))
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00002167 return !tcp_pause_early_retransmit(sk, flag);
Yuchung Chengeed530b2012-05-02 13:30:03 +00002168
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002169 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002170}
2171
Yuchung Cheng974c1232012-01-19 14:42:21 +00002172/* Detect loss in event "A" above by marking head of queue up as lost.
2173 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2174 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2175 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2176 * the maximum SACKed segments to pass before reaching this limit.
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002177 */
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002178static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002179{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002180 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002181 struct sk_buff *skb;
Neal Cardwelld88270e2016-01-25 14:01:53 -08002182 int cnt, oldcnt, lost;
Ilpo Järvinenc137f3d2008-04-07 22:32:38 -07002183 unsigned int mss;
Yuchung Cheng974c1232012-01-19 14:42:21 +00002184 /* Use SACK to deduce losses of new sequences sent during recovery */
2185 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002186
Ilpo Järvinen547b7922008-07-25 21:43:18 -07002187 WARN_ON(packets > tp->packets_out);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002188 if (tp->lost_skb_hint) {
2189 skb = tp->lost_skb_hint;
2190 cnt = tp->lost_cnt_hint;
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002191 /* Head already handled? */
2192 if (mark_head && skb != tcp_write_queue_head(sk))
2193 return;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002194 } else {
David S. Millerfe067e82007-03-07 12:12:44 -08002195 skb = tcp_write_queue_head(sk);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002196 cnt = 0;
2197 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002198
David S. Millerfe067e82007-03-07 12:12:44 -08002199 tcp_for_write_queue_from(skb, sk) {
2200 if (skb == tcp_send_head(sk))
2201 break;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002202 /* TODO: do this better */
2203 /* this is not the most efficient way to do this... */
2204 tp->lost_skb_hint = skb;
2205 tp->lost_cnt_hint = cnt;
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002206
Yuchung Cheng974c1232012-01-19 14:42:21 +00002207 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
Ilpo Järvinenc137f3d2008-04-07 22:32:38 -07002208 break;
2209
2210 oldcnt = cnt;
Ilpo Järvinenad1984e2008-01-30 20:06:02 -08002211 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002212 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2213 cnt += tcp_skb_pcount(skb);
2214
Ilpo Järvinenc137f3d2008-04-07 22:32:38 -07002215 if (cnt > packets) {
Yuchung Chengb3de7552010-09-24 13:22:06 +00002216 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
Neal Cardwellc0638c22012-03-02 21:36:51 +00002217 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
Yuchung Chengb3de7552010-09-24 13:22:06 +00002218 (oldcnt >= packets))
Ilpo Järvinenc137f3d2008-04-07 22:32:38 -07002219 break;
2220
Eric Dumazetf69ad292015-06-11 09:15:18 -07002221 mss = tcp_skb_mss(skb);
Neal Cardwelld88270e2016-01-25 14:01:53 -08002222 /* If needed, chop off the prefix to mark as lost. */
2223 lost = (packets - oldcnt) * mss;
2224 if (lost < skb->len &&
2225 tcp_fragment(sk, skb, lost, mss, GFP_ATOMIC) < 0)
Ilpo Järvinenc137f3d2008-04-07 22:32:38 -07002226 break;
2227 cnt = packets;
2228 }
2229
Ilpo Järvinen41ea36e2008-09-20 21:19:22 -07002230 tcp_skb_mark_lost(tp, skb);
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002231
2232 if (mark_head)
2233 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002234 }
Ilpo Järvinen005903b2007-08-09 14:44:16 +03002235 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002236}
2237
2238/* Account newly detected lost packet(s) */
2239
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002240static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002241{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002242 struct tcp_sock *tp = tcp_sk(sk);
2243
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002244 if (tcp_is_reno(tp)) {
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002245 tcp_mark_head_lost(sk, 1, 1);
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002246 } else if (tcp_is_fack(tp)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002247 int lost = tp->fackets_out - tp->reordering;
2248 if (lost <= 0)
2249 lost = 1;
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002250 tcp_mark_head_lost(sk, lost, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002251 } else {
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002252 int sacked_upto = tp->sacked_out - tp->reordering;
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002253 if (sacked_upto >= 0)
2254 tcp_mark_head_lost(sk, sacked_upto, 0);
2255 else if (fast_rexmit)
2256 tcp_mark_head_lost(sk, 1, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002257 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002258}
2259
Yuchung Cheng77c63122015-10-16 21:57:44 -07002260static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2261{
2262 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2263 before(tp->rx_opt.rcv_tsecr, when);
2264}
2265
Yuchung Cheng659a8ad2015-10-16 21:57:46 -07002266/* skb is spurious retransmitted if the returned timestamp echo
2267 * reply is prior to the skb transmission time
2268 */
2269static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2270 const struct sk_buff *skb)
2271{
2272 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2273 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb));
2274}
2275
Linus Torvalds1da177e2005-04-16 15:20:36 -07002276/* Nothing was retransmitted or returned timestamp is less
2277 * than timestamp of the first retransmission.
2278 */
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00002279static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002280{
2281 return !tp->retrans_stamp ||
Yuchung Cheng77c63122015-10-16 21:57:44 -07002282 tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002283}
2284
2285/* Undo procedures. */
2286
Marcelo Leitner1f37bf82014-11-04 17:15:08 -02002287/* We can clear retrans_stamp when there are no retransmissions in the
2288 * window. It would seem that it is trivially available for us in
2289 * tp->retrans_out, however, that kind of assumptions doesn't consider
2290 * what will happen if errors occur when sending retransmission for the
2291 * second time. ...It could the that such segment has only
2292 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2293 * the head skb is enough except for some reneging corner cases that
2294 * are not worth the effort.
2295 *
2296 * Main reason for all this complexity is the fact that connection dying
2297 * time now depends on the validity of the retrans_stamp, in particular,
2298 * that successive retransmissions of a segment must not advance
2299 * retrans_stamp under any conditions.
2300 */
2301static bool tcp_any_retrans_done(const struct sock *sk)
2302{
2303 const struct tcp_sock *tp = tcp_sk(sk);
2304 struct sk_buff *skb;
2305
2306 if (tp->retrans_out)
2307 return true;
2308
2309 skb = tcp_write_queue_head(sk);
2310 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2311 return true;
2312
2313 return false;
2314}
2315
Linus Torvalds1da177e2005-04-16 15:20:36 -07002316#if FASTRETRANS_DEBUG > 1
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002317static void DBGUNDO(struct sock *sk, const char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002318{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002319 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002320 struct inet_sock *inet = inet_sk(sk);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002321
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002322 if (sk->sk_family == AF_INET) {
Joe Perches91df42b2012-05-15 14:11:54 +00002323 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2324 msg,
2325 &inet->inet_daddr, ntohs(inet->inet_dport),
2326 tp->snd_cwnd, tcp_left_out(tp),
2327 tp->snd_ssthresh, tp->prior_ssthresh,
2328 tp->packets_out);
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002329 }
Eric Dumazetdfd56b82011-12-10 09:48:31 +00002330#if IS_ENABLED(CONFIG_IPV6)
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002331 else if (sk->sk_family == AF_INET6) {
2332 struct ipv6_pinfo *np = inet6_sk(sk);
Joe Perches91df42b2012-05-15 14:11:54 +00002333 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2334 msg,
2335 &np->daddr, ntohs(inet->inet_dport),
2336 tp->snd_cwnd, tcp_left_out(tp),
2337 tp->snd_ssthresh, tp->prior_ssthresh,
2338 tp->packets_out);
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002339 }
2340#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002341}
2342#else
2343#define DBGUNDO(x...) do { } while (0)
2344#endif
2345
Yuchung Cheng7026b912013-05-29 14:20:13 +00002346static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002347{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002348 struct tcp_sock *tp = tcp_sk(sk);
2349
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002350 if (unmark_loss) {
2351 struct sk_buff *skb;
2352
2353 tcp_for_write_queue(skb, sk) {
2354 if (skb == tcp_send_head(sk))
2355 break;
2356 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2357 }
2358 tp->lost_out = 0;
2359 tcp_clear_all_retrans_hints(tp);
2360 }
2361
Linus Torvalds1da177e2005-04-16 15:20:36 -07002362 if (tp->prior_ssthresh) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002363 const struct inet_connection_sock *icsk = inet_csk(sk);
2364
2365 if (icsk->icsk_ca_ops->undo_cwnd)
2366 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002367 else
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002368 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002369
Yuchung Cheng7026b912013-05-29 14:20:13 +00002370 if (tp->prior_ssthresh > tp->snd_ssthresh) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002371 tp->snd_ssthresh = tp->prior_ssthresh;
Florian Westphal735d3832014-09-29 13:08:30 +02002372 tcp_ecn_withdraw_cwr(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002373 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002374 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002375 tp->snd_cwnd_stamp = tcp_time_stamp;
Yuchung Cheng7026b912013-05-29 14:20:13 +00002376 tp->undo_marker = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002377}
2378
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00002379static inline bool tcp_may_undo(const struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002380{
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002381 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002382}
2383
2384/* People celebrate: "We love our President!" */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002385static bool tcp_try_undo_recovery(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002386{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002387 struct tcp_sock *tp = tcp_sk(sk);
2388
Linus Torvalds1da177e2005-04-16 15:20:36 -07002389 if (tcp_may_undo(tp)) {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07002390 int mib_idx;
2391
Linus Torvalds1da177e2005-04-16 15:20:36 -07002392 /* Happy end! We did not retransmit anything
2393 * or our original transmission succeeded.
2394 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002395 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
Yuchung Cheng7026b912013-05-29 14:20:13 +00002396 tcp_undo_cwnd_reduction(sk, false);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002397 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07002398 mib_idx = LINUX_MIB_TCPLOSSUNDO;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002399 else
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07002400 mib_idx = LINUX_MIB_TCPFULLUNDO;
2401
Eric Dumazetc10d9312016-04-29 14:16:47 -07002402 NET_INC_STATS(sock_net(sk), mib_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002403 }
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002404 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002405 /* Hold old state until something *above* high_seq
2406 * is ACKed. For Reno it is MUST to prevent false
2407 * fast retransmits (RFC2582). SACK TCP is safe. */
Marcelo Leitner1f37bf82014-11-04 17:15:08 -02002408 if (!tcp_any_retrans_done(sk))
2409 tp->retrans_stamp = 0;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002410 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002411 }
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002412 tcp_set_ca_state(sk, TCP_CA_Open);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002413 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002414}
2415
2416/* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
Yuchung Chengc7d9d6a2013-05-29 14:20:14 +00002417static bool tcp_try_undo_dsack(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002418{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002419 struct tcp_sock *tp = tcp_sk(sk);
2420
Linus Torvalds1da177e2005-04-16 15:20:36 -07002421 if (tp->undo_marker && !tp->undo_retrans) {
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002422 DBGUNDO(sk, "D-SACK");
Yuchung Cheng7026b912013-05-29 14:20:13 +00002423 tcp_undo_cwnd_reduction(sk, false);
Eric Dumazetc10d9312016-04-29 14:16:47 -07002424 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
Yuchung Chengc7d9d6a2013-05-29 14:20:14 +00002425 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002426 }
Yuchung Chengc7d9d6a2013-05-29 14:20:14 +00002427 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002428}
2429
Yuchung Chenge33099f2013-03-20 13:33:00 +00002430/* Undo during loss recovery after partial ACK or using F-RTO. */
2431static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002432{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002433 struct tcp_sock *tp = tcp_sk(sk);
2434
Yuchung Chenge33099f2013-03-20 13:33:00 +00002435 if (frto_undo || tcp_may_undo(tp)) {
Yuchung Cheng7026b912013-05-29 14:20:13 +00002436 tcp_undo_cwnd_reduction(sk, true);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002437
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002438 DBGUNDO(sk, "partial loss");
Eric Dumazetc10d9312016-04-29 14:16:47 -07002439 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
Yuchung Chenge33099f2013-03-20 13:33:00 +00002440 if (frto_undo)
Eric Dumazetc10d9312016-04-29 14:16:47 -07002441 NET_INC_STATS(sock_net(sk),
Eric Dumazet02a1d6e2016-04-27 16:44:39 -07002442 LINUX_MIB_TCPSPURIOUSRTOS);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002443 inet_csk(sk)->icsk_retransmits = 0;
Yuchung Chenge33099f2013-03-20 13:33:00 +00002444 if (frto_undo || tcp_is_sack(tp))
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002445 tcp_set_ca_state(sk, TCP_CA_Open);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002446 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002447 }
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002448 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002449}
2450
Yuchung Cheng37598242015-07-01 14:11:15 -07002451/* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002452 * It computes the number of packets to send (sndcnt) based on packets newly
2453 * delivered:
2454 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2455 * cwnd reductions across a full RTT.
Yuchung Cheng37598242015-07-01 14:11:15 -07002456 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2457 * But when the retransmits are acked without further losses, PRR
2458 * slow starts cwnd up to ssthresh to speed up the recovery.
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002459 */
Christoph Paasch5ee2c942014-07-14 16:58:32 +02002460static void tcp_init_cwnd_reduction(struct sock *sk)
Yuchung Cheng684bad12012-09-02 17:38:04 +00002461{
2462 struct tcp_sock *tp = tcp_sk(sk);
2463
2464 tp->high_seq = tp->snd_nxt;
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00002465 tp->tlp_high_seq = 0;
Yuchung Cheng684bad12012-09-02 17:38:04 +00002466 tp->snd_cwnd_cnt = 0;
2467 tp->prior_cwnd = tp->snd_cwnd;
2468 tp->prr_delivered = 0;
2469 tp->prr_out = 0;
Christoph Paasch5ee2c942014-07-14 16:58:32 +02002470 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
Florian Westphal735d3832014-09-29 13:08:30 +02002471 tcp_ecn_queue_cwr(tp);
Yuchung Cheng684bad12012-09-02 17:38:04 +00002472}
2473
Yuchung Cheng31ba0c12016-02-02 10:33:05 -08002474static void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked,
2475 int flag)
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002476{
2477 struct tcp_sock *tp = tcp_sk(sk);
2478 int sndcnt = 0;
2479 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2480
Yuchung Cheng8b8a3212016-01-06 12:42:38 -08002481 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2482 return;
2483
Yuchung Cheng684bad12012-09-02 17:38:04 +00002484 tp->prr_delivered += newly_acked_sacked;
Yuchung Cheng37598242015-07-01 14:11:15 -07002485 if (delta < 0) {
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002486 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2487 tp->prior_cwnd - 1;
2488 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
Yuchung Cheng37598242015-07-01 14:11:15 -07002489 } else if ((flag & FLAG_RETRANS_DATA_ACKED) &&
2490 !(flag & FLAG_LOST_RETRANS)) {
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002491 sndcnt = min_t(int, delta,
2492 max_t(int, tp->prr_delivered - tp->prr_out,
2493 newly_acked_sacked) + 1);
Yuchung Cheng37598242015-07-01 14:11:15 -07002494 } else {
2495 sndcnt = min(delta, newly_acked_sacked);
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002496 }
Yuchung Cheng31ba0c12016-02-02 10:33:05 -08002497 /* Force a fast retransmit upon entering fast recovery */
2498 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002499 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2500}
2501
Yuchung Cheng684bad12012-09-02 17:38:04 +00002502static inline void tcp_end_cwnd_reduction(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002503{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002504 struct tcp_sock *tp = tcp_sk(sk);
Nandita Dukkipatia262f0c2011-08-21 20:21:57 +00002505
Yuchung Cheng684bad12012-09-02 17:38:04 +00002506 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2507 if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR ||
2508 (tp->undo_marker && tp->snd_ssthresh < TCP_INFINITE_SSTHRESH)) {
2509 tp->snd_cwnd = tp->snd_ssthresh;
2510 tp->snd_cwnd_stamp = tcp_time_stamp;
Yuchung Cheng67d41202011-03-14 10:57:03 +00002511 }
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002512 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002513}
2514
Yuchung Cheng684bad12012-09-02 17:38:04 +00002515/* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
Christoph Paasch5ee2c942014-07-14 16:58:32 +02002516void tcp_enter_cwr(struct sock *sk)
Yuchung Cheng09484d12012-09-02 17:38:02 +00002517{
2518 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng09484d12012-09-02 17:38:02 +00002519
2520 tp->prior_ssthresh = 0;
Yuchung Cheng684bad12012-09-02 17:38:04 +00002521 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
Yuchung Cheng09484d12012-09-02 17:38:02 +00002522 tp->undo_marker = 0;
Christoph Paasch5ee2c942014-07-14 16:58:32 +02002523 tcp_init_cwnd_reduction(sk);
Yuchung Cheng09484d12012-09-02 17:38:02 +00002524 tcp_set_ca_state(sk, TCP_CA_CWR);
2525 }
2526}
Kenneth Klette Jonassen7782ad82015-06-10 19:08:16 +02002527EXPORT_SYMBOL(tcp_enter_cwr);
Yuchung Cheng09484d12012-09-02 17:38:02 +00002528
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07002529static void tcp_try_keep_open(struct sock *sk)
2530{
2531 struct tcp_sock *tp = tcp_sk(sk);
2532 int state = TCP_CA_Open;
2533
Neal Cardwellf6982042011-11-16 08:58:04 +00002534 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07002535 state = TCP_CA_Disorder;
2536
2537 if (inet_csk(sk)->icsk_ca_state != state) {
2538 tcp_set_ca_state(sk, state);
2539 tp->high_seq = tp->snd_nxt;
2540 }
2541}
2542
Yuchung Cheng31ba0c12016-02-02 10:33:05 -08002543static void tcp_try_to_open(struct sock *sk, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002544{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002545 struct tcp_sock *tp = tcp_sk(sk);
2546
Ilpo Järvinen86426c22007-08-09 14:45:17 +03002547 tcp_verify_left_out(tp);
2548
Yuchung Cheng9b441902013-03-20 13:32:58 +00002549 if (!tcp_any_retrans_done(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002550 tp->retrans_stamp = 0;
2551
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002552 if (flag & FLAG_ECE)
Christoph Paasch5ee2c942014-07-14 16:58:32 +02002553 tcp_enter_cwr(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002554
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002555 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07002556 tcp_try_keep_open(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002557 }
2558}
2559
John Heffner5d424d52006-03-20 17:53:41 -08002560static void tcp_mtup_probe_failed(struct sock *sk)
2561{
2562 struct inet_connection_sock *icsk = inet_csk(sk);
2563
2564 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2565 icsk->icsk_mtup.probe_size = 0;
Eric Dumazetc10d9312016-04-29 14:16:47 -07002566 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
John Heffner5d424d52006-03-20 17:53:41 -08002567}
2568
Ilpo Järvinen72211e92009-03-14 14:23:04 +00002569static void tcp_mtup_probe_success(struct sock *sk)
John Heffner5d424d52006-03-20 17:53:41 -08002570{
2571 struct tcp_sock *tp = tcp_sk(sk);
2572 struct inet_connection_sock *icsk = inet_csk(sk);
2573
2574 /* FIXME: breaks with very large cwnd */
2575 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2576 tp->snd_cwnd = tp->snd_cwnd *
2577 tcp_mss_to_mtu(sk, tp->mss_cache) /
2578 icsk->icsk_mtup.probe_size;
2579 tp->snd_cwnd_cnt = 0;
2580 tp->snd_cwnd_stamp = tcp_time_stamp;
John Heffner9c6d5e52010-10-06 21:18:02 -07002581 tp->snd_ssthresh = tcp_current_ssthresh(sk);
John Heffner5d424d52006-03-20 17:53:41 -08002582
2583 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2584 icsk->icsk_mtup.probe_size = 0;
2585 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
Eric Dumazetc10d9312016-04-29 14:16:47 -07002586 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
John Heffner5d424d52006-03-20 17:53:41 -08002587}
2588
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08002589/* Do a simple retransmit without using the backoff mechanisms in
2590 * tcp_timer. This is used for path mtu discovery.
2591 * The socket is already locked here.
2592 */
2593void tcp_simple_retransmit(struct sock *sk)
2594{
2595 const struct inet_connection_sock *icsk = inet_csk(sk);
2596 struct tcp_sock *tp = tcp_sk(sk);
2597 struct sk_buff *skb;
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00002598 unsigned int mss = tcp_current_mss(sk);
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08002599 u32 prior_lost = tp->lost_out;
2600
2601 tcp_for_write_queue(skb, sk) {
2602 if (skb == tcp_send_head(sk))
2603 break;
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08002604 if (tcp_skb_seglen(skb) > mss &&
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08002605 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2606 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2607 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2608 tp->retrans_out -= tcp_skb_pcount(skb);
2609 }
2610 tcp_skb_mark_lost_uncond_verify(tp, skb);
2611 }
2612 }
2613
2614 tcp_clear_retrans_hints_partial(tp);
2615
2616 if (prior_lost == tp->lost_out)
2617 return;
2618
2619 if (tcp_is_reno(tp))
2620 tcp_limit_reno_sacked(tp);
2621
2622 tcp_verify_left_out(tp);
2623
2624 /* Don't muck with the congestion window here.
2625 * Reason is that we do not increase amount of _data_
2626 * in network, but units changed and effective
2627 * cwnd/ssthresh really reduced now.
2628 */
2629 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2630 tp->high_seq = tp->snd_nxt;
2631 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2632 tp->prior_ssthresh = 0;
2633 tp->undo_marker = 0;
2634 tcp_set_ca_state(sk, TCP_CA_Loss);
2635 }
2636 tcp_xmit_retransmit_queue(sk);
2637}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00002638EXPORT_SYMBOL(tcp_simple_retransmit);
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08002639
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002640static void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2641{
2642 struct tcp_sock *tp = tcp_sk(sk);
2643 int mib_idx;
2644
2645 if (tcp_is_reno(tp))
2646 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2647 else
2648 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2649
Eric Dumazetc10d9312016-04-29 14:16:47 -07002650 NET_INC_STATS(sock_net(sk), mib_idx);
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002651
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002652 tp->prior_ssthresh = 0;
Yuchung Cheng989e04c2014-08-22 14:15:22 -07002653 tcp_init_undo(tp);
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002654
Yuchung Cheng291a00d2015-07-01 14:11:14 -07002655 if (!tcp_in_cwnd_reduction(sk)) {
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002656 if (!ece_ack)
2657 tp->prior_ssthresh = tcp_current_ssthresh(sk);
Christoph Paasch5ee2c942014-07-14 16:58:32 +02002658 tcp_init_cwnd_reduction(sk);
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002659 }
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002660 tcp_set_ca_state(sk, TCP_CA_Recovery);
2661}
2662
Yuchung Chengab42d9e2013-03-20 13:32:59 +00002663/* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2664 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2665 */
Yuchung Chenge662ca42016-02-02 10:33:04 -08002666static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack,
2667 int *rexmit)
Yuchung Chengab42d9e2013-03-20 13:32:59 +00002668{
Yuchung Chengab42d9e2013-03-20 13:32:59 +00002669 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Chenge33099f2013-03-20 13:33:00 +00002670 bool recovered = !before(tp->snd_una, tp->high_seq);
Yuchung Chengab42d9e2013-03-20 13:32:59 +00002671
Yuchung Chengda34ac72015-05-18 12:31:44 -07002672 if ((flag & FLAG_SND_UNA_ADVANCED) &&
2673 tcp_try_undo_loss(sk, false))
2674 return;
2675
Yuchung Chenge33099f2013-03-20 13:33:00 +00002676 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
Yuchung Cheng0cfa5c02014-05-30 15:25:59 -07002677 /* Step 3.b. A timeout is spurious if not all data are
2678 * lost, i.e., never-retransmitted data are (s)acked.
2679 */
Yuchung Chengda34ac72015-05-18 12:31:44 -07002680 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2681 tcp_try_undo_loss(sk, true))
Yuchung Chenge33099f2013-03-20 13:33:00 +00002682 return;
Yuchung Cheng0cfa5c02014-05-30 15:25:59 -07002683
Yuchung Chengb7b0ed92015-05-18 12:31:45 -07002684 if (after(tp->snd_nxt, tp->high_seq)) {
2685 if (flag & FLAG_DATA_SACKED || is_dupack)
2686 tp->frto = 0; /* Step 3.a. loss was real */
Yuchung Chenge33099f2013-03-20 13:33:00 +00002687 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2688 tp->high_seq = tp->snd_nxt;
Yuchung Chenge662ca42016-02-02 10:33:04 -08002689 /* Step 2.b. Try send new data (but deferred until cwnd
2690 * is updated in tcp_ack()). Otherwise fall back to
2691 * the conventional recovery.
2692 */
2693 if (tcp_send_head(sk) &&
2694 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2695 *rexmit = REXMIT_NEW;
2696 return;
2697 }
Yuchung Chenge33099f2013-03-20 13:33:00 +00002698 tp->frto = 0;
2699 }
2700 }
2701
2702 if (recovered) {
2703 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
Yuchung Chengab42d9e2013-03-20 13:32:59 +00002704 tcp_try_undo_recovery(sk);
2705 return;
2706 }
Yuchung Chenge33099f2013-03-20 13:33:00 +00002707 if (tcp_is_reno(tp)) {
2708 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2709 * delivered. Lower inflight to clock out (re)tranmissions.
2710 */
2711 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2712 tcp_add_reno_sack(sk);
2713 else if (flag & FLAG_SND_UNA_ADVANCED)
2714 tcp_reset_reno_sack(tp);
2715 }
Yuchung Chenge662ca42016-02-02 10:33:04 -08002716 *rexmit = REXMIT_LOST;
Yuchung Chengab42d9e2013-03-20 13:32:59 +00002717}
2718
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002719/* Undo during fast recovery after partial ACK. */
Yuchung Cheng31ba0c12016-02-02 10:33:05 -08002720static bool tcp_try_undo_partial(struct sock *sk, const int acked)
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002721{
2722 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002723
Yuchung Cheng7026b912013-05-29 14:20:13 +00002724 if (tp->undo_marker && tcp_packet_delayed(tp)) {
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002725 /* Plain luck! Hole if filled with delayed
2726 * packet, rather than with a retransmit.
2727 */
Yuchung Cheng7026b912013-05-29 14:20:13 +00002728 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2729
2730 /* We are getting evidence that the reordering degree is higher
2731 * than we realized. If there are no retransmits out then we
2732 * can undo. Otherwise we clock out new packets but do not
2733 * mark more packets lost or retransmit more.
2734 */
Yuchung Cheng31ba0c12016-02-02 10:33:05 -08002735 if (tp->retrans_out)
Yuchung Cheng7026b912013-05-29 14:20:13 +00002736 return true;
Yuchung Cheng7026b912013-05-29 14:20:13 +00002737
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002738 if (!tcp_any_retrans_done(sk))
2739 tp->retrans_stamp = 0;
2740
Yuchung Cheng7026b912013-05-29 14:20:13 +00002741 DBGUNDO(sk, "partial recovery");
2742 tcp_undo_cwnd_reduction(sk, true);
Eric Dumazetc10d9312016-04-29 14:16:47 -07002743 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
Yuchung Cheng7026b912013-05-29 14:20:13 +00002744 tcp_try_keep_open(sk);
2745 return true;
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002746 }
Yuchung Cheng7026b912013-05-29 14:20:13 +00002747 return false;
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002748}
2749
Linus Torvalds1da177e2005-04-16 15:20:36 -07002750/* Process an event, which can update packets-in-flight not trivially.
2751 * Main goal of this function is to calculate new estimate for left_out,
2752 * taking into account both packets sitting in receiver's buffer and
2753 * packets lost by network.
2754 *
Yuchung Cheng31ba0c12016-02-02 10:33:05 -08002755 * Besides that it updates the congestion state when packet loss or ECN
2756 * is detected. But it does not reduce the cwnd, it is done by the
2757 * congestion control later.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002758 *
2759 * It does _not_ decide what to send, it is made in function
2760 * tcp_xmit_retransmit_queue().
2761 */
Yuchung Cheng68049732013-05-29 14:20:11 +00002762static void tcp_fastretrans_alert(struct sock *sk, const int acked,
Yuchung Cheng31ba0c12016-02-02 10:33:05 -08002763 bool is_dupack, int *ack_flag, int *rexmit)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002764{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002765 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002766 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng31ba0c12016-02-02 10:33:05 -08002767 int fast_rexmit = 0, flag = *ack_flag;
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002768 bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002769 (tcp_fackets_out(tp) > tp->reordering));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002770
Ilpo Järvinen3ccd3132007-12-31 04:43:32 -08002771 if (WARN_ON(!tp->packets_out && tp->sacked_out))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002772 tp->sacked_out = 0;
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07002773 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002774 tp->fackets_out = 0;
2775
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002776 /* Now state machine starts.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002777 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002778 if (flag & FLAG_ECE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002779 tp->prior_ssthresh = 0;
2780
2781 /* B. In all the states check for reneging SACKs. */
Ilpo Järvinencadbd032007-12-31 04:49:21 -08002782 if (tcp_check_sack_reneging(sk, flag))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002783 return;
2784
Yuchung Cheng974c1232012-01-19 14:42:21 +00002785 /* C. Check consistency of the current state. */
Ilpo Järvinen005903b2007-08-09 14:44:16 +03002786 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002787
Yuchung Cheng974c1232012-01-19 14:42:21 +00002788 /* D. Check state exit conditions. State can be terminated
Linus Torvalds1da177e2005-04-16 15:20:36 -07002789 * when high_seq is ACKed. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002790 if (icsk->icsk_ca_state == TCP_CA_Open) {
Ilpo Järvinen547b7922008-07-25 21:43:18 -07002791 WARN_ON(tp->retrans_out != 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002792 tp->retrans_stamp = 0;
2793 } else if (!before(tp->snd_una, tp->high_seq)) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002794 switch (icsk->icsk_ca_state) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002795 case TCP_CA_CWR:
2796 /* CWR is to be held something *above* high_seq
2797 * is ACKed for CWR bit to reach receiver. */
2798 if (tp->snd_una != tp->high_seq) {
Yuchung Cheng684bad12012-09-02 17:38:04 +00002799 tcp_end_cwnd_reduction(sk);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002800 tcp_set_ca_state(sk, TCP_CA_Open);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002801 }
2802 break;
2803
Linus Torvalds1da177e2005-04-16 15:20:36 -07002804 case TCP_CA_Recovery:
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002805 if (tcp_is_reno(tp))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002806 tcp_reset_reno_sack(tp);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002807 if (tcp_try_undo_recovery(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002808 return;
Yuchung Cheng684bad12012-09-02 17:38:04 +00002809 tcp_end_cwnd_reduction(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002810 break;
2811 }
2812 }
2813
Yuchung Cheng4f41b1c2015-10-16 21:57:47 -07002814 /* Use RACK to detect loss */
2815 if (sysctl_tcp_recovery & TCP_RACK_LOST_RETRANS &&
Yuchung Cheng31ba0c12016-02-02 10:33:05 -08002816 tcp_rack_mark_lost(sk)) {
Yuchung Cheng4f41b1c2015-10-16 21:57:47 -07002817 flag |= FLAG_LOST_RETRANS;
Yuchung Cheng31ba0c12016-02-02 10:33:05 -08002818 *ack_flag |= FLAG_LOST_RETRANS;
2819 }
Yuchung Cheng4f41b1c2015-10-16 21:57:47 -07002820
Yuchung Cheng974c1232012-01-19 14:42:21 +00002821 /* E. Process state. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002822 switch (icsk->icsk_ca_state) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002823 case TCP_CA_Recovery:
Ilpo Järvinen2e605292007-08-02 19:46:58 -07002824 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002825 if (tcp_is_reno(tp) && is_dupack)
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002826 tcp_add_reno_sack(sk);
Yuchung Cheng7026b912013-05-29 14:20:13 +00002827 } else {
Yuchung Cheng31ba0c12016-02-02 10:33:05 -08002828 if (tcp_try_undo_partial(sk, acked))
Yuchung Cheng7026b912013-05-29 14:20:13 +00002829 return;
2830 /* Partial ACK arrived. Force fast retransmit. */
2831 do_lost = tcp_is_reno(tp) ||
2832 tcp_fackets_out(tp) > tp->reordering;
2833 }
Yuchung Chengc7d9d6a2013-05-29 14:20:14 +00002834 if (tcp_try_undo_dsack(sk)) {
2835 tcp_try_keep_open(sk);
2836 return;
2837 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002838 break;
2839 case TCP_CA_Loss:
Yuchung Chenge662ca42016-02-02 10:33:04 -08002840 tcp_process_loss(sk, flag, is_dupack, rexmit);
Yuchung Cheng291a00d2015-07-01 14:11:14 -07002841 if (icsk->icsk_ca_state != TCP_CA_Open &&
2842 !(flag & FLAG_LOST_RETRANS))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002843 return;
Yuchung Cheng291a00d2015-07-01 14:11:14 -07002844 /* Change state if cwnd is undone or retransmits are lost */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002845 default:
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002846 if (tcp_is_reno(tp)) {
Ilpo Järvinen2e605292007-08-02 19:46:58 -07002847 if (flag & FLAG_SND_UNA_ADVANCED)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002848 tcp_reset_reno_sack(tp);
2849 if (is_dupack)
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002850 tcp_add_reno_sack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002851 }
2852
Neal Cardwellf6982042011-11-16 08:58:04 +00002853 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002854 tcp_try_undo_dsack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002855
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00002856 if (!tcp_time_to_recover(sk, flag)) {
Yuchung Cheng31ba0c12016-02-02 10:33:05 -08002857 tcp_try_to_open(sk, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002858 return;
2859 }
2860
John Heffner5d424d52006-03-20 17:53:41 -08002861 /* MTU probe failure: don't reduce cwnd */
2862 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2863 icsk->icsk_mtup.probe_size &&
John Heffner0e7b1362006-03-20 21:32:58 -08002864 tp->snd_una == tp->mtu_probe.probe_seq_start) {
John Heffner5d424d52006-03-20 17:53:41 -08002865 tcp_mtup_probe_failed(sk);
2866 /* Restores the reduction we did in tcp_mtup_probe() */
2867 tp->snd_cwnd++;
2868 tcp_simple_retransmit(sk);
2869 return;
2870 }
2871
Linus Torvalds1da177e2005-04-16 15:20:36 -07002872 /* Otherwise enter Recovery state */
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002873 tcp_enter_recovery(sk, (flag & FLAG_ECE));
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002874 fast_rexmit = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002875 }
2876
Yuchung Cheng3e59cb02013-05-17 13:45:05 +00002877 if (do_lost)
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002878 tcp_update_scoreboard(sk, fast_rexmit);
Yuchung Chenge662ca42016-02-02 10:33:04 -08002879 *rexmit = REXMIT_LOST;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002880}
2881
Yuchung Chengf6722582015-10-16 21:57:42 -07002882/* Kathleen Nichols' algorithm for tracking the minimum value of
2883 * a data stream over some fixed time interval. (E.g., the minimum
2884 * RTT over the past five minutes.) It uses constant space and constant
2885 * time per update yet almost always delivers the same minimum as an
2886 * implementation that has to keep all the data in the window.
2887 *
2888 * The algorithm keeps track of the best, 2nd best & 3rd best min
2889 * values, maintaining an invariant that the measurement time of the
2890 * n'th best >= n-1'th best. It also makes sure that the three values
2891 * are widely separated in the time window since that bounds the worse
2892 * case error when that data is monotonically increasing over the window.
2893 *
2894 * Upon getting a new min, we can forget everything earlier because it
2895 * has no value - the new min is <= everything else in the window by
2896 * definition and it's the most recent. So we restart fresh on every new min
2897 * and overwrites 2nd & 3rd choices. The same property holds for 2nd & 3rd
2898 * best.
2899 */
2900static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us)
2901{
2902 const u32 now = tcp_time_stamp, wlen = sysctl_tcp_min_rtt_wlen * HZ;
2903 struct rtt_meas *m = tcp_sk(sk)->rtt_min;
Eric Dumazet37202282016-02-11 08:58:18 -08002904 struct rtt_meas rttm = {
2905 .rtt = likely(rtt_us) ? rtt_us : jiffies_to_usecs(1),
2906 .ts = now,
2907 };
Yuchung Chengf6722582015-10-16 21:57:42 -07002908 u32 elapsed;
2909
2910 /* Check if the new measurement updates the 1st, 2nd, or 3rd choices */
2911 if (unlikely(rttm.rtt <= m[0].rtt))
2912 m[0] = m[1] = m[2] = rttm;
2913 else if (rttm.rtt <= m[1].rtt)
2914 m[1] = m[2] = rttm;
2915 else if (rttm.rtt <= m[2].rtt)
2916 m[2] = rttm;
2917
2918 elapsed = now - m[0].ts;
2919 if (unlikely(elapsed > wlen)) {
2920 /* Passed entire window without a new min so make 2nd choice
2921 * the new min & 3rd choice the new 2nd. So forth and so on.
2922 */
2923 m[0] = m[1];
2924 m[1] = m[2];
2925 m[2] = rttm;
2926 if (now - m[0].ts > wlen) {
2927 m[0] = m[1];
2928 m[1] = rttm;
2929 if (now - m[0].ts > wlen)
2930 m[0] = rttm;
2931 }
2932 } else if (m[1].ts == m[0].ts && elapsed > wlen / 4) {
2933 /* Passed a quarter of the window without a new min so
2934 * take 2nd choice from the 2nd quarter of the window.
2935 */
2936 m[2] = m[1] = rttm;
2937 } else if (m[2].ts == m[1].ts && elapsed > wlen / 2) {
2938 /* Passed half the window without a new min so take the 3rd
2939 * choice from the last half of the window.
2940 */
2941 m[2] = rttm;
2942 }
2943}
2944
Yuchung Chenged084952013-07-22 16:20:48 -07002945static inline bool tcp_ack_update_rtt(struct sock *sk, const int flag,
Yuchung Chengf6722582015-10-16 21:57:42 -07002946 long seq_rtt_us, long sack_rtt_us,
2947 long ca_rtt_us)
Ilpo Järvinen41834b72008-12-05 22:43:26 -08002948{
Yuchung Cheng5b08e472013-07-22 16:20:46 -07002949 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen41834b72008-12-05 22:43:26 -08002950
Yuchung Cheng5b08e472013-07-22 16:20:46 -07002951 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2952 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2953 * Karn's algorithm forbids taking RTT if some retransmitted data
2954 * is acked (RFC6298).
2955 */
Eric Dumazet740b0f12014-02-26 14:02:48 -08002956 if (seq_rtt_us < 0)
2957 seq_rtt_us = sack_rtt_us;
Yuchung Chenged084952013-07-22 16:20:48 -07002958
Linus Torvalds1da177e2005-04-16 15:20:36 -07002959 /* RTTM Rule: A TSecr value received in a segment is used to
2960 * update the averaged RTT measurement only if the segment
2961 * acknowledges some new data, i.e., only if it advances the
2962 * left edge of the send window.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002963 * See draft-ietf-tcplw-high-performance-00, section 3.3.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002964 */
Eric Dumazet740b0f12014-02-26 14:02:48 -08002965 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
Yuchung Cheng2909d872013-10-24 08:55:25 -07002966 flag & FLAG_ACKED)
Yuchung Chengf6722582015-10-16 21:57:42 -07002967 seq_rtt_us = ca_rtt_us = jiffies_to_usecs(tcp_time_stamp -
2968 tp->rx_opt.rcv_tsecr);
Eric Dumazet740b0f12014-02-26 14:02:48 -08002969 if (seq_rtt_us < 0)
Yuchung Chenged084952013-07-22 16:20:48 -07002970 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002971
Yuchung Chengf6722582015-10-16 21:57:42 -07002972 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2973 * always taken together with ACK, SACK, or TS-opts. Any negative
2974 * values will be skipped with the seq_rtt_us < 0 check above.
2975 */
2976 tcp_update_rtt_min(sk, ca_rtt_us);
Eric Dumazet740b0f12014-02-26 14:02:48 -08002977 tcp_rtt_estimator(sk, seq_rtt_us);
Yuchung Cheng5b08e472013-07-22 16:20:46 -07002978 tcp_set_rto(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002979
Yuchung Cheng5b08e472013-07-22 16:20:46 -07002980 /* RFC6298: only reset backoff on valid RTT measurement. */
2981 inet_csk(sk)->icsk_backoff = 0;
Yuchung Chenged084952013-07-22 16:20:48 -07002982 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002983}
2984
Yuchung Cheng375fe022013-07-22 16:20:45 -07002985/* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
Yuchung Cheng0f1c28a2015-09-18 11:36:14 -07002986void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
Yuchung Cheng375fe022013-07-22 16:20:45 -07002987{
Yuchung Cheng0f1c28a2015-09-18 11:36:14 -07002988 long rtt_us = -1L;
Yuchung Cheng375fe022013-07-22 16:20:45 -07002989
Yuchung Cheng0f1c28a2015-09-18 11:36:14 -07002990 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack.v64) {
2991 struct skb_mstamp now;
Yuchung Chengbc15afa2013-10-24 08:44:25 -07002992
Yuchung Cheng0f1c28a2015-09-18 11:36:14 -07002993 skb_mstamp_get(&now);
2994 rtt_us = skb_mstamp_us_delta(&now, &tcp_rsk(req)->snt_synack);
2995 }
2996
Yuchung Chengf6722582015-10-16 21:57:42 -07002997 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us);
Yuchung Cheng375fe022013-07-22 16:20:45 -07002998}
2999
Yuchung Cheng0f1c28a2015-09-18 11:36:14 -07003000
Eric Dumazet24901552014-05-02 21:18:05 -07003001static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003002{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003003 const struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazet24901552014-05-02 21:18:05 -07003004
3005 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003006 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003007}
3008
Linus Torvalds1da177e2005-04-16 15:20:36 -07003009/* Restart timer after forward progress on connection.
3010 * RFC2988 recommends to restart timer to now+rto.
3011 */
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00003012void tcp_rearm_rto(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003013{
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00003014 const struct inet_connection_sock *icsk = inet_csk(sk);
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00003015 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003016
Jerry Chu168a8f52012-08-31 12:29:13 +00003017 /* If the retrans timer is currently being used by Fast Open
3018 * for SYN-ACK retrans purpose, stay put.
3019 */
3020 if (tp->fastopen_rsk)
3021 return;
3022
Linus Torvalds1da177e2005-04-16 15:20:36 -07003023 if (!tp->packets_out) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003024 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003025 } else {
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00003026 u32 rto = inet_csk(sk)->icsk_rto;
3027 /* Offset the time elapsed after installing regular RTO */
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00003028 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
3029 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00003030 struct sk_buff *skb = tcp_write_queue_head(sk);
Eric Dumazet7faee5c2014-09-05 15:33:33 -07003031 const u32 rto_time_stamp =
3032 tcp_skb_timestamp(skb) + rto;
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00003033 s32 delta = (s32)(rto_time_stamp - tcp_time_stamp);
3034 /* delta may not be positive if the socket is locked
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00003035 * when the retrans timer fires and is rescheduled.
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00003036 */
3037 if (delta > 0)
3038 rto = delta;
3039 }
3040 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3041 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003042 }
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00003043}
3044
3045/* This function is called when the delayed ER timer fires. TCP enters
3046 * fast recovery and performs fast-retransmit.
3047 */
3048void tcp_resume_early_retransmit(struct sock *sk)
3049{
3050 struct tcp_sock *tp = tcp_sk(sk);
3051
3052 tcp_rearm_rto(sk);
3053
3054 /* Stop if ER is disabled after the delayed ER timer is scheduled */
3055 if (!tp->do_early_retrans)
3056 return;
3057
3058 tcp_enter_recovery(sk, false);
3059 tcp_update_scoreboard(sk, 1);
3060 tcp_xmit_retransmit_queue(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003061}
3062
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003063/* If we get here, the whole TSO packet has not been acked. */
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003064static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003065{
3066 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003067 u32 packets_acked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003068
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003069 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003070
3071 packets_acked = tcp_skb_pcount(skb);
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003072 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003073 return 0;
3074 packets_acked -= tcp_skb_pcount(skb);
3075
3076 if (packets_acked) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003077 BUG_ON(tcp_skb_pcount(skb) == 0);
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003078 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003079 }
3080
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003081 return packets_acked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003082}
3083
Eric Dumazetad971f62014-10-11 15:17:29 -07003084static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3085 u32 prior_snd_una)
3086{
3087 const struct skb_shared_info *shinfo;
3088
3089 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
Soheil Hassas Yeganeh6b084922016-04-02 23:08:08 -04003090 if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
Eric Dumazetad971f62014-10-11 15:17:29 -07003091 return;
3092
3093 shinfo = skb_shinfo(skb);
Soheil Hassas Yeganeh0a2cf202016-04-27 23:39:01 -04003094 if (!before(shinfo->tskey, prior_snd_una) &&
Martin KaFai Lau479f85c2016-04-18 15:39:53 -07003095 before(shinfo->tskey, tcp_sk(sk)->snd_una))
Eric Dumazetad971f62014-10-11 15:17:29 -07003096 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3097}
3098
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003099/* Remove acknowledged frames from the retransmission queue. If our packet
3100 * is before the ack sequence we can discard it as it's confirmed to have
3101 * arrived at the other end.
3102 */
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07003103static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
Yuchung Cheng3ebd88712016-02-02 10:33:07 -08003104 u32 prior_snd_una, int *acked,
Kenneth Klette Jonassen31231a82015-05-01 01:10:58 +02003105 struct tcp_sacktag_state *sack)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003106{
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08003107 const struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazet740b0f12014-02-26 14:02:48 -08003108 struct skb_mstamp first_ackt, last_ackt, now;
3109 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen90638a02008-09-20 21:25:52 -07003110 u32 prior_sacked = tp->sacked_out;
Eric Dumazet740b0f12014-02-26 14:02:48 -08003111 u32 reord = tp->packets_out;
3112 bool fully_acked = true;
Kenneth Klette Jonassen31231a82015-05-01 01:10:58 +02003113 long sack_rtt_us = -1L;
Eric Dumazet740b0f12014-02-26 14:02:48 -08003114 long seq_rtt_us = -1L;
Kenneth Klette Jonassen31231a82015-05-01 01:10:58 +02003115 long ca_rtt_us = -1L;
Eric Dumazet740b0f12014-02-26 14:02:48 -08003116 struct sk_buff *skb;
3117 u32 pkts_acked = 0;
Lawrence Brakmo6f094b92016-06-08 21:16:44 -07003118 u32 last_in_flight = 0;
Yuchung Cheng2f715c12013-10-24 08:59:27 -07003119 bool rtt_update;
Eric Dumazet740b0f12014-02-26 14:02:48 -08003120 int flag = 0;
3121
3122 first_ackt.v64 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003123
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003124 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09003125 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003126 u8 sacked = scb->sacked;
Eric Dumazet740b0f12014-02-26 14:02:48 -08003127 u32 acked_pcount;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003128
Eric Dumazetad971f62014-10-11 15:17:29 -07003129 tcp_ack_tstamp(sk, skb, prior_snd_una);
Willem de Bruijn712a7222014-08-12 14:53:16 -04003130
Gavin McCullagh2072c222007-12-29 19:11:21 -08003131 /* Determine how many packets and what bytes were acked, tso and else */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003132 if (after(scb->end_seq, tp->snd_una)) {
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003133 if (tcp_skb_pcount(skb) == 1 ||
3134 !after(tp->snd_una, scb->seq))
3135 break;
3136
Ilpo Järvinen72018832007-12-30 04:37:55 -08003137 acked_pcount = tcp_tso_acked(sk, skb);
3138 if (!acked_pcount)
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003139 break;
3140
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003141 fully_acked = false;
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003142 } else {
Eric Dumazetad971f62014-10-11 15:17:29 -07003143 /* Speedup tcp_unlink_write_queue() and next loop */
3144 prefetchw(skb->next);
Ilpo Järvinen72018832007-12-30 04:37:55 -08003145 acked_pcount = tcp_skb_pcount(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003146 }
3147
Eric Dumazetad971f62014-10-11 15:17:29 -07003148 if (unlikely(sacked & TCPCB_RETRANS)) {
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003149 if (sacked & TCPCB_SACKED_RETRANS)
Ilpo Järvinen72018832007-12-30 04:37:55 -08003150 tp->retrans_out -= acked_pcount;
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003151 flag |= FLAG_RETRANS_DATA_ACKED;
Kenneth Klette Jonassen3d0d26c2015-04-11 02:17:49 +02003152 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
Eric Dumazet740b0f12014-02-26 14:02:48 -08003153 last_ackt = skb->skb_mstamp;
Eric Dumazet431a9122014-03-09 17:36:02 -07003154 WARN_ON_ONCE(last_ackt.v64 == 0);
Eric Dumazet740b0f12014-02-26 14:02:48 -08003155 if (!first_ackt.v64)
3156 first_ackt = last_ackt;
3157
Lawrence Brakmo6f094b92016-06-08 21:16:44 -07003158 last_in_flight = TCP_SKB_CB(skb)->tx.in_flight;
Kenneth Klette Jonassen3d0d26c2015-04-11 02:17:49 +02003159 reord = min(pkts_acked, reord);
3160 if (!after(scb->end_seq, tp->high_seq))
3161 flag |= FLAG_ORIG_SACK_ACKED;
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08003162 }
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003163
Yuchung Chengddf1af62016-02-02 10:33:06 -08003164 if (sacked & TCPCB_SACKED_ACKED) {
Ilpo Järvinen72018832007-12-30 04:37:55 -08003165 tp->sacked_out -= acked_pcount;
Yuchung Chengddf1af62016-02-02 10:33:06 -08003166 } else if (tcp_is_sack(tp)) {
3167 tp->delivered += acked_pcount;
3168 if (!tcp_skb_spurious_retrans(tp, skb))
3169 tcp_rack_advance(tp, &skb->skb_mstamp, sacked);
3170 }
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003171 if (sacked & TCPCB_LOST)
Ilpo Järvinen72018832007-12-30 04:37:55 -08003172 tp->lost_out -= acked_pcount;
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003173
Ilpo Järvinen72018832007-12-30 04:37:55 -08003174 tp->packets_out -= acked_pcount;
3175 pkts_acked += acked_pcount;
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003176
Ilpo Järvinen009a2e32007-09-20 11:34:38 -07003177 /* Initial outgoing SYN's get put onto the write_queue
3178 * just like anything else we transmit. It is not
3179 * true data, and if we misinform our callers that
3180 * this ACK acks real data, we will erroneously exit
3181 * connection startup slow start one packet too
3182 * quickly. This is severely frowned upon behavior.
3183 */
Eric Dumazetad971f62014-10-11 15:17:29 -07003184 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
Ilpo Järvinen009a2e32007-09-20 11:34:38 -07003185 flag |= FLAG_DATA_ACKED;
3186 } else {
3187 flag |= FLAG_SYN_ACKED;
3188 tp->retrans_stamp = 0;
3189 }
3190
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003191 if (!fully_acked)
3192 break;
3193
David S. Millerfe067e82007-03-07 12:12:44 -08003194 tcp_unlink_write_queue(skb, sk);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08003195 sk_wmem_free_skb(sk, skb);
Eric Dumazetad971f62014-10-11 15:17:29 -07003196 if (unlikely(skb == tp->retransmit_skb_hint))
Ilpo Järvinenef9da472008-09-20 21:25:15 -07003197 tp->retransmit_skb_hint = NULL;
Eric Dumazetad971f62014-10-11 15:17:29 -07003198 if (unlikely(skb == tp->lost_skb_hint))
Ilpo Järvinen90638a02008-09-20 21:25:52 -07003199 tp->lost_skb_hint = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003200 }
3201
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07003202 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3203 tp->snd_up = tp->snd_una;
3204
Ilpo Järvinencadbd032007-12-31 04:49:21 -08003205 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3206 flag |= FLAG_SACK_RENEGING;
3207
Eric Dumazet740b0f12014-02-26 14:02:48 -08003208 skb_mstamp_get(&now);
Yuchung Cheng9e45a3e2015-10-16 21:57:41 -07003209 if (likely(first_ackt.v64) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
Eric Dumazet740b0f12014-02-26 14:02:48 -08003210 seq_rtt_us = skb_mstamp_us_delta(&now, &first_ackt);
Kenneth Klette Jonassen31231a82015-05-01 01:10:58 +02003211 ca_rtt_us = skb_mstamp_us_delta(&now, &last_ackt);
3212 }
3213 if (sack->first_sackt.v64) {
3214 sack_rtt_us = skb_mstamp_us_delta(&now, &sack->first_sackt);
3215 ca_rtt_us = skb_mstamp_us_delta(&now, &sack->last_sackt);
Eric Dumazet740b0f12014-02-26 14:02:48 -08003216 }
3217
Yuchung Chengf6722582015-10-16 21:57:42 -07003218 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3219 ca_rtt_us);
Yuchung Chenged084952013-07-22 16:20:48 -07003220
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003221 if (flag & FLAG_ACKED) {
Yuchung Cheng2f715c12013-10-24 08:59:27 -07003222 tcp_rearm_rto(sk);
Ilpo Järvinen72211e92009-03-14 14:23:04 +00003223 if (unlikely(icsk->icsk_mtup.probe_size &&
3224 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3225 tcp_mtup_probe_success(sk);
3226 }
3227
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003228 if (tcp_is_reno(tp)) {
3229 tcp_remove_reno_sacks(sk, pkts_acked);
3230 } else {
Ilpo Järvinen59a08cb2009-02-28 04:44:28 +00003231 int delta;
3232
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003233 /* Non-retransmitted hole got filled? That's reordering */
3234 if (reord < prior_fackets)
3235 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
Ilpo Järvinen90638a02008-09-20 21:25:52 -07003236
Ilpo Järvinen59a08cb2009-02-28 04:44:28 +00003237 delta = tcp_is_fack(tp) ? pkts_acked :
3238 prior_sacked - tp->sacked_out;
3239 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003240 }
3241
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07003242 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08003243
Eric Dumazet740b0f12014-02-26 14:02:48 -08003244 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3245 sack_rtt_us > skb_mstamp_us_delta(&now, &skb->skb_mstamp)) {
Yuchung Cheng2f715c12013-10-24 08:59:27 -07003246 /* Do not re-arm RTO if the sack RTT is measured from data sent
3247 * after when the head was last (re)transmitted. Otherwise the
3248 * timeout may continue to extend in loss recovery.
3249 */
3250 tcp_rearm_rto(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003251 }
3252
Lawrence Brakmo756ee172016-05-11 10:02:13 -07003253 if (icsk->icsk_ca_ops->pkts_acked) {
3254 struct ack_sample sample = { .pkts_acked = pkts_acked,
Lawrence Brakmo6f094b92016-06-08 21:16:44 -07003255 .rtt_us = ca_rtt_us,
3256 .in_flight = last_in_flight };
Lawrence Brakmo756ee172016-05-11 10:02:13 -07003257
3258 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3259 }
Kenneth Klette Jonassen138998f2015-05-01 01:10:59 +02003260
Linus Torvalds1da177e2005-04-16 15:20:36 -07003261#if FASTRETRANS_DEBUG > 0
Ilpo Järvinen547b7922008-07-25 21:43:18 -07003262 WARN_ON((int)tp->sacked_out < 0);
3263 WARN_ON((int)tp->lost_out < 0);
3264 WARN_ON((int)tp->retrans_out < 0);
Ilpo Järvinene60402d2007-08-09 15:14:46 +03003265 if (!tp->packets_out && tcp_is_sack(tp)) {
Stephen Hemmingercfcabdc2007-10-09 01:59:42 -07003266 icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003267 if (tp->lost_out) {
Joe Perches91df42b2012-05-15 14:11:54 +00003268 pr_debug("Leak l=%u %d\n",
3269 tp->lost_out, icsk->icsk_ca_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003270 tp->lost_out = 0;
3271 }
3272 if (tp->sacked_out) {
Joe Perches91df42b2012-05-15 14:11:54 +00003273 pr_debug("Leak s=%u %d\n",
3274 tp->sacked_out, icsk->icsk_ca_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003275 tp->sacked_out = 0;
3276 }
3277 if (tp->retrans_out) {
Joe Perches91df42b2012-05-15 14:11:54 +00003278 pr_debug("Leak r=%u %d\n",
3279 tp->retrans_out, icsk->icsk_ca_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003280 tp->retrans_out = 0;
3281 }
3282 }
3283#endif
Yuchung Cheng3ebd88712016-02-02 10:33:07 -08003284 *acked = pkts_acked;
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003285 return flag;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003286}
3287
3288static void tcp_ack_probe(struct sock *sk)
3289{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003290 const struct tcp_sock *tp = tcp_sk(sk);
3291 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003292
3293 /* Was it a usable window open? */
3294
Ilpo Järvinen90840de2007-12-31 04:48:41 -08003295 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003296 icsk->icsk_backoff = 0;
3297 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003298 /* Socket must be waked up by subsequent tcp_data_snd_check().
3299 * This function is not for random using!
3300 */
3301 } else {
Eric Dumazet21c8fe92015-05-06 14:26:24 -07003302 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
Eric Dumazetfcdd1cf2014-09-22 13:19:44 -07003303
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003304 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
Eric Dumazetfcdd1cf2014-09-22 13:19:44 -07003305 when, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003306 }
3307}
3308
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00003309static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003310{
Eric Dumazeta02cec22010-09-22 20:43:57 +00003311 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3312 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003313}
3314
Yuchung Cheng0f7cc9a2013-08-21 17:29:23 -07003315/* Decide wheather to run the increase function of congestion control. */
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00003316static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003317{
Yuchung Cheng0f7cc9a2013-08-21 17:29:23 -07003318 /* If reordering is high then always grow cwnd whenever data is
3319 * delivered regardless of its ordering. Otherwise stay conservative
Yuchung Cheng16edfe72013-09-05 15:36:09 -07003320 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
Yuchung Cheng0f7cc9a2013-08-21 17:29:23 -07003321 * new SACK or ECE mark may first advance cwnd here and later reduce
3322 * cwnd in tcp_fastretrans_alert() based on more states.
3323 */
Nikolay Borisov1043e252016-02-03 09:46:52 +02003324 if (tcp_sk(sk)->reordering > sock_net(sk)->ipv4.sysctl_tcp_reordering)
Yuchung Cheng0f7cc9a2013-08-21 17:29:23 -07003325 return flag & FLAG_FORWARD_PROGRESS;
3326
Yuchung Cheng16edfe72013-09-05 15:36:09 -07003327 return flag & FLAG_DATA_ACKED;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003328}
3329
Yuchung Chengd452e6c2016-02-02 10:33:09 -08003330/* The "ultimate" congestion control function that aims to replace the rigid
3331 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3332 * It's called toward the end of processing an ACK with precise rate
3333 * information. All transmission or retransmission are delayed afterwards.
3334 */
3335static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3336 int flag)
3337{
3338 if (tcp_in_cwnd_reduction(sk)) {
3339 /* Reduce cwnd if state mandates */
3340 tcp_cwnd_reduction(sk, acked_sacked, flag);
3341 } else if (tcp_may_raise_cwnd(sk, flag)) {
3342 /* Advance cwnd if state allows */
3343 tcp_cong_avoid(sk, ack, acked_sacked);
3344 }
3345 tcp_update_pacing_rate(sk);
3346}
3347
Linus Torvalds1da177e2005-04-16 15:20:36 -07003348/* Check that window update is acceptable.
3349 * The function assumes that snd_una<=ack<=snd_next.
3350 */
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00003351static inline bool tcp_may_update_window(const struct tcp_sock *tp,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003352 const u32 ack, const u32 ack_seq,
3353 const u32 nwin)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003354{
Eric Dumazeta02cec22010-09-22 20:43:57 +00003355 return after(ack, tp->snd_una) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07003356 after(ack_seq, tp->snd_wl1) ||
Eric Dumazeta02cec22010-09-22 20:43:57 +00003357 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003358}
3359
Eric Dumazet0df48c22015-04-28 15:28:17 -07003360/* If we update tp->snd_una, also update tp->bytes_acked */
3361static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3362{
3363 u32 delta = ack - tp->snd_una;
3364
Eric Dumazet46cc6e42016-05-03 16:56:03 -07003365 sock_owned_by_me((struct sock *)tp);
3366 u64_stats_update_begin_raw(&tp->syncp);
Eric Dumazet0df48c22015-04-28 15:28:17 -07003367 tp->bytes_acked += delta;
Eric Dumazet46cc6e42016-05-03 16:56:03 -07003368 u64_stats_update_end_raw(&tp->syncp);
Eric Dumazet0df48c22015-04-28 15:28:17 -07003369 tp->snd_una = ack;
3370}
3371
Eric Dumazetbdd1f9e2015-04-28 15:28:18 -07003372/* If we update tp->rcv_nxt, also update tp->bytes_received */
3373static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3374{
3375 u32 delta = seq - tp->rcv_nxt;
3376
Eric Dumazet46cc6e42016-05-03 16:56:03 -07003377 sock_owned_by_me((struct sock *)tp);
3378 u64_stats_update_begin_raw(&tp->syncp);
Eric Dumazetbdd1f9e2015-04-28 15:28:18 -07003379 tp->bytes_received += delta;
Eric Dumazet46cc6e42016-05-03 16:56:03 -07003380 u64_stats_update_end_raw(&tp->syncp);
Eric Dumazetbdd1f9e2015-04-28 15:28:18 -07003381 tp->rcv_nxt = seq;
3382}
3383
Linus Torvalds1da177e2005-04-16 15:20:36 -07003384/* Update our send window.
3385 *
3386 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3387 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3388 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003389static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003390 u32 ack_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003391{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003392 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003393 int flag = 0;
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07003394 u32 nwin = ntohs(tcp_hdr(skb)->window);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003395
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07003396 if (likely(!tcp_hdr(skb)->syn))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003397 nwin <<= tp->rx_opt.snd_wscale;
3398
3399 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3400 flag |= FLAG_WIN_UPDATE;
Hantzis Fotisee7537b2009-03-02 22:42:02 -08003401 tcp_update_wl(tp, ack_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003402
3403 if (tp->snd_wnd != nwin) {
3404 tp->snd_wnd = nwin;
3405
3406 /* Note, it is the only place, where
3407 * fast path is recovered for sending TCP.
3408 */
Herbert Xu2ad41062005-10-27 18:47:46 +10003409 tp->pred_flags = 0;
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003410 tcp_fast_path_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003411
Eric Dumazet6f021c62015-08-21 12:30:00 -07003412 if (tcp_send_head(sk))
3413 tcp_slow_start_after_idle_check(sk);
3414
Linus Torvalds1da177e2005-04-16 15:20:36 -07003415 if (nwin > tp->max_window) {
3416 tp->max_window = nwin;
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08003417 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003418 }
3419 }
3420 }
3421
Eric Dumazet0df48c22015-04-28 15:28:17 -07003422 tcp_snd_una_update(tp, ack);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003423
3424 return flag;
3425}
3426
Jason Baron083ae302016-07-14 11:38:40 -04003427static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3428 u32 *last_oow_ack_time)
3429{
3430 if (*last_oow_ack_time) {
3431 s32 elapsed = (s32)(tcp_time_stamp - *last_oow_ack_time);
3432
3433 if (0 <= elapsed && elapsed < sysctl_tcp_invalid_ratelimit) {
3434 NET_INC_STATS(net, mib_idx);
3435 return true; /* rate-limited: don't send yet! */
3436 }
3437 }
3438
3439 *last_oow_ack_time = tcp_time_stamp;
3440
3441 return false; /* not rate-limited: go ahead, send dupack now! */
3442}
3443
Eric Dumazet7970ddc2015-03-16 21:06:20 -07003444/* Return true if we're currently rate-limiting out-of-window ACKs and
3445 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3446 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3447 * attacks that send repeated SYNs or ACKs for the same connection. To
3448 * do this, we do not send a duplicate SYNACK or ACK if the remote
3449 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3450 */
3451bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3452 int mib_idx, u32 *last_oow_ack_time)
3453{
3454 /* Data packets without SYNs are not likely part of an ACK loop. */
3455 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3456 !tcp_hdr(skb)->syn)
Jason Baron083ae302016-07-14 11:38:40 -04003457 return false;
Eric Dumazet7970ddc2015-03-16 21:06:20 -07003458
Jason Baron083ae302016-07-14 11:38:40 -04003459 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
Eric Dumazet7970ddc2015-03-16 21:06:20 -07003460}
3461
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003462/* RFC 5961 7 [ACK Throttling] */
Neal Cardwellf2b2c582015-02-06 16:04:40 -05003463static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003464{
3465 /* unprotected vars, we dont care of overwrites */
3466 static u32 challenge_timestamp;
3467 static unsigned int challenge_count;
Neal Cardwellf2b2c582015-02-06 16:04:40 -05003468 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet75ff39c2016-07-10 10:04:02 +02003469 u32 count, now;
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003470
Neal Cardwellf2b2c582015-02-06 16:04:40 -05003471 /* First check our per-socket dupack rate limit. */
Jason Baron083ae302016-07-14 11:38:40 -04003472 if (__tcp_oow_rate_limited(sock_net(sk),
3473 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3474 &tp->last_oow_ack_time))
Neal Cardwellf2b2c582015-02-06 16:04:40 -05003475 return;
3476
Eric Dumazet75ff39c2016-07-10 10:04:02 +02003477 /* Then check host-wide RFC 5961 rate limit. */
Neal Cardwellf2b2c582015-02-06 16:04:40 -05003478 now = jiffies / HZ;
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003479 if (now != challenge_timestamp) {
Eric Dumazet75ff39c2016-07-10 10:04:02 +02003480 u32 half = (sysctl_tcp_challenge_ack_limit + 1) >> 1;
3481
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003482 challenge_timestamp = now;
Eric Dumazet75ff39c2016-07-10 10:04:02 +02003483 WRITE_ONCE(challenge_count, half +
3484 prandom_u32_max(sysctl_tcp_challenge_ack_limit));
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003485 }
Eric Dumazet75ff39c2016-07-10 10:04:02 +02003486 count = READ_ONCE(challenge_count);
3487 if (count > 0) {
3488 WRITE_ONCE(challenge_count, count - 1);
Eric Dumazetc10d9312016-04-29 14:16:47 -07003489 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK);
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003490 tcp_send_ack(sk);
3491 }
3492}
3493
Eric Dumazet12fb3dd2013-04-19 07:19:48 +00003494static void tcp_store_ts_recent(struct tcp_sock *tp)
3495{
3496 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3497 tp->rx_opt.ts_recent_stamp = get_seconds();
3498}
3499
3500static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3501{
3502 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3503 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3504 * extra check below makes sure this can only happen
3505 * for pure ACK frames. -DaveM
3506 *
3507 * Not only, also it occurs for expired timestamps.
3508 */
3509
3510 if (tcp_paws_check(&tp->rx_opt, 0))
3511 tcp_store_ts_recent(tp);
3512 }
3513}
3514
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003515/* This routine deals with acks during a TLP episode.
Sébastien Barré08abdff2015-01-12 10:30:40 +01003516 * We mark the end of a TLP episode on receiving TLP dupack or when
3517 * ack is after tlp_high_seq.
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003518 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3519 */
3520static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3521{
3522 struct tcp_sock *tp = tcp_sk(sk);
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003523
Sébastien Barré08abdff2015-01-12 10:30:40 +01003524 if (before(ack, tp->tlp_high_seq))
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003525 return;
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003526
Sébastien Barré08abdff2015-01-12 10:30:40 +01003527 if (flag & FLAG_DSACKING_ACK) {
3528 /* This DSACK means original and TLP probe arrived; no loss */
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003529 tp->tlp_high_seq = 0;
Sébastien Barré08abdff2015-01-12 10:30:40 +01003530 } else if (after(ack, tp->tlp_high_seq)) {
3531 /* ACK advances: there was a loss, so reduce cwnd. Reset
3532 * tlp_high_seq in tcp_init_cwnd_reduction()
3533 */
3534 tcp_init_cwnd_reduction(sk);
3535 tcp_set_ca_state(sk, TCP_CA_CWR);
3536 tcp_end_cwnd_reduction(sk);
3537 tcp_try_keep_open(sk);
Eric Dumazetc10d9312016-04-29 14:16:47 -07003538 NET_INC_STATS(sock_net(sk),
Eric Dumazet02a1d6e2016-04-27 16:44:39 -07003539 LINUX_MIB_TCPLOSSPROBERECOVERY);
Sébastien Barré08abdff2015-01-12 10:30:40 +01003540 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3541 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3542 /* Pure dupack: original and TLP probe arrived; no loss */
3543 tp->tlp_high_seq = 0;
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003544 }
3545}
3546
Florian Westphal7354c8c2014-09-26 22:37:34 +02003547static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3548{
3549 const struct inet_connection_sock *icsk = inet_csk(sk);
3550
3551 if (icsk->icsk_ca_ops->in_ack_event)
3552 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3553}
3554
Yuchung Chenge662ca42016-02-02 10:33:04 -08003555/* Congestion control has updated the cwnd already. So if we're in
3556 * loss recovery then now we do any new sends (for FRTO) or
3557 * retransmits (for CA_Loss or CA_recovery) that make sense.
3558 */
3559static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3560{
3561 struct tcp_sock *tp = tcp_sk(sk);
3562
3563 if (rexmit == REXMIT_NONE)
3564 return;
3565
3566 if (unlikely(rexmit == 2)) {
3567 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3568 TCP_NAGLE_OFF);
3569 if (after(tp->snd_nxt, tp->high_seq))
3570 return;
3571 tp->frto = 0;
3572 }
3573 tcp_xmit_retransmit_queue(sk);
3574}
3575
Linus Torvalds1da177e2005-04-16 15:20:36 -07003576/* This routine deals with incoming acks, but not outgoing ones. */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003577static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003578{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003579 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003580 struct tcp_sock *tp = tcp_sk(sk);
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02003581 struct tcp_sacktag_state sack_state;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003582 u32 prior_snd_una = tp->snd_una;
3583 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3584 u32 ack = TCP_SKB_CB(skb)->ack_seq;
Neal Cardwell7d2b55f2011-11-16 08:58:01 +00003585 bool is_dupack = false;
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003586 u32 prior_fackets;
Nandita Dukkipati35f079e2013-05-21 15:12:07 +00003587 int prior_packets = tp->packets_out;
Yuchung Chengddf1af62016-02-02 10:33:06 -08003588 u32 prior_delivered = tp->delivered;
Yuchung Cheng68049732013-05-29 14:20:11 +00003589 int acked = 0; /* Number of packets newly acked */
Yuchung Chenge662ca42016-02-02 10:33:04 -08003590 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02003591
Kenneth Klette Jonassen31231a82015-05-01 01:10:58 +02003592 sack_state.first_sackt.v64 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003593
Eric Dumazetad971f62014-10-11 15:17:29 -07003594 /* We very likely will need to access write queue head. */
3595 prefetchw(sk->sk_write_queue.next);
3596
John Dykstra96e0bf42009-03-22 21:49:57 -07003597 /* If the ack is older than previous acks
Linus Torvalds1da177e2005-04-16 15:20:36 -07003598 * then we can probably ignore it.
3599 */
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003600 if (before(ack, prior_snd_una)) {
3601 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3602 if (before(ack, prior_snd_una - tp->max_window)) {
Neal Cardwellf2b2c582015-02-06 16:04:40 -05003603 tcp_send_challenge_ack(sk, skb);
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003604 return -1;
3605 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003606 goto old_ack;
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003607 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003608
John Dykstra96e0bf42009-03-22 21:49:57 -07003609 /* If the ack includes data we haven't sent yet, discard
3610 * this segment (RFC793 Section 3.9).
3611 */
3612 if (after(ack, tp->snd_nxt))
3613 goto invalid_ack;
3614
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00003615 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
3616 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00003617 tcp_rearm_rto(sk);
3618
Neal Cardwell0c9ab092014-08-14 16:13:07 -04003619 if (after(ack, prior_snd_una)) {
Ilpo Järvinen2e605292007-08-02 19:46:58 -07003620 flag |= FLAG_SND_UNA_ADVANCED;
Neal Cardwell0c9ab092014-08-14 16:13:07 -04003621 icsk->icsk_retransmits = 0;
3622 }
Ilpo Järvinen2e605292007-08-02 19:46:58 -07003623
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003624 prior_fackets = tp->fackets_out;
3625
Eric Dumazet12fb3dd2013-04-19 07:19:48 +00003626 /* ts_recent update must be made after we are sure that the packet
3627 * is in window.
3628 */
3629 if (flag & FLAG_UPDATE_TS_RECENT)
3630 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3631
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003632 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003633 /* Window is constant, pure forward advance.
3634 * No more checks are required.
3635 * Note, we use the fact that SND.UNA>=SND.WL2.
3636 */
Hantzis Fotisee7537b2009-03-02 22:42:02 -08003637 tcp_update_wl(tp, ack_seq);
Eric Dumazet0df48c22015-04-28 15:28:17 -07003638 tcp_snd_una_update(tp, ack);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003639 flag |= FLAG_WIN_UPDATE;
3640
Florian Westphal98900922014-09-26 22:37:35 +02003641 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
Stephen Hemminger317a76f2005-06-23 12:19:55 -07003642
Eric Dumazetc10d9312016-04-29 14:16:47 -07003643 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003644 } else {
Florian Westphal98900922014-09-26 22:37:35 +02003645 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3646
Linus Torvalds1da177e2005-04-16 15:20:36 -07003647 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3648 flag |= FLAG_DATA;
3649 else
Eric Dumazetc10d9312016-04-29 14:16:47 -07003650 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003651
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003652 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003653
3654 if (TCP_SKB_CB(skb)->sacked)
Yuchung Cheng59c9af42013-07-22 16:20:47 -07003655 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02003656 &sack_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003657
Florian Westphal735d3832014-09-29 13:08:30 +02003658 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003659 flag |= FLAG_ECE;
Florian Westphal98900922014-09-26 22:37:35 +02003660 ack_ev_flags |= CA_ACK_ECE;
3661 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003662
Florian Westphal98900922014-09-26 22:37:35 +02003663 if (flag & FLAG_WIN_UPDATE)
3664 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3665
3666 tcp_in_ack_event(sk, ack_ev_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003667 }
3668
3669 /* We passed data and got it acked, remove any soft error
3670 * log. Something worked...
3671 */
3672 sk->sk_err_soft = 0;
David S. Miller4b53fb62008-07-23 16:38:45 -07003673 icsk->icsk_probes_out = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003674 tp->rcv_tstamp = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003675 if (!prior_packets)
3676 goto no_queue;
3677
Linus Torvalds1da177e2005-04-16 15:20:36 -07003678 /* See if we can take anything off of the retransmit queue. */
Yuchung Cheng3ebd88712016-02-02 10:33:07 -08003679 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una, &acked,
Kenneth Klette Jonassen31231a82015-05-01 01:10:58 +02003680 &sack_state);
Nandita Dukkipatia262f0c2011-08-21 20:21:57 +00003681
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003682 if (tcp_ack_is_dubious(sk, flag)) {
Neal Cardwell7d2b55f2011-11-16 08:58:01 +00003683 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
Yuchung Cheng31ba0c12016-02-02 10:33:05 -08003684 tcp_fastretrans_alert(sk, acked, is_dupack, &flag, &rexmit);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003685 }
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003686 if (tp->tlp_high_seq)
3687 tcp_process_tlp_ack(sk, ack, flag);
3688
David S. Miller5110effe2012-07-02 02:21:03 -07003689 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) {
3690 struct dst_entry *dst = __sk_dst_get(sk);
3691 if (dst)
3692 dst_confirm(dst);
3693 }
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00003694
3695 if (icsk->icsk_pending == ICSK_TIME_RETRANS)
3696 tcp_schedule_loss_probe(sk);
Yuchung Chengd452e6c2016-02-02 10:33:09 -08003697 tcp_cong_control(sk, ack, tp->delivered - prior_delivered, flag);
Yuchung Chenge662ca42016-02-02 10:33:04 -08003698 tcp_xmit_recovery(sk, rexmit);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003699 return 1;
3700
3701no_queue:
Neal Cardwell5628adf2011-11-16 08:58:02 +00003702 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3703 if (flag & FLAG_DSACKING_ACK)
Yuchung Cheng31ba0c12016-02-02 10:33:05 -08003704 tcp_fastretrans_alert(sk, acked, is_dupack, &flag, &rexmit);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003705 /* If this ack opens up a zero window, clear backoff. It was
3706 * being used to time the probes, and is probably far higher than
3707 * it needs to be for normal retransmission.
3708 */
David S. Millerfe067e82007-03-07 12:12:44 -08003709 if (tcp_send_head(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003710 tcp_ack_probe(sk);
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003711
3712 if (tp->tlp_high_seq)
3713 tcp_process_tlp_ack(sk, ack, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003714 return 1;
3715
John Dykstra96e0bf42009-03-22 21:49:57 -07003716invalid_ack:
3717 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3718 return -1;
3719
Linus Torvalds1da177e2005-04-16 15:20:36 -07003720old_ack:
Neal Cardwelle95ae2f2011-11-16 08:58:03 +00003721 /* If data was SACKed, tag it and see if we should send more data.
3722 * If data was DSACKed, see if we can undo a cwnd reduction.
3723 */
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07003724 if (TCP_SKB_CB(skb)->sacked) {
Yuchung Cheng59c9af42013-07-22 16:20:47 -07003725 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02003726 &sack_state);
Yuchung Cheng31ba0c12016-02-02 10:33:05 -08003727 tcp_fastretrans_alert(sk, acked, is_dupack, &flag, &rexmit);
Yuchung Chenge662ca42016-02-02 10:33:04 -08003728 tcp_xmit_recovery(sk, rexmit);
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07003729 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003730
John Dykstra96e0bf42009-03-22 21:49:57 -07003731 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003732 return 0;
3733}
3734
Daniel Lee7f9b8382015-04-06 14:37:26 -07003735static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3736 bool syn, struct tcp_fastopen_cookie *foc,
3737 bool exp_opt)
3738{
3739 /* Valid only in SYN or SYN-ACK with an even length. */
3740 if (!foc || !syn || len < 0 || (len & 1))
3741 return;
3742
3743 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3744 len <= TCP_FASTOPEN_COOKIE_MAX)
3745 memcpy(foc->val, cookie, len);
3746 else if (len != 0)
3747 len = -1;
3748 foc->len = len;
3749 foc->exp = exp_opt;
3750}
3751
Linus Torvalds1da177e2005-04-16 15:20:36 -07003752/* Look for tcp options. Normally only called on SYN and SYNACK packets.
3753 * But, this can also be called on packets in the established flow when
3754 * the fast version below fails.
3755 */
Christoph Paasch1a2c6182013-03-17 08:23:34 +00003756void tcp_parse_options(const struct sk_buff *skb,
3757 struct tcp_options_received *opt_rx, int estab,
Yuchung Cheng2100c8d2012-07-19 06:43:05 +00003758 struct tcp_fastopen_cookie *foc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003759{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003760 const unsigned char *ptr;
3761 const struct tcphdr *th = tcp_hdr(skb);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003762 int length = (th->doff * 4) - sizeof(struct tcphdr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003763
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003764 ptr = (const unsigned char *)(th + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003765 opt_rx->saw_tstamp = 0;
3766
Stephen Hemminger2de979b2007-03-08 20:45:19 -08003767 while (length > 0) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003768 int opcode = *ptr++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003769 int opsize;
3770
3771 switch (opcode) {
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003772 case TCPOPT_EOL:
3773 return;
3774 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3775 length--;
3776 continue;
3777 default:
3778 opsize = *ptr++;
3779 if (opsize < 2) /* "silly options" */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003780 return;
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003781 if (opsize > length)
3782 return; /* don't parse partial options */
3783 switch (opcode) {
3784 case TCPOPT_MSS:
3785 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07003786 u16 in_mss = get_unaligned_be16(ptr);
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003787 if (in_mss) {
3788 if (opt_rx->user_mss &&
3789 opt_rx->user_mss < in_mss)
3790 in_mss = opt_rx->user_mss;
3791 opt_rx->mss_clamp = in_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003792 }
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07003793 }
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003794 break;
3795 case TCPOPT_WINDOW:
3796 if (opsize == TCPOLEN_WINDOW && th->syn &&
David S. Millerbb5b7c12009-12-15 20:56:42 -08003797 !estab && sysctl_tcp_window_scaling) {
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003798 __u8 snd_wscale = *(__u8 *)ptr;
3799 opt_rx->wscale_ok = 1;
3800 if (snd_wscale > 14) {
Joe Perchese87cc472012-05-13 21:56:26 +00003801 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3802 __func__,
3803 snd_wscale);
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003804 snd_wscale = 14;
3805 }
3806 opt_rx->snd_wscale = snd_wscale;
3807 }
3808 break;
3809 case TCPOPT_TIMESTAMP:
3810 if ((opsize == TCPOLEN_TIMESTAMP) &&
3811 ((estab && opt_rx->tstamp_ok) ||
David S. Millerbb5b7c12009-12-15 20:56:42 -08003812 (!estab && sysctl_tcp_timestamps))) {
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003813 opt_rx->saw_tstamp = 1;
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07003814 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3815 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003816 }
3817 break;
3818 case TCPOPT_SACK_PERM:
3819 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
David S. Millerbb5b7c12009-12-15 20:56:42 -08003820 !estab && sysctl_tcp_sack) {
Vijay Subramanianab562222011-12-20 13:23:24 +00003821 opt_rx->sack_ok = TCP_SACK_SEEN;
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003822 tcp_sack_reset(opt_rx);
3823 }
3824 break;
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07003825
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003826 case TCPOPT_SACK:
3827 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3828 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3829 opt_rx->sack_ok) {
3830 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3831 }
3832 break;
3833#ifdef CONFIG_TCP_MD5SIG
3834 case TCPOPT_MD5SIG:
3835 /*
3836 * The MD5 Hash has already been
3837 * checked (see tcp_v{4,6}_do_rcv()).
3838 */
3839 break;
3840#endif
Daniel Lee7f9b8382015-04-06 14:37:26 -07003841 case TCPOPT_FASTOPEN:
3842 tcp_parse_fastopen_option(
3843 opsize - TCPOLEN_FASTOPEN_BASE,
3844 ptr, th->syn, foc, false);
3845 break;
3846
Yuchung Cheng2100c8d2012-07-19 06:43:05 +00003847 case TCPOPT_EXP:
3848 /* Fast Open option shares code 254 using a
Daniel Lee7f9b8382015-04-06 14:37:26 -07003849 * 16 bits magic number.
Yuchung Cheng2100c8d2012-07-19 06:43:05 +00003850 */
Daniel Lee7f9b8382015-04-06 14:37:26 -07003851 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
3852 get_unaligned_be16(ptr) ==
3853 TCPOPT_FASTOPEN_MAGIC)
3854 tcp_parse_fastopen_option(opsize -
3855 TCPOLEN_EXP_FASTOPEN_BASE,
3856 ptr + 2, th->syn, foc, true);
Yuchung Cheng2100c8d2012-07-19 06:43:05 +00003857 break;
3858
3859 }
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003860 ptr += opsize-2;
3861 length -= opsize;
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07003862 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003863 }
3864}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00003865EXPORT_SYMBOL(tcp_parse_options);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003866
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003867static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
Ilpo Järvinena4356b22008-08-23 05:12:29 -07003868{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003869 const __be32 *ptr = (const __be32 *)(th + 1);
Ilpo Järvinena4356b22008-08-23 05:12:29 -07003870
3871 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3872 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3873 tp->rx_opt.saw_tstamp = 1;
3874 ++ptr;
3875 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3876 ++ptr;
Andrew Vagine3e12022013-08-27 12:21:55 +04003877 if (*ptr)
3878 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3879 else
3880 tp->rx_opt.rcv_tsecr = 0;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003881 return true;
Ilpo Järvinena4356b22008-08-23 05:12:29 -07003882 }
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003883 return false;
Ilpo Järvinena4356b22008-08-23 05:12:29 -07003884}
3885
Linus Torvalds1da177e2005-04-16 15:20:36 -07003886/* Fast parse options. This hopes to only see timestamps.
3887 * If it is wrong it falls back on tcp_parse_options().
3888 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003889static bool tcp_fast_parse_options(const struct sk_buff *skb,
Christoph Paasch1a2c6182013-03-17 08:23:34 +00003890 const struct tcphdr *th, struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003891{
William Allen Simpson4957faade2009-12-02 18:25:27 +00003892 /* In the spirit of fast parsing, compare doff directly to constant
3893 * values. Because equality is used, short doff can be ignored here.
3894 */
3895 if (th->doff == (sizeof(*th) / 4)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003896 tp->rx_opt.saw_tstamp = 0;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003897 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003898 } else if (tp->rx_opt.tstamp_ok &&
William Allen Simpson4957faade2009-12-02 18:25:27 +00003899 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
Ilpo Järvinena4356b22008-08-23 05:12:29 -07003900 if (tcp_parse_aligned_timestamp(tp, th))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003901 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003902 }
Andrey Vaginee684b62013-02-11 05:50:19 +00003903
Christoph Paasch1a2c6182013-03-17 08:23:34 +00003904 tcp_parse_options(skb, &tp->rx_opt, 1, NULL);
Andrew Vagine3e12022013-08-27 12:21:55 +04003905 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
Andrey Vaginee684b62013-02-11 05:50:19 +00003906 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3907
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003908 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003909}
3910
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003911#ifdef CONFIG_TCP_MD5SIG
3912/*
3913 * Parse MD5 Signature option
3914 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003915const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003916{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003917 int length = (th->doff << 2) - sizeof(*th);
3918 const u8 *ptr = (const u8 *)(th + 1);
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003919
3920 /* If the TCP option is too short, we can short cut */
3921 if (length < TCPOLEN_MD5SIG)
3922 return NULL;
3923
3924 while (length > 0) {
3925 int opcode = *ptr++;
3926 int opsize;
3927
Weilong Chena22318e2013-12-23 14:37:26 +08003928 switch (opcode) {
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003929 case TCPOPT_EOL:
3930 return NULL;
3931 case TCPOPT_NOP:
3932 length--;
3933 continue;
3934 default:
3935 opsize = *ptr++;
3936 if (opsize < 2 || opsize > length)
3937 return NULL;
3938 if (opcode == TCPOPT_MD5SIG)
Dmitry Popovba78e2d2010-08-07 20:24:28 -07003939 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003940 }
3941 ptr += opsize - 2;
3942 length -= opsize;
3943 }
3944 return NULL;
3945}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00003946EXPORT_SYMBOL(tcp_parse_md5sig_option);
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003947#endif
3948
Linus Torvalds1da177e2005-04-16 15:20:36 -07003949/* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3950 *
3951 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3952 * it can pass through stack. So, the following predicate verifies that
3953 * this segment is not used for anything but congestion avoidance or
3954 * fast retransmit. Moreover, we even are able to eliminate most of such
3955 * second order effects, if we apply some small "replay" window (~RTO)
3956 * to timestamp space.
3957 *
3958 * All these measures still do not guarantee that we reject wrapped ACKs
3959 * on networks with high bandwidth, when sequence space is recycled fastly,
3960 * but it guarantees that such events will be very rare and do not affect
3961 * connection seriously. This doesn't look nice, but alas, PAWS is really
3962 * buggy extension.
3963 *
3964 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3965 * states that events when retransmit arrives after original data are rare.
3966 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3967 * the biggest problem on large power networks even with minor reordering.
3968 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3969 * up to bandwidth of 18Gigabit/sec. 8) ]
3970 */
3971
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003972static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003973{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003974 const struct tcp_sock *tp = tcp_sk(sk);
3975 const struct tcphdr *th = tcp_hdr(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003976 u32 seq = TCP_SKB_CB(skb)->seq;
3977 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3978
3979 return (/* 1. Pure ACK with correct sequence number. */
3980 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3981
3982 /* 2. ... and duplicate ACK. */
3983 ack == tp->snd_una &&
3984
3985 /* 3. ... and does not update window. */
3986 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3987
3988 /* 4. ... and sits in replay window. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003989 (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 -07003990}
3991
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00003992static inline bool tcp_paws_discard(const struct sock *sk,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003993 const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003994{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003995 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinenc887e6d2009-03-14 14:23:03 +00003996
3997 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
3998 !tcp_disordered_ack(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003999}
4000
4001/* Check segment sequence number for validity.
4002 *
4003 * Segment controls are considered valid, if the segment
4004 * fits to the window after truncation to the window. Acceptability
4005 * of data (and SYN, FIN, of course) is checked separately.
4006 * See tcp_data_queue(), for example.
4007 *
4008 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4009 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4010 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4011 * (borrowed from freebsd)
4012 */
4013
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00004014static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004015{
4016 return !before(end_seq, tp->rcv_wup) &&
4017 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4018}
4019
4020/* When we get a reset we do this. */
Jerry Chu10467162012-08-31 12:29:11 +00004021void tcp_reset(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004022{
4023 /* We want the right error as BSD sees it (and indeed as we do). */
4024 switch (sk->sk_state) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004025 case TCP_SYN_SENT:
4026 sk->sk_err = ECONNREFUSED;
4027 break;
4028 case TCP_CLOSE_WAIT:
4029 sk->sk_err = EPIPE;
4030 break;
4031 case TCP_CLOSE:
4032 return;
4033 default:
4034 sk->sk_err = ECONNRESET;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004035 }
Tom Marshalla4d25802010-09-20 15:42:05 -07004036 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4037 smp_wmb();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004038
4039 if (!sock_flag(sk, SOCK_DEAD))
4040 sk->sk_error_report(sk);
4041
4042 tcp_done(sk);
4043}
4044
4045/*
4046 * Process the FIN bit. This now behaves as it is supposed to work
4047 * and the FIN takes effect when it is validly part of sequence
4048 * space. Not before when we get holes.
4049 *
4050 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4051 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4052 * TIME-WAIT)
4053 *
4054 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4055 * close and we go into CLOSING (and later onto TIME-WAIT)
4056 *
4057 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4058 */
Eric Dumazete3e17b72016-02-06 11:16:28 -08004059void tcp_fin(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004060{
4061 struct tcp_sock *tp = tcp_sk(sk);
4062
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004063 inet_csk_schedule_ack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004064
4065 sk->sk_shutdown |= RCV_SHUTDOWN;
4066 sock_set_flag(sk, SOCK_DONE);
4067
4068 switch (sk->sk_state) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004069 case TCP_SYN_RECV:
4070 case TCP_ESTABLISHED:
4071 /* Move to CLOSE_WAIT */
4072 tcp_set_state(sk, TCP_CLOSE_WAIT);
Jon Maxwell2251ae42015-07-08 10:12:28 +10004073 inet_csk(sk)->icsk_ack.pingpong = 1;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004074 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004075
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004076 case TCP_CLOSE_WAIT:
4077 case TCP_CLOSING:
4078 /* Received a retransmission of the FIN, do
4079 * nothing.
4080 */
4081 break;
4082 case TCP_LAST_ACK:
4083 /* RFC793: Remain in the LAST-ACK state. */
4084 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004085
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004086 case TCP_FIN_WAIT1:
4087 /* This case occurs when a simultaneous close
4088 * happens, we must ack the received FIN and
4089 * enter the CLOSING state.
4090 */
4091 tcp_send_ack(sk);
4092 tcp_set_state(sk, TCP_CLOSING);
4093 break;
4094 case TCP_FIN_WAIT2:
4095 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4096 tcp_send_ack(sk);
4097 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4098 break;
4099 default:
4100 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4101 * cases we should never reach this piece of code.
4102 */
Joe Perches058bd4d2012-03-11 18:36:11 +00004103 pr_err("%s: Impossible, sk->sk_state=%d\n",
Harvey Harrison0dc47872008-03-05 20:47:47 -08004104 __func__, sk->sk_state);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004105 break;
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07004106 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004107
4108 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4109 * Probably, we should reset in this case. For now drop them.
4110 */
4111 __skb_queue_purge(&tp->out_of_order_queue);
Ilpo Järvinene60402d2007-08-09 15:14:46 +03004112 if (tcp_is_sack(tp))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004113 tcp_sack_reset(&tp->rx_opt);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08004114 sk_mem_reclaim(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004115
4116 if (!sock_flag(sk, SOCK_DEAD)) {
4117 sk->sk_state_change(sk);
4118
4119 /* Do not send POLL_HUP for half duplex close. */
4120 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4121 sk->sk_state == TCP_CLOSE)
Pavel Emelyanov8d8ad9d2007-11-26 20:10:50 +08004122 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004123 else
Pavel Emelyanov8d8ad9d2007-11-26 20:10:50 +08004124 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004125 }
4126}
4127
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004128static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004129 u32 end_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004130{
4131 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4132 if (before(seq, sp->start_seq))
4133 sp->start_seq = seq;
4134 if (after(end_seq, sp->end_seq))
4135 sp->end_seq = end_seq;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004136 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004137 }
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004138 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004139}
4140
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004141static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004142{
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004143 struct tcp_sock *tp = tcp_sk(sk);
4144
David S. Millerbb5b7c12009-12-15 20:56:42 -08004145 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07004146 int mib_idx;
4147
Linus Torvalds1da177e2005-04-16 15:20:36 -07004148 if (before(seq, tp->rcv_nxt))
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07004149 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004150 else
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07004151 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4152
Eric Dumazetc10d9312016-04-29 14:16:47 -07004153 NET_INC_STATS(sock_net(sk), mib_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004154
4155 tp->rx_opt.dsack = 1;
4156 tp->duplicate_sack[0].start_seq = seq;
4157 tp->duplicate_sack[0].end_seq = end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004158 }
4159}
4160
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004161static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004162{
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004163 struct tcp_sock *tp = tcp_sk(sk);
4164
Linus Torvalds1da177e2005-04-16 15:20:36 -07004165 if (!tp->rx_opt.dsack)
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004166 tcp_dsack_set(sk, seq, end_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004167 else
4168 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4169}
4170
Eric Dumazetcf533ea2011-10-21 05:22:42 -04004171static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004172{
4173 struct tcp_sock *tp = tcp_sk(sk);
4174
4175 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4176 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
Eric Dumazetc10d9312016-04-29 14:16:47 -07004177 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004178 tcp_enter_quickack_mode(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004179
David S. Millerbb5b7c12009-12-15 20:56:42 -08004180 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004181 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4182
4183 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4184 end_seq = tp->rcv_nxt;
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004185 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004186 }
4187 }
4188
4189 tcp_send_ack(sk);
4190}
4191
4192/* These routines update the SACK block as out-of-order packets arrive or
4193 * in-order packets close up the sequence space.
4194 */
4195static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4196{
4197 int this_sack;
4198 struct tcp_sack_block *sp = &tp->selective_acks[0];
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004199 struct tcp_sack_block *swalk = sp + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004200
4201 /* See if the recent change to the first SACK eats into
4202 * or hits the sequence space of other SACK blocks, if so coalesce.
4203 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004204 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004205 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4206 int i;
4207
4208 /* Zap SWALK, by moving every further SACK up by one slot.
4209 * Decrease num_sacks.
4210 */
4211 tp->rx_opt.num_sacks--;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004212 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4213 sp[i] = sp[i + 1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07004214 continue;
4215 }
4216 this_sack++, swalk++;
4217 }
4218}
4219
Linus Torvalds1da177e2005-04-16 15:20:36 -07004220static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4221{
4222 struct tcp_sock *tp = tcp_sk(sk);
4223 struct tcp_sack_block *sp = &tp->selective_acks[0];
4224 int cur_sacks = tp->rx_opt.num_sacks;
4225 int this_sack;
4226
4227 if (!cur_sacks)
4228 goto new_sack;
4229
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004230 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004231 if (tcp_sack_extend(sp, seq, end_seq)) {
4232 /* Rotate this_sack to the first one. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004233 for (; this_sack > 0; this_sack--, sp--)
Ilpo Järvinena0bffff2009-03-21 13:36:17 -07004234 swap(*sp, *(sp - 1));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004235 if (cur_sacks > 1)
4236 tcp_sack_maybe_coalesce(tp);
4237 return;
4238 }
4239 }
4240
4241 /* Could not find an adjacent existing SACK, build a new one,
4242 * put it at the front, and shift everyone else down. We
4243 * always know there is at least one SACK present already here.
4244 *
4245 * If the sack array is full, forget about the last one.
4246 */
Adam Langley4389dde2008-07-19 00:07:02 -07004247 if (this_sack >= TCP_NUM_SACKS) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004248 this_sack--;
4249 tp->rx_opt.num_sacks--;
4250 sp--;
4251 }
Stephen Hemminger2de979b2007-03-08 20:45:19 -08004252 for (; this_sack > 0; this_sack--, sp--)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004253 *sp = *(sp - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004254
4255new_sack:
4256 /* Build the new head SACK, and we're done. */
4257 sp->start_seq = seq;
4258 sp->end_seq = end_seq;
4259 tp->rx_opt.num_sacks++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004260}
4261
4262/* RCV.NXT advances, some SACKs should be eaten. */
4263
4264static void tcp_sack_remove(struct tcp_sock *tp)
4265{
4266 struct tcp_sack_block *sp = &tp->selective_acks[0];
4267 int num_sacks = tp->rx_opt.num_sacks;
4268 int this_sack;
4269
4270 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
David S. Millerb03efcf2005-07-08 14:57:23 -07004271 if (skb_queue_empty(&tp->out_of_order_queue)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004272 tp->rx_opt.num_sacks = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004273 return;
4274 }
4275
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004276 for (this_sack = 0; this_sack < num_sacks;) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004277 /* Check if the start of the sack is covered by RCV.NXT. */
4278 if (!before(tp->rcv_nxt, sp->start_seq)) {
4279 int i;
4280
4281 /* RCV.NXT must cover all the block! */
Ilpo Järvinen547b7922008-07-25 21:43:18 -07004282 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004283
4284 /* Zap this SACK, by moving forward any other SACKS. */
Weilong Chena22318e2013-12-23 14:37:26 +08004285 for (i = this_sack+1; i < num_sacks; i++)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004286 tp->selective_acks[i-1] = tp->selective_acks[i];
4287 num_sacks--;
4288 continue;
4289 }
4290 this_sack++;
4291 sp++;
4292 }
Ilpo Järvinen5861f8e2009-03-14 14:23:01 +00004293 tp->rx_opt.num_sacks = num_sacks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004294}
4295
Eric Dumazet1402d362012-04-23 07:11:42 +00004296/**
4297 * tcp_try_coalesce - try to merge skb to prior one
4298 * @sk: socket
4299 * @to: prior buffer
4300 * @from: buffer to add in queue
Eric Dumazet923dd342012-05-02 07:55:58 +00004301 * @fragstolen: pointer to boolean
Eric Dumazet1402d362012-04-23 07:11:42 +00004302 *
4303 * Before queueing skb @from after @to, try to merge them
4304 * to reduce overall memory use and queue lengths, if cost is small.
4305 * Packets in ofo or receive queues can stay a long time.
4306 * Better try to coalesce them right now to avoid future collapses.
Eric Dumazet783c1752012-04-23 17:34:36 +00004307 * Returns true if caller should free @from instead of queueing it
Eric Dumazet1402d362012-04-23 07:11:42 +00004308 */
Eric Dumazet783c1752012-04-23 17:34:36 +00004309static bool tcp_try_coalesce(struct sock *sk,
4310 struct sk_buff *to,
Eric Dumazet329033f2012-04-27 00:38:33 +00004311 struct sk_buff *from,
4312 bool *fragstolen)
Eric Dumazet1402d362012-04-23 07:11:42 +00004313{
Eric Dumazetbad43ca2012-05-19 03:02:02 +00004314 int delta;
Eric Dumazet1402d362012-04-23 07:11:42 +00004315
Eric Dumazet329033f2012-04-27 00:38:33 +00004316 *fragstolen = false;
Alexander Duyck34a802a2012-05-02 21:19:09 +00004317
Eric Dumazet1ca7ee32012-05-23 17:51:37 +00004318 /* Its possible this segment overlaps with prior segment in queue */
4319 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4320 return false;
4321
Eric Dumazetbad43ca2012-05-19 03:02:02 +00004322 if (!skb_try_coalesce(to, from, fragstolen, &delta))
Eric Dumazet783c1752012-04-23 17:34:36 +00004323 return false;
Alexander Duyck34a802a2012-05-02 21:19:09 +00004324
Alexander Duyck34a802a2012-05-02 21:19:09 +00004325 atomic_add(delta, &sk->sk_rmem_alloc);
4326 sk_mem_charge(sk, delta);
Eric Dumazetc10d9312016-04-29 14:16:47 -07004327 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
Alexander Duyck34a802a2012-05-02 21:19:09 +00004328 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4329 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
Eric Dumazete93a0432014-09-15 04:19:52 -07004330 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
Alexander Duyck34a802a2012-05-02 21:19:09 +00004331 return true;
Eric Dumazet1402d362012-04-23 07:11:42 +00004332}
4333
Eric Dumazet532182c2016-04-01 08:52:19 -07004334static void tcp_drop(struct sock *sk, struct sk_buff *skb)
4335{
4336 sk_drops_add(sk, skb);
4337 __kfree_skb(skb);
4338}
4339
Linus Torvalds1da177e2005-04-16 15:20:36 -07004340/* This one checks to see if we can put data from the
4341 * out_of_order queue into the receive_queue.
4342 */
4343static void tcp_ofo_queue(struct sock *sk)
4344{
4345 struct tcp_sock *tp = tcp_sk(sk);
4346 __u32 dsack_high = tp->rcv_nxt;
Eric Dumazetbd1e75a2014-09-19 08:26:20 -07004347 struct sk_buff *skb, *tail;
4348 bool fragstolen, eaten;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004349
4350 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4351 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4352 break;
4353
4354 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4355 __u32 dsack = dsack_high;
4356 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4357 dsack_high = TCP_SKB_CB(skb)->end_seq;
4358 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4359 }
4360
Eric Dumazetbd1e75a2014-09-19 08:26:20 -07004361 __skb_unlink(skb, &tp->out_of_order_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004362 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4363 SOCK_DEBUG(sk, "ofo packet was already received\n");
Eric Dumazet532182c2016-04-01 08:52:19 -07004364 tcp_drop(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004365 continue;
4366 }
4367 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4368 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4369 TCP_SKB_CB(skb)->end_seq);
4370
Eric Dumazetbd1e75a2014-09-19 08:26:20 -07004371 tail = skb_peek_tail(&sk->sk_receive_queue);
4372 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
Eric Dumazetbdd1f9e2015-04-28 15:28:18 -07004373 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
Eric Dumazetbd1e75a2014-09-19 08:26:20 -07004374 if (!eaten)
4375 __skb_queue_tail(&sk->sk_receive_queue, skb);
4376 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004377 tcp_fin(sk);
Eric Dumazetbd1e75a2014-09-19 08:26:20 -07004378 if (eaten)
4379 kfree_skb_partial(skb, fragstolen);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004380 }
4381}
4382
4383static bool tcp_prune_ofo_queue(struct sock *sk);
4384static int tcp_prune_queue(struct sock *sk);
4385
4386static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4387 unsigned int size)
4388{
4389 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4390 !sk_rmem_schedule(sk, skb, size)) {
4391
4392 if (tcp_prune_queue(sk) < 0)
4393 return -1;
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004394
Linus Torvalds1da177e2005-04-16 15:20:36 -07004395 if (!sk_rmem_schedule(sk, skb, size)) {
4396 if (!tcp_prune_ofo_queue(sk))
4397 return -1;
Frans Popb1383382010-03-24 07:57:28 +00004398
David S. Miller8728b832005-08-09 19:25:21 -07004399 if (!sk_rmem_schedule(sk, skb, size))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004400 return -1;
4401 }
4402 }
4403 return 0;
4404}
4405
Eric Dumazete86b29192012-03-18 11:06:44 +00004406static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4407{
4408 struct tcp_sock *tp = tcp_sk(sk);
4409 struct sk_buff *skb1;
4410 u32 seq, end_seq;
4411
Florian Westphal735d3832014-09-29 13:08:30 +02004412 tcp_ecn_check_ce(tp, skb);
Eric Dumazete86b29192012-03-18 11:06:44 +00004413
Mel Gormanc76562b2012-07-31 16:44:41 -07004414 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
Eric Dumazetc10d9312016-04-29 14:16:47 -07004415 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
Eric Dumazet532182c2016-04-01 08:52:19 -07004416 tcp_drop(sk, skb);
Eric Dumazete86b29192012-03-18 11:06:44 +00004417 return;
4418 }
4419
4420 /* Disable header prediction. */
4421 tp->pred_flags = 0;
4422 inet_csk_schedule_ack(sk);
4423
Eric Dumazetc10d9312016-04-29 14:16:47 -07004424 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
Eric Dumazete86b29192012-03-18 11:06:44 +00004425 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4426 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4427
4428 skb1 = skb_peek_tail(&tp->out_of_order_queue);
4429 if (!skb1) {
4430 /* Initial out of order segment, build 1 SACK. */
4431 if (tcp_is_sack(tp)) {
4432 tp->rx_opt.num_sacks = 1;
4433 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4434 tp->selective_acks[0].end_seq =
4435 TCP_SKB_CB(skb)->end_seq;
4436 }
4437 __skb_queue_head(&tp->out_of_order_queue, skb);
4438 goto end;
4439 }
4440
4441 seq = TCP_SKB_CB(skb)->seq;
4442 end_seq = TCP_SKB_CB(skb)->end_seq;
4443
4444 if (seq == TCP_SKB_CB(skb1)->end_seq) {
Eric Dumazet329033f2012-04-27 00:38:33 +00004445 bool fragstolen;
4446
4447 if (!tcp_try_coalesce(sk, skb1, skb, &fragstolen)) {
Eric Dumazet1402d362012-04-23 07:11:42 +00004448 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4449 } else {
Eric Dumazet4e4f1fc2013-09-06 10:35:58 -07004450 tcp_grow_window(sk, skb);
Eric Dumazet923dd342012-05-02 07:55:58 +00004451 kfree_skb_partial(skb, fragstolen);
Eric Dumazetc8628152012-03-18 11:07:47 +00004452 skb = NULL;
Eric Dumazetc8628152012-03-18 11:07:47 +00004453 }
Eric Dumazete86b29192012-03-18 11:06:44 +00004454
4455 if (!tp->rx_opt.num_sacks ||
4456 tp->selective_acks[0].end_seq != seq)
4457 goto add_sack;
4458
4459 /* Common case: data arrive in order after hole. */
4460 tp->selective_acks[0].end_seq = end_seq;
4461 goto end;
4462 }
4463
4464 /* Find place to insert this segment. */
4465 while (1) {
4466 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4467 break;
4468 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4469 skb1 = NULL;
4470 break;
4471 }
4472 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4473 }
4474
4475 /* Do skb overlap to previous one? */
4476 if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4477 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4478 /* All the bits are present. Drop. */
Eric Dumazetc10d9312016-04-29 14:16:47 -07004479 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
Eric Dumazet532182c2016-04-01 08:52:19 -07004480 tcp_drop(sk, skb);
Eric Dumazete86b29192012-03-18 11:06:44 +00004481 skb = NULL;
4482 tcp_dsack_set(sk, seq, end_seq);
4483 goto add_sack;
4484 }
4485 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4486 /* Partial overlap. */
4487 tcp_dsack_set(sk, seq,
4488 TCP_SKB_CB(skb1)->end_seq);
4489 } else {
4490 if (skb_queue_is_first(&tp->out_of_order_queue,
4491 skb1))
4492 skb1 = NULL;
4493 else
4494 skb1 = skb_queue_prev(
4495 &tp->out_of_order_queue,
4496 skb1);
4497 }
4498 }
4499 if (!skb1)
4500 __skb_queue_head(&tp->out_of_order_queue, skb);
4501 else
4502 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4503
4504 /* And clean segments covered by new one as whole. */
4505 while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4506 skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4507
4508 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4509 break;
4510 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4511 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4512 end_seq);
4513 break;
4514 }
4515 __skb_unlink(skb1, &tp->out_of_order_queue);
4516 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4517 TCP_SKB_CB(skb1)->end_seq);
Eric Dumazetc10d9312016-04-29 14:16:47 -07004518 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
Eric Dumazet532182c2016-04-01 08:52:19 -07004519 tcp_drop(sk, skb1);
Eric Dumazete86b29192012-03-18 11:06:44 +00004520 }
4521
4522add_sack:
4523 if (tcp_is_sack(tp))
4524 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4525end:
Eric Dumazet4e4f1fc2013-09-06 10:35:58 -07004526 if (skb) {
4527 tcp_grow_window(sk, skb);
Eric Dumazete86b29192012-03-18 11:06:44 +00004528 skb_set_owner_r(skb, sk);
Eric Dumazet4e4f1fc2013-09-06 10:35:58 -07004529 }
Eric Dumazete86b29192012-03-18 11:06:44 +00004530}
4531
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004532static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
Eric Dumazetb081f852012-05-02 09:58:29 +00004533 bool *fragstolen)
4534{
4535 int eaten;
4536 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4537
4538 __skb_pull(skb, hdrlen);
4539 eaten = (tail &&
4540 tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0;
Eric Dumazetbdd1f9e2015-04-28 15:28:18 -07004541 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
Eric Dumazetb081f852012-05-02 09:58:29 +00004542 if (!eaten) {
4543 __skb_queue_tail(&sk->sk_receive_queue, skb);
4544 skb_set_owner_r(skb, sk);
4545 }
4546 return eaten;
4547}
Eric Dumazete86b29192012-03-18 11:06:44 +00004548
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004549int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4550{
Eric Dumazetcb934712014-09-17 03:14:42 -07004551 struct sk_buff *skb;
Eric Dumazet5d4c9bf2015-11-18 21:03:33 -08004552 int err = -ENOMEM;
4553 int data_len = 0;
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004554 bool fragstolen;
4555
Pavel Emelyanovc454e612012-10-29 05:05:33 +00004556 if (size == 0)
4557 return 0;
4558
Eric Dumazet5d4c9bf2015-11-18 21:03:33 -08004559 if (size > PAGE_SIZE) {
4560 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4561
4562 data_len = npages << PAGE_SHIFT;
4563 size = data_len + (size & ~PAGE_MASK);
4564 }
4565 skb = alloc_skb_with_frags(size - data_len, data_len,
4566 PAGE_ALLOC_COSTLY_ORDER,
4567 &err, sk->sk_allocation);
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004568 if (!skb)
4569 goto err;
4570
Eric Dumazet5d4c9bf2015-11-18 21:03:33 -08004571 skb_put(skb, size - data_len);
4572 skb->data_len = data_len;
4573 skb->len = size;
4574
Eric Dumazetcb934712014-09-17 03:14:42 -07004575 if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
Mel Gormanc76562b2012-07-31 16:44:41 -07004576 goto err_free;
4577
Eric Dumazet5d4c9bf2015-11-18 21:03:33 -08004578 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
4579 if (err)
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004580 goto err_free;
4581
4582 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4583 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4584 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4585
Eric Dumazetcb934712014-09-17 03:14:42 -07004586 if (tcp_queue_rcv(sk, skb, 0, &fragstolen)) {
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004587 WARN_ON_ONCE(fragstolen); /* should not happen */
4588 __kfree_skb(skb);
4589 }
4590 return size;
4591
4592err_free:
4593 kfree_skb(skb);
4594err:
Eric Dumazet5d4c9bf2015-11-18 21:03:33 -08004595 return err;
4596
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004597}
4598
Linus Torvalds1da177e2005-04-16 15:20:36 -07004599static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4600{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004601 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet329033f2012-04-27 00:38:33 +00004602 bool fragstolen = false;
Eric Dumazet532182c2016-04-01 08:52:19 -07004603 int eaten = -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004604
Eric Dumazet532182c2016-04-01 08:52:19 -07004605 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
4606 __kfree_skb(skb);
4607 return;
4608 }
Eric Dumazetf84af322010-04-28 15:31:51 -07004609 skb_dst_drop(skb);
Peter Pan(潘卫平)155c6e12014-09-24 22:17:02 +08004610 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004611
Florian Westphal735d3832014-09-29 13:08:30 +02004612 tcp_ecn_accept_cwr(tp, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004613
Ilpo Järvinen5861f8e2009-03-14 14:23:01 +00004614 tp->rx_opt.dsack = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004615
4616 /* Queue data for delivery to the user.
4617 * Packets in sequence go to the receive queue.
4618 * Out of sequence packets to the out_of_order_queue.
4619 */
4620 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4621 if (tcp_receive_window(tp) == 0)
4622 goto out_of_window;
4623
4624 /* Ok. In sequence. In window. */
4625 if (tp->ucopy.task == current &&
4626 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4627 sock_owned_by_user(sk) && !tp->urg_data) {
4628 int chunk = min_t(unsigned int, skb->len,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004629 tp->ucopy.len);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004630
4631 __set_current_state(TASK_RUNNING);
4632
Al Virof4362a22014-11-24 13:26:06 -05004633 if (!skb_copy_datagram_msg(skb, 0, tp->ucopy.msg, chunk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004634 tp->ucopy.len -= chunk;
4635 tp->copied_seq += chunk;
Jerry Chu44f53242011-01-25 13:46:30 -08004636 eaten = (chunk == skb->len);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004637 tcp_rcv_space_adjust(sk);
4638 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004639 }
4640
4641 if (eaten <= 0) {
4642queue_and_out:
Eric Dumazet76dfa602015-05-15 12:39:29 -07004643 if (eaten < 0) {
4644 if (skb_queue_len(&sk->sk_receive_queue) == 0)
4645 sk_forced_mem_schedule(sk, skb->truesize);
4646 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4647 goto drop;
4648 }
Eric Dumazetb081f852012-05-02 09:58:29 +00004649 eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004650 }
Eric Dumazetbdd1f9e2015-04-28 15:28:18 -07004651 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
Stephen Hemminger2de979b2007-03-08 20:45:19 -08004652 if (skb->len)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004653 tcp_event_data_recv(sk, skb);
Peter Pan(潘卫平)155c6e12014-09-24 22:17:02 +08004654 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
Eric Dumazet20c4cb72011-10-20 17:44:03 -04004655 tcp_fin(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004656
David S. Millerb03efcf2005-07-08 14:57:23 -07004657 if (!skb_queue_empty(&tp->out_of_order_queue)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004658 tcp_ofo_queue(sk);
4659
4660 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4661 * gap in queue is filled.
4662 */
David S. Millerb03efcf2005-07-08 14:57:23 -07004663 if (skb_queue_empty(&tp->out_of_order_queue))
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004664 inet_csk(sk)->icsk_ack.pingpong = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004665 }
4666
4667 if (tp->rx_opt.num_sacks)
4668 tcp_sack_remove(tp);
4669
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004670 tcp_fast_path_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004671
Eric Dumazet923dd342012-05-02 07:55:58 +00004672 if (eaten > 0)
4673 kfree_skb_partial(skb, fragstolen);
Eric Dumazet1d57f192012-09-17 12:51:39 +00004674 if (!sock_flag(sk, SOCK_DEAD))
David S. Miller676d2362014-04-11 16:15:36 -04004675 sk->sk_data_ready(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004676 return;
4677 }
4678
4679 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4680 /* A retransmit, 2nd most common case. Force an immediate ack. */
Eric Dumazetc10d9312016-04-29 14:16:47 -07004681 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004682 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004683
4684out_of_window:
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004685 tcp_enter_quickack_mode(sk);
4686 inet_csk_schedule_ack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004687drop:
Eric Dumazet532182c2016-04-01 08:52:19 -07004688 tcp_drop(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004689 return;
4690 }
4691
4692 /* Out of window. F.e. zero window probe. */
4693 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4694 goto out_of_window;
4695
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004696 tcp_enter_quickack_mode(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004697
4698 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4699 /* Partial packet, seq < rcv_next < end_seq */
4700 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4701 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4702 TCP_SKB_CB(skb)->end_seq);
4703
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004704 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004705
Linus Torvalds1da177e2005-04-16 15:20:36 -07004706 /* If window is closed, drop tail of packet. But after
4707 * remembering D-SACK for its head made in previous line.
4708 */
4709 if (!tcp_receive_window(tp))
4710 goto out_of_window;
4711 goto queue_and_out;
4712 }
4713
Eric Dumazete86b29192012-03-18 11:06:44 +00004714 tcp_data_queue_ofo(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004715}
4716
Ilpo Järvinen2cf46632008-08-23 05:11:41 -07004717static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4718 struct sk_buff_head *list)
4719{
David S. Miller91521942009-05-28 21:35:47 -07004720 struct sk_buff *next = NULL;
4721
4722 if (!skb_queue_is_last(list, skb))
4723 next = skb_queue_next(list, skb);
Ilpo Järvinen2cf46632008-08-23 05:11:41 -07004724
4725 __skb_unlink(skb, list);
4726 __kfree_skb(skb);
Eric Dumazetc10d9312016-04-29 14:16:47 -07004727 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
Ilpo Järvinen2cf46632008-08-23 05:11:41 -07004728
4729 return next;
4730}
4731
Linus Torvalds1da177e2005-04-16 15:20:36 -07004732/* Collapse contiguous sequence of skbs head..tail with
4733 * sequence numbers start..end.
David S. Miller91521942009-05-28 21:35:47 -07004734 *
4735 * If tail is NULL, this means until the end of the list.
4736 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07004737 * Segments with FIN/SYN are not collapsed (only because this
4738 * simplifies code)
4739 */
4740static void
David S. Miller8728b832005-08-09 19:25:21 -07004741tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4742 struct sk_buff *head, struct sk_buff *tail,
4743 u32 start, u32 end)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004744{
David S. Miller91521942009-05-28 21:35:47 -07004745 struct sk_buff *skb, *n;
4746 bool end_of_skbs;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004747
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08004748 /* First, check that queue is collapsible and find
Linus Torvalds1da177e2005-04-16 15:20:36 -07004749 * the point where collapsing can be useful. */
David S. Miller91521942009-05-28 21:35:47 -07004750 skb = head;
4751restart:
4752 end_of_skbs = true;
4753 skb_queue_walk_from_safe(list, skb, n) {
4754 if (skb == tail)
4755 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004756 /* No new bits? It is possible on ofo queue. */
4757 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
Ilpo Järvinen2cf46632008-08-23 05:11:41 -07004758 skb = tcp_collapse_one(sk, skb, list);
David S. Miller91521942009-05-28 21:35:47 -07004759 if (!skb)
4760 break;
4761 goto restart;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004762 }
4763
4764 /* The first skb to collapse is:
4765 * - not SYN/FIN and
4766 * - bloated or contains data before "start" or
4767 * overlaps to the next one.
4768 */
Eric Dumazete11ecdd2014-09-15 04:19:51 -07004769 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07004770 (tcp_win_from_space(skb->truesize) > skb->len ||
David S. Miller91521942009-05-28 21:35:47 -07004771 before(TCP_SKB_CB(skb)->seq, start))) {
4772 end_of_skbs = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004773 break;
David S. Miller91521942009-05-28 21:35:47 -07004774 }
4775
4776 if (!skb_queue_is_last(list, skb)) {
4777 struct sk_buff *next = skb_queue_next(list, skb);
4778 if (next != tail &&
4779 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4780 end_of_skbs = false;
4781 break;
4782 }
4783 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004784
4785 /* Decided to skip this, advance start seq. */
4786 start = TCP_SKB_CB(skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004787 }
Eric Dumazete11ecdd2014-09-15 04:19:51 -07004788 if (end_of_skbs ||
4789 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004790 return;
4791
4792 while (before(start, end)) {
Eric Dumazetb3d6cb92014-09-15 04:19:53 -07004793 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004794 struct sk_buff *nskb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004795
Eric Dumazetb3d6cb92014-09-15 04:19:53 -07004796 nskb = alloc_skb(copy, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004797 if (!nskb)
4798 return;
Arnaldo Carvalho de Meloc51957d2007-03-10 12:47:22 -03004799
Linus Torvalds1da177e2005-04-16 15:20:36 -07004800 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4801 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
David S. Miller43f59c82008-09-21 21:28:51 -07004802 __skb_queue_before(list, skb, nskb);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08004803 skb_set_owner_r(nskb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004804
4805 /* Copy data, releasing collapsed skbs. */
4806 while (copy > 0) {
4807 int offset = start - TCP_SKB_CB(skb)->seq;
4808 int size = TCP_SKB_CB(skb)->end_seq - start;
4809
Kris Katterjohn09a62662006-01-08 22:24:28 -08004810 BUG_ON(offset < 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004811 if (size > 0) {
4812 size = min(copy, size);
4813 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4814 BUG();
4815 TCP_SKB_CB(nskb)->end_seq += size;
4816 copy -= size;
4817 start += size;
4818 }
4819 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
Ilpo Järvinen2cf46632008-08-23 05:11:41 -07004820 skb = tcp_collapse_one(sk, skb, list);
David S. Miller91521942009-05-28 21:35:47 -07004821 if (!skb ||
4822 skb == tail ||
Eric Dumazete11ecdd2014-09-15 04:19:51 -07004823 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004824 return;
4825 }
4826 }
4827 }
4828}
4829
4830/* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4831 * and tcp_collapse() them until all the queue is collapsed.
4832 */
4833static void tcp_collapse_ofo_queue(struct sock *sk)
4834{
4835 struct tcp_sock *tp = tcp_sk(sk);
4836 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4837 struct sk_buff *head;
4838 u32 start, end;
4839
Ian Morris51456b22015-04-03 09:17:26 +01004840 if (!skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004841 return;
4842
4843 start = TCP_SKB_CB(skb)->seq;
4844 end = TCP_SKB_CB(skb)->end_seq;
4845 head = skb;
4846
4847 for (;;) {
David S. Miller91521942009-05-28 21:35:47 -07004848 struct sk_buff *next = NULL;
4849
4850 if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4851 next = skb_queue_next(&tp->out_of_order_queue, skb);
4852 skb = next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004853
4854 /* Segment is terminated when we see gap or when
4855 * we are at the end of all the queue. */
David S. Miller91521942009-05-28 21:35:47 -07004856 if (!skb ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07004857 after(TCP_SKB_CB(skb)->seq, end) ||
4858 before(TCP_SKB_CB(skb)->end_seq, start)) {
David S. Miller8728b832005-08-09 19:25:21 -07004859 tcp_collapse(sk, &tp->out_of_order_queue,
4860 head, skb, start, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004861 head = skb;
David S. Miller91521942009-05-28 21:35:47 -07004862 if (!skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004863 break;
4864 /* Start new segment */
4865 start = TCP_SKB_CB(skb)->seq;
4866 end = TCP_SKB_CB(skb)->end_seq;
4867 } else {
4868 if (before(TCP_SKB_CB(skb)->seq, start))
4869 start = TCP_SKB_CB(skb)->seq;
4870 if (after(TCP_SKB_CB(skb)->end_seq, end))
4871 end = TCP_SKB_CB(skb)->end_seq;
4872 }
4873 }
4874}
4875
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004876/*
4877 * Purge the out-of-order queue.
Vitaliy Gusev56f367b2008-04-15 20:26:34 -07004878 * Return true if queue was pruned.
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004879 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004880static bool tcp_prune_ofo_queue(struct sock *sk)
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004881{
4882 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004883 bool res = false;
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004884
4885 if (!skb_queue_empty(&tp->out_of_order_queue)) {
Eric Dumazetc10d9312016-04-29 14:16:47 -07004886 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004887 __skb_queue_purge(&tp->out_of_order_queue);
4888
4889 /* Reset SACK state. A conforming SACK implementation will
4890 * do the same at a timeout based retransmit. When a connection
4891 * is in a sad state like this, we care only about integrity
4892 * of the connection not performance.
4893 */
4894 if (tp->rx_opt.sack_ok)
4895 tcp_sack_reset(&tp->rx_opt);
4896 sk_mem_reclaim(sk);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004897 res = true;
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004898 }
Vitaliy Gusev56f367b2008-04-15 20:26:34 -07004899 return res;
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004900}
4901
Linus Torvalds1da177e2005-04-16 15:20:36 -07004902/* Reduce allocated memory if we can, trying to get
4903 * the socket within its memory limits again.
4904 *
4905 * Return less than zero if we should start dropping frames
4906 * until the socket owning process reads some of the data
4907 * to stabilize the situation.
4908 */
4909static int tcp_prune_queue(struct sock *sk)
4910{
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004911 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004912
4913 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4914
Eric Dumazetc10d9312016-04-29 14:16:47 -07004915 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004916
4917 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004918 tcp_clamp_window(sk);
Eric Dumazetb8da51e2015-05-15 12:39:27 -07004919 else if (tcp_under_memory_pressure(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004920 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4921
4922 tcp_collapse_ofo_queue(sk);
David S. Miller91521942009-05-28 21:35:47 -07004923 if (!skb_queue_empty(&sk->sk_receive_queue))
4924 tcp_collapse(sk, &sk->sk_receive_queue,
4925 skb_peek(&sk->sk_receive_queue),
4926 NULL,
4927 tp->copied_seq, tp->rcv_nxt);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08004928 sk_mem_reclaim(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004929
4930 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4931 return 0;
4932
4933 /* Collapsing did not help, destructive actions follow.
4934 * This must not ever occur. */
4935
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004936 tcp_prune_ofo_queue(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004937
4938 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4939 return 0;
4940
4941 /* If we are really being abused, tell the caller to silently
4942 * drop receive data on the floor. It will get retransmitted
4943 * and hopefully then we'll have sufficient space.
4944 */
Eric Dumazetc10d9312016-04-29 14:16:47 -07004945 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004946
4947 /* Massive buffer overcommit. */
4948 tp->pred_flags = 0;
4949 return -1;
4950}
4951
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004952static bool tcp_should_expand_sndbuf(const struct sock *sk)
David S. Miller0d9901d2005-07-05 15:21:10 -07004953{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04004954 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004955
David S. Miller0d9901d2005-07-05 15:21:10 -07004956 /* If the user specified a specific send buffer setting, do
4957 * not modify it.
4958 */
4959 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004960 return false;
David S. Miller0d9901d2005-07-05 15:21:10 -07004961
4962 /* If we are under global TCP memory pressure, do not expand. */
Eric Dumazetb8da51e2015-05-15 12:39:27 -07004963 if (tcp_under_memory_pressure(sk))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004964 return false;
David S. Miller0d9901d2005-07-05 15:21:10 -07004965
4966 /* If we are under soft global TCP memory pressure, do not expand. */
Glauber Costa180d8cd2011-12-11 21:47:02 +00004967 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004968 return false;
David S. Miller0d9901d2005-07-05 15:21:10 -07004969
4970 /* If we filled the congestion window, do not expand. */
Neal Cardwell65148902015-02-20 13:33:16 -05004971 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004972 return false;
David S. Miller0d9901d2005-07-05 15:21:10 -07004973
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004974 return true;
David S. Miller0d9901d2005-07-05 15:21:10 -07004975}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004976
4977/* When incoming ACK allowed to free some skb from write_queue,
4978 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4979 * on the exit from tcp input handler.
4980 *
4981 * PROBLEM: sndbuf expansion does not work well with largesend.
4982 */
4983static void tcp_new_space(struct sock *sk)
4984{
4985 struct tcp_sock *tp = tcp_sk(sk);
4986
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004987 if (tcp_should_expand_sndbuf(sk)) {
Eric Dumazet6ae70532013-10-01 10:23:44 -07004988 tcp_sndbuf_expand(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004989 tp->snd_cwnd_stamp = tcp_time_stamp;
4990 }
4991
4992 sk->sk_write_space(sk);
4993}
4994
Stephen Hemminger40efc6f2006-01-03 16:03:49 -08004995static void tcp_check_space(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004996{
4997 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4998 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
jbaron@akamai.com3c715122015-04-20 20:05:07 +00004999 /* pairs with tcp_poll() */
5000 smp_mb__after_atomic();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005001 if (sk->sk_socket &&
5002 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5003 tcp_new_space(sk);
5004 }
5005}
5006
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005007static inline void tcp_data_snd_check(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005008{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005009 tcp_push_pending_frames(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005010 tcp_check_space(sk);
5011}
5012
5013/*
5014 * Check if sending an ack is needed.
5015 */
5016static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5017{
5018 struct tcp_sock *tp = tcp_sk(sk);
5019
5020 /* More than one full frame received... */
Joe Perches9d4fb272009-11-23 10:41:23 -08005021 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07005022 /* ... and right edge of window advances far enough.
5023 * (tcp_recvmsg() will send ACK otherwise). Or...
5024 */
Joe Perches9d4fb272009-11-23 10:41:23 -08005025 __tcp_select_window(sk) >= tp->rcv_wnd) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07005026 /* We ACK each frame or... */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005027 tcp_in_quickack_mode(sk) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07005028 /* We have out of order data. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005029 (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005030 /* Then ack it now */
5031 tcp_send_ack(sk);
5032 } else {
5033 /* Else, send delayed ack. */
5034 tcp_send_delayed_ack(sk);
5035 }
5036}
5037
Stephen Hemminger40efc6f2006-01-03 16:03:49 -08005038static inline void tcp_ack_snd_check(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005039{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005040 if (!inet_csk_ack_scheduled(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005041 /* We sent a data segment already. */
5042 return;
5043 }
5044 __tcp_ack_snd_check(sk, 1);
5045}
5046
5047/*
5048 * This routine is only called when we have urgent data
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08005049 * signaled. Its the 'slow' part of tcp_urg. It could be
Linus Torvalds1da177e2005-04-16 15:20:36 -07005050 * moved inline now as tcp_urg is only called from one
5051 * place. We handle URGent data wrong. We have to - as
5052 * BSD still doesn't use the correction from RFC961.
5053 * For 1003.1g we should support a new option TCP_STDURG to permit
5054 * either form (or just set the sysctl tcp_stdurg).
5055 */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005056
Eric Dumazetcf533ea2011-10-21 05:22:42 -04005057static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005058{
5059 struct tcp_sock *tp = tcp_sk(sk);
5060 u32 ptr = ntohs(th->urg_ptr);
5061
5062 if (ptr && !sysctl_tcp_stdurg)
5063 ptr--;
5064 ptr += ntohl(th->seq);
5065
5066 /* Ignore urgent data that we've already seen and read. */
5067 if (after(tp->copied_seq, ptr))
5068 return;
5069
5070 /* Do not replay urg ptr.
5071 *
5072 * NOTE: interesting situation not covered by specs.
5073 * Misbehaving sender may send urg ptr, pointing to segment,
5074 * which we already have in ofo queue. We are not able to fetch
5075 * such data and will stay in TCP_URG_NOTYET until will be eaten
5076 * by recvmsg(). Seems, we are not obliged to handle such wicked
5077 * situations. But it is worth to think about possibility of some
5078 * DoSes using some hypothetical application level deadlock.
5079 */
5080 if (before(ptr, tp->rcv_nxt))
5081 return;
5082
5083 /* Do we already have a newer (or duplicate) urgent pointer? */
5084 if (tp->urg_data && !after(ptr, tp->urg_seq))
5085 return;
5086
5087 /* Tell the world about our new urgent pointer. */
5088 sk_send_sigurg(sk);
5089
5090 /* We may be adding urgent data when the last byte read was
5091 * urgent. To do this requires some care. We cannot just ignore
5092 * tp->copied_seq since we would read the last urgent byte again
5093 * as data, nor can we alter copied_seq until this data arrives
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08005094 * or we break the semantics of SIOCATMARK (and thus sockatmark())
Linus Torvalds1da177e2005-04-16 15:20:36 -07005095 *
5096 * NOTE. Double Dutch. Rendering to plain English: author of comment
5097 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5098 * and expect that both A and B disappear from stream. This is _wrong_.
5099 * Though this happens in BSD with high probability, this is occasional.
5100 * Any application relying on this is buggy. Note also, that fix "works"
5101 * only in this artificial test. Insert some normal data between A and B and we will
5102 * decline of BSD again. Verdict: it is better to remove to trap
5103 * buggy users.
5104 */
5105 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005106 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005107 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5108 tp->copied_seq++;
5109 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
David S. Miller8728b832005-08-09 19:25:21 -07005110 __skb_unlink(skb, &sk->sk_receive_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005111 __kfree_skb(skb);
5112 }
5113 }
5114
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005115 tp->urg_data = TCP_URG_NOTYET;
5116 tp->urg_seq = ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005117
5118 /* Disable header prediction. */
5119 tp->pred_flags = 0;
5120}
5121
5122/* This is the 'fast' part of urgent handling. */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04005123static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005124{
5125 struct tcp_sock *tp = tcp_sk(sk);
5126
5127 /* Check if we get a new urgent pointer - normally not. */
5128 if (th->urg)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005129 tcp_check_urg(sk, th);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005130
5131 /* Do we wait for any urgent data? - normally not... */
5132 if (tp->urg_data == TCP_URG_NOTYET) {
5133 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5134 th->syn;
5135
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005136 /* Is the urgent pointer pointing into this packet? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005137 if (ptr < skb->len) {
5138 u8 tmp;
5139 if (skb_copy_bits(skb, ptr, &tmp, 1))
5140 BUG();
5141 tp->urg_data = TCP_URG_VALID | tmp;
5142 if (!sock_flag(sk, SOCK_DEAD))
David S. Miller676d2362014-04-11 16:15:36 -04005143 sk->sk_data_ready(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005144 }
5145 }
5146}
5147
5148static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
5149{
5150 struct tcp_sock *tp = tcp_sk(sk);
5151 int chunk = skb->len - hlen;
5152 int err;
5153
Herbert Xu60476372007-04-09 11:59:39 -07005154 if (skb_csum_unnecessary(skb))
Al Virof4362a22014-11-24 13:26:06 -05005155 err = skb_copy_datagram_msg(skb, hlen, tp->ucopy.msg, chunk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005156 else
Al Virof4362a22014-11-24 13:26:06 -05005157 err = skb_copy_and_csum_datagram_msg(skb, hlen, tp->ucopy.msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005158
5159 if (!err) {
5160 tp->ucopy.len -= chunk;
5161 tp->copied_seq += chunk;
5162 tcp_rcv_space_adjust(sk);
5163 }
5164
Linus Torvalds1da177e2005-04-16 15:20:36 -07005165 return err;
5166}
5167
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005168/* Does PAWS and seqno based validation of an incoming segment, flags will
5169 * play significant role here.
5170 */
Eric Dumazet0c246042012-07-17 01:41:30 +00005171static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5172 const struct tcphdr *th, int syn_inerr)
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005173{
5174 struct tcp_sock *tp = tcp_sk(sk);
Pau Espin Pedrole00431b2016-06-07 16:30:34 +02005175 bool rst_seq_match = false;
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005176
5177 /* RFC1323: H1. Apply PAWS check first. */
Christoph Paasch1a2c6182013-03-17 08:23:34 +00005178 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005179 tcp_paws_discard(sk, skb)) {
5180 if (!th->rst) {
Eric Dumazetc10d9312016-04-29 14:16:47 -07005181 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
Neal Cardwellf2b2c582015-02-06 16:04:40 -05005182 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5183 LINUX_MIB_TCPACKSKIPPEDPAWS,
5184 &tp->last_oow_ack_time))
5185 tcp_send_dupack(sk, skb);
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005186 goto discard;
5187 }
5188 /* Reset is accepted even if it did not pass PAWS. */
5189 }
5190
5191 /* Step 1: check sequence number */
5192 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5193 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5194 * (RST) segments are validated by checking their SEQ-fields."
5195 * And page 69: "If an incoming segment is not acceptable,
5196 * an acknowledgment should be sent in reply (unless the RST
5197 * bit is set, if so drop the segment and return)".
5198 */
Eric Dumazete3715892012-07-17 12:29:30 +00005199 if (!th->rst) {
5200 if (th->syn)
5201 goto syn_challenge;
Neal Cardwellf2b2c582015-02-06 16:04:40 -05005202 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5203 LINUX_MIB_TCPACKSKIPPEDSEQ,
5204 &tp->last_oow_ack_time))
5205 tcp_send_dupack(sk, skb);
Eric Dumazete3715892012-07-17 12:29:30 +00005206 }
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005207 goto discard;
5208 }
5209
5210 /* Step 2: check RST bit */
5211 if (th->rst) {
Pau Espin Pedrole00431b2016-06-07 16:30:34 +02005212 /* RFC 5961 3.2 (extend to match against SACK too if available):
5213 * If seq num matches RCV.NXT or the right-most SACK block,
5214 * then
Eric Dumazet282f23c2012-07-17 10:13:05 +02005215 * RESET the connection
5216 * else
5217 * Send a challenge ACK
5218 */
Pau Espin Pedrole00431b2016-06-07 16:30:34 +02005219 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
5220 rst_seq_match = true;
5221 } else if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
5222 struct tcp_sack_block *sp = &tp->selective_acks[0];
5223 int max_sack = sp[0].end_seq;
5224 int this_sack;
5225
5226 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
5227 ++this_sack) {
5228 max_sack = after(sp[this_sack].end_seq,
5229 max_sack) ?
5230 sp[this_sack].end_seq : max_sack;
5231 }
5232
5233 if (TCP_SKB_CB(skb)->seq == max_sack)
5234 rst_seq_match = true;
5235 }
5236
5237 if (rst_seq_match)
Eric Dumazet282f23c2012-07-17 10:13:05 +02005238 tcp_reset(sk);
5239 else
Neal Cardwellf2b2c582015-02-06 16:04:40 -05005240 tcp_send_challenge_ack(sk, skb);
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005241 goto discard;
5242 }
5243
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005244 /* step 3: check security and precedence [ignored] */
5245
Eric Dumazet0c246042012-07-17 01:41:30 +00005246 /* step 4: Check for a SYN
Sowmini Varadhancd214532014-10-30 12:48:08 -04005247 * RFC 5961 4.2 : Send a challenge ack
Eric Dumazet0c246042012-07-17 01:41:30 +00005248 */
5249 if (th->syn) {
Eric Dumazete3715892012-07-17 12:29:30 +00005250syn_challenge:
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005251 if (syn_inerr)
Eric Dumazetc10d9312016-04-29 14:16:47 -07005252 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5253 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
Neal Cardwellf2b2c582015-02-06 16:04:40 -05005254 tcp_send_challenge_ack(sk, skb);
Eric Dumazet0c246042012-07-17 01:41:30 +00005255 goto discard;
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005256 }
5257
Eric Dumazet0c246042012-07-17 01:41:30 +00005258 return true;
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005259
5260discard:
Eric Dumazet532182c2016-04-01 08:52:19 -07005261 tcp_drop(sk, skb);
Eric Dumazet0c246042012-07-17 01:41:30 +00005262 return false;
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005263}
5264
Linus Torvalds1da177e2005-04-16 15:20:36 -07005265/*
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005266 * TCP receive function for the ESTABLISHED state.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005267 *
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005268 * It is split into a fast path and a slow path. The fast path is
Linus Torvalds1da177e2005-04-16 15:20:36 -07005269 * disabled when:
5270 * - A zero window was announced from us - zero window probing
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005271 * is only handled properly in the slow path.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005272 * - Out of order segments arrived.
5273 * - Urgent data is expected.
5274 * - There is no buffer space left
5275 * - Unexpected TCP flags/window values/header lengths are received
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005276 * (detected by checking the TCP header against pred_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005277 * - Data is sent in both directions. Fast path only supports pure senders
5278 * or pure receivers (this means either the sequence number or the ack
5279 * value must stay constant)
5280 * - Unexpected TCP option.
5281 *
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005282 * When these conditions are not satisfied it drops into a standard
Linus Torvalds1da177e2005-04-16 15:20:36 -07005283 * receive procedure patterned after RFC793 to handle all cases.
5284 * The first three cases are guaranteed by proper pred_flags setting,
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005285 * the rest is checked inline. Fast processing is turned on in
Linus Torvalds1da177e2005-04-16 15:20:36 -07005286 * tcp_data_queue when everything is OK.
5287 */
Vijay Subramanianc995ae22013-09-03 12:23:22 -07005288void tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5289 const struct tcphdr *th, unsigned int len)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005290{
5291 struct tcp_sock *tp = tcp_sk(sk);
5292
Ian Morris51456b22015-04-03 09:17:26 +01005293 if (unlikely(!sk->sk_rx_dst))
Eric Dumazet5d299f32012-08-06 05:09:33 +00005294 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005295 /*
5296 * Header prediction.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005297 * The code loosely follows the one in the famous
Linus Torvalds1da177e2005-04-16 15:20:36 -07005298 * "30 instruction TCP receive" Van Jacobson mail.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005299 *
5300 * Van's trick is to deposit buffers into socket queue
Linus Torvalds1da177e2005-04-16 15:20:36 -07005301 * on a device interrupt, to call tcp_recv function
5302 * on the receive process context and checksum and copy
5303 * the buffer to user space. smart...
5304 *
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005305 * Our current scheme is not silly either but we take the
Linus Torvalds1da177e2005-04-16 15:20:36 -07005306 * extra cost of the net_bh soft interrupt processing...
5307 * We do checksum and copy also but from device to kernel.
5308 */
5309
5310 tp->rx_opt.saw_tstamp = 0;
5311
5312 /* pred_flags is 0xS?10 << 16 + snd_wnd
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08005313 * if header_prediction is to be made
Linus Torvalds1da177e2005-04-16 15:20:36 -07005314 * 'S' will always be tp->tcp_header_len >> 2
5315 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005316 * turn it off (when there are holes in the receive
Linus Torvalds1da177e2005-04-16 15:20:36 -07005317 * space for instance)
5318 * PSH flag is ignored.
5319 */
5320
5321 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
John Dykstra96e0bf42009-03-22 21:49:57 -07005322 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5323 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005324 int tcp_header_len = tp->tcp_header_len;
5325
5326 /* Timestamp header prediction: tcp_header_len
5327 * is automatically equal to th->doff*4 due to pred_flags
5328 * match.
5329 */
5330
5331 /* Check timestamp */
5332 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005333 /* No? Slow path! */
Ilpo Järvinena4356b22008-08-23 05:12:29 -07005334 if (!tcp_parse_aligned_timestamp(tp, th))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005335 goto slow_path;
5336
Linus Torvalds1da177e2005-04-16 15:20:36 -07005337 /* If PAWS failed, check it more carefully in slow path */
5338 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5339 goto slow_path;
5340
5341 /* DO NOT update ts_recent here, if checksum fails
5342 * and timestamp was corrupted part, it will result
5343 * in a hung connection since we will drop all
5344 * future packets due to the PAWS test.
5345 */
5346 }
5347
5348 if (len <= tcp_header_len) {
5349 /* Bulk data transfer: sender */
5350 if (len == tcp_header_len) {
5351 /* Predicted packet is in window by definition.
5352 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5353 * Hence, check seq<=rcv_wup reduces to:
5354 */
5355 if (tcp_header_len ==
5356 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5357 tp->rcv_nxt == tp->rcv_wup)
5358 tcp_store_ts_recent(tp);
5359
Linus Torvalds1da177e2005-04-16 15:20:36 -07005360 /* We know that such packets are checksummed
5361 * on entry.
5362 */
5363 tcp_ack(sk, skb, 0);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005364 __kfree_skb(skb);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005365 tcp_data_snd_check(sk);
Vijay Subramanianc995ae22013-09-03 12:23:22 -07005366 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005367 } else { /* Header too small */
Eric Dumazetc10d9312016-04-29 14:16:47 -07005368 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005369 goto discard;
5370 }
5371 } else {
5372 int eaten = 0;
Eric Dumazetb081f852012-05-02 09:58:29 +00005373 bool fragstolen = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005374
Dan Williamsd27f9bc2013-12-30 11:37:15 -08005375 if (tp->ucopy.task == current &&
5376 tp->copied_seq == tp->rcv_nxt &&
5377 len - tcp_header_len <= tp->ucopy.len &&
5378 sock_owned_by_user(sk)) {
5379 __set_current_state(TASK_RUNNING);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005380
Dan Williamsd27f9bc2013-12-30 11:37:15 -08005381 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005382 /* Predicted packet is in window by definition.
5383 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5384 * Hence, check seq<=rcv_wup reduces to:
5385 */
5386 if (tcp_header_len ==
5387 (sizeof(struct tcphdr) +
5388 TCPOLEN_TSTAMP_ALIGNED) &&
5389 tp->rcv_nxt == tp->rcv_wup)
5390 tcp_store_ts_recent(tp);
5391
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005392 tcp_rcv_rtt_measure_ts(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005393
5394 __skb_pull(skb, tcp_header_len);
Eric Dumazetbdd1f9e2015-04-28 15:28:18 -07005395 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
Eric Dumazetc10d9312016-04-29 14:16:47 -07005396 NET_INC_STATS(sock_net(sk),
Eric Dumazet02a1d6e2016-04-27 16:44:39 -07005397 LINUX_MIB_TCPHPHITSTOUSER);
Dan Williamsd27f9bc2013-12-30 11:37:15 -08005398 eaten = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005399 }
5400 }
5401 if (!eaten) {
Eric Dumazetfb3477c2016-04-29 14:16:48 -07005402 if (tcp_checksum_complete(skb))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005403 goto csum_error;
5404
Neal Cardwellaab2b4b2013-03-04 06:23:05 +00005405 if ((int)skb->truesize > sk->sk_forward_alloc)
5406 goto step5;
5407
Linus Torvalds1da177e2005-04-16 15:20:36 -07005408 /* Predicted packet is in window by definition.
5409 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5410 * Hence, check seq<=rcv_wup reduces to:
5411 */
5412 if (tcp_header_len ==
5413 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5414 tp->rcv_nxt == tp->rcv_wup)
5415 tcp_store_ts_recent(tp);
5416
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005417 tcp_rcv_rtt_measure_ts(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005418
Eric Dumazetc10d9312016-04-29 14:16:47 -07005419 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005420
5421 /* Bulk data transfer: receiver */
Eric Dumazetb081f852012-05-02 09:58:29 +00005422 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5423 &fragstolen);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005424 }
5425
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005426 tcp_event_data_recv(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005427
5428 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5429 /* Well, only one small jumplet in fast path... */
5430 tcp_ack(sk, skb, FLAG_DATA);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005431 tcp_data_snd_check(sk);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005432 if (!inet_csk_ack_scheduled(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005433 goto no_ack;
5434 }
5435
Dan Williamsd27f9bc2013-12-30 11:37:15 -08005436 __tcp_ack_snd_check(sk, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005437no_ack:
5438 if (eaten)
Eric Dumazetb081f852012-05-02 09:58:29 +00005439 kfree_skb_partial(skb, fragstolen);
David S. Miller676d2362014-04-11 16:15:36 -04005440 sk->sk_data_ready(sk);
Vijay Subramanianc995ae22013-09-03 12:23:22 -07005441 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005442 }
5443 }
5444
5445slow_path:
Eric Dumazetfb3477c2016-04-29 14:16:48 -07005446 if (len < (th->doff << 2) || tcp_checksum_complete(skb))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005447 goto csum_error;
5448
Calvin Owens0c228e82014-11-20 15:09:53 -08005449 if (!th->ack && !th->rst && !th->syn)
Eric Dumazetc3ae62af2012-12-26 12:44:34 +00005450 goto discard;
5451
Linus Torvalds1da177e2005-04-16 15:20:36 -07005452 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07005453 * Standard slow path.
5454 */
5455
Eric Dumazet0c246042012-07-17 01:41:30 +00005456 if (!tcp_validate_incoming(sk, skb, th, 1))
Vijay Subramanianc995ae22013-09-03 12:23:22 -07005457 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005458
5459step5:
Eric Dumazet12fb3dd2013-04-19 07:19:48 +00005460 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
John Dykstra96e0bf42009-03-22 21:49:57 -07005461 goto discard;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005462
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005463 tcp_rcv_rtt_measure_ts(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005464
5465 /* Process urgent data. */
5466 tcp_urg(sk, skb, th);
5467
5468 /* step 7: process the segment text */
5469 tcp_data_queue(sk, skb);
5470
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005471 tcp_data_snd_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005472 tcp_ack_snd_check(sk);
Vijay Subramanianc995ae22013-09-03 12:23:22 -07005473 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005474
5475csum_error:
Eric Dumazetc10d9312016-04-29 14:16:47 -07005476 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
5477 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005478
5479discard:
Eric Dumazet532182c2016-04-01 08:52:19 -07005480 tcp_drop(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005481}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00005482EXPORT_SYMBOL(tcp_rcv_established);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005483
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005484void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5485{
5486 struct tcp_sock *tp = tcp_sk(sk);
5487 struct inet_connection_sock *icsk = inet_csk(sk);
5488
5489 tcp_set_state(sk, TCP_ESTABLISHED);
5490
Ian Morris00db4122015-04-03 09:17:27 +01005491 if (skb) {
Eric Dumazet5d299f32012-08-06 05:09:33 +00005492 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005493 security_inet_conn_established(sk, skb);
David S. Miller41063e92012-06-19 21:22:05 -07005494 }
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005495
5496 /* Make sure socket is routed, for correct metrics. */
5497 icsk->icsk_af_ops->rebuild_header(sk);
5498
5499 tcp_init_metrics(sk);
5500
5501 tcp_init_congestion_control(sk);
5502
5503 /* Prevent spurious tcp_cwnd_restart() on first data
5504 * packet.
5505 */
5506 tp->lsndtime = tcp_time_stamp;
5507
5508 tcp_init_buffer_space(sk);
5509
5510 if (sock_flag(sk, SOCK_KEEPOPEN))
5511 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5512
5513 if (!tp->rx_opt.snd_wscale)
5514 __tcp_fast_path_on(tp, tp->snd_wnd);
5515 else
5516 tp->pred_flags = 0;
5517
5518 if (!sock_flag(sk, SOCK_DEAD)) {
5519 sk->sk_state_change(sk);
5520 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5521 }
5522}
5523
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005524static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5525 struct tcp_fastopen_cookie *cookie)
5526{
5527 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng67da22d2012-07-19 06:43:11 +00005528 struct sk_buff *data = tp->syn_data ? tcp_write_queue_head(sk) : NULL;
Daniel Lee2646c832015-04-06 14:37:27 -07005529 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
5530 bool syn_drop = false;
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005531
5532 if (mss == tp->rx_opt.user_mss) {
5533 struct tcp_options_received opt;
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005534
5535 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5536 tcp_clear_options(&opt);
5537 opt.user_mss = opt.mss_clamp = 0;
Christoph Paasch1a2c6182013-03-17 08:23:34 +00005538 tcp_parse_options(synack, &opt, 0, NULL);
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005539 mss = opt.mss_clamp;
5540 }
5541
Daniel Lee2646c832015-04-06 14:37:27 -07005542 if (!tp->syn_fastopen) {
5543 /* Ignore an unsolicited cookie */
Yuchung Cheng67da22d2012-07-19 06:43:11 +00005544 cookie->len = -1;
Daniel Lee2646c832015-04-06 14:37:27 -07005545 } else if (tp->total_retrans) {
5546 /* SYN timed out and the SYN-ACK neither has a cookie nor
5547 * acknowledges data. Presumably the remote received only
5548 * the retransmitted (regular) SYNs: either the original
5549 * SYN-data or the corresponding SYN-ACK was dropped.
5550 */
5551 syn_drop = (cookie->len < 0 && data);
5552 } else if (cookie->len < 0 && !tp->syn_data) {
5553 /* We requested a cookie but didn't get it. If we did not use
5554 * the (old) exp opt format then try so next time (try_exp=1).
5555 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5556 */
5557 try_exp = tp->syn_fastopen_exp ? 2 : 1;
5558 }
Yuchung Cheng67da22d2012-07-19 06:43:11 +00005559
Daniel Lee2646c832015-04-06 14:37:27 -07005560 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005561
5562 if (data) { /* Retransmit unacked data in SYN */
Yuchung Cheng93b174a2012-12-06 08:45:32 +00005563 tcp_for_write_queue_from(data, sk) {
5564 if (data == tcp_send_head(sk) ||
Eric Dumazet10d3be52016-04-21 10:55:23 -07005565 __tcp_retransmit_skb(sk, data, 1))
Yuchung Cheng93b174a2012-12-06 08:45:32 +00005566 break;
5567 }
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005568 tcp_rearm_rto(sk);
Eric Dumazetc10d9312016-04-29 14:16:47 -07005569 NET_INC_STATS(sock_net(sk),
Eric Dumazet02a1d6e2016-04-27 16:44:39 -07005570 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005571 return true;
5572 }
Yuchung Cheng6f736012012-10-19 15:14:44 +00005573 tp->syn_data_acked = tp->syn_data;
Yuchung Chengf19c29e2014-03-03 12:31:36 -08005574 if (tp->syn_data_acked)
Eric Dumazetc10d9312016-04-29 14:16:47 -07005575 NET_INC_STATS(sock_net(sk),
Eric Dumazet02a1d6e2016-04-27 16:44:39 -07005576 LINUX_MIB_TCPFASTOPENACTIVE);
Eric Dumazet61d2bca2016-02-01 21:03:07 -08005577
5578 tcp_fastopen_add_skb(sk, synack);
5579
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005580 return false;
5581}
5582
Linus Torvalds1da177e2005-04-16 15:20:36 -07005583static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
Eric Dumazetbda07a62015-09-29 07:42:40 -07005584 const struct tcphdr *th)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005585{
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08005586 struct inet_connection_sock *icsk = inet_csk(sk);
William Allen Simpson4957faade2009-12-02 18:25:27 +00005587 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005588 struct tcp_fastopen_cookie foc = { .len = -1 };
William Allen Simpson4957faade2009-12-02 18:25:27 +00005589 int saved_clamp = tp->rx_opt.mss_clamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005590
Christoph Paasch1a2c6182013-03-17 08:23:34 +00005591 tcp_parse_options(skb, &tp->rx_opt, 0, &foc);
Andrew Vagine3e12022013-08-27 12:21:55 +04005592 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
Andrey Vaginee684b62013-02-11 05:50:19 +00005593 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005594
5595 if (th->ack) {
5596 /* rfc793:
5597 * "If the state is SYN-SENT then
5598 * first check the ACK bit
5599 * If the ACK bit is set
5600 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5601 * a reset (unless the RST bit is set, if so drop
5602 * the segment and return)"
Linus Torvalds1da177e2005-04-16 15:20:36 -07005603 */
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005604 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5605 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005606 goto reset_and_undo;
5607
5608 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5609 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5610 tcp_time_stamp)) {
Eric Dumazetc10d9312016-04-29 14:16:47 -07005611 NET_INC_STATS(sock_net(sk),
Eric Dumazet02a1d6e2016-04-27 16:44:39 -07005612 LINUX_MIB_PAWSACTIVEREJECTED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005613 goto reset_and_undo;
5614 }
5615
5616 /* Now ACK is acceptable.
5617 *
5618 * "If the RST bit is set
5619 * If the ACK was acceptable then signal the user "error:
5620 * connection reset", drop the segment, enter CLOSED state,
5621 * delete TCB, and return."
5622 */
5623
5624 if (th->rst) {
5625 tcp_reset(sk);
5626 goto discard;
5627 }
5628
5629 /* rfc793:
5630 * "fifth, if neither of the SYN or RST bits is set then
5631 * drop the segment and return."
5632 *
5633 * See note below!
5634 * --ANK(990513)
5635 */
5636 if (!th->syn)
5637 goto discard_and_undo;
5638
5639 /* rfc793:
5640 * "If the SYN bit is on ...
5641 * are acceptable then ...
5642 * (our SYN has been ACKed), change the connection
5643 * state to ESTABLISHED..."
5644 */
5645
Florian Westphal735d3832014-09-29 13:08:30 +02005646 tcp_ecn_rcv_synack(tp, th);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005647
Razvan Ghitulete1b3a6922012-08-14 16:30:20 +03005648 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005649 tcp_ack(sk, skb, FLAG_SLOWPATH);
5650
5651 /* Ok.. it's good. Set up sequence numbers and
5652 * move to established.
5653 */
5654 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5655 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5656
5657 /* RFC1323: The window in SYN & SYN/ACK segments is
5658 * never scaled.
5659 */
5660 tp->snd_wnd = ntohs(th->window);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005661
5662 if (!tp->rx_opt.wscale_ok) {
5663 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5664 tp->window_clamp = min(tp->window_clamp, 65535U);
5665 }
5666
5667 if (tp->rx_opt.saw_tstamp) {
5668 tp->rx_opt.tstamp_ok = 1;
5669 tp->tcp_header_len =
5670 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5671 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5672 tcp_store_ts_recent(tp);
5673 } else {
5674 tp->tcp_header_len = sizeof(struct tcphdr);
5675 }
5676
Ilpo Järvinene60402d2007-08-09 15:14:46 +03005677 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5678 tcp_enable_fack(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005679
John Heffner5d424d52006-03-20 17:53:41 -08005680 tcp_mtup_init(sk);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08005681 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005682 tcp_initialize_rcv_mss(sk);
5683
5684 /* Remember, tcp_poll() does not lock socket!
5685 * Change state from SYN-SENT only after copied_seq
5686 * is initialized. */
5687 tp->copied_seq = tp->rcv_nxt;
William Allen Simpson4957faade2009-12-02 18:25:27 +00005688
Ralf Baechlee16aa202006-12-07 00:11:33 -08005689 smp_mb();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005690
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005691 tcp_finish_connect(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005692
Yuchung Cheng67da22d2012-07-19 06:43:11 +00005693 if ((tp->syn_fastopen || tp->syn_data) &&
5694 tcp_rcv_fastopen_synack(sk, skb, &foc))
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005695 return -1;
5696
Arnaldo Carvalho de Melo295f7322005-08-09 20:11:56 -07005697 if (sk->sk_write_pending ||
5698 icsk->icsk_accept_queue.rskq_defer_accept ||
5699 icsk->icsk_ack.pingpong) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005700 /* Save one ACK. Data will be ready after
5701 * several ticks, if write_pending is set.
5702 *
5703 * It may be deleted, but with this feature tcpdumps
5704 * look so _wonderfully_ clever, that I was not able
5705 * to stand against the temptation 8) --ANK
5706 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005707 inet_csk_schedule_ack(sk);
Arnaldo Carvalho de Melo295f7322005-08-09 20:11:56 -07005708 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005709 tcp_enter_quickack_mode(sk);
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07005710 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5711 TCP_DELACK_MAX, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005712
5713discard:
Eric Dumazet532182c2016-04-01 08:52:19 -07005714 tcp_drop(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005715 return 0;
5716 } else {
5717 tcp_send_ack(sk);
5718 }
5719 return -1;
5720 }
5721
5722 /* No ACK in the segment */
5723
5724 if (th->rst) {
5725 /* rfc793:
5726 * "If the RST bit is set
5727 *
5728 * Otherwise (no ACK) drop the segment and return."
5729 */
5730
5731 goto discard_and_undo;
5732 }
5733
5734 /* PAWS check. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005735 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
Ilpo Järvinenc887e6d2009-03-14 14:23:03 +00005736 tcp_paws_reject(&tp->rx_opt, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005737 goto discard_and_undo;
5738
5739 if (th->syn) {
5740 /* We see SYN without ACK. It is attempt of
5741 * simultaneous connect with crossed SYNs.
5742 * Particularly, it can be connect to self.
5743 */
5744 tcp_set_state(sk, TCP_SYN_RECV);
5745
5746 if (tp->rx_opt.saw_tstamp) {
5747 tp->rx_opt.tstamp_ok = 1;
5748 tcp_store_ts_recent(tp);
5749 tp->tcp_header_len =
5750 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5751 } else {
5752 tp->tcp_header_len = sizeof(struct tcphdr);
5753 }
5754
5755 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
Eric Dumazet142a2e72015-11-26 08:18:14 -08005756 tp->copied_seq = tp->rcv_nxt;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005757 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5758
5759 /* RFC1323: The window in SYN & SYN/ACK segments is
5760 * never scaled.
5761 */
5762 tp->snd_wnd = ntohs(th->window);
5763 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5764 tp->max_window = tp->snd_wnd;
5765
Florian Westphal735d3832014-09-29 13:08:30 +02005766 tcp_ecn_rcv_syn(tp, th);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005767
John Heffner5d424d52006-03-20 17:53:41 -08005768 tcp_mtup_init(sk);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08005769 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005770 tcp_initialize_rcv_mss(sk);
5771
Linus Torvalds1da177e2005-04-16 15:20:36 -07005772 tcp_send_synack(sk);
5773#if 0
5774 /* Note, we could accept data and URG from this segment.
Jerry Chu168a8f52012-08-31 12:29:13 +00005775 * There are no obstacles to make this (except that we must
5776 * either change tcp_recvmsg() to prevent it from returning data
5777 * before 3WHS completes per RFC793, or employ TCP Fast Open).
Linus Torvalds1da177e2005-04-16 15:20:36 -07005778 *
5779 * However, if we ignore data in ACKless segments sometimes,
5780 * we have no reasons to accept it sometimes.
5781 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5782 * is not flawless. So, discard packet for sanity.
5783 * Uncomment this return to process the data.
5784 */
5785 return -1;
5786#else
5787 goto discard;
5788#endif
5789 }
5790 /* "fifth, if neither of the SYN or RST bits is set then
5791 * drop the segment and return."
5792 */
5793
5794discard_and_undo:
5795 tcp_clear_options(&tp->rx_opt);
5796 tp->rx_opt.mss_clamp = saved_clamp;
5797 goto discard;
5798
5799reset_and_undo:
5800 tcp_clear_options(&tp->rx_opt);
5801 tp->rx_opt.mss_clamp = saved_clamp;
5802 return 1;
5803}
5804
Linus Torvalds1da177e2005-04-16 15:20:36 -07005805/*
5806 * This function implements the receiving procedure of RFC 793 for
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005807 * all states except ESTABLISHED and TIME_WAIT.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005808 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5809 * address independent.
5810 */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005811
Eric Dumazet72ab4a82015-09-29 07:42:41 -07005812int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005813{
5814 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -08005815 struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazet72ab4a82015-09-29 07:42:41 -07005816 const struct tcphdr *th = tcp_hdr(skb);
Jerry Chu168a8f52012-08-31 12:29:13 +00005817 struct request_sock *req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005818 int queued = 0;
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005819 bool acceptable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005820
Linus Torvalds1da177e2005-04-16 15:20:36 -07005821 switch (sk->sk_state) {
5822 case TCP_CLOSE:
5823 goto discard;
5824
5825 case TCP_LISTEN:
Stephen Hemminger2de979b2007-03-08 20:45:19 -08005826 if (th->ack)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005827 return 1;
5828
Stephen Hemminger2de979b2007-03-08 20:45:19 -08005829 if (th->rst)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005830 goto discard;
5831
Stephen Hemminger2de979b2007-03-08 20:45:19 -08005832 if (th->syn) {
Eric Dumazetfdf5af02011-12-02 23:41:42 +00005833 if (th->fin)
5834 goto discard;
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -08005835 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005836 return 1;
5837
Eric Dumazet0aea76d2016-04-21 22:13:01 -07005838 consume_skb(skb);
Masayuki Nakagawafb7e2392007-01-23 20:15:06 -08005839 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005840 }
5841 goto discard;
5842
5843 case TCP_SYN_SENT:
Eric Dumazet8804b272016-04-13 22:05:40 -07005844 tp->rx_opt.saw_tstamp = 0;
Eric Dumazetbda07a62015-09-29 07:42:40 -07005845 queued = tcp_rcv_synsent_state_process(sk, skb, th);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005846 if (queued >= 0)
5847 return queued;
5848
5849 /* Do step6 onward by hand. */
5850 tcp_urg(sk, skb, th);
5851 __kfree_skb(skb);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005852 tcp_data_snd_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005853 return 0;
5854 }
5855
Eric Dumazet8804b272016-04-13 22:05:40 -07005856 tp->rx_opt.saw_tstamp = 0;
Jerry Chu168a8f52012-08-31 12:29:13 +00005857 req = tp->fastopen_rsk;
Ian Morris00db4122015-04-03 09:17:27 +01005858 if (req) {
Jerry Chu37561f62012-10-22 11:26:36 +00005859 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
Jerry Chu168a8f52012-08-31 12:29:13 +00005860 sk->sk_state != TCP_FIN_WAIT1);
5861
Ian Morris51456b22015-04-03 09:17:26 +01005862 if (!tcp_check_req(sk, skb, req, true))
Jerry Chu168a8f52012-08-31 12:29:13 +00005863 goto discard;
Neal Cardwelle69bebd2012-09-22 04:18:57 +00005864 }
Eric Dumazetc3ae62af2012-12-26 12:44:34 +00005865
Calvin Owens0c228e82014-11-20 15:09:53 -08005866 if (!th->ack && !th->rst && !th->syn)
Eric Dumazetc3ae62af2012-12-26 12:44:34 +00005867 goto discard;
5868
Neal Cardwelle69bebd2012-09-22 04:18:57 +00005869 if (!tcp_validate_incoming(sk, skb, th, 0))
Eric Dumazet0c246042012-07-17 01:41:30 +00005870 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005871
5872 /* step 5: check the ACK field */
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005873 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
5874 FLAG_UPDATE_TS_RECENT) > 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005875
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005876 switch (sk->sk_state) {
5877 case TCP_SYN_RECV:
Joe Perches61eb90032013-05-24 18:36:13 +00005878 if (!acceptable)
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005879 return 1;
Joe Perches61eb90032013-05-24 18:36:13 +00005880
Yuchung Cheng0f1c28a2015-09-18 11:36:14 -07005881 if (!tp->srtt_us)
5882 tcp_synack_rtt_meas(sk, req);
5883
Joe Perches61eb90032013-05-24 18:36:13 +00005884 /* Once we leave TCP_SYN_RECV, we no longer need req
5885 * so release it.
5886 */
5887 if (req) {
Joe Perches61eb90032013-05-24 18:36:13 +00005888 tp->total_retrans = req->num_retrans;
Joe Perches61eb90032013-05-24 18:36:13 +00005889 reqsk_fastopen_remove(sk, req, false);
5890 } else {
5891 /* Make sure socket is routed, for correct metrics. */
5892 icsk->icsk_af_ops->rebuild_header(sk);
5893 tcp_init_congestion_control(sk);
5894
5895 tcp_mtup_init(sk);
Joe Perches61eb90032013-05-24 18:36:13 +00005896 tp->copied_seq = tp->rcv_nxt;
Eric Dumazetb0983d32013-09-20 13:56:58 -07005897 tcp_init_buffer_space(sk);
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005898 }
Joe Perches61eb90032013-05-24 18:36:13 +00005899 smp_mb();
5900 tcp_set_state(sk, TCP_ESTABLISHED);
5901 sk->sk_state_change(sk);
5902
5903 /* Note, that this wakeup is only for marginal crossed SYN case.
5904 * Passively open sockets are not waked up, because
5905 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5906 */
5907 if (sk->sk_socket)
5908 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5909
5910 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5911 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
5912 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5913
5914 if (tp->rx_opt.tstamp_ok)
5915 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5916
5917 if (req) {
5918 /* Re-arm the timer because data may have been sent out.
5919 * This is similar to the regular data transmission case
5920 * when new data has just been ack'ed.
5921 *
5922 * (TFO) - we could try to be more aggressive and
5923 * retransmitting any data sooner based on when they
5924 * are sent out.
5925 */
5926 tcp_rearm_rto(sk);
5927 } else
5928 tcp_init_metrics(sk);
5929
Neal Cardwell02cf4eb2013-10-21 15:40:19 -04005930 tcp_update_pacing_rate(sk);
5931
Joe Perches61eb90032013-05-24 18:36:13 +00005932 /* Prevent spurious tcp_cwnd_restart() on first data packet */
5933 tp->lsndtime = tcp_time_stamp;
5934
5935 tcp_initialize_rcv_mss(sk);
5936 tcp_fast_path_on(tp);
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005937 break;
5938
Joe Perchesc48b22d2013-05-24 18:06:58 +00005939 case TCP_FIN_WAIT1: {
5940 struct dst_entry *dst;
5941 int tmo;
5942
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005943 /* If we enter the TCP_FIN_WAIT1 state and we are a
5944 * Fast Open socket and this is the first acceptable
5945 * ACK we have received, this would have acknowledged
5946 * our SYNACK so stop the SYNACK timer.
5947 */
Ian Morris00db4122015-04-03 09:17:27 +01005948 if (req) {
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005949 /* Return RST if ack_seq is invalid.
5950 * Note that RFC793 only says to generate a
5951 * DUPACK for it but for TCP Fast Open it seems
5952 * better to treat this case like TCP_SYN_RECV
5953 * above.
5954 */
5955 if (!acceptable)
5956 return 1;
5957 /* We no longer need the request sock. */
5958 reqsk_fastopen_remove(sk, req, false);
5959 tcp_rearm_rto(sk);
5960 }
Joe Perchesc48b22d2013-05-24 18:06:58 +00005961 if (tp->snd_una != tp->write_seq)
5962 break;
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005963
Joe Perchesc48b22d2013-05-24 18:06:58 +00005964 tcp_set_state(sk, TCP_FIN_WAIT2);
5965 sk->sk_shutdown |= SEND_SHUTDOWN;
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005966
Joe Perchesc48b22d2013-05-24 18:06:58 +00005967 dst = __sk_dst_get(sk);
5968 if (dst)
5969 dst_confirm(dst);
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005970
Joe Perchesc48b22d2013-05-24 18:06:58 +00005971 if (!sock_flag(sk, SOCK_DEAD)) {
5972 /* Wake up lingering close() */
5973 sk->sk_state_change(sk);
5974 break;
5975 }
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005976
Joe Perchesc48b22d2013-05-24 18:06:58 +00005977 if (tp->linger2 < 0 ||
5978 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5979 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5980 tcp_done(sk);
Eric Dumazetc10d9312016-04-29 14:16:47 -07005981 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
Joe Perchesc48b22d2013-05-24 18:06:58 +00005982 return 1;
5983 }
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005984
Joe Perchesc48b22d2013-05-24 18:06:58 +00005985 tmo = tcp_fin_time(sk);
5986 if (tmo > TCP_TIMEWAIT_LEN) {
5987 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5988 } else if (th->fin || sock_owned_by_user(sk)) {
5989 /* Bad case. We could lose such FIN otherwise.
5990 * It is not a big problem, but it looks confusing
5991 * and not so rare event. We still can lose it now,
5992 * if it spins in bh_lock_sock(), but it is really
5993 * marginal case.
5994 */
5995 inet_csk_reset_keepalive_timer(sk, tmo);
5996 } else {
5997 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5998 goto discard;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005999 }
Eric Dumazet1f6afc82013-05-24 15:03:54 +00006000 break;
Joe Perchesc48b22d2013-05-24 18:06:58 +00006001 }
Eric Dumazet1f6afc82013-05-24 15:03:54 +00006002
6003 case TCP_CLOSING:
6004 if (tp->snd_una == tp->write_seq) {
6005 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6006 goto discard;
6007 }
6008 break;
6009
6010 case TCP_LAST_ACK:
6011 if (tp->snd_una == tp->write_seq) {
6012 tcp_update_metrics(sk);
6013 tcp_done(sk);
6014 goto discard;
6015 }
6016 break;
Eric Dumazetc3ae62af2012-12-26 12:44:34 +00006017 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07006018
6019 /* step 6: check the URG bit */
6020 tcp_urg(sk, skb, th);
6021
6022 /* step 7: process the segment text */
6023 switch (sk->sk_state) {
6024 case TCP_CLOSE_WAIT:
6025 case TCP_CLOSING:
6026 case TCP_LAST_ACK:
6027 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
6028 break;
6029 case TCP_FIN_WAIT1:
6030 case TCP_FIN_WAIT2:
6031 /* RFC 793 says to queue data in these states,
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09006032 * RFC 1122 says we MUST send a reset.
Linus Torvalds1da177e2005-04-16 15:20:36 -07006033 * BSD 4.4 also does reset.
6034 */
6035 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6036 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6037 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
Eric Dumazetc10d9312016-04-29 14:16:47 -07006038 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006039 tcp_reset(sk);
6040 return 1;
6041 }
6042 }
6043 /* Fall through */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09006044 case TCP_ESTABLISHED:
Linus Torvalds1da177e2005-04-16 15:20:36 -07006045 tcp_data_queue(sk, skb);
6046 queued = 1;
6047 break;
6048 }
6049
6050 /* tcp_data could move socket to TIME-WAIT */
6051 if (sk->sk_state != TCP_CLOSE) {
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07006052 tcp_data_snd_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006053 tcp_ack_snd_check(sk);
6054 }
6055
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09006056 if (!queued) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07006057discard:
Eric Dumazet532182c2016-04-01 08:52:19 -07006058 tcp_drop(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006059 }
6060 return 0;
6061}
Linus Torvalds1da177e2005-04-16 15:20:36 -07006062EXPORT_SYMBOL(tcp_rcv_state_process);
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006063
6064static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6065{
6066 struct inet_request_sock *ireq = inet_rsk(req);
6067
6068 if (family == AF_INET)
Joe Perchesba7a46f2014-11-11 10:59:17 -08006069 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6070 &ireq->ir_rmt_addr, port);
Octavian Purdila4135ab82014-06-28 21:20:54 +03006071#if IS_ENABLED(CONFIG_IPV6)
6072 else if (family == AF_INET6)
Joe Perchesba7a46f2014-11-11 10:59:17 -08006073 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6074 &ireq->ir_v6_rmt_addr, port);
Octavian Purdila4135ab82014-06-28 21:20:54 +03006075#endif
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006076}
6077
Florian Westphald82bd122014-09-29 13:08:29 +02006078/* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6079 *
6080 * If we receive a SYN packet with these bits set, it means a
6081 * network is playing bad games with TOS bits. In order to
6082 * avoid possible false congestion notifications, we disable
stephen hemmingerf4e715c2014-10-29 16:05:06 -07006083 * TCP ECN negotiation.
Florian Westphald82bd122014-09-29 13:08:29 +02006084 *
6085 * Exception: tcp_ca wants ECN. This is required for DCTCP
Florian Westphal843c2fd2015-01-30 20:45:20 +01006086 * congestion control: Linux DCTCP asserts ECT on all packets,
6087 * including SYN, which is most optimal solution; however,
6088 * others, such as FreeBSD do not.
Florian Westphald82bd122014-09-29 13:08:29 +02006089 */
6090static void tcp_ecn_create_request(struct request_sock *req,
6091 const struct sk_buff *skb,
Florian Westphalf7b3bec2014-11-03 17:35:03 +01006092 const struct sock *listen_sk,
6093 const struct dst_entry *dst)
Florian Westphald82bd122014-09-29 13:08:29 +02006094{
6095 const struct tcphdr *th = tcp_hdr(skb);
6096 const struct net *net = sock_net(listen_sk);
6097 bool th_ecn = th->ece && th->cwr;
Florian Westphal843c2fd2015-01-30 20:45:20 +01006098 bool ect, ecn_ok;
Daniel Borkmannc3a8d942015-08-31 15:58:47 +02006099 u32 ecn_ok_dst;
Florian Westphald82bd122014-09-29 13:08:29 +02006100
6101 if (!th_ecn)
6102 return;
6103
6104 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
Daniel Borkmannc3a8d942015-08-31 15:58:47 +02006105 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6106 ecn_ok = net->ipv4.sysctl_tcp_ecn || ecn_ok_dst;
Florian Westphald82bd122014-09-29 13:08:29 +02006107
Daniel Borkmannc3a8d942015-08-31 15:58:47 +02006108 if ((!ect && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
6109 (ecn_ok_dst & DST_FEATURE_ECN_CA))
Florian Westphald82bd122014-09-29 13:08:29 +02006110 inet_rsk(req)->ecn_ok = 1;
6111}
6112
Eric Dumazet1bfc4432015-03-16 21:06:19 -07006113static void tcp_openreq_init(struct request_sock *req,
6114 const struct tcp_options_received *rx_opt,
6115 struct sk_buff *skb, const struct sock *sk)
6116{
6117 struct inet_request_sock *ireq = inet_rsk(req);
6118
Eric Dumazeted53d0a2015-10-08 19:33:23 -07006119 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
Eric Dumazet1bfc4432015-03-16 21:06:19 -07006120 req->cookie_ts = 0;
6121 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6122 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
Yuchung Cheng0f1c28a2015-09-18 11:36:14 -07006123 skb_mstamp_get(&tcp_rsk(req)->snt_synack);
Eric Dumazet1bfc4432015-03-16 21:06:19 -07006124 tcp_rsk(req)->last_oow_ack_time = 0;
6125 req->mss = rx_opt->mss_clamp;
6126 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6127 ireq->tstamp_ok = rx_opt->tstamp_ok;
6128 ireq->sack_ok = rx_opt->sack_ok;
6129 ireq->snd_wscale = rx_opt->snd_wscale;
6130 ireq->wscale_ok = rx_opt->wscale_ok;
6131 ireq->acked = 0;
6132 ireq->ecn_ok = 0;
6133 ireq->ir_rmt_port = tcp_hdr(skb)->source;
6134 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6135 ireq->ir_mark = inet_request_mark(sk, skb);
6136}
6137
Eric Dumazete49bb332015-03-17 18:32:27 -07006138struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
Eric Dumazeta1a53442015-10-04 21:08:11 -07006139 struct sock *sk_listener,
6140 bool attach_listener)
Eric Dumazete49bb332015-03-17 18:32:27 -07006141{
Eric Dumazeta1a53442015-10-04 21:08:11 -07006142 struct request_sock *req = reqsk_alloc(ops, sk_listener,
6143 attach_listener);
Eric Dumazete49bb332015-03-17 18:32:27 -07006144
6145 if (req) {
6146 struct inet_request_sock *ireq = inet_rsk(req);
6147
6148 kmemcheck_annotate_bitfield(ireq, flags);
6149 ireq->opt = NULL;
Huw Davies56ac42b2016-06-27 15:05:28 -04006150#if IS_ENABLED(CONFIG_IPV6)
6151 ireq->pktopts = NULL;
6152#endif
Eric Dumazete49bb332015-03-17 18:32:27 -07006153 atomic64_set(&ireq->ir_cookie, 0);
6154 ireq->ireq_state = TCP_NEW_SYN_RECV;
6155 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
Eric Dumazet0144a812015-03-24 21:45:56 -07006156 ireq->ireq_family = sk_listener->sk_family;
Eric Dumazete49bb332015-03-17 18:32:27 -07006157 }
6158
6159 return req;
6160}
6161EXPORT_SYMBOL(inet_reqsk_alloc);
6162
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006163/*
6164 * Return true if a syncookie should be sent
6165 */
Eric Dumazet2985aaa2015-09-29 07:42:51 -07006166static bool tcp_syn_flood_action(const struct sock *sk,
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006167 const struct sk_buff *skb,
6168 const char *proto)
6169{
Eric Dumazet8d2675f2015-10-02 11:43:25 -07006170 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006171 const char *msg = "Dropping request";
6172 bool want_cookie = false;
Nikolay Borisov12ed8242016-02-03 09:46:51 +02006173 struct net *net = sock_net(sk);
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006174
6175#ifdef CONFIG_SYN_COOKIES
Nikolay Borisov12ed8242016-02-03 09:46:51 +02006176 if (net->ipv4.sysctl_tcp_syncookies) {
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006177 msg = "Sending cookies";
6178 want_cookie = true;
Eric Dumazet02a1d6e2016-04-27 16:44:39 -07006179 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006180 } else
6181#endif
Eric Dumazet02a1d6e2016-04-27 16:44:39 -07006182 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006183
Eric Dumazet8d2675f2015-10-02 11:43:25 -07006184 if (!queue->synflood_warned &&
Nikolay Borisov12ed8242016-02-03 09:46:51 +02006185 net->ipv4.sysctl_tcp_syncookies != 2 &&
Eric Dumazet8d2675f2015-10-02 11:43:25 -07006186 xchg(&queue->synflood_warned, 1) == 0)
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006187 pr_info("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6188 proto, ntohs(tcp_hdr(skb)->dest), msg);
Eric Dumazet2985aaa2015-09-29 07:42:51 -07006189
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006190 return want_cookie;
6191}
6192
Eric Dumazetcd8ae852015-05-03 21:34:46 -07006193static void tcp_reqsk_record_syn(const struct sock *sk,
6194 struct request_sock *req,
6195 const struct sk_buff *skb)
6196{
6197 if (tcp_sk(sk)->save_syn) {
6198 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
6199 u32 *copy;
6200
6201 copy = kmalloc(len + sizeof(u32), GFP_ATOMIC);
6202 if (copy) {
6203 copy[0] = len;
6204 memcpy(&copy[1], skb_network_header(skb), len);
6205 req->saved_syn = copy;
6206 }
6207 }
6208}
6209
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006210int tcp_conn_request(struct request_sock_ops *rsk_ops,
6211 const struct tcp_request_sock_ops *af_ops,
6212 struct sock *sk, struct sk_buff *skb)
6213{
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006214 struct tcp_fastopen_cookie foc = { .len = -1 };
Eric Dumazet7c85af82015-09-24 17:16:05 -07006215 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6216 struct tcp_options_received tmp_opt;
6217 struct tcp_sock *tp = tcp_sk(sk);
Nikolay Borisov12ed8242016-02-03 09:46:51 +02006218 struct net *net = sock_net(sk);
Eric Dumazet7c85af82015-09-24 17:16:05 -07006219 struct sock *fastopen_sk = NULL;
6220 struct dst_entry *dst = NULL;
6221 struct request_sock *req;
6222 bool want_cookie = false;
6223 struct flowi fl;
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006224
6225 /* TW buckets are converted to open requests without
6226 * limitations, they conserve resources and peer is
6227 * evidently real one.
6228 */
Nikolay Borisov12ed8242016-02-03 09:46:51 +02006229 if ((net->ipv4.sysctl_tcp_syncookies == 2 ||
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006230 inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6231 want_cookie = tcp_syn_flood_action(sk, skb, rsk_ops->slab_name);
6232 if (!want_cookie)
6233 goto drop;
6234 }
6235
6236
6237 /* Accept backlog is full. If we have already queued enough
6238 * of warm entries in syn queue, drop request. It is better than
6239 * clogging syn queue with openreqs with exponentially increasing
6240 * timeout.
6241 */
6242 if (sk_acceptq_is_full(sk) && inet_csk_reqsk_queue_young(sk) > 1) {
Eric Dumazetc10d9312016-04-29 14:16:47 -07006243 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006244 goto drop;
6245 }
6246
Eric Dumazeta1a53442015-10-04 21:08:11 -07006247 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006248 if (!req)
6249 goto drop;
6250
6251 tcp_rsk(req)->af_specific = af_ops;
6252
6253 tcp_clear_options(&tmp_opt);
6254 tmp_opt.mss_clamp = af_ops->mss_clamp;
6255 tmp_opt.user_mss = tp->rx_opt.user_mss;
6256 tcp_parse_options(skb, &tmp_opt, 0, want_cookie ? NULL : &foc);
6257
6258 if (want_cookie && !tmp_opt.saw_tstamp)
6259 tcp_clear_options(&tmp_opt);
6260
6261 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6262 tcp_openreq_init(req, &tmp_opt, skb, sk);
6263
Eric Dumazet16f86162015-03-13 15:51:10 -07006264 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
David Ahern6dd9a142015-12-16 13:20:44 -08006265 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
Eric Dumazet16f86162015-03-13 15:51:10 -07006266
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006267 af_ops->init_req(req, sk, skb);
6268
6269 if (security_inet_conn_request(sk, skb, req))
6270 goto drop_and_free;
6271
Florian Westphalf7b3bec2014-11-03 17:35:03 +01006272 if (!want_cookie && !isn) {
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006273 /* VJ's idea. We save last timestamp seen
6274 * from the destination in peer table, when entering
6275 * state TIME-WAIT, and check against it before
6276 * accepting new connection request.
6277 *
6278 * If "isn" is not zero, this request hit alive
6279 * timewait bucket, so that all the necessary checks
6280 * are made in the function processing timewait state.
6281 */
Hannes Frederic Sowaa26552a2014-08-14 22:06:12 +02006282 if (tcp_death_row.sysctl_tw_recycle) {
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006283 bool strict;
6284
6285 dst = af_ops->route_req(sk, &fl, req, &strict);
Hannes Frederic Sowaa26552a2014-08-14 22:06:12 +02006286
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006287 if (dst && strict &&
Hannes Frederic Sowaa26552a2014-08-14 22:06:12 +02006288 !tcp_peer_is_proven(req, dst, true,
6289 tmp_opt.saw_tstamp)) {
Eric Dumazetc10d9312016-04-29 14:16:47 -07006290 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSPASSIVEREJECTED);
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006291 goto drop_and_release;
6292 }
6293 }
6294 /* Kill the following clause, if you dislike this way. */
Nikolay Borisov12ed8242016-02-03 09:46:51 +02006295 else if (!net->ipv4.sysctl_tcp_syncookies &&
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006296 (sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6297 (sysctl_max_syn_backlog >> 2)) &&
Hannes Frederic Sowaa26552a2014-08-14 22:06:12 +02006298 !tcp_peer_is_proven(req, dst, false,
6299 tmp_opt.saw_tstamp)) {
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006300 /* Without syncookies last quarter of
6301 * backlog is filled with destinations,
6302 * proven to be alive.
6303 * It means that we continue to communicate
6304 * to destinations, already remembered
6305 * to the moment of synflood.
6306 */
6307 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6308 rsk_ops->family);
6309 goto drop_and_release;
6310 }
6311
6312 isn = af_ops->init_seq(skb);
6313 }
6314 if (!dst) {
6315 dst = af_ops->route_req(sk, &fl, req, NULL);
6316 if (!dst)
6317 goto drop_and_free;
6318 }
6319
Florian Westphalf7b3bec2014-11-03 17:35:03 +01006320 tcp_ecn_create_request(req, skb, sk, dst);
6321
6322 if (want_cookie) {
6323 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6324 req->cookie_ts = tmp_opt.tstamp_ok;
6325 if (!tmp_opt.tstamp_ok)
6326 inet_rsk(req)->ecn_ok = 0;
6327 }
6328
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006329 tcp_rsk(req)->snt_isn = isn;
Eric Dumazet58d607d2015-09-15 15:24:20 -07006330 tcp_rsk(req)->txhash = net_tx_rndhash();
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006331 tcp_openreq_init_rwin(req, sk, dst);
Eric Dumazetca6fb062015-10-02 11:43:35 -07006332 if (!want_cookie) {
Eric Dumazetca6fb062015-10-02 11:43:35 -07006333 tcp_reqsk_record_syn(sk, req, skb);
Eric Dumazet7656d842015-10-04 21:08:07 -07006334 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
Eric Dumazetca6fb062015-10-02 11:43:35 -07006335 }
Eric Dumazet7c85af82015-09-24 17:16:05 -07006336 if (fastopen_sk) {
Eric Dumazetca6fb062015-10-02 11:43:35 -07006337 af_ops->send_synack(fastopen_sk, dst, &fl, req,
Eric Dumazetb3d05142016-04-13 22:05:39 -07006338 &foc, TCP_SYNACK_FASTOPEN);
Eric Dumazet7656d842015-10-04 21:08:07 -07006339 /* Add the child socket directly into the accept queue */
6340 inet_csk_reqsk_queue_add(sk, req, fastopen_sk);
6341 sk->sk_data_ready(sk);
6342 bh_unlock_sock(fastopen_sk);
Eric Dumazet7c85af82015-09-24 17:16:05 -07006343 sock_put(fastopen_sk);
6344 } else {
Eric Dumazet9439ce02015-03-17 18:32:29 -07006345 tcp_rsk(req)->tfo_listener = false;
Eric Dumazetca6fb062015-10-02 11:43:35 -07006346 if (!want_cookie)
6347 inet_csk_reqsk_queue_hash_add(sk, req, TCP_TIMEOUT_INIT);
Eric Dumazetb3d05142016-04-13 22:05:39 -07006348 af_ops->send_synack(sk, dst, &fl, req, &foc,
6349 !want_cookie ? TCP_SYNACK_NORMAL :
6350 TCP_SYNACK_COOKIE);
Eric Dumazet9caad862016-04-01 08:52:20 -07006351 if (want_cookie) {
6352 reqsk_free(req);
6353 return 0;
6354 }
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006355 }
Eric Dumazetca6fb062015-10-02 11:43:35 -07006356 reqsk_put(req);
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006357 return 0;
6358
6359drop_and_release:
6360 dst_release(dst);
6361drop_and_free:
6362 reqsk_free(req);
6363drop:
Eric Dumazet9caad862016-04-01 08:52:20 -07006364 tcp_listendrop(sk);
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006365 return 0;
6366}
6367EXPORT_SYMBOL(tcp_conn_request);