blob: a13692560e637b215f56c906e3302e5a19a0bc44 [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>
Satoru SATOH5ffc02a2008-05-04 22:14:42 -070071#include <net/dst.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070072#include <net/tcp.h>
73#include <net/inet_common.h>
74#include <linux/ipsec.h>
75#include <asm/unaligned.h>
Chris Leech1a2449a2006-05-23 18:05:53 -070076#include <net/netdma.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070077
Brian Haleyab32ea52006-09-22 14:15:41 -070078int sysctl_tcp_timestamps __read_mostly = 1;
79int sysctl_tcp_window_scaling __read_mostly = 1;
80int sysctl_tcp_sack __read_mostly = 1;
81int sysctl_tcp_fack __read_mostly = 1;
82int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
Eric Dumazet4bc2f182010-07-09 21:22:10 +000083EXPORT_SYMBOL(sysctl_tcp_reordering);
Ilpo Järvinen255cac92009-05-04 11:07:36 -070084int sysctl_tcp_ecn __read_mostly = 2;
Eric Dumazet4bc2f182010-07-09 21:22:10 +000085EXPORT_SYMBOL(sysctl_tcp_ecn);
Brian Haleyab32ea52006-09-22 14:15:41 -070086int sysctl_tcp_dsack __read_mostly = 1;
87int sysctl_tcp_app_win __read_mostly = 31;
Eric Dumazetb49960a2012-05-02 02:28:41 +000088int sysctl_tcp_adv_win_scale __read_mostly = 1;
Eric Dumazet4bc2f182010-07-09 21:22:10 +000089EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
Linus Torvalds1da177e2005-04-16 15:20:36 -070090
Eric Dumazet282f23c2012-07-17 10:13:05 +020091/* rfc5961 challenge ack rate limiting */
92int sysctl_tcp_challenge_ack_limit = 100;
93
Brian Haleyab32ea52006-09-22 14:15:41 -070094int sysctl_tcp_stdurg __read_mostly;
95int sysctl_tcp_rfc1337 __read_mostly;
96int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
Ilpo Järvinenc96fd3d2007-09-20 11:36:37 -070097int sysctl_tcp_frto __read_mostly = 2;
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -080098int sysctl_tcp_frto_response __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -070099
Andreas Petlund7e380172010-02-18 04:48:19 +0000100int sysctl_tcp_thin_dupack __read_mostly;
101
Brian Haleyab32ea52006-09-22 14:15:41 -0700102int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
103int sysctl_tcp_abc __read_mostly;
Yuchung Chengeed530b2012-05-02 13:30:03 +0000104int sysctl_tcp_early_retrans __read_mostly = 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700105
Linus Torvalds1da177e2005-04-16 15:20:36 -0700106#define FLAG_DATA 0x01 /* Incoming frame contained data. */
107#define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
108#define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
109#define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
110#define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
111#define FLAG_DATA_SACKED 0x20 /* New SACK. */
112#define FLAG_ECE 0x40 /* ECE in this ACK */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700113#define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
Ilpo Järvinen4dc26652007-02-21 23:16:11 -0800114#define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */
Ilpo Järvinen2e605292007-08-02 19:46:58 -0700115#define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
Ryousei Takano564262c12007-10-25 23:03:52 -0700116#define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
Ilpo Järvinen009a2e32007-09-20 11:34:38 -0700117#define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */
Ilpo Järvinencadbd032007-12-31 04:49:21 -0800118#define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700119
120#define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
121#define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
122#define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
123#define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
Ilpo Järvinen2e605292007-08-02 19:46:58 -0700124#define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700125
Linus Torvalds1da177e2005-04-16 15:20:36 -0700126#define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700127#define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700128
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900129/* Adapt the MSS value used to make delayed ack decision to the
Linus Torvalds1da177e2005-04-16 15:20:36 -0700130 * real world.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900131 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800132static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700133{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700134 struct inet_connection_sock *icsk = inet_csk(sk);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900135 const unsigned int lss = icsk->icsk_ack.last_seg_size;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700136 unsigned int len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700137
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900138 icsk->icsk_ack.last_seg_size = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700139
140 /* skb->len may jitter because of SACKs, even if peer
141 * sends good full-sized frames.
142 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800143 len = skb_shinfo(skb)->gso_size ? : skb->len;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700144 if (len >= icsk->icsk_ack.rcv_mss) {
145 icsk->icsk_ack.rcv_mss = len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700146 } else {
147 /* Otherwise, we make more careful check taking into account,
148 * that SACKs block is variable.
149 *
150 * "len" is invariant segment length, including TCP header.
151 */
Arnaldo Carvalho de Melo9c702202007-04-25 18:04:18 -0700152 len += skb->data - skb_transport_header(skb);
William Allen Simpsonbee7ca92009-11-10 09:51:18 +0000153 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -0700154 /* If PSH is not set, packet should be
155 * full sized, provided peer TCP is not badly broken.
156 * This observation (if it is correct 8)) allows
157 * to handle super-low mtu links fairly.
158 */
159 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -0700160 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700161 /* Subtract also invariant (if peer is RFC compliant),
162 * tcp header plus fixed timestamp option length.
163 * Resulting "len" is MSS free of SACK jitter.
164 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700165 len -= tcp_sk(sk)->tcp_header_len;
166 icsk->icsk_ack.last_seg_size = len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700167 if (len == lss) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700168 icsk->icsk_ack.rcv_mss = len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700169 return;
170 }
171 }
Alexey Kuznetsov1ef96962006-09-19 12:52:50 -0700172 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
173 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700174 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700175 }
176}
177
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700178static void tcp_incr_quickack(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700179{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700180 struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazet95c96172012-04-15 05:58:06 +0000181 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700182
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800183 if (quickacks == 0)
184 quickacks = 2;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700185 if (quickacks > icsk->icsk_ack.quick)
186 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700187}
188
stephen hemminger1b9f4092010-09-28 19:30:14 +0000189static void tcp_enter_quickack_mode(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700190{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700191 struct inet_connection_sock *icsk = inet_csk(sk);
192 tcp_incr_quickack(sk);
193 icsk->icsk_ack.pingpong = 0;
194 icsk->icsk_ack.ato = TCP_ATO_MIN;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700195}
196
197/* Send ACKs quickly, if "quick" count is not exhausted
198 * and the session is not interactive.
199 */
200
Eric Dumazeta2a385d2012-05-16 23:15:34 +0000201static inline bool tcp_in_quickack_mode(const struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700202{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700203 const struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazeta2a385d2012-05-16 23:15:34 +0000204
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700205 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700206}
207
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700208static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
209{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800210 if (tp->ecn_flags & TCP_ECN_OK)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700211 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
212}
213
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400214static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700215{
216 if (tcp_hdr(skb)->cwr)
217 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
218}
219
220static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
221{
222 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
223}
224
Eric Dumazet7a269ff2011-09-22 20:02:19 +0000225static inline void TCP_ECN_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700226{
Eric Dumazet7a269ff2011-09-22 20:02:19 +0000227 if (!(tp->ecn_flags & TCP_ECN_OK))
228 return;
229
Eric Dumazetb82d1bb2011-09-27 02:20:08 -0400230 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
Eric Dumazet7a269ff2011-09-22 20:02:19 +0000231 case INET_ECN_NOT_ECT:
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700232 /* Funny extension: if ECT is not set on a segment,
Eric Dumazet7a269ff2011-09-22 20:02:19 +0000233 * and we already seen ECT on a previous segment,
234 * it is probably a retransmit.
235 */
236 if (tp->ecn_flags & TCP_ECN_SEEN)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700237 tcp_enter_quickack_mode((struct sock *)tp);
Eric Dumazet7a269ff2011-09-22 20:02:19 +0000238 break;
239 case INET_ECN_CE:
Eric Dumazetaae06bf2012-08-06 11:04:43 +0000240 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
241 /* Better not delay acks, sender can have a very low cwnd */
242 tcp_enter_quickack_mode((struct sock *)tp);
243 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
244 }
Eric Dumazet7a269ff2011-09-22 20:02:19 +0000245 /* fallinto */
246 default:
247 tp->ecn_flags |= TCP_ECN_SEEN;
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700248 }
249}
250
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400251static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700252{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800253 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700254 tp->ecn_flags &= ~TCP_ECN_OK;
255}
256
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400257static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700258{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800259 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700260 tp->ecn_flags &= ~TCP_ECN_OK;
261}
262
Eric Dumazeta2a385d2012-05-16 23:15:34 +0000263static bool TCP_ECN_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700264{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800265 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
Eric Dumazeta2a385d2012-05-16 23:15:34 +0000266 return true;
267 return false;
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700268}
269
Linus Torvalds1da177e2005-04-16 15:20:36 -0700270/* Buffer size and advertised window tuning.
271 *
272 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
273 */
274
275static void tcp_fixup_sndbuf(struct sock *sk)
276{
Eric Dumazet87fb4b72011-10-13 07:28:54 +0000277 int sndmem = SKB_TRUESIZE(tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700278
Eric Dumazet06a59ec2011-10-13 18:24:42 +0000279 sndmem *= TCP_INIT_CWND;
280 if (sk->sk_sndbuf < sndmem)
281 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700282}
283
284/* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
285 *
286 * All tcp_full_space() is split to two parts: "network" buffer, allocated
287 * forward and advertised in receiver window (tp->rcv_wnd) and
288 * "application buffer", required to isolate scheduling/application
289 * latencies from network.
290 * window_clamp is maximal advertised window. It can be less than
291 * tcp_full_space(), in this case tcp_full_space() - window_clamp
292 * is reserved for "application" buffer. The less window_clamp is
293 * the smoother our behaviour from viewpoint of network, but the lower
294 * throughput and the higher sensitivity of the connection to losses. 8)
295 *
296 * rcv_ssthresh is more strict window_clamp used at "slow start"
297 * phase to predict further behaviour of this connection.
298 * It is used for two goals:
299 * - to enforce header prediction at sender, even when application
300 * requires some significant "application buffer". It is check #1.
301 * - to prevent pruning of receive queue because of misprediction
302 * of receiver window. Check #2.
303 *
304 * The scheme does not work when sender sends good segments opening
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800305 * window and then starts to feed us spaghetti. But it should work
Linus Torvalds1da177e2005-04-16 15:20:36 -0700306 * in common situations. Otherwise, we have to rely on queue collapsing.
307 */
308
309/* Slow part of check#2. */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700310static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700311{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700312 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700313 /* Optimize this! */
Eric Dumazetdfd4f0a2007-12-21 06:07:53 -0800314 int truesize = tcp_win_from_space(skb->truesize) >> 1;
315 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700316
317 while (tp->rcv_ssthresh <= window) {
318 if (truesize <= skb->len)
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700319 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700320
321 truesize >>= 1;
322 window >>= 1;
323 }
324 return 0;
325}
326
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400327static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700328{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700329 struct tcp_sock *tp = tcp_sk(sk);
330
Linus Torvalds1da177e2005-04-16 15:20:36 -0700331 /* Check #1 */
332 if (tp->rcv_ssthresh < tp->window_clamp &&
333 (int)tp->rcv_ssthresh < tcp_space(sk) &&
Glauber Costa180d8cd2011-12-11 21:47:02 +0000334 !sk_under_memory_pressure(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700335 int incr;
336
337 /* Check #2. Increase window, if skb with such overhead
338 * will fit to rcvbuf in future.
339 */
340 if (tcp_win_from_space(skb->truesize) <= skb->len)
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800341 incr = 2 * tp->advmss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700342 else
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700343 incr = __tcp_grow_window(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700344
345 if (incr) {
Eric Dumazet4d846f02012-04-16 23:28:07 +0000346 incr = max_t(int, incr, 2 * skb->len);
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800347 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
348 tp->window_clamp);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700349 inet_csk(sk)->icsk_ack.quick |= 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700350 }
351 }
352}
353
354/* 3. Tuning rcvbuf, when connection enters established state. */
355
356static void tcp_fixup_rcvbuf(struct sock *sk)
357{
Eric Dumazete9266a02011-10-20 16:53:56 -0400358 u32 mss = tcp_sk(sk)->advmss;
359 u32 icwnd = TCP_DEFAULT_INIT_RCVWND;
360 int rcvmem;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700361
Eric Dumazete9266a02011-10-20 16:53:56 -0400362 /* Limit to 10 segments if mss <= 1460,
363 * or 14600/mss segments, with a minimum of two segments.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700364 */
Eric Dumazete9266a02011-10-20 16:53:56 -0400365 if (mss > 1460)
366 icwnd = max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2);
367
368 rcvmem = SKB_TRUESIZE(mss + MAX_TCP_HEADER);
369 while (tcp_win_from_space(rcvmem) < mss)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700370 rcvmem += 128;
Eric Dumazete9266a02011-10-20 16:53:56 -0400371
372 rcvmem *= icwnd;
373
374 if (sk->sk_rcvbuf < rcvmem)
375 sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700376}
377
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800378/* 4. Try to fixup all. It is made immediately after connection enters
Linus Torvalds1da177e2005-04-16 15:20:36 -0700379 * established state.
380 */
Jerry Chu10467162012-08-31 12:29:11 +0000381void tcp_init_buffer_space(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700382{
383 struct tcp_sock *tp = tcp_sk(sk);
384 int maxwin;
385
386 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
387 tcp_fixup_rcvbuf(sk);
388 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
389 tcp_fixup_sndbuf(sk);
390
391 tp->rcvq_space.space = tp->rcv_wnd;
392
393 maxwin = tcp_full_space(sk);
394
395 if (tp->window_clamp >= maxwin) {
396 tp->window_clamp = maxwin;
397
398 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
399 tp->window_clamp = max(maxwin -
400 (maxwin >> sysctl_tcp_app_win),
401 4 * tp->advmss);
402 }
403
404 /* Force reservation of one segment. */
405 if (sysctl_tcp_app_win &&
406 tp->window_clamp > 2 * tp->advmss &&
407 tp->window_clamp + tp->advmss > maxwin)
408 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
409
410 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
411 tp->snd_cwnd_stamp = tcp_time_stamp;
412}
413
Linus Torvalds1da177e2005-04-16 15:20:36 -0700414/* 5. Recalculate window clamp after socket hit its memory bounds. */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700415static void tcp_clamp_window(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700416{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700417 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300418 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700419
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300420 icsk->icsk_ack.quick = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700421
John Heffner326f36e2005-11-10 17:11:48 -0800422 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
423 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
Glauber Costa180d8cd2011-12-11 21:47:02 +0000424 !sk_under_memory_pressure(sk) &&
425 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
John Heffner326f36e2005-11-10 17:11:48 -0800426 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
427 sysctl_tcp_rmem[2]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700428 }
John Heffner326f36e2005-11-10 17:11:48 -0800429 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800430 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700431}
432
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800433/* Initialize RCV_MSS value.
434 * RCV_MSS is an our guess about MSS used by the peer.
435 * We haven't any direct information about the MSS.
436 * It's better to underestimate the RCV_MSS rather than overestimate.
437 * Overestimations make us ACKing less frequently than needed.
438 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
439 */
440void tcp_initialize_rcv_mss(struct sock *sk)
441{
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400442 const struct tcp_sock *tp = tcp_sk(sk);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800443 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
444
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800445 hint = min(hint, tp->rcv_wnd / 2);
William Allen Simpsonbee7ca92009-11-10 09:51:18 +0000446 hint = min(hint, TCP_MSS_DEFAULT);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800447 hint = max(hint, TCP_MIN_MSS);
448
449 inet_csk(sk)->icsk_ack.rcv_mss = hint;
450}
Eric Dumazet4bc2f182010-07-09 21:22:10 +0000451EXPORT_SYMBOL(tcp_initialize_rcv_mss);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800452
Linus Torvalds1da177e2005-04-16 15:20:36 -0700453/* Receiver "autotuning" code.
454 *
455 * The algorithm for RTT estimation w/o timestamps is based on
456 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
Justin P. Mattock631dd1a2010-10-18 11:03:14 +0200457 * <http://public.lanl.gov/radiant/pubs.html#DRS>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700458 *
459 * More detail on this code can be found at
Justin P. Mattock631dd1a2010-10-18 11:03:14 +0200460 * <http://staff.psc.edu/jheffner/>,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700461 * though this reference is out of date. A new paper
462 * is pending.
463 */
464static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
465{
466 u32 new_sample = tp->rcv_rtt_est.rtt;
467 long m = sample;
468
469 if (m == 0)
470 m = 1;
471
472 if (new_sample != 0) {
473 /* If we sample in larger samples in the non-timestamp
474 * case, we could grossly overestimate the RTT especially
475 * with chatty applications or bulk transfer apps which
476 * are stalled on filesystem I/O.
477 *
478 * Also, since we are only going for a minimum in the
Stephen Hemminger31f34262005-11-15 15:17:10 -0800479 * non-timestamp case, we do not smooth things out
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800480 * else with timestamps disabled convergence takes too
Linus Torvalds1da177e2005-04-16 15:20:36 -0700481 * long.
482 */
483 if (!win_dep) {
484 m -= (new_sample >> 3);
485 new_sample += m;
Neal Cardwell18a223e2012-04-10 07:59:20 +0000486 } else {
487 m <<= 3;
488 if (m < new_sample)
489 new_sample = m;
490 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700491 } else {
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800492 /* No previous measure. */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700493 new_sample = m << 3;
494 }
495
496 if (tp->rcv_rtt_est.rtt != new_sample)
497 tp->rcv_rtt_est.rtt = new_sample;
498}
499
500static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
501{
502 if (tp->rcv_rtt_est.time == 0)
503 goto new_measure;
504 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
505 return;
Neal Cardwell651913c2012-04-27 11:29:37 -0400506 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700507
508new_measure:
509 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
510 tp->rcv_rtt_est.time = tcp_time_stamp;
511}
512
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800513static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
514 const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700515{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700516 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700517 if (tp->rx_opt.rcv_tsecr &&
518 (TCP_SKB_CB(skb)->end_seq -
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700519 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700520 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
521}
522
523/*
524 * This function should be called every time data is copied to user space.
525 * It calculates the appropriate TCP receive buffer space.
526 */
527void tcp_rcv_space_adjust(struct sock *sk)
528{
529 struct tcp_sock *tp = tcp_sk(sk);
530 int time;
531 int space;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900532
Linus Torvalds1da177e2005-04-16 15:20:36 -0700533 if (tp->rcvq_space.time == 0)
534 goto new_measure;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900535
Linus Torvalds1da177e2005-04-16 15:20:36 -0700536 time = tcp_time_stamp - tp->rcvq_space.time;
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800537 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700538 return;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900539
Linus Torvalds1da177e2005-04-16 15:20:36 -0700540 space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
541
542 space = max(tp->rcvq_space.space, space);
543
544 if (tp->rcvq_space.space != space) {
545 int rcvmem;
546
547 tp->rcvq_space.space = space;
548
John Heffner6fcf9412006-02-09 17:06:57 -0800549 if (sysctl_tcp_moderate_rcvbuf &&
550 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700551 int new_clamp = space;
552
553 /* Receive space grows, normalize in order to
554 * take into account packet headers and sk_buff
555 * structure overhead.
556 */
557 space /= tp->advmss;
558 if (!space)
559 space = 1;
Eric Dumazet87fb4b72011-10-13 07:28:54 +0000560 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700561 while (tcp_win_from_space(rcvmem) < tp->advmss)
562 rcvmem += 128;
563 space *= rcvmem;
564 space = min(space, sysctl_tcp_rmem[2]);
565 if (space > sk->sk_rcvbuf) {
566 sk->sk_rcvbuf = space;
567
568 /* Make the window clamp follow along. */
569 tp->window_clamp = new_clamp;
570 }
571 }
572 }
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900573
Linus Torvalds1da177e2005-04-16 15:20:36 -0700574new_measure:
575 tp->rcvq_space.seq = tp->copied_seq;
576 tp->rcvq_space.time = tcp_time_stamp;
577}
578
579/* There is something which you must keep in mind when you analyze the
580 * behavior of the tp->ato delayed ack timeout interval. When a
581 * connection starts up, we want to ack as quickly as possible. The
582 * problem is that "good" TCP's do slow start at the beginning of data
583 * transmission. The means that until we send the first few ACK's the
584 * sender will sit on his end and only queue most of his data, because
585 * he can only send snd_cwnd unacked packets at any given time. For
586 * each ACK we send, he increments snd_cwnd and transmits more of his
587 * queue. -DaveM
588 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700589static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700590{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700591 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700592 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700593 u32 now;
594
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700595 inet_csk_schedule_ack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700596
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700597 tcp_measure_rcv_mss(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700598
599 tcp_rcv_rtt_measure(tp);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900600
Linus Torvalds1da177e2005-04-16 15:20:36 -0700601 now = tcp_time_stamp;
602
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700603 if (!icsk->icsk_ack.ato) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700604 /* The _first_ data packet received, initialize
605 * delayed ACK engine.
606 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700607 tcp_incr_quickack(sk);
608 icsk->icsk_ack.ato = TCP_ATO_MIN;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700609 } else {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700610 int m = now - icsk->icsk_ack.lrcvtime;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700611
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800612 if (m <= TCP_ATO_MIN / 2) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700613 /* The fastest case is the first. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700614 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
615 } else if (m < icsk->icsk_ack.ato) {
616 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
617 if (icsk->icsk_ack.ato > icsk->icsk_rto)
618 icsk->icsk_ack.ato = icsk->icsk_rto;
619 } else if (m > icsk->icsk_rto) {
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800620 /* Too long gap. Apparently sender failed to
Linus Torvalds1da177e2005-04-16 15:20:36 -0700621 * restart window, so that we send ACKs quickly.
622 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700623 tcp_incr_quickack(sk);
Hideo Aoki3ab224b2007-12-31 00:11:19 -0800624 sk_mem_reclaim(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700625 }
626 }
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700627 icsk->icsk_ack.lrcvtime = now;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700628
629 TCP_ECN_check_ce(tp, skb);
630
631 if (skb->len >= 128)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700632 tcp_grow_window(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700633}
634
Linus Torvalds1da177e2005-04-16 15:20:36 -0700635/* Called to compute a smoothed rtt estimate. The data fed to this
636 * routine either comes from timestamps, or from segments that were
637 * known _not_ to have been retransmitted [see Karn/Partridge
638 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
639 * piece by Van Jacobson.
640 * NOTE: the next three routines used to be one big routine.
641 * To save cycles in the RFC 1323 implementation it was better to break
642 * it up into three procedures. -- erics
643 */
Stephen Hemminger2d2abba2005-11-10 16:56:12 -0800644static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700645{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300646 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700647 long m = mrtt; /* RTT */
648
Linus Torvalds1da177e2005-04-16 15:20:36 -0700649 /* The following amusing code comes from Jacobson's
650 * article in SIGCOMM '88. Note that rtt and mdev
651 * are scaled versions of rtt and mean deviation.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900652 * This is designed to be as fast as possible
Linus Torvalds1da177e2005-04-16 15:20:36 -0700653 * m stands for "measurement".
654 *
655 * On a 1990 paper the rto value is changed to:
656 * RTO = rtt + 4 * mdev
657 *
658 * Funny. This algorithm seems to be very broken.
659 * These formulae increase RTO, when it should be decreased, increase
Stephen Hemminger31f34262005-11-15 15:17:10 -0800660 * too slowly, when it should be increased quickly, decrease too quickly
Linus Torvalds1da177e2005-04-16 15:20:36 -0700661 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
662 * does not matter how to _calculate_ it. Seems, it was trap
663 * that VJ failed to avoid. 8)
664 */
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800665 if (m == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700666 m = 1;
667 if (tp->srtt != 0) {
668 m -= (tp->srtt >> 3); /* m is now error in rtt est */
669 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
670 if (m < 0) {
671 m = -m; /* m is now abs(error) */
672 m -= (tp->mdev >> 2); /* similar update on mdev */
673 /* This is similar to one of Eifel findings.
674 * Eifel blocks mdev updates when rtt decreases.
675 * This solution is a bit different: we use finer gain
676 * for mdev in this case (alpha*beta).
677 * Like Eifel it also prevents growth of rto,
678 * but also it limits too fast rto decreases,
679 * happening in pure Eifel.
680 */
681 if (m > 0)
682 m >>= 3;
683 } else {
684 m -= (tp->mdev >> 2); /* similar update on mdev */
685 }
686 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
687 if (tp->mdev > tp->mdev_max) {
688 tp->mdev_max = tp->mdev;
689 if (tp->mdev_max > tp->rttvar)
690 tp->rttvar = tp->mdev_max;
691 }
692 if (after(tp->snd_una, tp->rtt_seq)) {
693 if (tp->mdev_max < tp->rttvar)
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800694 tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700695 tp->rtt_seq = tp->snd_nxt;
David S. Miller05bb1fa2007-08-30 22:10:28 -0700696 tp->mdev_max = tcp_rto_min(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700697 }
698 } else {
699 /* no previous measure. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800700 tp->srtt = m << 3; /* take the measured time to be rtt */
701 tp->mdev = m << 1; /* make sure rto = 3*rtt */
David S. Miller05bb1fa2007-08-30 22:10:28 -0700702 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700703 tp->rtt_seq = tp->snd_nxt;
704 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700705}
706
707/* Calculate rto without backoff. This is the second half of Van Jacobson's
708 * routine referred to above.
709 */
David S. Miller4aabd8e2012-07-09 16:07:30 -0700710void tcp_set_rto(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700711{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700712 const struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700713 /* Old crap is replaced with new one. 8)
714 *
715 * More seriously:
716 * 1. If rtt variance happened to be less 50msec, it is hallucination.
717 * It cannot be less due to utterly erratic ACK generation made
718 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
719 * to do with delayed acks, because at cwnd>2 true delack timeout
720 * is invisible. Actually, Linux-2.4 also generates erratic
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800721 * ACKs in some circumstances.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700722 */
Damian Lukowskif1ecd5d2009-08-26 00:16:31 +0000723 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700724
725 /* 2. Fixups made earlier cannot be right.
726 * If we do not estimate RTO correctly without them,
727 * all the algo is pure shit and should be replaced
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800728 * with correct one. It is exactly, which we pretend to do.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700729 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700730
Ilpo Järvinenee6aac52008-12-05 22:43:08 -0800731 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
732 * guarantees that rto is higher.
733 */
Damian Lukowskif1ecd5d2009-08-26 00:16:31 +0000734 tcp_bound_rto(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700735}
736
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400737__u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700738{
739 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
740
Gerrit Renker22b71c82010-08-29 19:23:12 +0000741 if (!cwnd)
David S. Miller442b9632011-02-02 17:05:11 -0800742 cwnd = TCP_INIT_CWND;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700743 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
744}
745
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300746/*
747 * Packet counting of FACK is based on in-order assumptions, therefore TCP
748 * disables it when reordering is detected
749 */
David S. Miller4aabd8e2012-07-09 16:07:30 -0700750void tcp_disable_fack(struct tcp_sock *tp)
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300751{
Ilpo Järvinen85cc3912007-11-15 19:39:31 -0800752 /* RFC3517 uses different metric in lost marker => reset on change */
753 if (tcp_is_fack(tp))
754 tp->lost_skb_hint = NULL;
Vijay Subramanianab562222011-12-20 13:23:24 +0000755 tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300756}
757
Ryousei Takano564262c12007-10-25 23:03:52 -0700758/* Take a notice that peer is sending D-SACKs */
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300759static void tcp_dsack_seen(struct tcp_sock *tp)
760{
Vijay Subramanianab562222011-12-20 13:23:24 +0000761 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300762}
763
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300764static void tcp_update_reordering(struct sock *sk, const int metric,
765 const int ts)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700766{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300767 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700768 if (metric > tp->reordering) {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -0700769 int mib_idx;
770
Linus Torvalds1da177e2005-04-16 15:20:36 -0700771 tp->reordering = min(TCP_MAX_REORDERING, metric);
772
773 /* This exciting event is worth to be remembered. 8) */
774 if (ts)
Pavel Emelyanov40b215e2008-07-03 01:05:41 -0700775 mib_idx = LINUX_MIB_TCPTSREORDER;
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300776 else if (tcp_is_reno(tp))
Pavel Emelyanov40b215e2008-07-03 01:05:41 -0700777 mib_idx = LINUX_MIB_TCPRENOREORDER;
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300778 else if (tcp_is_fack(tp))
Pavel Emelyanov40b215e2008-07-03 01:05:41 -0700779 mib_idx = LINUX_MIB_TCPFACKREORDER;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700780 else
Pavel Emelyanov40b215e2008-07-03 01:05:41 -0700781 mib_idx = LINUX_MIB_TCPSACKREORDER;
782
Pavel Emelyanovde0744a2008-07-16 20:31:16 -0700783 NET_INC_STATS_BH(sock_net(sk), mib_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700784#if FASTRETRANS_DEBUG > 1
Joe Perches91df42b2012-05-15 14:11:54 +0000785 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
786 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
787 tp->reordering,
788 tp->fackets_out,
789 tp->sacked_out,
790 tp->undo_marker ? tp->undo_retrans : 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700791#endif
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300792 tcp_disable_fack(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700793 }
Yuchung Chengeed530b2012-05-02 13:30:03 +0000794
795 if (metric > 0)
796 tcp_disable_early_retrans(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700797}
798
Ilpo Järvinen006f5822008-09-20 21:20:20 -0700799/* This must be called before lost_out is incremented */
Ilpo Järvinenc8c213f2008-09-20 21:18:55 -0700800static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
801{
Ilpo Järvinen006f5822008-09-20 21:20:20 -0700802 if ((tp->retransmit_skb_hint == NULL) ||
Ilpo Järvinenc8c213f2008-09-20 21:18:55 -0700803 before(TCP_SKB_CB(skb)->seq,
804 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
Ilpo Järvinen006f5822008-09-20 21:20:20 -0700805 tp->retransmit_skb_hint = skb;
806
807 if (!tp->lost_out ||
808 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
809 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
Ilpo Järvinenc8c213f2008-09-20 21:18:55 -0700810}
811
Ilpo Järvinen41ea36e2008-09-20 21:19:22 -0700812static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
813{
814 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
815 tcp_verify_retransmit_hint(tp, skb);
816
817 tp->lost_out += tcp_skb_pcount(skb);
818 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
819 }
820}
821
Ilpo Järvinene1aa6802008-11-24 21:11:55 -0800822static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
823 struct sk_buff *skb)
Ilpo Järvinen006f5822008-09-20 21:20:20 -0700824{
825 tcp_verify_retransmit_hint(tp, skb);
826
827 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
828 tp->lost_out += tcp_skb_pcount(skb);
829 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
830 }
831}
832
Linus Torvalds1da177e2005-04-16 15:20:36 -0700833/* This procedure tags the retransmission queue when SACKs arrive.
834 *
835 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
836 * Packets in queue with these bits set are counted in variables
837 * sacked_out, retrans_out and lost_out, correspondingly.
838 *
839 * Valid combinations are:
840 * Tag InFlight Description
841 * 0 1 - orig segment is in flight.
842 * S 0 - nothing flies, orig reached receiver.
843 * L 0 - nothing flies, orig lost by net.
844 * R 2 - both orig and retransmit are in flight.
845 * L|R 1 - orig is lost, retransmit is in flight.
846 * S|R 1 - orig reached receiver, retrans is still in flight.
847 * (L|S|R is logically valid, it could occur when L|R is sacked,
848 * but it is equivalent to plain S and code short-curcuits it to S.
849 * L|S is logically invalid, it would mean -1 packet in flight 8))
850 *
851 * These 6 states form finite state machine, controlled by the following events:
852 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
853 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
Yuchung Cheng974c1232012-01-19 14:42:21 +0000854 * 3. Loss detection event of two flavors:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700855 * A. Scoreboard estimator decided the packet is lost.
856 * A'. Reno "three dupacks" marks head of queue lost.
Yuchung Cheng974c1232012-01-19 14:42:21 +0000857 * A''. Its FACK modification, head until snd.fack is lost.
858 * B. SACK arrives sacking SND.NXT at the moment, when the
Linus Torvalds1da177e2005-04-16 15:20:36 -0700859 * segment was retransmitted.
860 * 4. D-SACK added new rule: D-SACK changes any tag to S.
861 *
862 * It is pleasant to note, that state diagram turns out to be commutative,
863 * so that we are allowed not to be bothered by order of our actions,
864 * when multiple events arrive simultaneously. (see the function below).
865 *
866 * Reordering detection.
867 * --------------------
868 * Reordering metric is maximal distance, which a packet can be displaced
869 * in packet stream. With SACKs we can estimate it:
870 *
871 * 1. SACK fills old hole and the corresponding segment was not
872 * ever retransmitted -> reordering. Alas, we cannot use it
873 * when segment was retransmitted.
874 * 2. The last flaw is solved with D-SACK. D-SACK arrives
875 * for retransmitted and already SACKed segment -> reordering..
876 * Both of these heuristics are not used in Loss state, when we cannot
877 * account for retransmits accurately.
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -0700878 *
879 * SACK block validation.
880 * ----------------------
881 *
882 * SACK block range validation checks that the received SACK block fits to
883 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
884 * Note that SND.UNA is not included to the range though being valid because
Ilpo Järvinen0e835332007-10-01 15:28:17 -0700885 * it means that the receiver is rather inconsistent with itself reporting
886 * SACK reneging when it should advance SND.UNA. Such SACK block this is
887 * perfectly valid, however, in light of RFC2018 which explicitly states
888 * that "SACK block MUST reflect the newest segment. Even if the newest
889 * segment is going to be discarded ...", not that it looks very clever
890 * in case of head skb. Due to potentional receiver driven attacks, we
891 * choose to avoid immediate execution of a walk in write queue due to
892 * reneging and defer head skb's loss recovery to standard loss recovery
893 * procedure that will eventually trigger (nothing forbids us doing this).
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -0700894 *
895 * Implements also blockage to start_seq wrap-around. Problem lies in the
896 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
897 * there's no guarantee that it will be before snd_nxt (n). The problem
898 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
899 * wrap (s_w):
900 *
901 * <- outs wnd -> <- wrapzone ->
902 * u e n u_w e_w s n_w
903 * | | | | | | |
904 * |<------------+------+----- TCP seqno space --------------+---------->|
905 * ...-- <2^31 ->| |<--------...
906 * ...---- >2^31 ------>| |<--------...
907 *
908 * Current code wouldn't be vulnerable but it's better still to discard such
909 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
910 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
911 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
912 * equal to the ideal case (infinite seqno space without wrap caused issues).
913 *
914 * With D-SACK the lower bound is extended to cover sequence space below
915 * SND.UNA down to undo_marker, which is the last point of interest. Yet
Ryousei Takano564262c12007-10-25 23:03:52 -0700916 * again, D-SACK block must not to go across snd_una (for the same reason as
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -0700917 * for the normal SACK blocks, explained above). But there all simplicity
918 * ends, TCP might receive valid D-SACKs below that. As long as they reside
919 * fully below undo_marker they do not affect behavior in anyway and can
920 * therefore be safely ignored. In rare cases (which are more or less
921 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
922 * fragmentation and packet reordering past skb's retransmission. To consider
923 * them correctly, the acceptable range must be extended even more though
924 * the exact amount is rather hard to quantify. However, tp->max_window can
925 * be used as an exaggerated estimate.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700926 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +0000927static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
928 u32 start_seq, u32 end_seq)
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -0700929{
930 /* Too far in future, or reversed (interpretation is ambiguous) */
931 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
Eric Dumazeta2a385d2012-05-16 23:15:34 +0000932 return false;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -0700933
934 /* Nasty start_seq wrap-around check (see comments above) */
935 if (!before(start_seq, tp->snd_nxt))
Eric Dumazeta2a385d2012-05-16 23:15:34 +0000936 return false;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -0700937
Ryousei Takano564262c12007-10-25 23:03:52 -0700938 /* In outstanding window? ...This is valid exit for D-SACKs too.
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -0700939 * start_seq == snd_una is non-sensical (see comments above)
940 */
941 if (after(start_seq, tp->snd_una))
Eric Dumazeta2a385d2012-05-16 23:15:34 +0000942 return true;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -0700943
944 if (!is_dsack || !tp->undo_marker)
Eric Dumazeta2a385d2012-05-16 23:15:34 +0000945 return false;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -0700946
947 /* ...Then it's D-SACK, and must reside below snd_una completely */
Zheng Yanf779b2d2011-09-18 22:37:34 -0400948 if (after(end_seq, tp->snd_una))
Eric Dumazeta2a385d2012-05-16 23:15:34 +0000949 return false;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -0700950
951 if (!before(start_seq, tp->undo_marker))
Eric Dumazeta2a385d2012-05-16 23:15:34 +0000952 return true;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -0700953
954 /* Too old */
955 if (!after(end_seq, tp->undo_marker))
Eric Dumazeta2a385d2012-05-16 23:15:34 +0000956 return false;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -0700957
958 /* Undo_marker boundary crossing (overestimates a lot). Known already:
959 * start_seq < undo_marker and end_seq >= undo_marker.
960 */
961 return !before(start_seq, end_seq - tp->max_window);
962}
963
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -0700964/* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
Yuchung Cheng974c1232012-01-19 14:42:21 +0000965 * Event "B". Later note: FACK people cheated me again 8), we have to account
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -0700966 * for reordering! Ugly, but should help.
Ilpo Järvinenf785a8e2007-10-11 17:35:41 -0700967 *
968 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
969 * less than what is now known to be received by the other end (derived from
Ilpo Järvinen9f58f3b2007-11-15 19:42:54 -0800970 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
971 * retransmitted skbs to avoid some costly processing per ACKs.
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -0700972 */
Ilpo Järvinen407ef1d2007-12-02 00:47:57 +0200973static void tcp_mark_lost_retrans(struct sock *sk)
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -0700974{
Ilpo Järvinen9f58f3b2007-11-15 19:42:54 -0800975 const struct inet_connection_sock *icsk = inet_csk(sk);
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -0700976 struct tcp_sock *tp = tcp_sk(sk);
977 struct sk_buff *skb;
Ilpo Järvinenf785a8e2007-10-11 17:35:41 -0700978 int cnt = 0;
Ilpo Järvinendf2e0142007-10-18 05:07:57 -0700979 u32 new_low_seq = tp->snd_nxt;
Ilpo Järvinen6859d492007-12-02 00:48:06 +0200980 u32 received_upto = tcp_highest_sack_seq(tp);
Ilpo Järvinen9f58f3b2007-11-15 19:42:54 -0800981
982 if (!tcp_is_fack(tp) || !tp->retrans_out ||
983 !after(received_upto, tp->lost_retrans_low) ||
984 icsk->icsk_ca_state != TCP_CA_Recovery)
Ilpo Järvinen407ef1d2007-12-02 00:47:57 +0200985 return;
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -0700986
987 tcp_for_write_queue(skb, sk) {
988 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
989
990 if (skb == tcp_send_head(sk))
991 break;
Ilpo Järvinenf785a8e2007-10-11 17:35:41 -0700992 if (cnt == tp->retrans_out)
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -0700993 break;
994 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
995 continue;
996
Ilpo Järvinenf785a8e2007-10-11 17:35:41 -0700997 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
998 continue;
999
Ilpo Järvinend0af4162009-02-28 04:44:32 +00001000 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1001 * constraint here (see above) but figuring out that at
1002 * least tp->reordering SACK blocks reside between ack_seq
1003 * and received_upto is not easy task to do cheaply with
1004 * the available datastructures.
1005 *
1006 * Whether FACK should check here for tp->reordering segs
1007 * in-between one could argue for either way (it would be
1008 * rather simple to implement as we could count fack_count
1009 * during the walk and do tp->fackets_out - fack_count).
1010 */
1011 if (after(received_upto, ack_seq)) {
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001012 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1013 tp->retrans_out -= tcp_skb_pcount(skb);
1014
Ilpo Järvinen006f5822008-09-20 21:20:20 -07001015 tcp_skb_mark_lost_uncond_verify(tp, skb);
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07001016 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
Ilpo Järvinenf785a8e2007-10-11 17:35:41 -07001017 } else {
Ilpo Järvinendf2e0142007-10-18 05:07:57 -07001018 if (before(ack_seq, new_low_seq))
Ilpo Järvinenb08d6cb2007-10-11 17:36:13 -07001019 new_low_seq = ack_seq;
Ilpo Järvinenf785a8e2007-10-11 17:35:41 -07001020 cnt += tcp_skb_pcount(skb);
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001021 }
1022 }
Ilpo Järvinenb08d6cb2007-10-11 17:36:13 -07001023
1024 if (tp->retrans_out)
1025 tp->lost_retrans_low = new_low_seq;
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001026}
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001027
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001028static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1029 struct tcp_sack_block_wire *sp, int num_sacks,
1030 u32 prior_snd_una)
David S. Millerd06e0212007-06-18 22:43:06 -07001031{
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07001032 struct tcp_sock *tp = tcp_sk(sk);
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07001033 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1034 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001035 bool dup_sack = false;
David S. Millerd06e0212007-06-18 22:43:06 -07001036
1037 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001038 dup_sack = true;
Ilpo Järvinene60402d2007-08-09 15:14:46 +03001039 tcp_dsack_seen(tp);
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07001040 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
David S. Millerd06e0212007-06-18 22:43:06 -07001041 } else if (num_sacks > 1) {
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07001042 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1043 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
David S. Millerd06e0212007-06-18 22:43:06 -07001044
1045 if (!after(end_seq_0, end_seq_1) &&
1046 !before(start_seq_0, start_seq_1)) {
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001047 dup_sack = true;
Ilpo Järvinene60402d2007-08-09 15:14:46 +03001048 tcp_dsack_seen(tp);
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07001049 NET_INC_STATS_BH(sock_net(sk),
1050 LINUX_MIB_TCPDSACKOFORECV);
David S. Millerd06e0212007-06-18 22:43:06 -07001051 }
1052 }
1053
1054 /* D-SACK for already forgotten data... Do dumb counting. */
Yuchung Chengc24f6912011-02-07 12:57:04 +00001055 if (dup_sack && tp->undo_marker && tp->undo_retrans &&
David S. Millerd06e0212007-06-18 22:43:06 -07001056 !after(end_seq_0, prior_snd_una) &&
1057 after(end_seq_0, tp->undo_marker))
1058 tp->undo_retrans--;
1059
1060 return dup_sack;
1061}
1062
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001063struct tcp_sacktag_state {
1064 int reord;
1065 int fack_count;
1066 int flag;
1067};
1068
Ilpo Järvinend1935942007-10-11 17:34:25 -07001069/* Check if skb is fully within the SACK block. In presence of GSO skbs,
1070 * the incoming SACK may not exactly match but we can find smaller MSS
1071 * aligned portion of it that matches. Therefore we might need to fragment
1072 * which may fail and creates some hassle (caller must handle error case
1073 * returns).
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001074 *
1075 * FIXME: this could be merged to shift decision code
Ilpo Järvinend1935942007-10-11 17:34:25 -07001076 */
Adrian Bunk0f79efd2007-10-26 03:57:36 -07001077static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001078 u32 start_seq, u32 end_seq)
Ilpo Järvinend1935942007-10-11 17:34:25 -07001079{
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001080 int err;
1081 bool in_sack;
Ilpo Järvinend1935942007-10-11 17:34:25 -07001082 unsigned int pkt_len;
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001083 unsigned int mss;
Ilpo Järvinend1935942007-10-11 17:34:25 -07001084
1085 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1086 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1087
1088 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1089 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001090 mss = tcp_skb_mss(skb);
Ilpo Järvinend1935942007-10-11 17:34:25 -07001091 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1092
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001093 if (!in_sack) {
Ilpo Järvinend1935942007-10-11 17:34:25 -07001094 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001095 if (pkt_len < mss)
1096 pkt_len = mss;
1097 } else {
Ilpo Järvinend1935942007-10-11 17:34:25 -07001098 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001099 if (pkt_len < mss)
1100 return -EINVAL;
1101 }
1102
1103 /* Round if necessary so that SACKs cover only full MSSes
1104 * and/or the remaining small portion (if present)
1105 */
1106 if (pkt_len > mss) {
1107 unsigned int new_len = (pkt_len / mss) * mss;
1108 if (!in_sack && new_len < pkt_len) {
1109 new_len += mss;
1110 if (new_len > skb->len)
1111 return 0;
1112 }
1113 pkt_len = new_len;
1114 }
1115 err = tcp_fragment(sk, skb, pkt_len, mss);
Ilpo Järvinend1935942007-10-11 17:34:25 -07001116 if (err < 0)
1117 return err;
1118 }
1119
1120 return in_sack;
1121}
1122
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001123/* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1124static u8 tcp_sacktag_one(struct sock *sk,
1125 struct tcp_sacktag_state *state, u8 sacked,
1126 u32 start_seq, u32 end_seq,
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001127 bool dup_sack, int pcount)
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001128{
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001129 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001130 int fack_count = state->fack_count;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001131
1132 /* Account D-SACK for retransmitted packet. */
1133 if (dup_sack && (sacked & TCPCB_RETRANS)) {
Yuchung Chengc24f6912011-02-07 12:57:04 +00001134 if (tp->undo_marker && tp->undo_retrans &&
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001135 after(end_seq, tp->undo_marker))
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001136 tp->undo_retrans--;
Ilpo Järvinenede9f3b2007-12-02 00:47:58 +02001137 if (sacked & TCPCB_SACKED_ACKED)
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001138 state->reord = min(fack_count, state->reord);
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001139 }
1140
1141 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001142 if (!after(end_seq, tp->snd_una))
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001143 return sacked;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001144
1145 if (!(sacked & TCPCB_SACKED_ACKED)) {
1146 if (sacked & TCPCB_SACKED_RETRANS) {
1147 /* If the segment is not tagged as lost,
1148 * we do not clear RETRANS, believing
1149 * that retransmission is still in flight.
1150 */
1151 if (sacked & TCPCB_LOST) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001152 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001153 tp->lost_out -= pcount;
1154 tp->retrans_out -= pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001155 }
1156 } else {
1157 if (!(sacked & TCPCB_RETRANS)) {
1158 /* New sack for not retransmitted frame,
1159 * which was in hole. It is reordering.
1160 */
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001161 if (before(start_seq,
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001162 tcp_highest_sack_seq(tp)))
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001163 state->reord = min(fack_count,
1164 state->reord);
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001165
1166 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001167 if (!after(end_seq, tp->frto_highmark))
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001168 state->flag |= FLAG_ONLY_ORIG_SACKED;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001169 }
1170
1171 if (sacked & TCPCB_LOST) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001172 sacked &= ~TCPCB_LOST;
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001173 tp->lost_out -= pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001174 }
1175 }
1176
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001177 sacked |= TCPCB_SACKED_ACKED;
1178 state->flag |= FLAG_DATA_SACKED;
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001179 tp->sacked_out += pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001180
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001181 fack_count += pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001182
1183 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1184 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001185 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001186 tp->lost_cnt_hint += pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001187
1188 if (fack_count > tp->fackets_out)
1189 tp->fackets_out = fack_count;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001190 }
1191
1192 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1193 * frames and clear it. undo_retrans is decreased above, L|R frames
1194 * are accounted above as well.
1195 */
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001196 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1197 sacked &= ~TCPCB_SACKED_RETRANS;
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001198 tp->retrans_out -= pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001199 }
1200
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001201 return sacked;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001202}
1203
Neal Cardwelldaef52b2012-02-12 18:37:10 +00001204/* Shift newly-SACKed bytes from this skb to the immediately previous
1205 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1206 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001207static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1208 struct tcp_sacktag_state *state,
1209 unsigned int pcount, int shifted, int mss,
1210 bool dup_sack)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001211{
1212 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen50133162008-12-05 22:42:41 -08001213 struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
Neal Cardwelldaef52b2012-02-12 18:37:10 +00001214 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1215 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001216
1217 BUG_ON(!pcount);
1218
Neal Cardwell4c90d3b2012-02-26 10:06:19 +00001219 /* Adjust counters and hints for the newly sacked sequence
1220 * range but discard the return value since prev is already
1221 * marked. We must tag the range first because the seq
1222 * advancement below implicitly advances
1223 * tcp_highest_sack_seq() when skb is highest_sack.
1224 */
1225 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1226 start_seq, end_seq, dup_sack, pcount);
1227
1228 if (skb == tp->lost_skb_hint)
Neal Cardwell0af2a0d2012-02-13 20:22:08 +00001229 tp->lost_cnt_hint += pcount;
1230
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001231 TCP_SKB_CB(prev)->end_seq += shifted;
1232 TCP_SKB_CB(skb)->seq += shifted;
1233
1234 skb_shinfo(prev)->gso_segs += pcount;
1235 BUG_ON(skb_shinfo(skb)->gso_segs < pcount);
1236 skb_shinfo(skb)->gso_segs -= pcount;
1237
1238 /* When we're adding to gso_segs == 1, gso_size will be zero,
1239 * in theory this shouldn't be necessary but as long as DSACK
1240 * code can come after this skb later on it's better to keep
1241 * setting gso_size to something.
1242 */
1243 if (!skb_shinfo(prev)->gso_size) {
1244 skb_shinfo(prev)->gso_size = mss;
1245 skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1246 }
1247
1248 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1249 if (skb_shinfo(skb)->gso_segs <= 1) {
1250 skb_shinfo(skb)->gso_size = 0;
1251 skb_shinfo(skb)->gso_type = 0;
1252 }
1253
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001254 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1255 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1256
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001257 if (skb->len > 0) {
1258 BUG_ON(!tcp_skb_pcount(skb));
Ilpo Järvinen111cc8b2008-11-24 21:27:22 -08001259 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001260 return false;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001261 }
1262
1263 /* Whole SKB was eaten :-) */
1264
Ilpo Järvinen92ee76b2008-11-24 21:26:56 -08001265 if (skb == tp->retransmit_skb_hint)
1266 tp->retransmit_skb_hint = prev;
1267 if (skb == tp->scoreboard_skb_hint)
1268 tp->scoreboard_skb_hint = prev;
1269 if (skb == tp->lost_skb_hint) {
1270 tp->lost_skb_hint = prev;
1271 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1272 }
1273
Eric Dumazet4de075e2011-09-27 13:25:05 -04001274 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(prev)->tcp_flags;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001275 if (skb == tcp_highest_sack(sk))
1276 tcp_advance_highest_sack(sk, skb);
1277
1278 tcp_unlink_write_queue(skb, sk);
1279 sk_wmem_free_skb(sk, skb);
1280
Ilpo Järvinen111cc8b2008-11-24 21:27:22 -08001281 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1282
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001283 return true;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001284}
1285
1286/* I wish gso_size would have a bit more sane initialization than
1287 * something-or-zero which complicates things
1288 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001289static int tcp_skb_seglen(const struct sk_buff *skb)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001290{
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001291 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001292}
1293
1294/* Shifting pages past head area doesn't work */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001295static int skb_can_shift(const struct sk_buff *skb)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001296{
1297 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1298}
1299
1300/* Try collapsing SACK blocks spanning across multiple skbs to a single
1301 * skb.
1302 */
1303static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001304 struct tcp_sacktag_state *state,
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001305 u32 start_seq, u32 end_seq,
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001306 bool dup_sack)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001307{
1308 struct tcp_sock *tp = tcp_sk(sk);
1309 struct sk_buff *prev;
1310 int mss;
1311 int pcount = 0;
1312 int len;
1313 int in_sack;
1314
1315 if (!sk_can_gso(sk))
1316 goto fallback;
1317
1318 /* Normally R but no L won't result in plain S */
1319 if (!dup_sack &&
Ilpo Järvinen9969ca52008-12-05 22:41:06 -08001320 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001321 goto fallback;
1322 if (!skb_can_shift(skb))
1323 goto fallback;
1324 /* This frame is about to be dropped (was ACKed). */
1325 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1326 goto fallback;
1327
1328 /* Can only happen with delayed DSACK + discard craziness */
1329 if (unlikely(skb == tcp_write_queue_head(sk)))
1330 goto fallback;
1331 prev = tcp_write_queue_prev(sk, skb);
1332
1333 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1334 goto fallback;
1335
1336 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1337 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1338
1339 if (in_sack) {
1340 len = skb->len;
1341 pcount = tcp_skb_pcount(skb);
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001342 mss = tcp_skb_seglen(skb);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001343
1344 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1345 * drop this restriction as unnecessary
1346 */
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001347 if (mss != tcp_skb_seglen(prev))
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001348 goto fallback;
1349 } else {
1350 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1351 goto noop;
1352 /* CHECKME: This is non-MSS split case only?, this will
1353 * cause skipped skbs due to advancing loop btw, original
1354 * has that feature too
1355 */
1356 if (tcp_skb_pcount(skb) <= 1)
1357 goto noop;
1358
1359 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1360 if (!in_sack) {
1361 /* TODO: head merge to next could be attempted here
1362 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1363 * though it might not be worth of the additional hassle
1364 *
1365 * ...we can probably just fallback to what was done
1366 * previously. We could try merging non-SACKed ones
1367 * as well but it probably isn't going to buy off
1368 * because later SACKs might again split them, and
1369 * it would make skb timestamp tracking considerably
1370 * harder problem.
1371 */
1372 goto fallback;
1373 }
1374
1375 len = end_seq - TCP_SKB_CB(skb)->seq;
1376 BUG_ON(len < 0);
1377 BUG_ON(len > skb->len);
1378
1379 /* MSS boundaries should be honoured or else pcount will
1380 * severely break even though it makes things bit trickier.
1381 * Optimize common case to avoid most of the divides
1382 */
1383 mss = tcp_skb_mss(skb);
1384
1385 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1386 * drop this restriction as unnecessary
1387 */
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001388 if (mss != tcp_skb_seglen(prev))
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001389 goto fallback;
1390
1391 if (len == mss) {
1392 pcount = 1;
1393 } else if (len < mss) {
1394 goto noop;
1395 } else {
1396 pcount = len / mss;
1397 len = pcount * mss;
1398 }
1399 }
1400
Neal Cardwell4648dc92012-03-05 19:35:04 +00001401 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1402 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1403 goto fallback;
1404
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001405 if (!skb_shift(prev, skb, len))
1406 goto fallback;
Ilpo Järvinen9ec06ff2009-03-01 00:21:36 -08001407 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001408 goto out;
1409
1410 /* Hole filled allows collapsing with the next as well, this is very
1411 * useful when hole on every nth skb pattern happens
1412 */
1413 if (prev == tcp_write_queue_tail(sk))
1414 goto out;
1415 skb = tcp_write_queue_next(sk, prev);
1416
Ilpo Järvinenf0bc52f2008-12-05 22:40:47 -08001417 if (!skb_can_shift(skb) ||
1418 (skb == tcp_send_head(sk)) ||
1419 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001420 (mss != tcp_skb_seglen(skb)))
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001421 goto out;
1422
1423 len = skb->len;
1424 if (skb_shift(prev, skb, len)) {
1425 pcount += tcp_skb_pcount(skb);
Ilpo Järvinen9ec06ff2009-03-01 00:21:36 -08001426 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001427 }
1428
1429out:
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001430 state->fack_count += pcount;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001431 return prev;
1432
1433noop:
1434 return skb;
1435
1436fallback:
Ilpo Järvinen111cc8b2008-11-24 21:27:22 -08001437 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001438 return NULL;
1439}
1440
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001441static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1442 struct tcp_sack_block *next_dup,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001443 struct tcp_sacktag_state *state,
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001444 u32 start_seq, u32 end_seq,
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001445 bool dup_sack_in)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001446{
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001447 struct tcp_sock *tp = tcp_sk(sk);
1448 struct sk_buff *tmp;
1449
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001450 tcp_for_write_queue_from(skb, sk) {
1451 int in_sack = 0;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001452 bool dup_sack = dup_sack_in;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001453
1454 if (skb == tcp_send_head(sk))
1455 break;
1456
1457 /* queue is in-order => we can short-circuit the walk early */
1458 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1459 break;
1460
1461 if ((next_dup != NULL) &&
1462 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1463 in_sack = tcp_match_skb_to_sack(sk, skb,
1464 next_dup->start_seq,
1465 next_dup->end_seq);
1466 if (in_sack > 0)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001467 dup_sack = true;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001468 }
1469
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001470 /* skb reference here is a bit tricky to get right, since
1471 * shifting can eat and free both this skb and the next,
1472 * so not even _safe variant of the loop is enough.
1473 */
1474 if (in_sack <= 0) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001475 tmp = tcp_shift_skb_data(sk, skb, state,
1476 start_seq, end_seq, dup_sack);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001477 if (tmp != NULL) {
1478 if (tmp != skb) {
1479 skb = tmp;
1480 continue;
1481 }
1482
1483 in_sack = 0;
1484 } else {
1485 in_sack = tcp_match_skb_to_sack(sk, skb,
1486 start_seq,
1487 end_seq);
1488 }
1489 }
1490
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001491 if (unlikely(in_sack < 0))
1492 break;
1493
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001494 if (in_sack) {
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001495 TCP_SKB_CB(skb)->sacked =
1496 tcp_sacktag_one(sk,
1497 state,
1498 TCP_SKB_CB(skb)->sacked,
1499 TCP_SKB_CB(skb)->seq,
1500 TCP_SKB_CB(skb)->end_seq,
1501 dup_sack,
1502 tcp_skb_pcount(skb));
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001503
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001504 if (!before(TCP_SKB_CB(skb)->seq,
1505 tcp_highest_sack_seq(tp)))
1506 tcp_advance_highest_sack(sk, skb);
1507 }
1508
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001509 state->fack_count += tcp_skb_pcount(skb);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001510 }
1511 return skb;
1512}
1513
1514/* Avoid all extra work that is being done by sacktag while walking in
1515 * a normal way
1516 */
1517static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001518 struct tcp_sacktag_state *state,
1519 u32 skip_to_seq)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001520{
1521 tcp_for_write_queue_from(skb, sk) {
1522 if (skb == tcp_send_head(sk))
1523 break;
1524
Ilpo Järvinene8bae272008-11-24 21:12:28 -08001525 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001526 break;
Ilpo Järvinend152a7d2008-03-03 12:10:16 -08001527
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001528 state->fack_count += tcp_skb_pcount(skb);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001529 }
1530 return skb;
1531}
1532
1533static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1534 struct sock *sk,
1535 struct tcp_sack_block *next_dup,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001536 struct tcp_sacktag_state *state,
1537 u32 skip_to_seq)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001538{
1539 if (next_dup == NULL)
1540 return skb;
1541
1542 if (before(next_dup->start_seq, skip_to_seq)) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001543 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1544 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1545 next_dup->start_seq, next_dup->end_seq,
1546 1);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001547 }
1548
1549 return skb;
1550}
1551
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001552static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001553{
1554 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1555}
1556
Linus Torvalds1da177e2005-04-16 15:20:36 -07001557static int
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001558tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001559 u32 prior_snd_una)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001560{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001561 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001562 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001563 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1564 TCP_SKB_CB(ack_skb)->sacked);
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001565 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
Adam Langley4389dde2008-07-19 00:07:02 -07001566 struct tcp_sack_block sp[TCP_NUM_SACKS];
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001567 struct tcp_sack_block *cache;
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001568 struct tcp_sacktag_state state;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001569 struct sk_buff *skb;
Adam Langley4389dde2008-07-19 00:07:02 -07001570 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001571 int used_sacks;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001572 bool found_dup_sack = false;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001573 int i, j;
Baruch Evenfda03fb2007-02-04 23:35:57 -08001574 int first_sack_index;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001575
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001576 state.flag = 0;
1577 state.reord = tp->packets_out;
1578
Ilpo Järvinend738cd82007-03-24 21:03:23 -07001579 if (!tp->sacked_out) {
Ilpo Järvinende83c052007-10-07 23:37:55 -07001580 if (WARN_ON(tp->fackets_out))
1581 tp->fackets_out = 0;
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001582 tcp_highest_sack_reset(sk);
Ilpo Järvinend738cd82007-03-24 21:03:23 -07001583 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001584
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07001585 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
David S. Millerd06e0212007-06-18 22:43:06 -07001586 num_sacks, prior_snd_una);
1587 if (found_dup_sack)
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001588 state.flag |= FLAG_DSACKING_ACK;
Baruch Even6f746512007-02-04 23:36:42 -08001589
1590 /* Eliminate too old ACKs, but take into
1591 * account more or less fresh ones, they can
1592 * contain valid SACK info.
1593 */
1594 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1595 return 0;
1596
Ilpo Järvinen96a2d412007-11-14 15:47:18 -08001597 if (!tp->packets_out)
1598 goto out;
1599
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001600 used_sacks = 0;
1601 first_sack_index = 0;
1602 for (i = 0; i < num_sacks; i++) {
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001603 bool dup_sack = !i && found_dup_sack;
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001604
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07001605 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1606 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001607
1608 if (!tcp_is_sackblock_valid(tp, dup_sack,
1609 sp[used_sacks].start_seq,
1610 sp[used_sacks].end_seq)) {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001611 int mib_idx;
1612
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001613 if (dup_sack) {
1614 if (!tp->undo_marker)
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001615 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001616 else
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001617 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001618 } else {
1619 /* Don't count olds caused by ACK reordering */
1620 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1621 !after(sp[used_sacks].end_seq, tp->snd_una))
1622 continue;
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001623 mib_idx = LINUX_MIB_TCPSACKDISCARD;
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001624 }
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001625
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07001626 NET_INC_STATS_BH(sock_net(sk), mib_idx);
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001627 if (i == 0)
1628 first_sack_index = -1;
1629 continue;
1630 }
1631
1632 /* Ignore very old stuff early */
1633 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1634 continue;
1635
1636 used_sacks++;
1637 }
1638
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001639 /* order SACK blocks to allow in order walk of the retrans queue */
1640 for (i = used_sacks - 1; i > 0; i--) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001641 for (j = 0; j < i; j++) {
1642 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
Ilpo Järvinena0bffff2009-03-21 13:36:17 -07001643 swap(sp[j], sp[j + 1]);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001644
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001645 /* Track where the first SACK block goes to */
1646 if (j == first_sack_index)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001647 first_sack_index = j + 1;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001648 }
1649 }
1650 }
1651
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001652 skb = tcp_write_queue_head(sk);
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001653 state.fack_count = 0;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001654 i = 0;
1655
1656 if (!tp->sacked_out) {
1657 /* It's already past, so skip checking against it */
1658 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1659 } else {
1660 cache = tp->recv_sack_cache;
1661 /* Skip empty blocks in at head of the cache */
1662 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1663 !cache->end_seq)
1664 cache++;
Baruch Evenfda03fb2007-02-04 23:35:57 -08001665 }
1666
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001667 while (i < used_sacks) {
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001668 u32 start_seq = sp[i].start_seq;
1669 u32 end_seq = sp[i].end_seq;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001670 bool dup_sack = (found_dup_sack && (i == first_sack_index));
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001671 struct tcp_sack_block *next_dup = NULL;
Ilpo Järvinene56d6cd2007-11-01 00:09:37 -07001672
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001673 if (found_dup_sack && ((i + 1) == first_sack_index))
1674 next_dup = &sp[i + 1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001675
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001676 /* Skip too early cached blocks */
1677 while (tcp_sack_cache_ok(tp, cache) &&
1678 !before(start_seq, cache->end_seq))
1679 cache++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001680
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001681 /* Can skip some work by looking recv_sack_cache? */
1682 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1683 after(end_seq, cache->start_seq)) {
David S. Millerfe067e82007-03-07 12:12:44 -08001684
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001685 /* Head todo? */
1686 if (before(start_seq, cache->start_seq)) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001687 skb = tcp_sacktag_skip(skb, sk, &state,
1688 start_seq);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001689 skb = tcp_sacktag_walk(skb, sk, next_dup,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001690 &state,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001691 start_seq,
1692 cache->start_seq,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001693 dup_sack);
Baruch Evenfda03fb2007-02-04 23:35:57 -08001694 }
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001695
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001696 /* Rest of the block already fully processed? */
Ilpo Järvinen20de20b2007-11-16 16:17:05 -08001697 if (!after(end_seq, cache->end_seq))
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001698 goto advance_sp;
Ilpo Järvinen20de20b2007-11-16 16:17:05 -08001699
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001700 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001701 &state,
1702 cache->end_seq);
Ilpo Järvinene56d6cd2007-11-01 00:09:37 -07001703
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001704 /* ...tail remains todo... */
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001705 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
Ilpo Järvinen20de20b2007-11-16 16:17:05 -08001706 /* ...but better entrypoint exists! */
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001707 skb = tcp_highest_sack(sk);
1708 if (skb == NULL)
1709 break;
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001710 state.fack_count = tp->fackets_out;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001711 cache++;
1712 goto walk;
1713 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001714
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001715 skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001716 /* Check overlap against next cached too (past this one already) */
1717 cache++;
1718 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001719 }
Ilpo Järvinenfbd52eb2007-11-10 21:24:19 -08001720
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001721 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1722 skb = tcp_highest_sack(sk);
1723 if (skb == NULL)
1724 break;
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001725 state.fack_count = tp->fackets_out;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001726 }
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001727 skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001728
1729walk:
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001730 skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1731 start_seq, end_seq, dup_sack);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001732
1733advance_sp:
Ilpo Järvinenfbd52eb2007-11-10 21:24:19 -08001734 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1735 * due to in-order walk
1736 */
1737 if (after(end_seq, tp->frto_highmark))
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001738 state.flag &= ~FLAG_ONLY_ORIG_SACKED;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001739
1740 i++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001741 }
1742
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001743 /* Clear the head of the cache sack blocks so we can skip it next time */
1744 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1745 tp->recv_sack_cache[i].start_seq = 0;
1746 tp->recv_sack_cache[i].end_seq = 0;
1747 }
1748 for (j = 0; j < used_sacks; j++)
1749 tp->recv_sack_cache[i++] = sp[j];
1750
Ilpo Järvinen407ef1d2007-12-02 00:47:57 +02001751 tcp_mark_lost_retrans(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001752
Ilpo Järvinen86426c22007-08-09 14:45:17 +03001753 tcp_verify_left_out(tp);
1754
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001755 if ((state.reord < tp->fackets_out) &&
Ilpo Järvinenf5771112007-11-15 19:35:11 -08001756 ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) &&
Ilpo Järvinenc5e7af02007-02-23 16:22:06 -08001757 (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001758 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001759
Ilpo Järvinen96a2d412007-11-14 15:47:18 -08001760out:
1761
Linus Torvalds1da177e2005-04-16 15:20:36 -07001762#if FASTRETRANS_DEBUG > 0
Ilpo Järvinen547b7922008-07-25 21:43:18 -07001763 WARN_ON((int)tp->sacked_out < 0);
1764 WARN_ON((int)tp->lost_out < 0);
1765 WARN_ON((int)tp->retrans_out < 0);
1766 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001767#endif
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001768 return state.flag;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001769}
1770
Ilpo Järvinen882beba2008-04-07 22:33:07 -07001771/* Limits sacked_out so that sum with lost_out isn't ever larger than
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001772 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
Ilpo Järvinen30935cf2007-02-21 23:01:36 -08001773 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001774static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001775{
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001776 u32 holes;
1777
1778 holes = max(tp->lost_out, 1U);
1779 holes = min(holes, tp->packets_out);
1780
1781 if ((tp->sacked_out + holes) > tp->packets_out) {
1782 tp->sacked_out = tp->packets_out - holes;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001783 return true;
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001784 }
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001785 return false;
Ilpo Järvinen882beba2008-04-07 22:33:07 -07001786}
1787
1788/* If we receive more dupacks than we expected counting segments
1789 * in assumption of absent reordering, interpret this as reordering.
1790 * The only another reason could be bug in receiver TCP.
1791 */
1792static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1793{
1794 struct tcp_sock *tp = tcp_sk(sk);
1795 if (tcp_limit_reno_sacked(tp))
1796 tcp_update_reordering(sk, tp->packets_out + addend, 0);
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001797}
1798
1799/* Emulate SACKs for SACKless connection: account for a new dupack. */
1800
1801static void tcp_add_reno_sack(struct sock *sk)
1802{
1803 struct tcp_sock *tp = tcp_sk(sk);
1804 tp->sacked_out++;
1805 tcp_check_reno_reordering(sk, 0);
Ilpo Järvinen005903b2007-08-09 14:44:16 +03001806 tcp_verify_left_out(tp);
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001807}
1808
1809/* Account for ACK, ACKing some data in Reno Recovery phase. */
1810
1811static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1812{
1813 struct tcp_sock *tp = tcp_sk(sk);
1814
1815 if (acked > 0) {
1816 /* One ACK acked hole. The rest eat duplicate ACKs. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001817 if (acked - 1 >= tp->sacked_out)
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001818 tp->sacked_out = 0;
1819 else
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001820 tp->sacked_out -= acked - 1;
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001821 }
1822 tcp_check_reno_reordering(sk, acked);
Ilpo Järvinen005903b2007-08-09 14:44:16 +03001823 tcp_verify_left_out(tp);
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001824}
1825
1826static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1827{
1828 tp->sacked_out = 0;
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001829}
1830
Ilpo Järvinen62ab2222008-05-08 01:09:11 -07001831static int tcp_is_sackfrto(const struct tcp_sock *tp)
1832{
1833 return (sysctl_tcp_frto == 0x2) && !tcp_is_reno(tp);
1834}
1835
Ilpo Järvinen95eacd22007-10-01 15:27:42 -07001836/* F-RTO can only be used if TCP has never retransmitted anything other than
1837 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1838 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001839bool tcp_use_frto(struct sock *sk)
Ilpo Järvinenbdaae172007-02-21 22:59:58 -08001840{
1841 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen6adb4f72008-04-07 22:33:57 -07001842 const struct inet_connection_sock *icsk = inet_csk(sk);
Ilpo Järvinen46d0de42007-02-21 23:10:39 -08001843 struct sk_buff *skb;
Ilpo Järvinenbdaae172007-02-21 22:59:58 -08001844
Ilpo Järvinen575ee712007-04-30 00:39:55 -07001845 if (!sysctl_tcp_frto)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001846 return false;
Ilpo Järvinen46d0de42007-02-21 23:10:39 -08001847
Ilpo Järvinen6adb4f72008-04-07 22:33:57 -07001848 /* MTU probe and F-RTO won't really play nicely along currently */
1849 if (icsk->icsk_mtup.probe_size)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001850 return false;
Ilpo Järvinen6adb4f72008-04-07 22:33:57 -07001851
Ilpo Järvinen62ab2222008-05-08 01:09:11 -07001852 if (tcp_is_sackfrto(tp))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001853 return true;
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08001854
Ilpo Järvinen46d0de42007-02-21 23:10:39 -08001855 /* Avoid expensive walking of rexmit queue if possible */
1856 if (tp->retrans_out > 1)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001857 return false;
Ilpo Järvinen46d0de42007-02-21 23:10:39 -08001858
David S. Millerfe067e82007-03-07 12:12:44 -08001859 skb = tcp_write_queue_head(sk);
David S. Miller28e34872008-09-23 02:51:41 -07001860 if (tcp_skb_is_last(sk, skb))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001861 return true;
David S. Millerfe067e82007-03-07 12:12:44 -08001862 skb = tcp_write_queue_next(sk, skb); /* Skips head */
1863 tcp_for_write_queue_from(skb, sk) {
1864 if (skb == tcp_send_head(sk))
1865 break;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001866 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001867 return false;
Ilpo Järvinen46d0de42007-02-21 23:10:39 -08001868 /* Short-circuit when first non-SACKed skb has been checked */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001869 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
Ilpo Järvinen46d0de42007-02-21 23:10:39 -08001870 break;
1871 }
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001872 return true;
Ilpo Järvinenbdaae172007-02-21 22:59:58 -08001873}
1874
Ilpo Järvinen30935cf2007-02-21 23:01:36 -08001875/* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1876 * recovery a bit and use heuristics in tcp_process_frto() to detect if
Ilpo Järvinend1a54c62007-02-21 23:11:57 -08001877 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1878 * keep retrans_out counting accurate (with SACK F-RTO, other than head
1879 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1880 * bits are handled if the Loss state is really to be entered (in
1881 * tcp_enter_frto_loss).
Ilpo Järvinen7487c482007-02-21 23:02:30 -08001882 *
1883 * Do like tcp_enter_loss() would; when RTO expires the second time it
1884 * does:
1885 * "Reduce ssthresh if it has not yet been made inside this window."
Linus Torvalds1da177e2005-04-16 15:20:36 -07001886 */
1887void tcp_enter_frto(struct sock *sk)
1888{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001889 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001890 struct tcp_sock *tp = tcp_sk(sk);
1891 struct sk_buff *skb;
1892
Ilpo Järvinen7487c482007-02-21 23:02:30 -08001893 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001894 tp->snd_una == tp->high_seq ||
Ilpo Järvinen7487c482007-02-21 23:02:30 -08001895 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
1896 !icsk->icsk_retransmits)) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001897 tp->prior_ssthresh = tcp_current_ssthresh(sk);
Ilpo Järvinen66e93e42007-02-21 23:13:47 -08001898 /* Our state is too optimistic in ssthresh() call because cwnd
Ryousei Takano564262c12007-10-25 23:03:52 -07001899 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
Ilpo Järvinen66e93e42007-02-21 23:13:47 -08001900 * recovery has not yet completed. Pattern would be this: RTO,
1901 * Cumulative ACK, RTO (2xRTO for the same segment does not end
1902 * up here twice).
1903 * RFC4138 should be more specific on what to do, even though
1904 * RTO is quite unlikely to occur after the first Cumulative ACK
1905 * due to back-off and complexity of triggering events ...
1906 */
1907 if (tp->frto_counter) {
1908 u32 stored_cwnd;
1909 stored_cwnd = tp->snd_cwnd;
1910 tp->snd_cwnd = 2;
1911 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1912 tp->snd_cwnd = stored_cwnd;
1913 } else {
1914 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1915 }
1916 /* ... in theory, cong.control module could do "any tricks" in
1917 * ssthresh(), which means that ca_state, lost bits and lost_out
1918 * counter would have to be faked before the call occurs. We
1919 * consider that too expensive, unlikely and hacky, so modules
1920 * using these in ssthresh() must deal these incompatibility
1921 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1922 */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001923 tcp_ca_event(sk, CA_EVENT_FRTO);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001924 }
1925
Linus Torvalds1da177e2005-04-16 15:20:36 -07001926 tp->undo_marker = tp->snd_una;
1927 tp->undo_retrans = 0;
1928
David S. Millerfe067e82007-03-07 12:12:44 -08001929 skb = tcp_write_queue_head(sk);
Ilpo Järvinen009a2e32007-09-20 11:34:38 -07001930 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1931 tp->undo_marker = 0;
Ilpo Järvinend1a54c62007-02-21 23:11:57 -08001932 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
Ilpo Järvinen522e7542007-02-21 22:54:52 -08001933 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
Ilpo Järvinend1a54c62007-02-21 23:11:57 -08001934 tp->retrans_out -= tcp_skb_pcount(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001935 }
Ilpo Järvinen005903b2007-08-09 14:44:16 +03001936 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001937
Ilpo Järvinen746aa322007-11-13 21:01:23 -08001938 /* Too bad if TCP was application limited */
1939 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
1940
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08001941 /* Earlier loss recovery underway (see RFC4138; Appendix B).
1942 * The last condition is necessary at least in tp->frto_counter case.
1943 */
Ilpo Järvinen62ab2222008-05-08 01:09:11 -07001944 if (tcp_is_sackfrto(tp) && (tp->frto_counter ||
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08001945 ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
1946 after(tp->high_seq, tp->snd_una)) {
1947 tp->frto_highmark = tp->high_seq;
1948 } else {
1949 tp->frto_highmark = tp->snd_nxt;
1950 }
Ilpo Järvinen7b0eb222007-02-21 23:03:35 -08001951 tcp_set_ca_state(sk, TCP_CA_Disorder);
1952 tp->high_seq = tp->snd_nxt;
Ilpo Järvinen7487c482007-02-21 23:02:30 -08001953 tp->frto_counter = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001954}
1955
1956/* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1957 * which indicates that we should follow the traditional RTO recovery,
1958 * i.e. mark everything lost and do go-back-N retransmission.
1959 */
Ilpo Järvinend1a54c62007-02-21 23:11:57 -08001960static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001961{
1962 struct tcp_sock *tp = tcp_sk(sk);
1963 struct sk_buff *skb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001964
Linus Torvalds1da177e2005-04-16 15:20:36 -07001965 tp->lost_out = 0;
Ilpo Järvinend1a54c62007-02-21 23:11:57 -08001966 tp->retrans_out = 0;
Ilpo Järvinene60402d2007-08-09 15:14:46 +03001967 if (tcp_is_reno(tp))
Ilpo Järvinen9bff40f2007-05-27 01:53:49 -07001968 tcp_reset_reno_sack(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001969
David S. Millerfe067e82007-03-07 12:12:44 -08001970 tcp_for_write_queue(skb, sk) {
1971 if (skb == tcp_send_head(sk))
1972 break;
Ilpo Järvinen23aeeec2007-11-13 21:03:13 -08001973
1974 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
Ilpo Järvinend1a54c62007-02-21 23:11:57 -08001975 /*
1976 * Count the retransmission made on RTO correctly (only when
1977 * waiting for the first ACK and did not get it)...
1978 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001979 if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) {
Ilpo Järvinen0a9f2a42007-07-15 00:19:29 -07001980 /* For some reason this R-bit might get cleared? */
1981 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1982 tp->retrans_out += tcp_skb_pcount(skb);
Ilpo Järvinend1a54c62007-02-21 23:11:57 -08001983 /* ...enter this if branch just for the first segment */
1984 flag |= FLAG_DATA_ACKED;
1985 } else {
Ilpo Järvinen009a2e32007-09-20 11:34:38 -07001986 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1987 tp->undo_marker = 0;
Ilpo Järvinen23aeeec2007-11-13 21:03:13 -08001988 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
Ilpo Järvinend1a54c62007-02-21 23:11:57 -08001989 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001990
Ilpo Järvinen79d445162008-05-13 02:54:19 -07001991 /* Marking forward transmissions that were made after RTO lost
1992 * can cause unnecessary retransmissions in some scenarios,
1993 * SACK blocks will mitigate that in some but not in all cases.
1994 * We used to not mark them but it was causing break-ups with
1995 * receivers that do only in-order receival.
1996 *
1997 * TODO: we could detect presence of such receiver and select
1998 * different behavior per flow.
1999 */
2000 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
Ilpo Järvinen9bff40f2007-05-27 01:53:49 -07002001 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2002 tp->lost_out += tcp_skb_pcount(skb);
Ilpo Järvinen006f5822008-09-20 21:20:20 -07002003 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002004 }
2005 }
Ilpo Järvinen005903b2007-08-09 14:44:16 +03002006 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002007
Ilpo Järvinen95c49222007-02-21 23:05:18 -08002008 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002009 tp->snd_cwnd_cnt = 0;
2010 tp->snd_cwnd_stamp = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002011 tp->frto_counter = 0;
Ilpo Järvinen16e90682007-10-11 17:32:31 -07002012 tp->bytes_acked = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002013
2014 tp->reordering = min_t(unsigned int, tp->reordering,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002015 sysctl_tcp_reordering);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002016 tcp_set_ca_state(sk, TCP_CA_Loss);
Ilpo Järvinen79d445162008-05-13 02:54:19 -07002017 tp->high_seq = tp->snd_nxt;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002018 TCP_ECN_queue_cwr(tp);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002019
Ilpo Järvinen64edc272008-09-20 21:18:32 -07002020 tcp_clear_all_retrans_hints(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002021}
2022
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07002023static void tcp_clear_retrans_partial(struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002024{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002025 tp->retrans_out = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002026 tp->lost_out = 0;
2027
2028 tp->undo_marker = 0;
2029 tp->undo_retrans = 0;
2030}
2031
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07002032void tcp_clear_retrans(struct tcp_sock *tp)
2033{
2034 tcp_clear_retrans_partial(tp);
2035
2036 tp->fackets_out = 0;
2037 tp->sacked_out = 0;
2038}
2039
Linus Torvalds1da177e2005-04-16 15:20:36 -07002040/* Enter Loss state. If "how" is not zero, forget all SACK information
2041 * and reset tags completely, otherwise preserve SACKs. If receiver
2042 * dropped its ofo queue, we will know this due to reneging detection.
2043 */
2044void tcp_enter_loss(struct sock *sk, int how)
2045{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002046 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002047 struct tcp_sock *tp = tcp_sk(sk);
2048 struct sk_buff *skb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002049
2050 /* Reduce ssthresh if it has not yet been made inside this window. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002051 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
2052 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2053 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2054 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2055 tcp_ca_event(sk, CA_EVENT_LOSS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002056 }
2057 tp->snd_cwnd = 1;
2058 tp->snd_cwnd_cnt = 0;
2059 tp->snd_cwnd_stamp = tcp_time_stamp;
2060
Stephen Hemminger9772efb2005-11-10 17:09:53 -08002061 tp->bytes_acked = 0;
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07002062 tcp_clear_retrans_partial(tp);
2063
2064 if (tcp_is_reno(tp))
2065 tcp_reset_reno_sack(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002066
Ilpo Järvinenb7689202007-09-20 11:37:19 -07002067 if (!how) {
2068 /* Push undo marker, if it was plain RTO and nothing
2069 * was retransmitted. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002070 tp->undo_marker = tp->snd_una;
Ilpo Järvinenb7689202007-09-20 11:37:19 -07002071 } else {
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07002072 tp->sacked_out = 0;
2073 tp->fackets_out = 0;
Ilpo Järvinenb7689202007-09-20 11:37:19 -07002074 }
Ilpo Järvinen64edc272008-09-20 21:18:32 -07002075 tcp_clear_all_retrans_hints(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002076
David S. Millerfe067e82007-03-07 12:12:44 -08002077 tcp_for_write_queue(skb, sk) {
2078 if (skb == tcp_send_head(sk))
2079 break;
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07002080
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002081 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002082 tp->undo_marker = 0;
2083 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
2084 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
2085 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2086 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2087 tp->lost_out += tcp_skb_pcount(skb);
Ilpo Järvinen006f5822008-09-20 21:20:20 -07002088 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002089 }
2090 }
Ilpo Järvinen005903b2007-08-09 14:44:16 +03002091 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002092
2093 tp->reordering = min_t(unsigned int, tp->reordering,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002094 sysctl_tcp_reordering);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002095 tcp_set_ca_state(sk, TCP_CA_Loss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002096 tp->high_seq = tp->snd_nxt;
2097 TCP_ECN_queue_cwr(tp);
Ryousei Takano564262c12007-10-25 23:03:52 -07002098 /* Abort F-RTO algorithm if one is in progress */
Ilpo Järvinen580e5722007-05-19 13:56:57 -07002099 tp->frto_counter = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002100}
2101
Ilpo Järvinencadbd032007-12-31 04:49:21 -08002102/* If ACK arrived pointing to a remembered SACK, it means that our
2103 * remembered SACKs do not reflect real state of receiver i.e.
2104 * receiver _host_ is heavily congested (or buggy).
2105 *
2106 * Do processing similar to RTO timeout.
2107 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002108static bool tcp_check_sack_reneging(struct sock *sk, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002109{
Ilpo Järvinencadbd032007-12-31 04:49:21 -08002110 if (flag & FLAG_SACK_RENEGING) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002111 struct inet_connection_sock *icsk = inet_csk(sk);
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07002112 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002113
2114 tcp_enter_loss(sk, 1);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002115 icsk->icsk_retransmits++;
David S. Millerfe067e82007-03-07 12:12:44 -08002116 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002117 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002118 icsk->icsk_rto, TCP_RTO_MAX);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002119 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002120 }
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002121 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002122}
2123
Eric Dumazetcf533ea2011-10-21 05:22:42 -04002124static inline int tcp_fackets_out(const struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002125{
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002126 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002127}
2128
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002129/* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2130 * counter when SACK is enabled (without SACK, sacked_out is used for
2131 * that purpose).
2132 *
2133 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2134 * segments up to the highest received SACK block so far and holes in
2135 * between them.
2136 *
2137 * With reordering, holes may still be in flight, so RFC3517 recovery
2138 * uses pure sacked_out (total number of SACKed segments) even though
2139 * it violates the RFC that uses duplicate ACKs, often these are equal
2140 * but when e.g. out-of-window ACKs or packet duplication occurs,
2141 * they differ. Since neither occurs due to loss, TCP should really
2142 * ignore them.
2143 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04002144static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002145{
2146 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2147}
2148
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00002149static bool tcp_pause_early_retransmit(struct sock *sk, int flag)
2150{
2151 struct tcp_sock *tp = tcp_sk(sk);
2152 unsigned long delay;
2153
2154 /* Delay early retransmit and entering fast recovery for
2155 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
2156 * available, or RTO is scheduled to fire first.
2157 */
2158 if (sysctl_tcp_early_retrans < 2 || (flag & FLAG_ECE) || !tp->srtt)
2159 return false;
2160
2161 delay = max_t(unsigned long, (tp->srtt >> 5), msecs_to_jiffies(2));
2162 if (!time_after(inet_csk(sk)->icsk_timeout, (jiffies + delay)))
2163 return false;
2164
2165 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, delay, TCP_RTO_MAX);
2166 tp->early_retrans_delayed = 1;
2167 return true;
2168}
2169
Eric Dumazetcf533ea2011-10-21 05:22:42 -04002170static inline int tcp_skb_timedout(const struct sock *sk,
2171 const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002172{
Eric Dumazeta02cec22010-09-22 20:43:57 +00002173 return tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002174}
2175
Eric Dumazetcf533ea2011-10-21 05:22:42 -04002176static inline int tcp_head_timedout(const struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002177{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04002178 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002179
Linus Torvalds1da177e2005-04-16 15:20:36 -07002180 return tp->packets_out &&
David S. Millerfe067e82007-03-07 12:12:44 -08002181 tcp_skb_timedout(sk, tcp_write_queue_head(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002182}
2183
2184/* Linux NewReno/SACK/FACK/ECN state machine.
2185 * --------------------------------------
2186 *
2187 * "Open" Normal state, no dubious events, fast path.
2188 * "Disorder" In all the respects it is "Open",
2189 * but requires a bit more attention. It is entered when
2190 * we see some SACKs or dupacks. It is split of "Open"
2191 * mainly to move some processing from fast path to slow one.
2192 * "CWR" CWND was reduced due to some Congestion Notification event.
2193 * It can be ECN, ICMP source quench, local device congestion.
2194 * "Recovery" CWND was reduced, we are fast-retransmitting.
2195 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2196 *
2197 * tcp_fastretrans_alert() is entered:
2198 * - each incoming ACK, if state is not "Open"
2199 * - when arrived ACK is unusual, namely:
2200 * * SACK
2201 * * Duplicate ACK.
2202 * * ECN ECE.
2203 *
2204 * Counting packets in flight is pretty simple.
2205 *
2206 * in_flight = packets_out - left_out + retrans_out
2207 *
2208 * packets_out is SND.NXT-SND.UNA counted in packets.
2209 *
2210 * retrans_out is number of retransmitted segments.
2211 *
2212 * left_out is number of segments left network, but not ACKed yet.
2213 *
2214 * left_out = sacked_out + lost_out
2215 *
2216 * sacked_out: Packets, which arrived to receiver out of order
2217 * and hence not ACKed. With SACKs this number is simply
2218 * amount of SACKed data. Even without SACKs
2219 * it is easy to give pretty reliable estimate of this number,
2220 * counting duplicate ACKs.
2221 *
2222 * lost_out: Packets lost by network. TCP has no explicit
2223 * "loss notification" feedback from network (for now).
2224 * It means that this number can be only _guessed_.
2225 * Actually, it is the heuristics to predict lossage that
2226 * distinguishes different algorithms.
2227 *
2228 * F.e. after RTO, when all the queue is considered as lost,
2229 * lost_out = packets_out and in_flight = retrans_out.
2230 *
2231 * Essentially, we have now two algorithms counting
2232 * lost packets.
2233 *
2234 * FACK: It is the simplest heuristics. As soon as we decided
2235 * that something is lost, we decide that _all_ not SACKed
2236 * packets until the most forward SACK are lost. I.e.
2237 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2238 * It is absolutely correct estimate, if network does not reorder
2239 * packets. And it loses any connection to reality when reordering
2240 * takes place. We use FACK by default until reordering
2241 * is suspected on the path to this destination.
2242 *
2243 * NewReno: when Recovery is entered, we assume that one segment
2244 * is lost (classic Reno). While we are in Recovery and
2245 * a partial ACK arrives, we assume that one more packet
2246 * is lost (NewReno). This heuristics are the same in NewReno
2247 * and SACK.
2248 *
2249 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2250 * deflation etc. CWND is real congestion window, never inflated, changes
2251 * only according to classic VJ rules.
2252 *
2253 * Really tricky (and requiring careful tuning) part of algorithm
2254 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2255 * The first determines the moment _when_ we should reduce CWND and,
2256 * hence, slow down forward transmission. In fact, it determines the moment
2257 * when we decide that hole is caused by loss, rather than by a reorder.
2258 *
2259 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2260 * holes, caused by lost packets.
2261 *
2262 * And the most logically complicated part of algorithm is undo
2263 * heuristics. We detect false retransmits due to both too early
2264 * fast retransmit (reordering) and underestimated RTO, analyzing
2265 * timestamps and D-SACKs. When we detect that some segments were
2266 * retransmitted by mistake and CWND reduction was wrong, we undo
2267 * window reduction and abort recovery phase. This logic is hidden
2268 * inside several functions named tcp_try_undo_<something>.
2269 */
2270
2271/* This function decides, when we should leave Disordered state
2272 * and enter Recovery phase, reducing congestion window.
2273 *
2274 * Main question: may we further continue forward transmission
2275 * with the same cwnd?
2276 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002277static bool tcp_time_to_recover(struct sock *sk, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002278{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002279 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002280 __u32 packets_out;
2281
Ryousei Takano564262c12007-10-25 23:03:52 -07002282 /* Do not perform any recovery during F-RTO algorithm */
Ilpo Järvinen52c63f12007-02-21 23:06:52 -08002283 if (tp->frto_counter)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002284 return false;
Ilpo Järvinen52c63f12007-02-21 23:06:52 -08002285
Linus Torvalds1da177e2005-04-16 15:20:36 -07002286 /* Trick#1: The loss is proven. */
2287 if (tp->lost_out)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002288 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002289
2290 /* Not-A-Trick#2 : Classic rule... */
Andreas Petlundea84e552009-10-27 03:27:21 +00002291 if (tcp_dupack_heuristics(tp) > tp->reordering)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002292 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002293
2294 /* Trick#3 : when we use RFC2988 timer restart, fast
2295 * retransmit can be triggered by timeout of queue head.
2296 */
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002297 if (tcp_is_fack(tp) && tcp_head_timedout(sk))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002298 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002299
2300 /* Trick#4: It is still not OK... But will it be useful to delay
2301 * recovery more?
2302 */
2303 packets_out = tp->packets_out;
2304 if (packets_out <= tp->reordering &&
2305 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002306 !tcp_may_send_now(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002307 /* We have nothing to send. This connection is limited
2308 * either by receiver window or by application.
2309 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002310 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002311 }
2312
Andreas Petlund7e380172010-02-18 04:48:19 +00002313 /* If a thin stream is detected, retransmit after first
2314 * received dupack. Employ only if SACK is supported in order
2315 * to avoid possible corner-case series of spurious retransmissions
2316 * Use only if there are no unsent data.
2317 */
2318 if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2319 tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2320 tcp_is_sack(tp) && !tcp_send_head(sk))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002321 return true;
Andreas Petlund7e380172010-02-18 04:48:19 +00002322
Yuchung Chengeed530b2012-05-02 13:30:03 +00002323 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious
2324 * retransmissions due to small network reorderings, we implement
2325 * Mitigation A.3 in the RFC and delay the retransmission for a short
2326 * interval if appropriate.
2327 */
2328 if (tp->do_early_retrans && !tp->retrans_out && tp->sacked_out &&
2329 (tp->packets_out == (tp->sacked_out + 1) && tp->packets_out < 4) &&
2330 !tcp_may_send_now(sk))
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00002331 return !tcp_pause_early_retransmit(sk, flag);
Yuchung Chengeed530b2012-05-02 13:30:03 +00002332
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002333 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002334}
2335
Ilpo Järvinen7363a5b2009-02-28 04:44:33 +00002336/* New heuristics: it is possible only after we switched to restart timer
2337 * each time when something is ACKed. Hence, we can detect timed out packets
2338 * during fast retransmit without falling to slow start.
2339 *
2340 * Usefulness of this as is very questionable, since we should know which of
2341 * the segments is the next to timeout which is relatively expensive to find
2342 * in general case unless we add some data structure just for that. The
2343 * current approach certainly won't find the right one too often and when it
2344 * finally does find _something_ it usually marks large part of the window
2345 * right away (because a retransmission with a larger timestamp blocks the
2346 * loop from advancing). -ij
2347 */
2348static void tcp_timeout_skbs(struct sock *sk)
2349{
2350 struct tcp_sock *tp = tcp_sk(sk);
2351 struct sk_buff *skb;
2352
2353 if (!tcp_is_fack(tp) || !tcp_head_timedout(sk))
2354 return;
2355
2356 skb = tp->scoreboard_skb_hint;
2357 if (tp->scoreboard_skb_hint == NULL)
2358 skb = tcp_write_queue_head(sk);
2359
2360 tcp_for_write_queue_from(skb, sk) {
2361 if (skb == tcp_send_head(sk))
2362 break;
2363 if (!tcp_skb_timedout(sk, skb))
2364 break;
2365
2366 tcp_skb_mark_lost(tp, skb);
2367 }
2368
2369 tp->scoreboard_skb_hint = skb;
2370
2371 tcp_verify_left_out(tp);
2372}
2373
Yuchung Cheng974c1232012-01-19 14:42:21 +00002374/* Detect loss in event "A" above by marking head of queue up as lost.
2375 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2376 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2377 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2378 * the maximum SACKed segments to pass before reaching this limit.
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002379 */
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002380static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002381{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002382 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002383 struct sk_buff *skb;
Ilpo Järvinenc137f3d2008-04-07 22:32:38 -07002384 int cnt, oldcnt;
2385 int err;
2386 unsigned int mss;
Yuchung Cheng974c1232012-01-19 14:42:21 +00002387 /* Use SACK to deduce losses of new sequences sent during recovery */
2388 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002389
Ilpo Järvinen547b7922008-07-25 21:43:18 -07002390 WARN_ON(packets > tp->packets_out);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002391 if (tp->lost_skb_hint) {
2392 skb = tp->lost_skb_hint;
2393 cnt = tp->lost_cnt_hint;
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002394 /* Head already handled? */
2395 if (mark_head && skb != tcp_write_queue_head(sk))
2396 return;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002397 } else {
David S. Millerfe067e82007-03-07 12:12:44 -08002398 skb = tcp_write_queue_head(sk);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002399 cnt = 0;
2400 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002401
David S. Millerfe067e82007-03-07 12:12:44 -08002402 tcp_for_write_queue_from(skb, sk) {
2403 if (skb == tcp_send_head(sk))
2404 break;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002405 /* TODO: do this better */
2406 /* this is not the most efficient way to do this... */
2407 tp->lost_skb_hint = skb;
2408 tp->lost_cnt_hint = cnt;
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002409
Yuchung Cheng974c1232012-01-19 14:42:21 +00002410 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
Ilpo Järvinenc137f3d2008-04-07 22:32:38 -07002411 break;
2412
2413 oldcnt = cnt;
Ilpo Järvinenad1984e2008-01-30 20:06:02 -08002414 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002415 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2416 cnt += tcp_skb_pcount(skb);
2417
Ilpo Järvinenc137f3d2008-04-07 22:32:38 -07002418 if (cnt > packets) {
Yuchung Chengb3de7552010-09-24 13:22:06 +00002419 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
Neal Cardwellc0638c22012-03-02 21:36:51 +00002420 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
Yuchung Chengb3de7552010-09-24 13:22:06 +00002421 (oldcnt >= packets))
Ilpo Järvinenc137f3d2008-04-07 22:32:38 -07002422 break;
2423
2424 mss = skb_shinfo(skb)->gso_size;
2425 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2426 if (err < 0)
2427 break;
2428 cnt = packets;
2429 }
2430
Ilpo Järvinen41ea36e2008-09-20 21:19:22 -07002431 tcp_skb_mark_lost(tp, skb);
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002432
2433 if (mark_head)
2434 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002435 }
Ilpo Järvinen005903b2007-08-09 14:44:16 +03002436 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002437}
2438
2439/* Account newly detected lost packet(s) */
2440
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002441static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002442{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002443 struct tcp_sock *tp = tcp_sk(sk);
2444
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002445 if (tcp_is_reno(tp)) {
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002446 tcp_mark_head_lost(sk, 1, 1);
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002447 } else if (tcp_is_fack(tp)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002448 int lost = tp->fackets_out - tp->reordering;
2449 if (lost <= 0)
2450 lost = 1;
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002451 tcp_mark_head_lost(sk, lost, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002452 } else {
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002453 int sacked_upto = tp->sacked_out - tp->reordering;
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002454 if (sacked_upto >= 0)
2455 tcp_mark_head_lost(sk, sacked_upto, 0);
2456 else if (fast_rexmit)
2457 tcp_mark_head_lost(sk, 1, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002458 }
2459
Ilpo Järvinen7363a5b2009-02-28 04:44:33 +00002460 tcp_timeout_skbs(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002461}
2462
2463/* CWND moderation, preventing bursts due to too big ACKs
2464 * in dubious situations.
2465 */
2466static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2467{
2468 tp->snd_cwnd = min(tp->snd_cwnd,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002469 tcp_packets_in_flight(tp) + tcp_max_burst(tp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002470 tp->snd_cwnd_stamp = tcp_time_stamp;
2471}
2472
Linus Torvalds1da177e2005-04-16 15:20:36 -07002473/* Nothing was retransmitted or returned timestamp is less
2474 * than timestamp of the first retransmission.
2475 */
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00002476static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002477{
2478 return !tp->retrans_stamp ||
2479 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
Arnd Hannemannd7ee1472008-04-21 14:46:22 -07002480 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002481}
2482
2483/* Undo procedures. */
2484
2485#if FASTRETRANS_DEBUG > 1
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002486static void DBGUNDO(struct sock *sk, const char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002487{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002488 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002489 struct inet_sock *inet = inet_sk(sk);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002490
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002491 if (sk->sk_family == AF_INET) {
Joe Perches91df42b2012-05-15 14:11:54 +00002492 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2493 msg,
2494 &inet->inet_daddr, ntohs(inet->inet_dport),
2495 tp->snd_cwnd, tcp_left_out(tp),
2496 tp->snd_ssthresh, tp->prior_ssthresh,
2497 tp->packets_out);
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002498 }
Eric Dumazetdfd56b82011-12-10 09:48:31 +00002499#if IS_ENABLED(CONFIG_IPV6)
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002500 else if (sk->sk_family == AF_INET6) {
2501 struct ipv6_pinfo *np = inet6_sk(sk);
Joe Perches91df42b2012-05-15 14:11:54 +00002502 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2503 msg,
2504 &np->daddr, ntohs(inet->inet_dport),
2505 tp->snd_cwnd, tcp_left_out(tp),
2506 tp->snd_ssthresh, tp->prior_ssthresh,
2507 tp->packets_out);
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002508 }
2509#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002510}
2511#else
2512#define DBGUNDO(x...) do { } while (0)
2513#endif
2514
David S. Millerf6152732011-03-22 19:37:11 -07002515static void tcp_undo_cwr(struct sock *sk, const bool undo_ssthresh)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002516{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002517 struct tcp_sock *tp = tcp_sk(sk);
2518
Linus Torvalds1da177e2005-04-16 15:20:36 -07002519 if (tp->prior_ssthresh) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002520 const struct inet_connection_sock *icsk = inet_csk(sk);
2521
2522 if (icsk->icsk_ca_ops->undo_cwnd)
2523 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002524 else
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002525 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002526
Yuchung Cheng67d41202011-03-14 10:57:03 +00002527 if (undo_ssthresh && tp->prior_ssthresh > tp->snd_ssthresh) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002528 tp->snd_ssthresh = tp->prior_ssthresh;
2529 TCP_ECN_withdraw_cwr(tp);
2530 }
2531 } else {
2532 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2533 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002534 tp->snd_cwnd_stamp = tcp_time_stamp;
2535}
2536
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00002537static inline bool tcp_may_undo(const struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002538{
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002539 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002540}
2541
2542/* People celebrate: "We love our President!" */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002543static bool tcp_try_undo_recovery(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002544{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002545 struct tcp_sock *tp = tcp_sk(sk);
2546
Linus Torvalds1da177e2005-04-16 15:20:36 -07002547 if (tcp_may_undo(tp)) {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07002548 int mib_idx;
2549
Linus Torvalds1da177e2005-04-16 15:20:36 -07002550 /* Happy end! We did not retransmit anything
2551 * or our original transmission succeeded.
2552 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002553 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
David S. Millerf6152732011-03-22 19:37:11 -07002554 tcp_undo_cwr(sk, true);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002555 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07002556 mib_idx = LINUX_MIB_TCPLOSSUNDO;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002557 else
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07002558 mib_idx = LINUX_MIB_TCPFULLUNDO;
2559
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07002560 NET_INC_STATS_BH(sock_net(sk), mib_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002561 tp->undo_marker = 0;
2562 }
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002563 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002564 /* Hold old state until something *above* high_seq
2565 * is ACKed. For Reno it is MUST to prevent false
2566 * fast retransmits (RFC2582). SACK TCP is safe. */
2567 tcp_moderate_cwnd(tp);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002568 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002569 }
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002570 tcp_set_ca_state(sk, TCP_CA_Open);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002571 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002572}
2573
2574/* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002575static void tcp_try_undo_dsack(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002576{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002577 struct tcp_sock *tp = tcp_sk(sk);
2578
Linus Torvalds1da177e2005-04-16 15:20:36 -07002579 if (tp->undo_marker && !tp->undo_retrans) {
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002580 DBGUNDO(sk, "D-SACK");
David S. Millerf6152732011-03-22 19:37:11 -07002581 tcp_undo_cwr(sk, true);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002582 tp->undo_marker = 0;
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07002583 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002584 }
2585}
2586
Ilpo Järvinen77722b12009-12-08 20:54:11 -08002587/* We can clear retrans_stamp when there are no retransmissions in the
2588 * window. It would seem that it is trivially available for us in
2589 * tp->retrans_out, however, that kind of assumptions doesn't consider
2590 * what will happen if errors occur when sending retransmission for the
2591 * second time. ...It could the that such segment has only
2592 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2593 * the head skb is enough except for some reneging corner cases that
2594 * are not worth the effort.
2595 *
2596 * Main reason for all this complexity is the fact that connection dying
2597 * time now depends on the validity of the retrans_stamp, in particular,
2598 * that successive retransmissions of a segment must not advance
2599 * retrans_stamp under any conditions.
2600 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002601static bool tcp_any_retrans_done(const struct sock *sk)
Ilpo Järvinen77722b12009-12-08 20:54:11 -08002602{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04002603 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen77722b12009-12-08 20:54:11 -08002604 struct sk_buff *skb;
2605
2606 if (tp->retrans_out)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002607 return true;
Ilpo Järvinen77722b12009-12-08 20:54:11 -08002608
2609 skb = tcp_write_queue_head(sk);
2610 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002611 return true;
Ilpo Järvinen77722b12009-12-08 20:54:11 -08002612
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002613 return false;
Ilpo Järvinen77722b12009-12-08 20:54:11 -08002614}
2615
Linus Torvalds1da177e2005-04-16 15:20:36 -07002616/* Undo during fast recovery after partial ACK. */
2617
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002618static int tcp_try_undo_partial(struct sock *sk, int acked)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002619{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002620 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002621 /* Partial ACK arrived. Force Hoe's retransmit. */
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002622 int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002623
2624 if (tcp_may_undo(tp)) {
2625 /* Plain luck! Hole if filled with delayed
2626 * packet, rather than with a retransmit.
2627 */
Ilpo Järvinen77722b12009-12-08 20:54:11 -08002628 if (!tcp_any_retrans_done(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002629 tp->retrans_stamp = 0;
2630
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002631 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002632
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002633 DBGUNDO(sk, "Hoe");
David S. Millerf6152732011-03-22 19:37:11 -07002634 tcp_undo_cwr(sk, false);
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07002635 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002636
2637 /* So... Do not make Hoe's retransmit yet.
2638 * If the first packet was delayed, the rest
2639 * ones are most probably delayed as well.
2640 */
2641 failed = 0;
2642 }
2643 return failed;
2644}
2645
2646/* Undo during loss recovery after partial ACK. */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002647static bool tcp_try_undo_loss(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002648{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002649 struct tcp_sock *tp = tcp_sk(sk);
2650
Linus Torvalds1da177e2005-04-16 15:20:36 -07002651 if (tcp_may_undo(tp)) {
2652 struct sk_buff *skb;
David S. Millerfe067e82007-03-07 12:12:44 -08002653 tcp_for_write_queue(skb, sk) {
2654 if (skb == tcp_send_head(sk))
2655 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002656 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2657 }
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002658
Ilpo Järvinen5af4ec22007-09-20 11:30:48 -07002659 tcp_clear_all_retrans_hints(tp);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002660
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002661 DBGUNDO(sk, "partial loss");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002662 tp->lost_out = 0;
David S. Millerf6152732011-03-22 19:37:11 -07002663 tcp_undo_cwr(sk, true);
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07002664 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002665 inet_csk(sk)->icsk_retransmits = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002666 tp->undo_marker = 0;
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002667 if (tcp_is_sack(tp))
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002668 tcp_set_ca_state(sk, TCP_CA_Open);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002669 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002670 }
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002671 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002672}
2673
Yuchung Cheng684bad12012-09-02 17:38:04 +00002674/* The cwnd reduction in CWR and Recovery use the PRR algorithm
2675 * https://datatracker.ietf.org/doc/draft-ietf-tcpm-proportional-rate-reduction/
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002676 * It computes the number of packets to send (sndcnt) based on packets newly
2677 * delivered:
2678 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2679 * cwnd reductions across a full RTT.
2680 * 2) If packets in flight is lower than ssthresh (such as due to excess
2681 * losses and/or application stalls), do not perform any further cwnd
2682 * reductions, but instead slow start up to ssthresh.
2683 */
Yuchung Cheng684bad12012-09-02 17:38:04 +00002684static void tcp_init_cwnd_reduction(struct sock *sk, const bool set_ssthresh)
2685{
2686 struct tcp_sock *tp = tcp_sk(sk);
2687
2688 tp->high_seq = tp->snd_nxt;
2689 tp->bytes_acked = 0;
2690 tp->snd_cwnd_cnt = 0;
2691 tp->prior_cwnd = tp->snd_cwnd;
2692 tp->prr_delivered = 0;
2693 tp->prr_out = 0;
2694 if (set_ssthresh)
2695 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2696 TCP_ECN_queue_cwr(tp);
2697}
2698
2699static void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked,
2700 int fast_rexmit)
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002701{
2702 struct tcp_sock *tp = tcp_sk(sk);
2703 int sndcnt = 0;
2704 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2705
Yuchung Cheng684bad12012-09-02 17:38:04 +00002706 tp->prr_delivered += newly_acked_sacked;
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002707 if (tcp_packets_in_flight(tp) > tp->snd_ssthresh) {
2708 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2709 tp->prior_cwnd - 1;
2710 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2711 } else {
2712 sndcnt = min_t(int, delta,
2713 max_t(int, tp->prr_delivered - tp->prr_out,
2714 newly_acked_sacked) + 1);
2715 }
2716
2717 sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0));
2718 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2719}
2720
Yuchung Cheng684bad12012-09-02 17:38:04 +00002721static inline void tcp_end_cwnd_reduction(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002722{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002723 struct tcp_sock *tp = tcp_sk(sk);
Nandita Dukkipatia262f0c2011-08-21 20:21:57 +00002724
Yuchung Cheng684bad12012-09-02 17:38:04 +00002725 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2726 if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR ||
2727 (tp->undo_marker && tp->snd_ssthresh < TCP_INFINITE_SSTHRESH)) {
2728 tp->snd_cwnd = tp->snd_ssthresh;
2729 tp->snd_cwnd_stamp = tcp_time_stamp;
Yuchung Cheng67d41202011-03-14 10:57:03 +00002730 }
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002731 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002732}
2733
Yuchung Cheng684bad12012-09-02 17:38:04 +00002734/* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
Yuchung Cheng09484d12012-09-02 17:38:02 +00002735void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
2736{
2737 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng09484d12012-09-02 17:38:02 +00002738
2739 tp->prior_ssthresh = 0;
2740 tp->bytes_acked = 0;
Yuchung Cheng684bad12012-09-02 17:38:04 +00002741 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
Yuchung Cheng09484d12012-09-02 17:38:02 +00002742 tp->undo_marker = 0;
Yuchung Cheng684bad12012-09-02 17:38:04 +00002743 tcp_init_cwnd_reduction(sk, set_ssthresh);
Yuchung Cheng09484d12012-09-02 17:38:02 +00002744 tcp_set_ca_state(sk, TCP_CA_CWR);
2745 }
2746}
2747
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07002748static void tcp_try_keep_open(struct sock *sk)
2749{
2750 struct tcp_sock *tp = tcp_sk(sk);
2751 int state = TCP_CA_Open;
2752
Neal Cardwellf6982042011-11-16 08:58:04 +00002753 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07002754 state = TCP_CA_Disorder;
2755
2756 if (inet_csk(sk)->icsk_ca_state != state) {
2757 tcp_set_ca_state(sk, state);
2758 tp->high_seq = tp->snd_nxt;
2759 }
2760}
2761
Yuchung Cheng684bad12012-09-02 17:38:04 +00002762static void tcp_try_to_open(struct sock *sk, int flag, int newly_acked_sacked)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002763{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002764 struct tcp_sock *tp = tcp_sk(sk);
2765
Ilpo Järvinen86426c22007-08-09 14:45:17 +03002766 tcp_verify_left_out(tp);
2767
Ilpo Järvinen77722b12009-12-08 20:54:11 -08002768 if (!tp->frto_counter && !tcp_any_retrans_done(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002769 tp->retrans_stamp = 0;
2770
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002771 if (flag & FLAG_ECE)
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -08002772 tcp_enter_cwr(sk, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002773
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002774 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07002775 tcp_try_keep_open(sk);
Neal Cardwell8cd6d612011-11-16 08:58:05 +00002776 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
2777 tcp_moderate_cwnd(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002778 } else {
Yuchung Cheng684bad12012-09-02 17:38:04 +00002779 tcp_cwnd_reduction(sk, newly_acked_sacked, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002780 }
2781}
2782
John Heffner5d424d52006-03-20 17:53:41 -08002783static void tcp_mtup_probe_failed(struct sock *sk)
2784{
2785 struct inet_connection_sock *icsk = inet_csk(sk);
2786
2787 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2788 icsk->icsk_mtup.probe_size = 0;
2789}
2790
Ilpo Järvinen72211e92009-03-14 14:23:04 +00002791static void tcp_mtup_probe_success(struct sock *sk)
John Heffner5d424d52006-03-20 17:53:41 -08002792{
2793 struct tcp_sock *tp = tcp_sk(sk);
2794 struct inet_connection_sock *icsk = inet_csk(sk);
2795
2796 /* FIXME: breaks with very large cwnd */
2797 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2798 tp->snd_cwnd = tp->snd_cwnd *
2799 tcp_mss_to_mtu(sk, tp->mss_cache) /
2800 icsk->icsk_mtup.probe_size;
2801 tp->snd_cwnd_cnt = 0;
2802 tp->snd_cwnd_stamp = tcp_time_stamp;
John Heffner9c6d5e52010-10-06 21:18:02 -07002803 tp->snd_ssthresh = tcp_current_ssthresh(sk);
John Heffner5d424d52006-03-20 17:53:41 -08002804
2805 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2806 icsk->icsk_mtup.probe_size = 0;
2807 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2808}
2809
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08002810/* Do a simple retransmit without using the backoff mechanisms in
2811 * tcp_timer. This is used for path mtu discovery.
2812 * The socket is already locked here.
2813 */
2814void tcp_simple_retransmit(struct sock *sk)
2815{
2816 const struct inet_connection_sock *icsk = inet_csk(sk);
2817 struct tcp_sock *tp = tcp_sk(sk);
2818 struct sk_buff *skb;
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00002819 unsigned int mss = tcp_current_mss(sk);
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08002820 u32 prior_lost = tp->lost_out;
2821
2822 tcp_for_write_queue(skb, sk) {
2823 if (skb == tcp_send_head(sk))
2824 break;
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08002825 if (tcp_skb_seglen(skb) > mss &&
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08002826 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2827 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2828 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2829 tp->retrans_out -= tcp_skb_pcount(skb);
2830 }
2831 tcp_skb_mark_lost_uncond_verify(tp, skb);
2832 }
2833 }
2834
2835 tcp_clear_retrans_hints_partial(tp);
2836
2837 if (prior_lost == tp->lost_out)
2838 return;
2839
2840 if (tcp_is_reno(tp))
2841 tcp_limit_reno_sacked(tp);
2842
2843 tcp_verify_left_out(tp);
2844
2845 /* Don't muck with the congestion window here.
2846 * Reason is that we do not increase amount of _data_
2847 * in network, but units changed and effective
2848 * cwnd/ssthresh really reduced now.
2849 */
2850 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2851 tp->high_seq = tp->snd_nxt;
2852 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2853 tp->prior_ssthresh = 0;
2854 tp->undo_marker = 0;
2855 tcp_set_ca_state(sk, TCP_CA_Loss);
2856 }
2857 tcp_xmit_retransmit_queue(sk);
2858}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00002859EXPORT_SYMBOL(tcp_simple_retransmit);
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08002860
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002861static void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2862{
2863 struct tcp_sock *tp = tcp_sk(sk);
2864 int mib_idx;
2865
2866 if (tcp_is_reno(tp))
2867 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2868 else
2869 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2870
2871 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2872
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002873 tp->prior_ssthresh = 0;
2874 tp->undo_marker = tp->snd_una;
2875 tp->undo_retrans = tp->retrans_out;
2876
2877 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2878 if (!ece_ack)
2879 tp->prior_ssthresh = tcp_current_ssthresh(sk);
Yuchung Cheng684bad12012-09-02 17:38:04 +00002880 tcp_init_cwnd_reduction(sk, true);
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002881 }
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002882 tcp_set_ca_state(sk, TCP_CA_Recovery);
2883}
2884
Linus Torvalds1da177e2005-04-16 15:20:36 -07002885/* Process an event, which can update packets-in-flight not trivially.
2886 * Main goal of this function is to calculate new estimate for left_out,
2887 * taking into account both packets sitting in receiver's buffer and
2888 * packets lost by network.
2889 *
2890 * Besides that it does CWND reduction, when packet loss is detected
2891 * and changes state of machine.
2892 *
2893 * It does _not_ decide what to send, it is made in function
2894 * tcp_xmit_retransmit_queue().
2895 */
Nandita Dukkipatia262f0c2011-08-21 20:21:57 +00002896static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked,
Yuchung Cheng7c4a56f2012-08-23 07:05:17 +00002897 int prior_sacked, bool is_dupack,
Neal Cardwell7d2b55f2011-11-16 08:58:01 +00002898 int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002899{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002900 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002901 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002902 int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002903 (tcp_fackets_out(tp) > tp->reordering));
Yuchung Cheng7c4a56f2012-08-23 07:05:17 +00002904 int newly_acked_sacked = 0;
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002905 int fast_rexmit = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002906
Ilpo Järvinen3ccd3132007-12-31 04:43:32 -08002907 if (WARN_ON(!tp->packets_out && tp->sacked_out))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002908 tp->sacked_out = 0;
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07002909 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002910 tp->fackets_out = 0;
2911
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002912 /* Now state machine starts.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002913 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002914 if (flag & FLAG_ECE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002915 tp->prior_ssthresh = 0;
2916
2917 /* B. In all the states check for reneging SACKs. */
Ilpo Järvinencadbd032007-12-31 04:49:21 -08002918 if (tcp_check_sack_reneging(sk, flag))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002919 return;
2920
Yuchung Cheng974c1232012-01-19 14:42:21 +00002921 /* C. Check consistency of the current state. */
Ilpo Järvinen005903b2007-08-09 14:44:16 +03002922 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002923
Yuchung Cheng974c1232012-01-19 14:42:21 +00002924 /* D. Check state exit conditions. State can be terminated
Linus Torvalds1da177e2005-04-16 15:20:36 -07002925 * when high_seq is ACKed. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002926 if (icsk->icsk_ca_state == TCP_CA_Open) {
Ilpo Järvinen547b7922008-07-25 21:43:18 -07002927 WARN_ON(tp->retrans_out != 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002928 tp->retrans_stamp = 0;
2929 } else if (!before(tp->snd_una, tp->high_seq)) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002930 switch (icsk->icsk_ca_state) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002931 case TCP_CA_Loss:
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002932 icsk->icsk_retransmits = 0;
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002933 if (tcp_try_undo_recovery(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002934 return;
2935 break;
2936
2937 case TCP_CA_CWR:
2938 /* CWR is to be held something *above* high_seq
2939 * is ACKed for CWR bit to reach receiver. */
2940 if (tp->snd_una != tp->high_seq) {
Yuchung Cheng684bad12012-09-02 17:38:04 +00002941 tcp_end_cwnd_reduction(sk);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002942 tcp_set_ca_state(sk, TCP_CA_Open);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002943 }
2944 break;
2945
Linus Torvalds1da177e2005-04-16 15:20:36 -07002946 case TCP_CA_Recovery:
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002947 if (tcp_is_reno(tp))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002948 tcp_reset_reno_sack(tp);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002949 if (tcp_try_undo_recovery(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002950 return;
Yuchung Cheng684bad12012-09-02 17:38:04 +00002951 tcp_end_cwnd_reduction(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002952 break;
2953 }
2954 }
2955
Yuchung Cheng974c1232012-01-19 14:42:21 +00002956 /* E. Process state. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002957 switch (icsk->icsk_ca_state) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002958 case TCP_CA_Recovery:
Ilpo Järvinen2e605292007-08-02 19:46:58 -07002959 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002960 if (tcp_is_reno(tp) && is_dupack)
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002961 tcp_add_reno_sack(sk);
Ilpo Järvinen1b6d427b2007-08-09 14:53:36 +03002962 } else
2963 do_lost = tcp_try_undo_partial(sk, pkts_acked);
Yuchung Cheng7c4a56f2012-08-23 07:05:17 +00002964 newly_acked_sacked = pkts_acked + tp->sacked_out - prior_sacked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002965 break;
2966 case TCP_CA_Loss:
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002967 if (flag & FLAG_DATA_ACKED)
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002968 icsk->icsk_retransmits = 0;
Ilpo Järvinen882beba2008-04-07 22:33:07 -07002969 if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED)
2970 tcp_reset_reno_sack(tp);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002971 if (!tcp_try_undo_loss(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002972 tcp_moderate_cwnd(tp);
2973 tcp_xmit_retransmit_queue(sk);
2974 return;
2975 }
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002976 if (icsk->icsk_ca_state != TCP_CA_Open)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002977 return;
2978 /* Loss is undone; fall through to processing in Open state. */
2979 default:
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002980 if (tcp_is_reno(tp)) {
Ilpo Järvinen2e605292007-08-02 19:46:58 -07002981 if (flag & FLAG_SND_UNA_ADVANCED)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002982 tcp_reset_reno_sack(tp);
2983 if (is_dupack)
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002984 tcp_add_reno_sack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002985 }
Yuchung Cheng7c4a56f2012-08-23 07:05:17 +00002986 newly_acked_sacked = pkts_acked + tp->sacked_out - prior_sacked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002987
Neal Cardwellf6982042011-11-16 08:58:04 +00002988 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002989 tcp_try_undo_dsack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002990
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00002991 if (!tcp_time_to_recover(sk, flag)) {
Yuchung Cheng684bad12012-09-02 17:38:04 +00002992 tcp_try_to_open(sk, flag, newly_acked_sacked);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002993 return;
2994 }
2995
John Heffner5d424d52006-03-20 17:53:41 -08002996 /* MTU probe failure: don't reduce cwnd */
2997 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2998 icsk->icsk_mtup.probe_size &&
John Heffner0e7b1362006-03-20 21:32:58 -08002999 tp->snd_una == tp->mtu_probe.probe_seq_start) {
John Heffner5d424d52006-03-20 17:53:41 -08003000 tcp_mtup_probe_failed(sk);
3001 /* Restores the reduction we did in tcp_mtup_probe() */
3002 tp->snd_cwnd++;
3003 tcp_simple_retransmit(sk);
3004 return;
3005 }
3006
Linus Torvalds1da177e2005-04-16 15:20:36 -07003007 /* Otherwise enter Recovery state */
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00003008 tcp_enter_recovery(sk, (flag & FLAG_ECE));
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08003009 fast_rexmit = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003010 }
3011
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08003012 if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk)))
3013 tcp_update_scoreboard(sk, fast_rexmit);
Yuchung Cheng684bad12012-09-02 17:38:04 +00003014 tcp_cwnd_reduction(sk, newly_acked_sacked, fast_rexmit);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003015 tcp_xmit_retransmit_queue(sk);
3016}
3017
Jerry Chu9ad7c042011-06-08 11:08:38 +00003018void tcp_valid_rtt_meas(struct sock *sk, u32 seq_rtt)
Ilpo Järvinen41834b72008-12-05 22:43:26 -08003019{
3020 tcp_rtt_estimator(sk, seq_rtt);
3021 tcp_set_rto(sk);
3022 inet_csk(sk)->icsk_backoff = 0;
3023}
Jerry Chu9ad7c042011-06-08 11:08:38 +00003024EXPORT_SYMBOL(tcp_valid_rtt_meas);
Ilpo Järvinen41834b72008-12-05 22:43:26 -08003025
Linus Torvalds1da177e2005-04-16 15:20:36 -07003026/* Read draft-ietf-tcplw-high-performance before mucking
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08003027 * with this code. (Supersedes RFC1323)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003028 */
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08003029static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003030{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003031 /* RTTM Rule: A TSecr value received in a segment is used to
3032 * update the averaged RTT measurement only if the segment
3033 * acknowledges some new data, i.e., only if it advances the
3034 * left edge of the send window.
3035 *
3036 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3037 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
3038 *
3039 * Changed: reset backoff as soon as we see the first valid sample.
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08003040 * If we do not, we get strongly overestimated rto. With timestamps
Linus Torvalds1da177e2005-04-16 15:20:36 -07003041 * samples are accepted even from very old segments: f.e., when rtt=1
3042 * increases to 8, we retransmit 5 times and after 8 seconds delayed
3043 * answer arrives rto becomes 120 seconds! If at least one of segments
3044 * in window is lost... Voila. --ANK (010210)
3045 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003046 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen41834b72008-12-05 22:43:26 -08003047
3048 tcp_valid_rtt_meas(sk, tcp_time_stamp - tp->rx_opt.rcv_tsecr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003049}
3050
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08003051static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003052{
3053 /* We don't have a timestamp. Can only use
3054 * packets that are not retransmitted to determine
3055 * rtt estimates. Also, we must not reset the
3056 * backoff for rto until we get a non-retransmitted
3057 * packet. This allows us to deal with a situation
3058 * where the network delay has increased suddenly.
3059 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
3060 */
3061
3062 if (flag & FLAG_RETRANS_DATA_ACKED)
3063 return;
3064
Ilpo Järvinen41834b72008-12-05 22:43:26 -08003065 tcp_valid_rtt_meas(sk, seq_rtt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003066}
3067
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003068static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08003069 const s32 seq_rtt)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003070{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003071 const struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003072 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
3073 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08003074 tcp_ack_saw_tstamp(sk, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003075 else if (seq_rtt >= 0)
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08003076 tcp_ack_no_tstamp(sk, seq_rtt, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003077}
3078
Ilpo Järvinenc3a05c62007-12-02 00:47:59 +02003079static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003080{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003081 const struct inet_connection_sock *icsk = inet_csk(sk);
Ilpo Järvinenc3a05c62007-12-02 00:47:59 +02003082 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003083 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003084}
3085
Linus Torvalds1da177e2005-04-16 15:20:36 -07003086/* Restart timer after forward progress on connection.
3087 * RFC2988 recommends to restart timer to now+rto.
3088 */
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00003089void tcp_rearm_rto(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003090{
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00003091 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003092
Jerry Chu168a8f52012-08-31 12:29:13 +00003093 /* If the retrans timer is currently being used by Fast Open
3094 * for SYN-ACK retrans purpose, stay put.
3095 */
3096 if (tp->fastopen_rsk)
3097 return;
3098
Linus Torvalds1da177e2005-04-16 15:20:36 -07003099 if (!tp->packets_out) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003100 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003101 } else {
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00003102 u32 rto = inet_csk(sk)->icsk_rto;
3103 /* Offset the time elapsed after installing regular RTO */
3104 if (tp->early_retrans_delayed) {
3105 struct sk_buff *skb = tcp_write_queue_head(sk);
3106 const u32 rto_time_stamp = TCP_SKB_CB(skb)->when + rto;
3107 s32 delta = (s32)(rto_time_stamp - tcp_time_stamp);
3108 /* delta may not be positive if the socket is locked
3109 * when the delayed ER timer fires and is rescheduled.
3110 */
3111 if (delta > 0)
3112 rto = delta;
3113 }
3114 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3115 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003116 }
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00003117 tp->early_retrans_delayed = 0;
3118}
3119
3120/* This function is called when the delayed ER timer fires. TCP enters
3121 * fast recovery and performs fast-retransmit.
3122 */
3123void tcp_resume_early_retransmit(struct sock *sk)
3124{
3125 struct tcp_sock *tp = tcp_sk(sk);
3126
3127 tcp_rearm_rto(sk);
3128
3129 /* Stop if ER is disabled after the delayed ER timer is scheduled */
3130 if (!tp->do_early_retrans)
3131 return;
3132
3133 tcp_enter_recovery(sk, false);
3134 tcp_update_scoreboard(sk, 1);
3135 tcp_xmit_retransmit_queue(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003136}
3137
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003138/* If we get here, the whole TSO packet has not been acked. */
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003139static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003140{
3141 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003142 u32 packets_acked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003143
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003144 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003145
3146 packets_acked = tcp_skb_pcount(skb);
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003147 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003148 return 0;
3149 packets_acked -= tcp_skb_pcount(skb);
3150
3151 if (packets_acked) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003152 BUG_ON(tcp_skb_pcount(skb) == 0);
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003153 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003154 }
3155
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003156 return packets_acked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003157}
3158
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003159/* Remove acknowledged frames from the retransmission queue. If our packet
3160 * is before the ack sequence we can discard it as it's confirmed to have
3161 * arrived at the other end.
3162 */
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07003163static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3164 u32 prior_snd_una)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003165{
3166 struct tcp_sock *tp = tcp_sk(sk);
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08003167 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003168 struct sk_buff *skb;
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003169 u32 now = tcp_time_stamp;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003170 int fully_acked = true;
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003171 int flag = 0;
Ilpo Järvinen72018832007-12-30 04:37:55 -08003172 u32 pkts_acked = 0;
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003173 u32 reord = tp->packets_out;
Ilpo Järvinen90638a02008-09-20 21:25:52 -07003174 u32 prior_sacked = tp->sacked_out;
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003175 s32 seq_rtt = -1;
Gavin McCullagh2072c222007-12-29 19:11:21 -08003176 s32 ca_seq_rtt = -1;
Ilpo Järvinenb9ce2042007-06-15 15:08:43 -07003177 ktime_t last_ackt = net_invalid_timestamp();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003178
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003179 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09003180 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
Ilpo Järvinen72018832007-12-30 04:37:55 -08003181 u32 acked_pcount;
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003182 u8 sacked = scb->sacked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003183
Gavin McCullagh2072c222007-12-29 19:11:21 -08003184 /* Determine how many packets and what bytes were acked, tso and else */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003185 if (after(scb->end_seq, tp->snd_una)) {
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003186 if (tcp_skb_pcount(skb) == 1 ||
3187 !after(tp->snd_una, scb->seq))
3188 break;
3189
Ilpo Järvinen72018832007-12-30 04:37:55 -08003190 acked_pcount = tcp_tso_acked(sk, skb);
3191 if (!acked_pcount)
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003192 break;
3193
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003194 fully_acked = false;
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003195 } else {
Ilpo Järvinen72018832007-12-30 04:37:55 -08003196 acked_pcount = tcp_skb_pcount(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003197 }
3198
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003199 if (sacked & TCPCB_RETRANS) {
3200 if (sacked & TCPCB_SACKED_RETRANS)
Ilpo Järvinen72018832007-12-30 04:37:55 -08003201 tp->retrans_out -= acked_pcount;
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003202 flag |= FLAG_RETRANS_DATA_ACKED;
3203 ca_seq_rtt = -1;
3204 seq_rtt = -1;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003205 if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1))
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003206 flag |= FLAG_NONHEAD_RETRANS_ACKED;
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003207 } else {
Gavin McCullagh2072c222007-12-29 19:11:21 -08003208 ca_seq_rtt = now - scb->when;
3209 last_ackt = skb->tstamp;
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003210 if (seq_rtt < 0) {
Gavin McCullagh2072c222007-12-29 19:11:21 -08003211 seq_rtt = ca_seq_rtt;
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003212 }
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003213 if (!(sacked & TCPCB_SACKED_ACKED))
Ilpo Järvinen72018832007-12-30 04:37:55 -08003214 reord = min(pkts_acked, reord);
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08003215 }
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003216
3217 if (sacked & TCPCB_SACKED_ACKED)
Ilpo Järvinen72018832007-12-30 04:37:55 -08003218 tp->sacked_out -= acked_pcount;
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003219 if (sacked & TCPCB_LOST)
Ilpo Järvinen72018832007-12-30 04:37:55 -08003220 tp->lost_out -= acked_pcount;
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003221
Ilpo Järvinen72018832007-12-30 04:37:55 -08003222 tp->packets_out -= acked_pcount;
3223 pkts_acked += acked_pcount;
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003224
Ilpo Järvinen009a2e32007-09-20 11:34:38 -07003225 /* Initial outgoing SYN's get put onto the write_queue
3226 * just like anything else we transmit. It is not
3227 * true data, and if we misinform our callers that
3228 * this ACK acks real data, we will erroneously exit
3229 * connection startup slow start one packet too
3230 * quickly. This is severely frowned upon behavior.
3231 */
Eric Dumazet4de075e2011-09-27 13:25:05 -04003232 if (!(scb->tcp_flags & TCPHDR_SYN)) {
Ilpo Järvinen009a2e32007-09-20 11:34:38 -07003233 flag |= FLAG_DATA_ACKED;
3234 } else {
3235 flag |= FLAG_SYN_ACKED;
3236 tp->retrans_stamp = 0;
3237 }
3238
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003239 if (!fully_acked)
3240 break;
3241
David S. Millerfe067e82007-03-07 12:12:44 -08003242 tcp_unlink_write_queue(skb, sk);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08003243 sk_wmem_free_skb(sk, skb);
Ilpo Järvinen90638a02008-09-20 21:25:52 -07003244 tp->scoreboard_skb_hint = NULL;
Ilpo Järvinenef9da472008-09-20 21:25:15 -07003245 if (skb == tp->retransmit_skb_hint)
3246 tp->retransmit_skb_hint = NULL;
Ilpo Järvinen90638a02008-09-20 21:25:52 -07003247 if (skb == tp->lost_skb_hint)
3248 tp->lost_skb_hint = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003249 }
3250
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07003251 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3252 tp->snd_up = tp->snd_una;
3253
Ilpo Järvinencadbd032007-12-31 04:49:21 -08003254 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3255 flag |= FLAG_SACK_RENEGING;
3256
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003257 if (flag & FLAG_ACKED) {
Stephen Hemminger164891a2007-04-23 22:26:16 -07003258 const struct tcp_congestion_ops *ca_ops
3259 = inet_csk(sk)->icsk_ca_ops;
3260
Ilpo Järvinen72211e92009-03-14 14:23:04 +00003261 if (unlikely(icsk->icsk_mtup.probe_size &&
3262 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3263 tcp_mtup_probe_success(sk);
3264 }
3265
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003266 tcp_ack_update_rtt(sk, flag, seq_rtt);
Ilpo Järvinen6728e7d2007-08-24 22:53:26 -07003267 tcp_rearm_rto(sk);
Stephen Hemminger317a76f2005-06-23 12:19:55 -07003268
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003269 if (tcp_is_reno(tp)) {
3270 tcp_remove_reno_sacks(sk, pkts_acked);
3271 } else {
Ilpo Järvinen59a08cb2009-02-28 04:44:28 +00003272 int delta;
3273
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003274 /* Non-retransmitted hole got filled? That's reordering */
3275 if (reord < prior_fackets)
3276 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
Ilpo Järvinen90638a02008-09-20 21:25:52 -07003277
Ilpo Järvinen59a08cb2009-02-28 04:44:28 +00003278 delta = tcp_is_fack(tp) ? pkts_acked :
3279 prior_sacked - tp->sacked_out;
3280 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003281 }
3282
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07003283 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08003284
Stephen Hemminger30cfd0b2007-07-25 23:49:34 -07003285 if (ca_ops->pkts_acked) {
3286 s32 rtt_us = -1;
Ilpo Järvinenb9ce2042007-06-15 15:08:43 -07003287
Stephen Hemminger30cfd0b2007-07-25 23:49:34 -07003288 /* Is the ACK triggering packet unambiguous? */
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003289 if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
Stephen Hemminger30cfd0b2007-07-25 23:49:34 -07003290 /* High resolution needed and available? */
3291 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
3292 !ktime_equal(last_ackt,
3293 net_invalid_timestamp()))
3294 rtt_us = ktime_us_delta(ktime_get_real(),
3295 last_ackt);
stephen hemmingerfebf0812011-03-14 07:52:12 +00003296 else if (ca_seq_rtt >= 0)
Gavin McCullagh2072c222007-12-29 19:11:21 -08003297 rtt_us = jiffies_to_usecs(ca_seq_rtt);
Stephen Hemminger30cfd0b2007-07-25 23:49:34 -07003298 }
3299
3300 ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
3301 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003302 }
3303
3304#if FASTRETRANS_DEBUG > 0
Ilpo Järvinen547b7922008-07-25 21:43:18 -07003305 WARN_ON((int)tp->sacked_out < 0);
3306 WARN_ON((int)tp->lost_out < 0);
3307 WARN_ON((int)tp->retrans_out < 0);
Ilpo Järvinene60402d2007-08-09 15:14:46 +03003308 if (!tp->packets_out && tcp_is_sack(tp)) {
Stephen Hemmingercfcabdc2007-10-09 01:59:42 -07003309 icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003310 if (tp->lost_out) {
Joe Perches91df42b2012-05-15 14:11:54 +00003311 pr_debug("Leak l=%u %d\n",
3312 tp->lost_out, icsk->icsk_ca_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003313 tp->lost_out = 0;
3314 }
3315 if (tp->sacked_out) {
Joe Perches91df42b2012-05-15 14:11:54 +00003316 pr_debug("Leak s=%u %d\n",
3317 tp->sacked_out, icsk->icsk_ca_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003318 tp->sacked_out = 0;
3319 }
3320 if (tp->retrans_out) {
Joe Perches91df42b2012-05-15 14:11:54 +00003321 pr_debug("Leak r=%u %d\n",
3322 tp->retrans_out, icsk->icsk_ca_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003323 tp->retrans_out = 0;
3324 }
3325 }
3326#endif
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003327 return flag;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003328}
3329
3330static void tcp_ack_probe(struct sock *sk)
3331{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003332 const struct tcp_sock *tp = tcp_sk(sk);
3333 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003334
3335 /* Was it a usable window open? */
3336
Ilpo Järvinen90840de2007-12-31 04:48:41 -08003337 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003338 icsk->icsk_backoff = 0;
3339 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003340 /* Socket must be waked up by subsequent tcp_data_snd_check().
3341 * This function is not for random using!
3342 */
3343 } else {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003344 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07003345 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3346 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003347 }
3348}
3349
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00003350static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003351{
Eric Dumazeta02cec22010-09-22 20:43:57 +00003352 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3353 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003354}
3355
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00003356static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003357{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003358 const struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003359 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
Yuchung Cheng684bad12012-09-02 17:38:04 +00003360 !tcp_in_cwnd_reduction(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003361}
3362
3363/* Check that window update is acceptable.
3364 * The function assumes that snd_una<=ack<=snd_next.
3365 */
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00003366static inline bool tcp_may_update_window(const struct tcp_sock *tp,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003367 const u32 ack, const u32 ack_seq,
3368 const u32 nwin)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003369{
Eric Dumazeta02cec22010-09-22 20:43:57 +00003370 return after(ack, tp->snd_una) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07003371 after(ack_seq, tp->snd_wl1) ||
Eric Dumazeta02cec22010-09-22 20:43:57 +00003372 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003373}
3374
3375/* Update our send window.
3376 *
3377 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3378 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3379 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003380static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003381 u32 ack_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003382{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003383 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003384 int flag = 0;
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07003385 u32 nwin = ntohs(tcp_hdr(skb)->window);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003386
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07003387 if (likely(!tcp_hdr(skb)->syn))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003388 nwin <<= tp->rx_opt.snd_wscale;
3389
3390 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3391 flag |= FLAG_WIN_UPDATE;
Hantzis Fotisee7537b2009-03-02 22:42:02 -08003392 tcp_update_wl(tp, ack_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003393
3394 if (tp->snd_wnd != nwin) {
3395 tp->snd_wnd = nwin;
3396
3397 /* Note, it is the only place, where
3398 * fast path is recovered for sending TCP.
3399 */
Herbert Xu2ad41062005-10-27 18:47:46 +10003400 tp->pred_flags = 0;
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003401 tcp_fast_path_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003402
3403 if (nwin > tp->max_window) {
3404 tp->max_window = nwin;
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08003405 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003406 }
3407 }
3408 }
3409
3410 tp->snd_una = ack;
3411
3412 return flag;
3413}
3414
Ilpo Järvinen9ead9a12007-02-21 22:56:19 -08003415/* A very conservative spurious RTO response algorithm: reduce cwnd and
3416 * continue in congestion avoidance.
3417 */
3418static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
3419{
3420 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
Ilpo Järvinenaa8b6a72007-02-21 23:06:03 -08003421 tp->snd_cwnd_cnt = 0;
Ilpo Järvinen16e90682007-10-11 17:32:31 -07003422 tp->bytes_acked = 0;
Ilpo Järvinen46323652007-05-19 13:56:23 -07003423 TCP_ECN_queue_cwr(tp);
Ilpo Järvinen9ead9a12007-02-21 22:56:19 -08003424 tcp_moderate_cwnd(tp);
3425}
3426
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -08003427/* A conservative spurious RTO response algorithm: reduce cwnd using
Yuchung Cheng684bad12012-09-02 17:38:04 +00003428 * PRR and continue in congestion avoidance.
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -08003429 */
Yuchung Cheng684bad12012-09-02 17:38:04 +00003430static void tcp_cwr_spur_to_response(struct sock *sk)
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -08003431{
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -08003432 tcp_enter_cwr(sk, 0);
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -08003433}
3434
Ilpo Järvinene317f6f2007-03-02 13:34:19 -08003435static void tcp_undo_spur_to_response(struct sock *sk, int flag)
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -08003436{
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003437 if (flag & FLAG_ECE)
Yuchung Cheng684bad12012-09-02 17:38:04 +00003438 tcp_cwr_spur_to_response(sk);
Ilpo Järvinene317f6f2007-03-02 13:34:19 -08003439 else
David S. Millerf6152732011-03-22 19:37:11 -07003440 tcp_undo_cwr(sk, true);
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -08003441}
3442
Ilpo Järvinen30935cf2007-02-21 23:01:36 -08003443/* F-RTO spurious RTO detection algorithm (RFC4138)
3444 *
Ilpo Järvinen6408d202007-02-21 23:04:11 -08003445 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3446 * comments). State (ACK number) is kept in frto_counter. When ACK advances
3447 * window (but not to or beyond highest sequence sent before RTO):
Ilpo Järvinen30935cf2007-02-21 23:01:36 -08003448 * On First ACK, send two new segments out.
3449 * On Second ACK, RTO was likely spurious. Do spurious response (response
3450 * algorithm is not part of the F-RTO detection algorithm
3451 * given in RFC4138 but can be selected separately).
3452 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
Ilpo Järvinend551e452007-04-30 00:42:20 -07003453 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3454 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3455 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
Ilpo Järvinen30935cf2007-02-21 23:01:36 -08003456 *
3457 * Rationale: if the RTO was spurious, new ACKs should arrive from the
3458 * original window even after we transmit two new data segments.
3459 *
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08003460 * SACK version:
3461 * on first step, wait until first cumulative ACK arrives, then move to
3462 * the second step. In second step, the next ACK decides.
3463 *
Ilpo Järvinen30935cf2007-02-21 23:01:36 -08003464 * F-RTO is implemented (mainly) in four functions:
3465 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
3466 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3467 * called when tcp_use_frto() showed green light
3468 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3469 * - tcp_enter_frto_loss() is called if there is not enough evidence
3470 * to prove that the RTO is indeed spurious. It transfers the control
3471 * from F-RTO to the conventional RTO recovery
3472 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003473static bool tcp_process_frto(struct sock *sk, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003474{
3475 struct tcp_sock *tp = tcp_sk(sk);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09003476
Ilpo Järvinen005903b2007-08-09 14:44:16 +03003477 tcp_verify_left_out(tp);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09003478
Ilpo Järvinen7487c482007-02-21 23:02:30 -08003479 /* Duplicate the behavior from Loss state (fastretrans_alert) */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003480 if (flag & FLAG_DATA_ACKED)
Ilpo Järvinen7487c482007-02-21 23:02:30 -08003481 inet_csk(sk)->icsk_retransmits = 0;
3482
Ilpo Järvinen009a2e32007-09-20 11:34:38 -07003483 if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||
3484 ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
3485 tp->undo_marker = 0;
3486
Ilpo Järvinen95c49222007-02-21 23:05:18 -08003487 if (!before(tp->snd_una, tp->frto_highmark)) {
Ilpo Järvinend551e452007-04-30 00:42:20 -07003488 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003489 return true;
Ilpo Järvinen95c49222007-02-21 23:05:18 -08003490 }
3491
Ilpo Järvinen62ab2222008-05-08 01:09:11 -07003492 if (!tcp_is_sackfrto(tp)) {
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08003493 /* RFC4138 shortcoming in step 2; should also have case c):
3494 * ACK isn't duplicate nor advances window, e.g., opposite dir
3495 * data, winupdate
3496 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003497 if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003498 return true;
Ilpo Järvinen6408d202007-02-21 23:04:11 -08003499
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003500 if (!(flag & FLAG_DATA_ACKED)) {
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08003501 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
3502 flag);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003503 return true;
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08003504 }
3505 } else {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003506 if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08003507 /* Prevent sending of new data. */
3508 tp->snd_cwnd = min(tp->snd_cwnd,
3509 tcp_packets_in_flight(tp));
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003510 return true;
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08003511 }
3512
Ilpo Järvinend551e452007-04-30 00:42:20 -07003513 if ((tp->frto_counter >= 2) &&
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003514 (!(flag & FLAG_FORWARD_PROGRESS) ||
3515 ((flag & FLAG_DATA_SACKED) &&
3516 !(flag & FLAG_ONLY_ORIG_SACKED)))) {
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08003517 /* RFC4138 shortcoming (see comment above) */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003518 if (!(flag & FLAG_FORWARD_PROGRESS) &&
3519 (flag & FLAG_NOT_DUP))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003520 return true;
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08003521
3522 tcp_enter_frto_loss(sk, 3, flag);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003523 return true;
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08003524 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003525 }
3526
3527 if (tp->frto_counter == 1) {
Ilpo Järvinen3e6f0492007-12-05 02:20:21 -08003528 /* tcp_may_send_now needs to see updated state */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003529 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
Ilpo Järvinen94d0ea72007-02-21 23:07:27 -08003530 tp->frto_counter = 2;
Ilpo Järvinen3e6f0492007-12-05 02:20:21 -08003531
3532 if (!tcp_may_send_now(sk))
3533 tcp_enter_frto_loss(sk, 2, flag);
3534
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003535 return true;
Ilpo Järvinend551e452007-04-30 00:42:20 -07003536 } else {
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -08003537 switch (sysctl_tcp_frto_response) {
3538 case 2:
Ilpo Järvinene317f6f2007-03-02 13:34:19 -08003539 tcp_undo_spur_to_response(sk, flag);
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -08003540 break;
3541 case 1:
3542 tcp_conservative_spur_to_response(tp);
3543 break;
3544 default:
Yuchung Cheng684bad12012-09-02 17:38:04 +00003545 tcp_cwr_spur_to_response(sk);
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -08003546 break;
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07003547 }
Ilpo Järvinen94d0ea72007-02-21 23:07:27 -08003548 tp->frto_counter = 0;
Ilpo Järvinen009a2e32007-09-20 11:34:38 -07003549 tp->undo_marker = 0;
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07003550 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSPURIOUSRTOS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003551 }
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003552 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003553}
3554
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003555/* RFC 5961 7 [ACK Throttling] */
3556static void tcp_send_challenge_ack(struct sock *sk)
3557{
3558 /* unprotected vars, we dont care of overwrites */
3559 static u32 challenge_timestamp;
3560 static unsigned int challenge_count;
3561 u32 now = jiffies / HZ;
3562
3563 if (now != challenge_timestamp) {
3564 challenge_timestamp = now;
3565 challenge_count = 0;
3566 }
3567 if (++challenge_count <= sysctl_tcp_challenge_ack_limit) {
3568 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK);
3569 tcp_send_ack(sk);
3570 }
3571}
3572
Linus Torvalds1da177e2005-04-16 15:20:36 -07003573/* This routine deals with incoming acks, but not outgoing ones. */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003574static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003575{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003576 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003577 struct tcp_sock *tp = tcp_sk(sk);
3578 u32 prior_snd_una = tp->snd_una;
3579 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3580 u32 ack = TCP_SKB_CB(skb)->ack_seq;
Neal Cardwell7d2b55f2011-11-16 08:58:01 +00003581 bool is_dupack = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003582 u32 prior_in_flight;
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003583 u32 prior_fackets;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003584 int prior_packets;
Nandita Dukkipatia262f0c2011-08-21 20:21:57 +00003585 int prior_sacked = tp->sacked_out;
Neal Cardwell7d2b55f2011-11-16 08:58:01 +00003586 int pkts_acked = 0;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003587 bool frto_cwnd = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003588
John Dykstra96e0bf42009-03-22 21:49:57 -07003589 /* If the ack is older than previous acks
Linus Torvalds1da177e2005-04-16 15:20:36 -07003590 * then we can probably ignore it.
3591 */
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003592 if (before(ack, prior_snd_una)) {
3593 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3594 if (before(ack, prior_snd_una - tp->max_window)) {
3595 tcp_send_challenge_ack(sk);
3596 return -1;
3597 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003598 goto old_ack;
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003599 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003600
John Dykstra96e0bf42009-03-22 21:49:57 -07003601 /* If the ack includes data we haven't sent yet, discard
3602 * this segment (RFC793 Section 3.9).
3603 */
3604 if (after(ack, tp->snd_nxt))
3605 goto invalid_ack;
3606
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00003607 if (tp->early_retrans_delayed)
3608 tcp_rearm_rto(sk);
3609
Ilpo Järvinen2e605292007-08-02 19:46:58 -07003610 if (after(ack, prior_snd_una))
3611 flag |= FLAG_SND_UNA_ADVANCED;
3612
Daikichi Osuga3fdf3f02006-08-29 02:01:44 -07003613 if (sysctl_tcp_abc) {
3614 if (icsk->icsk_ca_state < TCP_CA_CWR)
3615 tp->bytes_acked += ack - prior_snd_una;
3616 else if (icsk->icsk_ca_state == TCP_CA_Loss)
3617 /* we assume just one segment left network */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003618 tp->bytes_acked += min(ack - prior_snd_una,
3619 tp->mss_cache);
Daikichi Osuga3fdf3f02006-08-29 02:01:44 -07003620 }
Stephen Hemminger9772efb2005-11-10 17:09:53 -08003621
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003622 prior_fackets = tp->fackets_out;
Ilpo Järvinen52d34082007-12-05 02:21:35 -08003623 prior_in_flight = tcp_packets_in_flight(tp);
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003624
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003625 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003626 /* Window is constant, pure forward advance.
3627 * No more checks are required.
3628 * Note, we use the fact that SND.UNA>=SND.WL2.
3629 */
Hantzis Fotisee7537b2009-03-02 22:42:02 -08003630 tcp_update_wl(tp, ack_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003631 tp->snd_una = ack;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003632 flag |= FLAG_WIN_UPDATE;
3633
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003634 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
Stephen Hemminger317a76f2005-06-23 12:19:55 -07003635
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07003636 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003637 } else {
3638 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3639 flag |= FLAG_DATA;
3640 else
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07003641 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003642
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003643 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003644
3645 if (TCP_SKB_CB(skb)->sacked)
3646 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3647
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07003648 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003649 flag |= FLAG_ECE;
3650
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003651 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003652 }
3653
3654 /* We passed data and got it acked, remove any soft error
3655 * log. Something worked...
3656 */
3657 sk->sk_err_soft = 0;
David S. Miller4b53fb62008-07-23 16:38:45 -07003658 icsk->icsk_probes_out = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003659 tp->rcv_tstamp = tcp_time_stamp;
3660 prior_packets = tp->packets_out;
3661 if (!prior_packets)
3662 goto no_queue;
3663
Linus Torvalds1da177e2005-04-16 15:20:36 -07003664 /* See if we can take anything off of the retransmit queue. */
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07003665 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003666
Neal Cardwell7d2b55f2011-11-16 08:58:01 +00003667 pkts_acked = prior_packets - tp->packets_out;
Nandita Dukkipatia262f0c2011-08-21 20:21:57 +00003668
Ilpo Järvinene1cd8f72007-11-14 15:55:09 -08003669 if (tp->frto_counter)
3670 frto_cwnd = tcp_process_frto(sk, flag);
Ilpo Järvinen3de96472007-10-01 15:28:48 -07003671 /* Guarantee sacktag reordering detection against wrap-arounds */
3672 if (before(tp->frto_highmark, tp->snd_una))
3673 tp->frto_highmark = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003674
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003675 if (tcp_ack_is_dubious(sk, flag)) {
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08003676 /* Advance CWND, if state allows this. */
Ilpo Järvinen7c9a4a52007-02-21 23:08:34 -08003677 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
3678 tcp_may_raise_cwnd(sk, flag))
Ilpo Järvinenc3a05c62007-12-02 00:47:59 +02003679 tcp_cong_avoid(sk, ack, prior_in_flight);
Neal Cardwell7d2b55f2011-11-16 08:58:01 +00003680 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
Yuchung Cheng7c4a56f2012-08-23 07:05:17 +00003681 tcp_fastretrans_alert(sk, pkts_acked, prior_sacked,
Neal Cardwell7d2b55f2011-11-16 08:58:01 +00003682 is_dupack, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003683 } else {
Ilpo Järvinen7c9a4a52007-02-21 23:08:34 -08003684 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
Ilpo Järvinenc3a05c62007-12-02 00:47:59 +02003685 tcp_cong_avoid(sk, ack, prior_in_flight);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003686 }
3687
David S. Miller5110effe2012-07-02 02:21:03 -07003688 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) {
3689 struct dst_entry *dst = __sk_dst_get(sk);
3690 if (dst)
3691 dst_confirm(dst);
3692 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003693 return 1;
3694
3695no_queue:
Neal Cardwell5628adf2011-11-16 08:58:02 +00003696 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3697 if (flag & FLAG_DSACKING_ACK)
Yuchung Cheng7c4a56f2012-08-23 07:05:17 +00003698 tcp_fastretrans_alert(sk, pkts_acked, prior_sacked,
Neal Cardwell5628adf2011-11-16 08:58:02 +00003699 is_dupack, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003700 /* If this ack opens up a zero window, clear backoff. It was
3701 * being used to time the probes, and is probably far higher than
3702 * it needs to be for normal retransmission.
3703 */
David S. Millerfe067e82007-03-07 12:12:44 -08003704 if (tcp_send_head(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003705 tcp_ack_probe(sk);
3706 return 1;
3707
John Dykstra96e0bf42009-03-22 21:49:57 -07003708invalid_ack:
3709 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3710 return -1;
3711
Linus Torvalds1da177e2005-04-16 15:20:36 -07003712old_ack:
Neal Cardwelle95ae2f2011-11-16 08:58:03 +00003713 /* If data was SACKed, tag it and see if we should send more data.
3714 * If data was DSACKed, see if we can undo a cwnd reduction.
3715 */
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07003716 if (TCP_SKB_CB(skb)->sacked) {
Neal Cardwelle95ae2f2011-11-16 08:58:03 +00003717 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
Yuchung Cheng7c4a56f2012-08-23 07:05:17 +00003718 tcp_fastretrans_alert(sk, pkts_acked, prior_sacked,
Neal Cardwelle95ae2f2011-11-16 08:58:03 +00003719 is_dupack, flag);
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07003720 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003721
John Dykstra96e0bf42009-03-22 21:49:57 -07003722 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003723 return 0;
3724}
3725
Linus Torvalds1da177e2005-04-16 15:20:36 -07003726/* Look for tcp options. Normally only called on SYN and SYNACK packets.
3727 * But, this can also be called on packets in the established flow when
3728 * the fast version below fails.
3729 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003730void tcp_parse_options(const struct sk_buff *skb, struct tcp_options_received *opt_rx,
Yuchung Cheng2100c8d2012-07-19 06:43:05 +00003731 const u8 **hvpp, int estab,
3732 struct tcp_fastopen_cookie *foc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003733{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003734 const unsigned char *ptr;
3735 const struct tcphdr *th = tcp_hdr(skb);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003736 int length = (th->doff * 4) - sizeof(struct tcphdr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003737
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003738 ptr = (const unsigned char *)(th + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003739 opt_rx->saw_tstamp = 0;
3740
Stephen Hemminger2de979b2007-03-08 20:45:19 -08003741 while (length > 0) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003742 int opcode = *ptr++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003743 int opsize;
3744
3745 switch (opcode) {
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003746 case TCPOPT_EOL:
3747 return;
3748 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3749 length--;
3750 continue;
3751 default:
3752 opsize = *ptr++;
3753 if (opsize < 2) /* "silly options" */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003754 return;
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003755 if (opsize > length)
3756 return; /* don't parse partial options */
3757 switch (opcode) {
3758 case TCPOPT_MSS:
3759 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07003760 u16 in_mss = get_unaligned_be16(ptr);
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003761 if (in_mss) {
3762 if (opt_rx->user_mss &&
3763 opt_rx->user_mss < in_mss)
3764 in_mss = opt_rx->user_mss;
3765 opt_rx->mss_clamp = in_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003766 }
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07003767 }
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003768 break;
3769 case TCPOPT_WINDOW:
3770 if (opsize == TCPOLEN_WINDOW && th->syn &&
David S. Millerbb5b7c12009-12-15 20:56:42 -08003771 !estab && sysctl_tcp_window_scaling) {
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003772 __u8 snd_wscale = *(__u8 *)ptr;
3773 opt_rx->wscale_ok = 1;
3774 if (snd_wscale > 14) {
Joe Perchese87cc472012-05-13 21:56:26 +00003775 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3776 __func__,
3777 snd_wscale);
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003778 snd_wscale = 14;
3779 }
3780 opt_rx->snd_wscale = snd_wscale;
3781 }
3782 break;
3783 case TCPOPT_TIMESTAMP:
3784 if ((opsize == TCPOLEN_TIMESTAMP) &&
3785 ((estab && opt_rx->tstamp_ok) ||
David S. Millerbb5b7c12009-12-15 20:56:42 -08003786 (!estab && sysctl_tcp_timestamps))) {
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003787 opt_rx->saw_tstamp = 1;
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07003788 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3789 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003790 }
3791 break;
3792 case TCPOPT_SACK_PERM:
3793 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
David S. Millerbb5b7c12009-12-15 20:56:42 -08003794 !estab && sysctl_tcp_sack) {
Vijay Subramanianab562222011-12-20 13:23:24 +00003795 opt_rx->sack_ok = TCP_SACK_SEEN;
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003796 tcp_sack_reset(opt_rx);
3797 }
3798 break;
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07003799
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003800 case TCPOPT_SACK:
3801 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3802 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3803 opt_rx->sack_ok) {
3804 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3805 }
3806 break;
3807#ifdef CONFIG_TCP_MD5SIG
3808 case TCPOPT_MD5SIG:
3809 /*
3810 * The MD5 Hash has already been
3811 * checked (see tcp_v{4,6}_do_rcv()).
3812 */
3813 break;
3814#endif
William Allen Simpson4957faade2009-12-02 18:25:27 +00003815 case TCPOPT_COOKIE:
3816 /* This option is variable length.
3817 */
3818 switch (opsize) {
3819 case TCPOLEN_COOKIE_BASE:
3820 /* not yet implemented */
3821 break;
3822 case TCPOLEN_COOKIE_PAIR:
3823 /* not yet implemented */
3824 break;
3825 case TCPOLEN_COOKIE_MIN+0:
3826 case TCPOLEN_COOKIE_MIN+2:
3827 case TCPOLEN_COOKIE_MIN+4:
3828 case TCPOLEN_COOKIE_MIN+6:
3829 case TCPOLEN_COOKIE_MAX:
3830 /* 16-bit multiple */
3831 opt_rx->cookie_plus = opsize;
3832 *hvpp = ptr;
Joe Perchesccbd6a52010-05-14 10:58:26 +00003833 break;
William Allen Simpson4957faade2009-12-02 18:25:27 +00003834 default:
3835 /* ignore option */
3836 break;
Joe Perchesccbd6a52010-05-14 10:58:26 +00003837 }
William Allen Simpson4957faade2009-12-02 18:25:27 +00003838 break;
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003839
Yuchung Cheng2100c8d2012-07-19 06:43:05 +00003840 case TCPOPT_EXP:
3841 /* Fast Open option shares code 254 using a
3842 * 16 bits magic number. It's valid only in
3843 * SYN or SYN-ACK with an even size.
3844 */
3845 if (opsize < TCPOLEN_EXP_FASTOPEN_BASE ||
3846 get_unaligned_be16(ptr) != TCPOPT_FASTOPEN_MAGIC ||
3847 foc == NULL || !th->syn || (opsize & 1))
3848 break;
3849 foc->len = opsize - TCPOLEN_EXP_FASTOPEN_BASE;
3850 if (foc->len >= TCP_FASTOPEN_COOKIE_MIN &&
3851 foc->len <= TCP_FASTOPEN_COOKIE_MAX)
3852 memcpy(foc->val, ptr + 2, foc->len);
3853 else if (foc->len != 0)
3854 foc->len = -1;
3855 break;
3856
3857 }
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003858 ptr += opsize-2;
3859 length -= opsize;
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07003860 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003861 }
3862}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00003863EXPORT_SYMBOL(tcp_parse_options);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003864
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003865static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
Ilpo Järvinena4356b22008-08-23 05:12:29 -07003866{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003867 const __be32 *ptr = (const __be32 *)(th + 1);
Ilpo Järvinena4356b22008-08-23 05:12:29 -07003868
3869 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3870 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3871 tp->rx_opt.saw_tstamp = 1;
3872 ++ptr;
3873 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3874 ++ptr;
3875 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003876 return true;
Ilpo Järvinena4356b22008-08-23 05:12:29 -07003877 }
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003878 return false;
Ilpo Järvinena4356b22008-08-23 05:12:29 -07003879}
3880
Linus Torvalds1da177e2005-04-16 15:20:36 -07003881/* Fast parse options. This hopes to only see timestamps.
3882 * If it is wrong it falls back on tcp_parse_options().
3883 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003884static bool tcp_fast_parse_options(const struct sk_buff *skb,
3885 const struct tcphdr *th,
3886 struct tcp_sock *tp, const u8 **hvpp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003887{
William Allen Simpson4957faade2009-12-02 18:25:27 +00003888 /* In the spirit of fast parsing, compare doff directly to constant
3889 * values. Because equality is used, short doff can be ignored here.
3890 */
3891 if (th->doff == (sizeof(*th) / 4)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003892 tp->rx_opt.saw_tstamp = 0;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003893 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003894 } else if (tp->rx_opt.tstamp_ok &&
William Allen Simpson4957faade2009-12-02 18:25:27 +00003895 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
Ilpo Järvinena4356b22008-08-23 05:12:29 -07003896 if (tcp_parse_aligned_timestamp(tp, th))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003897 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003898 }
Yuchung Cheng2100c8d2012-07-19 06:43:05 +00003899 tcp_parse_options(skb, &tp->rx_opt, hvpp, 1, NULL);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003900 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003901}
3902
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003903#ifdef CONFIG_TCP_MD5SIG
3904/*
3905 * Parse MD5 Signature option
3906 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003907const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003908{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003909 int length = (th->doff << 2) - sizeof(*th);
3910 const u8 *ptr = (const u8 *)(th + 1);
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003911
3912 /* If the TCP option is too short, we can short cut */
3913 if (length < TCPOLEN_MD5SIG)
3914 return NULL;
3915
3916 while (length > 0) {
3917 int opcode = *ptr++;
3918 int opsize;
3919
3920 switch(opcode) {
3921 case TCPOPT_EOL:
3922 return NULL;
3923 case TCPOPT_NOP:
3924 length--;
3925 continue;
3926 default:
3927 opsize = *ptr++;
3928 if (opsize < 2 || opsize > length)
3929 return NULL;
3930 if (opcode == TCPOPT_MD5SIG)
Dmitry Popovba78e2d2010-08-07 20:24:28 -07003931 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003932 }
3933 ptr += opsize - 2;
3934 length -= opsize;
3935 }
3936 return NULL;
3937}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00003938EXPORT_SYMBOL(tcp_parse_md5sig_option);
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003939#endif
3940
Linus Torvalds1da177e2005-04-16 15:20:36 -07003941static inline void tcp_store_ts_recent(struct tcp_sock *tp)
3942{
3943 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
James Morris9d729f72007-03-04 16:12:44 -08003944 tp->rx_opt.ts_recent_stamp = get_seconds();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003945}
3946
3947static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3948{
3949 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3950 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3951 * extra check below makes sure this can only happen
3952 * for pure ACK frames. -DaveM
3953 *
3954 * Not only, also it occurs for expired timestamps.
3955 */
3956
Ilpo Järvinenc887e6d2009-03-14 14:23:03 +00003957 if (tcp_paws_check(&tp->rx_opt, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003958 tcp_store_ts_recent(tp);
3959 }
3960}
3961
3962/* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3963 *
3964 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3965 * it can pass through stack. So, the following predicate verifies that
3966 * this segment is not used for anything but congestion avoidance or
3967 * fast retransmit. Moreover, we even are able to eliminate most of such
3968 * second order effects, if we apply some small "replay" window (~RTO)
3969 * to timestamp space.
3970 *
3971 * All these measures still do not guarantee that we reject wrapped ACKs
3972 * on networks with high bandwidth, when sequence space is recycled fastly,
3973 * but it guarantees that such events will be very rare and do not affect
3974 * connection seriously. This doesn't look nice, but alas, PAWS is really
3975 * buggy extension.
3976 *
3977 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3978 * states that events when retransmit arrives after original data are rare.
3979 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3980 * the biggest problem on large power networks even with minor reordering.
3981 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3982 * up to bandwidth of 18Gigabit/sec. 8) ]
3983 */
3984
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003985static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003986{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003987 const struct tcp_sock *tp = tcp_sk(sk);
3988 const struct tcphdr *th = tcp_hdr(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003989 u32 seq = TCP_SKB_CB(skb)->seq;
3990 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3991
3992 return (/* 1. Pure ACK with correct sequence number. */
3993 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3994
3995 /* 2. ... and duplicate ACK. */
3996 ack == tp->snd_una &&
3997
3998 /* 3. ... and does not update window. */
3999 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4000
4001 /* 4. ... and sits in replay window. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004002 (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 -07004003}
4004
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00004005static inline bool tcp_paws_discard(const struct sock *sk,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004006 const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004007{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004008 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinenc887e6d2009-03-14 14:23:03 +00004009
4010 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4011 !tcp_disordered_ack(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004012}
4013
4014/* Check segment sequence number for validity.
4015 *
4016 * Segment controls are considered valid, if the segment
4017 * fits to the window after truncation to the window. Acceptability
4018 * of data (and SYN, FIN, of course) is checked separately.
4019 * See tcp_data_queue(), for example.
4020 *
4021 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4022 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4023 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4024 * (borrowed from freebsd)
4025 */
4026
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00004027static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004028{
4029 return !before(end_seq, tp->rcv_wup) &&
4030 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4031}
4032
4033/* When we get a reset we do this. */
Jerry Chu10467162012-08-31 12:29:11 +00004034void tcp_reset(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004035{
4036 /* We want the right error as BSD sees it (and indeed as we do). */
4037 switch (sk->sk_state) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004038 case TCP_SYN_SENT:
4039 sk->sk_err = ECONNREFUSED;
4040 break;
4041 case TCP_CLOSE_WAIT:
4042 sk->sk_err = EPIPE;
4043 break;
4044 case TCP_CLOSE:
4045 return;
4046 default:
4047 sk->sk_err = ECONNRESET;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004048 }
Tom Marshalla4d25802010-09-20 15:42:05 -07004049 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4050 smp_wmb();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004051
4052 if (!sock_flag(sk, SOCK_DEAD))
4053 sk->sk_error_report(sk);
4054
4055 tcp_done(sk);
4056}
4057
4058/*
4059 * Process the FIN bit. This now behaves as it is supposed to work
4060 * and the FIN takes effect when it is validly part of sequence
4061 * space. Not before when we get holes.
4062 *
4063 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4064 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4065 * TIME-WAIT)
4066 *
4067 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4068 * close and we go into CLOSING (and later onto TIME-WAIT)
4069 *
4070 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4071 */
Eric Dumazet20c4cb72011-10-20 17:44:03 -04004072static void tcp_fin(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004073{
4074 struct tcp_sock *tp = tcp_sk(sk);
4075
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004076 inet_csk_schedule_ack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004077
4078 sk->sk_shutdown |= RCV_SHUTDOWN;
4079 sock_set_flag(sk, SOCK_DONE);
4080
4081 switch (sk->sk_state) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004082 case TCP_SYN_RECV:
4083 case TCP_ESTABLISHED:
4084 /* Move to CLOSE_WAIT */
4085 tcp_set_state(sk, TCP_CLOSE_WAIT);
4086 inet_csk(sk)->icsk_ack.pingpong = 1;
4087 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004088
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004089 case TCP_CLOSE_WAIT:
4090 case TCP_CLOSING:
4091 /* Received a retransmission of the FIN, do
4092 * nothing.
4093 */
4094 break;
4095 case TCP_LAST_ACK:
4096 /* RFC793: Remain in the LAST-ACK state. */
4097 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004098
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004099 case TCP_FIN_WAIT1:
4100 /* This case occurs when a simultaneous close
4101 * happens, we must ack the received FIN and
4102 * enter the CLOSING state.
4103 */
4104 tcp_send_ack(sk);
4105 tcp_set_state(sk, TCP_CLOSING);
4106 break;
4107 case TCP_FIN_WAIT2:
4108 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4109 tcp_send_ack(sk);
4110 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4111 break;
4112 default:
4113 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4114 * cases we should never reach this piece of code.
4115 */
Joe Perches058bd4d2012-03-11 18:36:11 +00004116 pr_err("%s: Impossible, sk->sk_state=%d\n",
Harvey Harrison0dc47872008-03-05 20:47:47 -08004117 __func__, sk->sk_state);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004118 break;
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07004119 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004120
4121 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4122 * Probably, we should reset in this case. For now drop them.
4123 */
4124 __skb_queue_purge(&tp->out_of_order_queue);
Ilpo Järvinene60402d2007-08-09 15:14:46 +03004125 if (tcp_is_sack(tp))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004126 tcp_sack_reset(&tp->rx_opt);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08004127 sk_mem_reclaim(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004128
4129 if (!sock_flag(sk, SOCK_DEAD)) {
4130 sk->sk_state_change(sk);
4131
4132 /* Do not send POLL_HUP for half duplex close. */
4133 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4134 sk->sk_state == TCP_CLOSE)
Pavel Emelyanov8d8ad9d2007-11-26 20:10:50 +08004135 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004136 else
Pavel Emelyanov8d8ad9d2007-11-26 20:10:50 +08004137 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004138 }
4139}
4140
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004141static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004142 u32 end_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004143{
4144 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4145 if (before(seq, sp->start_seq))
4146 sp->start_seq = seq;
4147 if (after(end_seq, sp->end_seq))
4148 sp->end_seq = end_seq;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004149 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004150 }
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004151 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004152}
4153
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004154static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004155{
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004156 struct tcp_sock *tp = tcp_sk(sk);
4157
David S. Millerbb5b7c12009-12-15 20:56:42 -08004158 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07004159 int mib_idx;
4160
Linus Torvalds1da177e2005-04-16 15:20:36 -07004161 if (before(seq, tp->rcv_nxt))
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07004162 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004163 else
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07004164 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4165
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07004166 NET_INC_STATS_BH(sock_net(sk), mib_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004167
4168 tp->rx_opt.dsack = 1;
4169 tp->duplicate_sack[0].start_seq = seq;
4170 tp->duplicate_sack[0].end_seq = end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004171 }
4172}
4173
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004174static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004175{
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004176 struct tcp_sock *tp = tcp_sk(sk);
4177
Linus Torvalds1da177e2005-04-16 15:20:36 -07004178 if (!tp->rx_opt.dsack)
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004179 tcp_dsack_set(sk, seq, end_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004180 else
4181 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4182}
4183
Eric Dumazetcf533ea2011-10-21 05:22:42 -04004184static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004185{
4186 struct tcp_sock *tp = tcp_sk(sk);
4187
4188 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4189 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07004190 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004191 tcp_enter_quickack_mode(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004192
David S. Millerbb5b7c12009-12-15 20:56:42 -08004193 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004194 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4195
4196 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4197 end_seq = tp->rcv_nxt;
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004198 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004199 }
4200 }
4201
4202 tcp_send_ack(sk);
4203}
4204
4205/* These routines update the SACK block as out-of-order packets arrive or
4206 * in-order packets close up the sequence space.
4207 */
4208static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4209{
4210 int this_sack;
4211 struct tcp_sack_block *sp = &tp->selective_acks[0];
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004212 struct tcp_sack_block *swalk = sp + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004213
4214 /* See if the recent change to the first SACK eats into
4215 * or hits the sequence space of other SACK blocks, if so coalesce.
4216 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004217 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004218 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4219 int i;
4220
4221 /* Zap SWALK, by moving every further SACK up by one slot.
4222 * Decrease num_sacks.
4223 */
4224 tp->rx_opt.num_sacks--;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004225 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4226 sp[i] = sp[i + 1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07004227 continue;
4228 }
4229 this_sack++, swalk++;
4230 }
4231}
4232
Linus Torvalds1da177e2005-04-16 15:20:36 -07004233static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4234{
4235 struct tcp_sock *tp = tcp_sk(sk);
4236 struct tcp_sack_block *sp = &tp->selective_acks[0];
4237 int cur_sacks = tp->rx_opt.num_sacks;
4238 int this_sack;
4239
4240 if (!cur_sacks)
4241 goto new_sack;
4242
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004243 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004244 if (tcp_sack_extend(sp, seq, end_seq)) {
4245 /* Rotate this_sack to the first one. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004246 for (; this_sack > 0; this_sack--, sp--)
Ilpo Järvinena0bffff2009-03-21 13:36:17 -07004247 swap(*sp, *(sp - 1));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004248 if (cur_sacks > 1)
4249 tcp_sack_maybe_coalesce(tp);
4250 return;
4251 }
4252 }
4253
4254 /* Could not find an adjacent existing SACK, build a new one,
4255 * put it at the front, and shift everyone else down. We
4256 * always know there is at least one SACK present already here.
4257 *
4258 * If the sack array is full, forget about the last one.
4259 */
Adam Langley4389dde2008-07-19 00:07:02 -07004260 if (this_sack >= TCP_NUM_SACKS) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004261 this_sack--;
4262 tp->rx_opt.num_sacks--;
4263 sp--;
4264 }
Stephen Hemminger2de979b2007-03-08 20:45:19 -08004265 for (; this_sack > 0; this_sack--, sp--)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004266 *sp = *(sp - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004267
4268new_sack:
4269 /* Build the new head SACK, and we're done. */
4270 sp->start_seq = seq;
4271 sp->end_seq = end_seq;
4272 tp->rx_opt.num_sacks++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004273}
4274
4275/* RCV.NXT advances, some SACKs should be eaten. */
4276
4277static void tcp_sack_remove(struct tcp_sock *tp)
4278{
4279 struct tcp_sack_block *sp = &tp->selective_acks[0];
4280 int num_sacks = tp->rx_opt.num_sacks;
4281 int this_sack;
4282
4283 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
David S. Millerb03efcf2005-07-08 14:57:23 -07004284 if (skb_queue_empty(&tp->out_of_order_queue)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004285 tp->rx_opt.num_sacks = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004286 return;
4287 }
4288
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004289 for (this_sack = 0; this_sack < num_sacks;) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004290 /* Check if the start of the sack is covered by RCV.NXT. */
4291 if (!before(tp->rcv_nxt, sp->start_seq)) {
4292 int i;
4293
4294 /* RCV.NXT must cover all the block! */
Ilpo Järvinen547b7922008-07-25 21:43:18 -07004295 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004296
4297 /* Zap this SACK, by moving forward any other SACKS. */
4298 for (i=this_sack+1; i < num_sacks; i++)
4299 tp->selective_acks[i-1] = tp->selective_acks[i];
4300 num_sacks--;
4301 continue;
4302 }
4303 this_sack++;
4304 sp++;
4305 }
Ilpo Järvinen5861f8e2009-03-14 14:23:01 +00004306 tp->rx_opt.num_sacks = num_sacks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004307}
4308
4309/* This one checks to see if we can put data from the
4310 * out_of_order queue into the receive_queue.
4311 */
4312static void tcp_ofo_queue(struct sock *sk)
4313{
4314 struct tcp_sock *tp = tcp_sk(sk);
4315 __u32 dsack_high = tp->rcv_nxt;
4316 struct sk_buff *skb;
4317
4318 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4319 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4320 break;
4321
4322 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4323 __u32 dsack = dsack_high;
4324 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4325 dsack_high = TCP_SKB_CB(skb)->end_seq;
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004326 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004327 }
4328
4329 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
Frans Popb1383382010-03-24 07:57:28 +00004330 SOCK_DEBUG(sk, "ofo packet was already received\n");
David S. Miller8728b832005-08-09 19:25:21 -07004331 __skb_unlink(skb, &tp->out_of_order_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004332 __kfree_skb(skb);
4333 continue;
4334 }
4335 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4336 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4337 TCP_SKB_CB(skb)->end_seq);
4338
David S. Miller8728b832005-08-09 19:25:21 -07004339 __skb_unlink(skb, &tp->out_of_order_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004340 __skb_queue_tail(&sk->sk_receive_queue, skb);
4341 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07004342 if (tcp_hdr(skb)->fin)
Eric Dumazet20c4cb72011-10-20 17:44:03 -04004343 tcp_fin(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004344 }
4345}
4346
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004347static bool tcp_prune_ofo_queue(struct sock *sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004348static int tcp_prune_queue(struct sock *sk);
4349
Mel Gormanc76562b2012-07-31 16:44:41 -07004350static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4351 unsigned int size)
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004352{
4353 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
Mel Gormanc76562b2012-07-31 16:44:41 -07004354 !sk_rmem_schedule(sk, skb, size)) {
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004355
4356 if (tcp_prune_queue(sk) < 0)
4357 return -1;
4358
Mel Gormanc76562b2012-07-31 16:44:41 -07004359 if (!sk_rmem_schedule(sk, skb, size)) {
Vitaliy Gusev56f367b2008-04-15 20:26:34 -07004360 if (!tcp_prune_ofo_queue(sk))
4361 return -1;
4362
Mel Gormanc76562b2012-07-31 16:44:41 -07004363 if (!sk_rmem_schedule(sk, skb, size))
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004364 return -1;
4365 }
4366 }
4367 return 0;
4368}
4369
Eric Dumazet1402d362012-04-23 07:11:42 +00004370/**
4371 * tcp_try_coalesce - try to merge skb to prior one
4372 * @sk: socket
4373 * @to: prior buffer
4374 * @from: buffer to add in queue
Eric Dumazet923dd342012-05-02 07:55:58 +00004375 * @fragstolen: pointer to boolean
Eric Dumazet1402d362012-04-23 07:11:42 +00004376 *
4377 * Before queueing skb @from after @to, try to merge them
4378 * to reduce overall memory use and queue lengths, if cost is small.
4379 * Packets in ofo or receive queues can stay a long time.
4380 * Better try to coalesce them right now to avoid future collapses.
Eric Dumazet783c1752012-04-23 17:34:36 +00004381 * Returns true if caller should free @from instead of queueing it
Eric Dumazet1402d362012-04-23 07:11:42 +00004382 */
Eric Dumazet783c1752012-04-23 17:34:36 +00004383static bool tcp_try_coalesce(struct sock *sk,
4384 struct sk_buff *to,
Eric Dumazet329033f2012-04-27 00:38:33 +00004385 struct sk_buff *from,
4386 bool *fragstolen)
Eric Dumazet1402d362012-04-23 07:11:42 +00004387{
Eric Dumazetbad43ca2012-05-19 03:02:02 +00004388 int delta;
Eric Dumazet1402d362012-04-23 07:11:42 +00004389
Eric Dumazet329033f2012-04-27 00:38:33 +00004390 *fragstolen = false;
Alexander Duyck34a802a2012-05-02 21:19:09 +00004391
Eric Dumazetbad43ca2012-05-19 03:02:02 +00004392 if (tcp_hdr(from)->fin)
4393 return false;
Eric Dumazet1ca7ee32012-05-23 17:51:37 +00004394
4395 /* Its possible this segment overlaps with prior segment in queue */
4396 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4397 return false;
4398
Eric Dumazetbad43ca2012-05-19 03:02:02 +00004399 if (!skb_try_coalesce(to, from, fragstolen, &delta))
Eric Dumazet783c1752012-04-23 17:34:36 +00004400 return false;
Alexander Duyck34a802a2012-05-02 21:19:09 +00004401
Alexander Duyck34a802a2012-05-02 21:19:09 +00004402 atomic_add(delta, &sk->sk_rmem_alloc);
4403 sk_mem_charge(sk, delta);
Alexander Duyck34a802a2012-05-02 21:19:09 +00004404 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4405 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4406 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4407 return true;
Eric Dumazet1402d362012-04-23 07:11:42 +00004408}
4409
Eric Dumazete86b29192012-03-18 11:06:44 +00004410static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4411{
4412 struct tcp_sock *tp = tcp_sk(sk);
4413 struct sk_buff *skb1;
4414 u32 seq, end_seq;
4415
4416 TCP_ECN_check_ce(tp, skb);
4417
Mel Gormanc76562b2012-07-31 16:44:41 -07004418 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
Eric Dumazeta6df1ae2012-07-16 01:41:36 +00004419 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFODROP);
Eric Dumazete86b29192012-03-18 11:06:44 +00004420 __kfree_skb(skb);
4421 return;
4422 }
4423
4424 /* Disable header prediction. */
4425 tp->pred_flags = 0;
4426 inet_csk_schedule_ack(sk);
4427
Eric Dumazeta6df1ae2012-07-16 01:41:36 +00004428 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
Eric Dumazete86b29192012-03-18 11:06:44 +00004429 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4430 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4431
4432 skb1 = skb_peek_tail(&tp->out_of_order_queue);
4433 if (!skb1) {
4434 /* Initial out of order segment, build 1 SACK. */
4435 if (tcp_is_sack(tp)) {
4436 tp->rx_opt.num_sacks = 1;
4437 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4438 tp->selective_acks[0].end_seq =
4439 TCP_SKB_CB(skb)->end_seq;
4440 }
4441 __skb_queue_head(&tp->out_of_order_queue, skb);
4442 goto end;
4443 }
4444
4445 seq = TCP_SKB_CB(skb)->seq;
4446 end_seq = TCP_SKB_CB(skb)->end_seq;
4447
4448 if (seq == TCP_SKB_CB(skb1)->end_seq) {
Eric Dumazet329033f2012-04-27 00:38:33 +00004449 bool fragstolen;
4450
4451 if (!tcp_try_coalesce(sk, skb1, skb, &fragstolen)) {
Eric Dumazet1402d362012-04-23 07:11:42 +00004452 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4453 } else {
Eric Dumazet923dd342012-05-02 07:55:58 +00004454 kfree_skb_partial(skb, fragstolen);
Eric Dumazetc8628152012-03-18 11:07:47 +00004455 skb = NULL;
Eric Dumazetc8628152012-03-18 11:07:47 +00004456 }
Eric Dumazete86b29192012-03-18 11:06:44 +00004457
4458 if (!tp->rx_opt.num_sacks ||
4459 tp->selective_acks[0].end_seq != seq)
4460 goto add_sack;
4461
4462 /* Common case: data arrive in order after hole. */
4463 tp->selective_acks[0].end_seq = end_seq;
4464 goto end;
4465 }
4466
4467 /* Find place to insert this segment. */
4468 while (1) {
4469 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4470 break;
4471 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4472 skb1 = NULL;
4473 break;
4474 }
4475 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4476 }
4477
4478 /* Do skb overlap to previous one? */
4479 if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4480 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4481 /* All the bits are present. Drop. */
Eric Dumazeta6df1ae2012-07-16 01:41:36 +00004482 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
Eric Dumazete86b29192012-03-18 11:06:44 +00004483 __kfree_skb(skb);
4484 skb = NULL;
4485 tcp_dsack_set(sk, seq, end_seq);
4486 goto add_sack;
4487 }
4488 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4489 /* Partial overlap. */
4490 tcp_dsack_set(sk, seq,
4491 TCP_SKB_CB(skb1)->end_seq);
4492 } else {
4493 if (skb_queue_is_first(&tp->out_of_order_queue,
4494 skb1))
4495 skb1 = NULL;
4496 else
4497 skb1 = skb_queue_prev(
4498 &tp->out_of_order_queue,
4499 skb1);
4500 }
4501 }
4502 if (!skb1)
4503 __skb_queue_head(&tp->out_of_order_queue, skb);
4504 else
4505 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4506
4507 /* And clean segments covered by new one as whole. */
4508 while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4509 skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4510
4511 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4512 break;
4513 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4514 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4515 end_seq);
4516 break;
4517 }
4518 __skb_unlink(skb1, &tp->out_of_order_queue);
4519 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4520 TCP_SKB_CB(skb1)->end_seq);
Eric Dumazeta6df1ae2012-07-16 01:41:36 +00004521 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
Eric Dumazete86b29192012-03-18 11:06:44 +00004522 __kfree_skb(skb1);
4523 }
4524
4525add_sack:
4526 if (tcp_is_sack(tp))
4527 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4528end:
4529 if (skb)
4530 skb_set_owner_r(skb, sk);
4531}
4532
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004533static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
Eric Dumazetb081f852012-05-02 09:58:29 +00004534 bool *fragstolen)
4535{
4536 int eaten;
4537 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4538
4539 __skb_pull(skb, hdrlen);
4540 eaten = (tail &&
4541 tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0;
4542 tcp_sk(sk)->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4543 if (!eaten) {
4544 __skb_queue_tail(&sk->sk_receive_queue, skb);
4545 skb_set_owner_r(skb, sk);
4546 }
4547 return eaten;
4548}
Eric Dumazete86b29192012-03-18 11:06:44 +00004549
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004550int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4551{
Mel Gormanc76562b2012-07-31 16:44:41 -07004552 struct sk_buff *skb = NULL;
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004553 struct tcphdr *th;
4554 bool fragstolen;
4555
Pavel Emelyanovc454e612012-10-29 05:05:33 +00004556 if (size == 0)
4557 return 0;
4558
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004559 skb = alloc_skb(size + sizeof(*th), sk->sk_allocation);
4560 if (!skb)
4561 goto err;
4562
Mel Gormanc76562b2012-07-31 16:44:41 -07004563 if (tcp_try_rmem_schedule(sk, skb, size + sizeof(*th)))
4564 goto err_free;
4565
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004566 th = (struct tcphdr *)skb_put(skb, sizeof(*th));
4567 skb_reset_transport_header(skb);
4568 memset(th, 0, sizeof(*th));
4569
4570 if (memcpy_fromiovec(skb_put(skb, size), msg->msg_iov, size))
4571 goto err_free;
4572
4573 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4574 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4575 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4576
4577 if (tcp_queue_rcv(sk, skb, sizeof(*th), &fragstolen)) {
4578 WARN_ON_ONCE(fragstolen); /* should not happen */
4579 __kfree_skb(skb);
4580 }
4581 return size;
4582
4583err_free:
4584 kfree_skb(skb);
4585err:
4586 return -ENOMEM;
4587}
4588
Linus Torvalds1da177e2005-04-16 15:20:36 -07004589static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4590{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04004591 const struct tcphdr *th = tcp_hdr(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004592 struct tcp_sock *tp = tcp_sk(sk);
4593 int eaten = -1;
Eric Dumazet329033f2012-04-27 00:38:33 +00004594 bool fragstolen = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004595
4596 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4597 goto drop;
4598
Eric Dumazetf84af322010-04-28 15:31:51 -07004599 skb_dst_drop(skb);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004600 __skb_pull(skb, th->doff * 4);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004601
4602 TCP_ECN_accept_cwr(tp, skb);
4603
Ilpo Järvinen5861f8e2009-03-14 14:23:01 +00004604 tp->rx_opt.dsack = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004605
4606 /* Queue data for delivery to the user.
4607 * Packets in sequence go to the receive queue.
4608 * Out of sequence packets to the out_of_order_queue.
4609 */
4610 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4611 if (tcp_receive_window(tp) == 0)
4612 goto out_of_window;
4613
4614 /* Ok. In sequence. In window. */
4615 if (tp->ucopy.task == current &&
4616 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4617 sock_owned_by_user(sk) && !tp->urg_data) {
4618 int chunk = min_t(unsigned int, skb->len,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004619 tp->ucopy.len);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004620
4621 __set_current_state(TASK_RUNNING);
4622
4623 local_bh_enable();
4624 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
4625 tp->ucopy.len -= chunk;
4626 tp->copied_seq += chunk;
Jerry Chu44f53242011-01-25 13:46:30 -08004627 eaten = (chunk == skb->len);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004628 tcp_rcv_space_adjust(sk);
4629 }
4630 local_bh_disable();
4631 }
4632
4633 if (eaten <= 0) {
4634queue_and_out:
4635 if (eaten < 0 &&
Mel Gormanc76562b2012-07-31 16:44:41 -07004636 tcp_try_rmem_schedule(sk, skb, skb->truesize))
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004637 goto drop;
4638
Eric Dumazetb081f852012-05-02 09:58:29 +00004639 eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004640 }
4641 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
Stephen Hemminger2de979b2007-03-08 20:45:19 -08004642 if (skb->len)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004643 tcp_event_data_recv(sk, skb);
Stephen Hemminger2de979b2007-03-08 20:45:19 -08004644 if (th->fin)
Eric Dumazet20c4cb72011-10-20 17:44:03 -04004645 tcp_fin(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004646
David S. Millerb03efcf2005-07-08 14:57:23 -07004647 if (!skb_queue_empty(&tp->out_of_order_queue)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004648 tcp_ofo_queue(sk);
4649
4650 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4651 * gap in queue is filled.
4652 */
David S. Millerb03efcf2005-07-08 14:57:23 -07004653 if (skb_queue_empty(&tp->out_of_order_queue))
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004654 inet_csk(sk)->icsk_ack.pingpong = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004655 }
4656
4657 if (tp->rx_opt.num_sacks)
4658 tcp_sack_remove(tp);
4659
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004660 tcp_fast_path_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004661
Eric Dumazet923dd342012-05-02 07:55:58 +00004662 if (eaten > 0)
4663 kfree_skb_partial(skb, fragstolen);
Eric Dumazet1d57f192012-09-17 12:51:39 +00004664 if (!sock_flag(sk, SOCK_DEAD))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004665 sk->sk_data_ready(sk, 0);
4666 return;
4667 }
4668
4669 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4670 /* A retransmit, 2nd most common case. Force an immediate ack. */
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07004671 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004672 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004673
4674out_of_window:
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004675 tcp_enter_quickack_mode(sk);
4676 inet_csk_schedule_ack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004677drop:
4678 __kfree_skb(skb);
4679 return;
4680 }
4681
4682 /* Out of window. F.e. zero window probe. */
4683 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4684 goto out_of_window;
4685
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004686 tcp_enter_quickack_mode(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004687
4688 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4689 /* Partial packet, seq < rcv_next < end_seq */
4690 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4691 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4692 TCP_SKB_CB(skb)->end_seq);
4693
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004694 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004695
Linus Torvalds1da177e2005-04-16 15:20:36 -07004696 /* If window is closed, drop tail of packet. But after
4697 * remembering D-SACK for its head made in previous line.
4698 */
4699 if (!tcp_receive_window(tp))
4700 goto out_of_window;
4701 goto queue_and_out;
4702 }
4703
Eric Dumazete86b29192012-03-18 11:06:44 +00004704 tcp_data_queue_ofo(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004705}
4706
Ilpo Järvinen2cf46632008-08-23 05:11:41 -07004707static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4708 struct sk_buff_head *list)
4709{
David S. Miller91521942009-05-28 21:35:47 -07004710 struct sk_buff *next = NULL;
4711
4712 if (!skb_queue_is_last(list, skb))
4713 next = skb_queue_next(list, skb);
Ilpo Järvinen2cf46632008-08-23 05:11:41 -07004714
4715 __skb_unlink(skb, list);
4716 __kfree_skb(skb);
4717 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4718
4719 return next;
4720}
4721
Linus Torvalds1da177e2005-04-16 15:20:36 -07004722/* Collapse contiguous sequence of skbs head..tail with
4723 * sequence numbers start..end.
David S. Miller91521942009-05-28 21:35:47 -07004724 *
4725 * If tail is NULL, this means until the end of the list.
4726 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07004727 * Segments with FIN/SYN are not collapsed (only because this
4728 * simplifies code)
4729 */
4730static void
David S. Miller8728b832005-08-09 19:25:21 -07004731tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4732 struct sk_buff *head, struct sk_buff *tail,
4733 u32 start, u32 end)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004734{
David S. Miller91521942009-05-28 21:35:47 -07004735 struct sk_buff *skb, *n;
4736 bool end_of_skbs;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004737
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08004738 /* First, check that queue is collapsible and find
Linus Torvalds1da177e2005-04-16 15:20:36 -07004739 * the point where collapsing can be useful. */
David S. Miller91521942009-05-28 21:35:47 -07004740 skb = head;
4741restart:
4742 end_of_skbs = true;
4743 skb_queue_walk_from_safe(list, skb, n) {
4744 if (skb == tail)
4745 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004746 /* No new bits? It is possible on ofo queue. */
4747 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
Ilpo Järvinen2cf46632008-08-23 05:11:41 -07004748 skb = tcp_collapse_one(sk, skb, list);
David S. Miller91521942009-05-28 21:35:47 -07004749 if (!skb)
4750 break;
4751 goto restart;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004752 }
4753
4754 /* The first skb to collapse is:
4755 * - not SYN/FIN and
4756 * - bloated or contains data before "start" or
4757 * overlaps to the next one.
4758 */
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07004759 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07004760 (tcp_win_from_space(skb->truesize) > skb->len ||
David S. Miller91521942009-05-28 21:35:47 -07004761 before(TCP_SKB_CB(skb)->seq, start))) {
4762 end_of_skbs = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004763 break;
David S. Miller91521942009-05-28 21:35:47 -07004764 }
4765
4766 if (!skb_queue_is_last(list, skb)) {
4767 struct sk_buff *next = skb_queue_next(list, skb);
4768 if (next != tail &&
4769 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4770 end_of_skbs = false;
4771 break;
4772 }
4773 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004774
4775 /* Decided to skip this, advance start seq. */
4776 start = TCP_SKB_CB(skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004777 }
David S. Miller91521942009-05-28 21:35:47 -07004778 if (end_of_skbs || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004779 return;
4780
4781 while (before(start, end)) {
4782 struct sk_buff *nskb;
Chuck Leverc2636b42007-10-23 21:07:32 -07004783 unsigned int header = skb_headroom(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004784 int copy = SKB_MAX_ORDER(header, 0);
4785
4786 /* Too big header? This can happen with IPv6. */
4787 if (copy < 0)
4788 return;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004789 if (end - start < copy)
4790 copy = end - start;
4791 nskb = alloc_skb(copy + header, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004792 if (!nskb)
4793 return;
Arnaldo Carvalho de Meloc51957d2007-03-10 12:47:22 -03004794
Arnaldo Carvalho de Melo98e399f2007-03-19 15:33:04 -07004795 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
Arnaldo Carvalho de Melo9c702202007-04-25 18:04:18 -07004796 skb_set_network_header(nskb, (skb_network_header(skb) -
4797 skb->head));
4798 skb_set_transport_header(nskb, (skb_transport_header(skb) -
4799 skb->head));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004800 skb_reserve(nskb, header);
4801 memcpy(nskb->head, skb->head, header);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004802 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4803 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
David S. Miller43f59c82008-09-21 21:28:51 -07004804 __skb_queue_before(list, skb, nskb);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08004805 skb_set_owner_r(nskb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004806
4807 /* Copy data, releasing collapsed skbs. */
4808 while (copy > 0) {
4809 int offset = start - TCP_SKB_CB(skb)->seq;
4810 int size = TCP_SKB_CB(skb)->end_seq - start;
4811
Kris Katterjohn09a62662006-01-08 22:24:28 -08004812 BUG_ON(offset < 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004813 if (size > 0) {
4814 size = min(copy, size);
4815 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4816 BUG();
4817 TCP_SKB_CB(nskb)->end_seq += size;
4818 copy -= size;
4819 start += size;
4820 }
4821 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
Ilpo Järvinen2cf46632008-08-23 05:11:41 -07004822 skb = tcp_collapse_one(sk, skb, list);
David S. Miller91521942009-05-28 21:35:47 -07004823 if (!skb ||
4824 skb == tail ||
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07004825 tcp_hdr(skb)->syn ||
4826 tcp_hdr(skb)->fin)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004827 return;
4828 }
4829 }
4830 }
4831}
4832
4833/* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4834 * and tcp_collapse() them until all the queue is collapsed.
4835 */
4836static void tcp_collapse_ofo_queue(struct sock *sk)
4837{
4838 struct tcp_sock *tp = tcp_sk(sk);
4839 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4840 struct sk_buff *head;
4841 u32 start, end;
4842
4843 if (skb == NULL)
4844 return;
4845
4846 start = TCP_SKB_CB(skb)->seq;
4847 end = TCP_SKB_CB(skb)->end_seq;
4848 head = skb;
4849
4850 for (;;) {
David S. Miller91521942009-05-28 21:35:47 -07004851 struct sk_buff *next = NULL;
4852
4853 if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4854 next = skb_queue_next(&tp->out_of_order_queue, skb);
4855 skb = next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004856
4857 /* Segment is terminated when we see gap or when
4858 * we are at the end of all the queue. */
David S. Miller91521942009-05-28 21:35:47 -07004859 if (!skb ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07004860 after(TCP_SKB_CB(skb)->seq, end) ||
4861 before(TCP_SKB_CB(skb)->end_seq, start)) {
David S. Miller8728b832005-08-09 19:25:21 -07004862 tcp_collapse(sk, &tp->out_of_order_queue,
4863 head, skb, start, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004864 head = skb;
David S. Miller91521942009-05-28 21:35:47 -07004865 if (!skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004866 break;
4867 /* Start new segment */
4868 start = TCP_SKB_CB(skb)->seq;
4869 end = TCP_SKB_CB(skb)->end_seq;
4870 } else {
4871 if (before(TCP_SKB_CB(skb)->seq, start))
4872 start = TCP_SKB_CB(skb)->seq;
4873 if (after(TCP_SKB_CB(skb)->end_seq, end))
4874 end = TCP_SKB_CB(skb)->end_seq;
4875 }
4876 }
4877}
4878
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004879/*
4880 * Purge the out-of-order queue.
Vitaliy Gusev56f367b2008-04-15 20:26:34 -07004881 * Return true if queue was pruned.
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004882 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004883static bool tcp_prune_ofo_queue(struct sock *sk)
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004884{
4885 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004886 bool res = false;
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004887
4888 if (!skb_queue_empty(&tp->out_of_order_queue)) {
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07004889 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004890 __skb_queue_purge(&tp->out_of_order_queue);
4891
4892 /* Reset SACK state. A conforming SACK implementation will
4893 * do the same at a timeout based retransmit. When a connection
4894 * is in a sad state like this, we care only about integrity
4895 * of the connection not performance.
4896 */
4897 if (tp->rx_opt.sack_ok)
4898 tcp_sack_reset(&tp->rx_opt);
4899 sk_mem_reclaim(sk);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004900 res = true;
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004901 }
Vitaliy Gusev56f367b2008-04-15 20:26:34 -07004902 return res;
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004903}
4904
Linus Torvalds1da177e2005-04-16 15:20:36 -07004905/* Reduce allocated memory if we can, trying to get
4906 * the socket within its memory limits again.
4907 *
4908 * Return less than zero if we should start dropping frames
4909 * until the socket owning process reads some of the data
4910 * to stabilize the situation.
4911 */
4912static int tcp_prune_queue(struct sock *sk)
4913{
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004914 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004915
4916 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4917
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07004918 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004919
4920 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004921 tcp_clamp_window(sk);
Glauber Costa180d8cd2011-12-11 21:47:02 +00004922 else if (sk_under_memory_pressure(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004923 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4924
4925 tcp_collapse_ofo_queue(sk);
David S. Miller91521942009-05-28 21:35:47 -07004926 if (!skb_queue_empty(&sk->sk_receive_queue))
4927 tcp_collapse(sk, &sk->sk_receive_queue,
4928 skb_peek(&sk->sk_receive_queue),
4929 NULL,
4930 tp->copied_seq, tp->rcv_nxt);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08004931 sk_mem_reclaim(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004932
4933 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4934 return 0;
4935
4936 /* Collapsing did not help, destructive actions follow.
4937 * This must not ever occur. */
4938
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004939 tcp_prune_ofo_queue(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004940
4941 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4942 return 0;
4943
4944 /* If we are really being abused, tell the caller to silently
4945 * drop receive data on the floor. It will get retransmitted
4946 * and hopefully then we'll have sufficient space.
4947 */
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07004948 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004949
4950 /* Massive buffer overcommit. */
4951 tp->pred_flags = 0;
4952 return -1;
4953}
4954
Linus Torvalds1da177e2005-04-16 15:20:36 -07004955/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4956 * As additional protections, we do not touch cwnd in retransmission phases,
4957 * and if application hit its sndbuf limit recently.
4958 */
4959void tcp_cwnd_application_limited(struct sock *sk)
4960{
4961 struct tcp_sock *tp = tcp_sk(sk);
4962
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03004963 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07004964 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4965 /* Limited by application or receiver window. */
Ilpo Järvinend254bcd2006-08-04 16:57:42 -07004966 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4967 u32 win_used = max(tp->snd_cwnd_used, init_win);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004968 if (win_used < tp->snd_cwnd) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03004969 tp->snd_ssthresh = tcp_current_ssthresh(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004970 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4971 }
4972 tp->snd_cwnd_used = 0;
4973 }
4974 tp->snd_cwnd_stamp = tcp_time_stamp;
4975}
4976
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004977static bool tcp_should_expand_sndbuf(const struct sock *sk)
David S. Miller0d9901d2005-07-05 15:21:10 -07004978{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04004979 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004980
David S. Miller0d9901d2005-07-05 15:21:10 -07004981 /* If the user specified a specific send buffer setting, do
4982 * not modify it.
4983 */
4984 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004985 return false;
David S. Miller0d9901d2005-07-05 15:21:10 -07004986
4987 /* If we are under global TCP memory pressure, do not expand. */
Glauber Costa180d8cd2011-12-11 21:47:02 +00004988 if (sk_under_memory_pressure(sk))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004989 return false;
David S. Miller0d9901d2005-07-05 15:21:10 -07004990
4991 /* If we are under soft global TCP memory pressure, do not expand. */
Glauber Costa180d8cd2011-12-11 21:47:02 +00004992 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004993 return false;
David S. Miller0d9901d2005-07-05 15:21:10 -07004994
4995 /* If we filled the congestion window, do not expand. */
4996 if (tp->packets_out >= tp->snd_cwnd)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004997 return false;
David S. Miller0d9901d2005-07-05 15:21:10 -07004998
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004999 return true;
David S. Miller0d9901d2005-07-05 15:21:10 -07005000}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005001
5002/* When incoming ACK allowed to free some skb from write_queue,
5003 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
5004 * on the exit from tcp input handler.
5005 *
5006 * PROBLEM: sndbuf expansion does not work well with largesend.
5007 */
5008static void tcp_new_space(struct sock *sk)
5009{
5010 struct tcp_sock *tp = tcp_sk(sk);
5011
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005012 if (tcp_should_expand_sndbuf(sk)) {
Eric Dumazet87fb4b72011-10-13 07:28:54 +00005013 int sndmem = SKB_TRUESIZE(max_t(u32,
5014 tp->rx_opt.mss_clamp,
5015 tp->mss_cache) +
5016 MAX_TCP_HEADER);
Ilpo Järvinen4a7e5602008-10-07 14:43:31 -07005017 int demanded = max_t(unsigned int, tp->snd_cwnd,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005018 tp->reordering + 1);
5019 sndmem *= 2 * demanded;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005020 if (sndmem > sk->sk_sndbuf)
5021 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
5022 tp->snd_cwnd_stamp = tcp_time_stamp;
5023 }
5024
5025 sk->sk_write_space(sk);
5026}
5027
Stephen Hemminger40efc6f2006-01-03 16:03:49 -08005028static void tcp_check_space(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005029{
5030 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
5031 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
5032 if (sk->sk_socket &&
5033 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5034 tcp_new_space(sk);
5035 }
5036}
5037
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005038static inline void tcp_data_snd_check(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005039{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005040 tcp_push_pending_frames(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005041 tcp_check_space(sk);
5042}
5043
5044/*
5045 * Check if sending an ack is needed.
5046 */
5047static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5048{
5049 struct tcp_sock *tp = tcp_sk(sk);
5050
5051 /* More than one full frame received... */
Joe Perches9d4fb272009-11-23 10:41:23 -08005052 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07005053 /* ... and right edge of window advances far enough.
5054 * (tcp_recvmsg() will send ACK otherwise). Or...
5055 */
Joe Perches9d4fb272009-11-23 10:41:23 -08005056 __tcp_select_window(sk) >= tp->rcv_wnd) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07005057 /* We ACK each frame or... */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005058 tcp_in_quickack_mode(sk) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07005059 /* We have out of order data. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005060 (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005061 /* Then ack it now */
5062 tcp_send_ack(sk);
5063 } else {
5064 /* Else, send delayed ack. */
5065 tcp_send_delayed_ack(sk);
5066 }
5067}
5068
Stephen Hemminger40efc6f2006-01-03 16:03:49 -08005069static inline void tcp_ack_snd_check(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005070{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005071 if (!inet_csk_ack_scheduled(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005072 /* We sent a data segment already. */
5073 return;
5074 }
5075 __tcp_ack_snd_check(sk, 1);
5076}
5077
5078/*
5079 * This routine is only called when we have urgent data
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08005080 * signaled. Its the 'slow' part of tcp_urg. It could be
Linus Torvalds1da177e2005-04-16 15:20:36 -07005081 * moved inline now as tcp_urg is only called from one
5082 * place. We handle URGent data wrong. We have to - as
5083 * BSD still doesn't use the correction from RFC961.
5084 * For 1003.1g we should support a new option TCP_STDURG to permit
5085 * either form (or just set the sysctl tcp_stdurg).
5086 */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005087
Eric Dumazetcf533ea2011-10-21 05:22:42 -04005088static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005089{
5090 struct tcp_sock *tp = tcp_sk(sk);
5091 u32 ptr = ntohs(th->urg_ptr);
5092
5093 if (ptr && !sysctl_tcp_stdurg)
5094 ptr--;
5095 ptr += ntohl(th->seq);
5096
5097 /* Ignore urgent data that we've already seen and read. */
5098 if (after(tp->copied_seq, ptr))
5099 return;
5100
5101 /* Do not replay urg ptr.
5102 *
5103 * NOTE: interesting situation not covered by specs.
5104 * Misbehaving sender may send urg ptr, pointing to segment,
5105 * which we already have in ofo queue. We are not able to fetch
5106 * such data and will stay in TCP_URG_NOTYET until will be eaten
5107 * by recvmsg(). Seems, we are not obliged to handle such wicked
5108 * situations. But it is worth to think about possibility of some
5109 * DoSes using some hypothetical application level deadlock.
5110 */
5111 if (before(ptr, tp->rcv_nxt))
5112 return;
5113
5114 /* Do we already have a newer (or duplicate) urgent pointer? */
5115 if (tp->urg_data && !after(ptr, tp->urg_seq))
5116 return;
5117
5118 /* Tell the world about our new urgent pointer. */
5119 sk_send_sigurg(sk);
5120
5121 /* We may be adding urgent data when the last byte read was
5122 * urgent. To do this requires some care. We cannot just ignore
5123 * tp->copied_seq since we would read the last urgent byte again
5124 * as data, nor can we alter copied_seq until this data arrives
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08005125 * or we break the semantics of SIOCATMARK (and thus sockatmark())
Linus Torvalds1da177e2005-04-16 15:20:36 -07005126 *
5127 * NOTE. Double Dutch. Rendering to plain English: author of comment
5128 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5129 * and expect that both A and B disappear from stream. This is _wrong_.
5130 * Though this happens in BSD with high probability, this is occasional.
5131 * Any application relying on this is buggy. Note also, that fix "works"
5132 * only in this artificial test. Insert some normal data between A and B and we will
5133 * decline of BSD again. Verdict: it is better to remove to trap
5134 * buggy users.
5135 */
5136 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005137 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005138 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5139 tp->copied_seq++;
5140 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
David S. Miller8728b832005-08-09 19:25:21 -07005141 __skb_unlink(skb, &sk->sk_receive_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005142 __kfree_skb(skb);
5143 }
5144 }
5145
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005146 tp->urg_data = TCP_URG_NOTYET;
5147 tp->urg_seq = ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005148
5149 /* Disable header prediction. */
5150 tp->pred_flags = 0;
5151}
5152
5153/* This is the 'fast' part of urgent handling. */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04005154static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005155{
5156 struct tcp_sock *tp = tcp_sk(sk);
5157
5158 /* Check if we get a new urgent pointer - normally not. */
5159 if (th->urg)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005160 tcp_check_urg(sk, th);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005161
5162 /* Do we wait for any urgent data? - normally not... */
5163 if (tp->urg_data == TCP_URG_NOTYET) {
5164 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5165 th->syn;
5166
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005167 /* Is the urgent pointer pointing into this packet? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005168 if (ptr < skb->len) {
5169 u8 tmp;
5170 if (skb_copy_bits(skb, ptr, &tmp, 1))
5171 BUG();
5172 tp->urg_data = TCP_URG_VALID | tmp;
5173 if (!sock_flag(sk, SOCK_DEAD))
5174 sk->sk_data_ready(sk, 0);
5175 }
5176 }
5177}
5178
5179static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
5180{
5181 struct tcp_sock *tp = tcp_sk(sk);
5182 int chunk = skb->len - hlen;
5183 int err;
5184
5185 local_bh_enable();
Herbert Xu60476372007-04-09 11:59:39 -07005186 if (skb_csum_unnecessary(skb))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005187 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
5188 else
5189 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
5190 tp->ucopy.iov);
5191
5192 if (!err) {
5193 tp->ucopy.len -= chunk;
5194 tp->copied_seq += chunk;
5195 tcp_rcv_space_adjust(sk);
5196 }
5197
5198 local_bh_disable();
5199 return err;
5200}
5201
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005202static __sum16 __tcp_checksum_complete_user(struct sock *sk,
5203 struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005204{
Al Virob51655b2006-11-14 21:40:42 -08005205 __sum16 result;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005206
5207 if (sock_owned_by_user(sk)) {
5208 local_bh_enable();
5209 result = __tcp_checksum_complete(skb);
5210 local_bh_disable();
5211 } else {
5212 result = __tcp_checksum_complete(skb);
5213 }
5214 return result;
5215}
5216
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00005217static inline bool tcp_checksum_complete_user(struct sock *sk,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005218 struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005219{
Herbert Xu60476372007-04-09 11:59:39 -07005220 return !skb_csum_unnecessary(skb) &&
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005221 __tcp_checksum_complete_user(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005222}
5223
Chris Leech1a2449a2006-05-23 18:05:53 -07005224#ifdef CONFIG_NET_DMA
Eric Dumazeta2a385d2012-05-16 23:15:34 +00005225static bool tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005226 int hlen)
Chris Leech1a2449a2006-05-23 18:05:53 -07005227{
5228 struct tcp_sock *tp = tcp_sk(sk);
5229 int chunk = skb->len - hlen;
5230 int dma_cookie;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00005231 bool copied_early = false;
Chris Leech1a2449a2006-05-23 18:05:53 -07005232
5233 if (tp->ucopy.wakeup)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00005234 return false;
Chris Leech1a2449a2006-05-23 18:05:53 -07005235
5236 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
Dave Jianga2bd1142012-04-04 16:10:46 -07005237 tp->ucopy.dma_chan = net_dma_find_channel();
Chris Leech1a2449a2006-05-23 18:05:53 -07005238
Herbert Xu60476372007-04-09 11:59:39 -07005239 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
Chris Leech1a2449a2006-05-23 18:05:53 -07005240
5241 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005242 skb, hlen,
5243 tp->ucopy.iov, chunk,
5244 tp->ucopy.pinned_list);
Chris Leech1a2449a2006-05-23 18:05:53 -07005245
5246 if (dma_cookie < 0)
5247 goto out;
5248
5249 tp->ucopy.dma_cookie = dma_cookie;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00005250 copied_early = true;
Chris Leech1a2449a2006-05-23 18:05:53 -07005251
5252 tp->ucopy.len -= chunk;
5253 tp->copied_seq += chunk;
5254 tcp_rcv_space_adjust(sk);
5255
5256 if ((tp->ucopy.len == 0) ||
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07005257 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
Chris Leech1a2449a2006-05-23 18:05:53 -07005258 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
5259 tp->ucopy.wakeup = 1;
5260 sk->sk_data_ready(sk, 0);
5261 }
5262 } else if (chunk > 0) {
5263 tp->ucopy.wakeup = 1;
5264 sk->sk_data_ready(sk, 0);
5265 }
5266out:
5267 return copied_early;
5268}
5269#endif /* CONFIG_NET_DMA */
5270
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005271/* Does PAWS and seqno based validation of an incoming segment, flags will
5272 * play significant role here.
5273 */
Eric Dumazet0c246042012-07-17 01:41:30 +00005274static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5275 const struct tcphdr *th, int syn_inerr)
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005276{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04005277 const u8 *hash_location;
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005278 struct tcp_sock *tp = tcp_sk(sk);
5279
5280 /* RFC1323: H1. Apply PAWS check first. */
William Allen Simpson4957faade2009-12-02 18:25:27 +00005281 if (tcp_fast_parse_options(skb, th, tp, &hash_location) &&
5282 tp->rx_opt.saw_tstamp &&
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005283 tcp_paws_discard(sk, skb)) {
5284 if (!th->rst) {
5285 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5286 tcp_send_dupack(sk, skb);
5287 goto discard;
5288 }
5289 /* Reset is accepted even if it did not pass PAWS. */
5290 }
5291
5292 /* Step 1: check sequence number */
5293 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5294 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5295 * (RST) segments are validated by checking their SEQ-fields."
5296 * And page 69: "If an incoming segment is not acceptable,
5297 * an acknowledgment should be sent in reply (unless the RST
5298 * bit is set, if so drop the segment and return)".
5299 */
Eric Dumazete3715892012-07-17 12:29:30 +00005300 if (!th->rst) {
5301 if (th->syn)
5302 goto syn_challenge;
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005303 tcp_send_dupack(sk, skb);
Eric Dumazete3715892012-07-17 12:29:30 +00005304 }
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005305 goto discard;
5306 }
5307
5308 /* Step 2: check RST bit */
5309 if (th->rst) {
Eric Dumazet282f23c2012-07-17 10:13:05 +02005310 /* RFC 5961 3.2 :
5311 * If sequence number exactly matches RCV.NXT, then
5312 * RESET the connection
5313 * else
5314 * Send a challenge ACK
5315 */
5316 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt)
5317 tcp_reset(sk);
5318 else
5319 tcp_send_challenge_ack(sk);
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005320 goto discard;
5321 }
5322
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005323 /* step 3: check security and precedence [ignored] */
5324
Eric Dumazet0c246042012-07-17 01:41:30 +00005325 /* step 4: Check for a SYN
5326 * RFC 5691 4.2 : Send a challenge ack
5327 */
5328 if (th->syn) {
Eric Dumazete3715892012-07-17 12:29:30 +00005329syn_challenge:
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005330 if (syn_inerr)
5331 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
Eric Dumazet0c246042012-07-17 01:41:30 +00005332 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5333 tcp_send_challenge_ack(sk);
5334 goto discard;
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005335 }
5336
Eric Dumazet0c246042012-07-17 01:41:30 +00005337 return true;
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005338
5339discard:
5340 __kfree_skb(skb);
Eric Dumazet0c246042012-07-17 01:41:30 +00005341 return false;
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005342}
5343
Linus Torvalds1da177e2005-04-16 15:20:36 -07005344/*
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005345 * TCP receive function for the ESTABLISHED state.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005346 *
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005347 * It is split into a fast path and a slow path. The fast path is
Linus Torvalds1da177e2005-04-16 15:20:36 -07005348 * disabled when:
5349 * - A zero window was announced from us - zero window probing
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005350 * is only handled properly in the slow path.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005351 * - Out of order segments arrived.
5352 * - Urgent data is expected.
5353 * - There is no buffer space left
5354 * - Unexpected TCP flags/window values/header lengths are received
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005355 * (detected by checking the TCP header against pred_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005356 * - Data is sent in both directions. Fast path only supports pure senders
5357 * or pure receivers (this means either the sequence number or the ack
5358 * value must stay constant)
5359 * - Unexpected TCP option.
5360 *
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005361 * When these conditions are not satisfied it drops into a standard
Linus Torvalds1da177e2005-04-16 15:20:36 -07005362 * receive procedure patterned after RFC793 to handle all cases.
5363 * The first three cases are guaranteed by proper pred_flags setting,
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005364 * the rest is checked inline. Fast processing is turned on in
Linus Torvalds1da177e2005-04-16 15:20:36 -07005365 * tcp_data_queue when everything is OK.
5366 */
5367int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
Eric Dumazetcf533ea2011-10-21 05:22:42 -04005368 const struct tcphdr *th, unsigned int len)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005369{
5370 struct tcp_sock *tp = tcp_sk(sk);
5371
Eric Dumazet5d299f32012-08-06 05:09:33 +00005372 if (unlikely(sk->sk_rx_dst == NULL))
5373 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005374 /*
5375 * Header prediction.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005376 * The code loosely follows the one in the famous
Linus Torvalds1da177e2005-04-16 15:20:36 -07005377 * "30 instruction TCP receive" Van Jacobson mail.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005378 *
5379 * Van's trick is to deposit buffers into socket queue
Linus Torvalds1da177e2005-04-16 15:20:36 -07005380 * on a device interrupt, to call tcp_recv function
5381 * on the receive process context and checksum and copy
5382 * the buffer to user space. smart...
5383 *
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005384 * Our current scheme is not silly either but we take the
Linus Torvalds1da177e2005-04-16 15:20:36 -07005385 * extra cost of the net_bh soft interrupt processing...
5386 * We do checksum and copy also but from device to kernel.
5387 */
5388
5389 tp->rx_opt.saw_tstamp = 0;
5390
5391 /* pred_flags is 0xS?10 << 16 + snd_wnd
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08005392 * if header_prediction is to be made
Linus Torvalds1da177e2005-04-16 15:20:36 -07005393 * 'S' will always be tp->tcp_header_len >> 2
5394 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005395 * turn it off (when there are holes in the receive
Linus Torvalds1da177e2005-04-16 15:20:36 -07005396 * space for instance)
5397 * PSH flag is ignored.
5398 */
5399
5400 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
John Dykstra96e0bf42009-03-22 21:49:57 -07005401 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5402 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005403 int tcp_header_len = tp->tcp_header_len;
5404
5405 /* Timestamp header prediction: tcp_header_len
5406 * is automatically equal to th->doff*4 due to pred_flags
5407 * match.
5408 */
5409
5410 /* Check timestamp */
5411 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005412 /* No? Slow path! */
Ilpo Järvinena4356b22008-08-23 05:12:29 -07005413 if (!tcp_parse_aligned_timestamp(tp, th))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005414 goto slow_path;
5415
Linus Torvalds1da177e2005-04-16 15:20:36 -07005416 /* If PAWS failed, check it more carefully in slow path */
5417 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5418 goto slow_path;
5419
5420 /* DO NOT update ts_recent here, if checksum fails
5421 * and timestamp was corrupted part, it will result
5422 * in a hung connection since we will drop all
5423 * future packets due to the PAWS test.
5424 */
5425 }
5426
5427 if (len <= tcp_header_len) {
5428 /* Bulk data transfer: sender */
5429 if (len == tcp_header_len) {
5430 /* Predicted packet is in window by definition.
5431 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5432 * Hence, check seq<=rcv_wup reduces to:
5433 */
5434 if (tcp_header_len ==
5435 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5436 tp->rcv_nxt == tp->rcv_wup)
5437 tcp_store_ts_recent(tp);
5438
Linus Torvalds1da177e2005-04-16 15:20:36 -07005439 /* We know that such packets are checksummed
5440 * on entry.
5441 */
5442 tcp_ack(sk, skb, 0);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005443 __kfree_skb(skb);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005444 tcp_data_snd_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005445 return 0;
5446 } else { /* Header too small */
Pavel Emelyanov63231bd2008-07-16 20:22:25 -07005447 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005448 goto discard;
5449 }
5450 } else {
5451 int eaten = 0;
Chris Leech1a2449a2006-05-23 18:05:53 -07005452 int copied_early = 0;
Eric Dumazetb081f852012-05-02 09:58:29 +00005453 bool fragstolen = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005454
Chris Leech1a2449a2006-05-23 18:05:53 -07005455 if (tp->copied_seq == tp->rcv_nxt &&
5456 len - tcp_header_len <= tp->ucopy.len) {
5457#ifdef CONFIG_NET_DMA
Jiri Kosina59ea33a2012-07-27 10:38:50 +00005458 if (tp->ucopy.task == current &&
5459 sock_owned_by_user(sk) &&
5460 tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
Chris Leech1a2449a2006-05-23 18:05:53 -07005461 copied_early = 1;
5462 eaten = 1;
5463 }
5464#endif
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005465 if (tp->ucopy.task == current &&
5466 sock_owned_by_user(sk) && !copied_early) {
Chris Leech1a2449a2006-05-23 18:05:53 -07005467 __set_current_state(TASK_RUNNING);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005468
Chris Leech1a2449a2006-05-23 18:05:53 -07005469 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
5470 eaten = 1;
5471 }
5472 if (eaten) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005473 /* Predicted packet is in window by definition.
5474 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5475 * Hence, check seq<=rcv_wup reduces to:
5476 */
5477 if (tcp_header_len ==
5478 (sizeof(struct tcphdr) +
5479 TCPOLEN_TSTAMP_ALIGNED) &&
5480 tp->rcv_nxt == tp->rcv_wup)
5481 tcp_store_ts_recent(tp);
5482
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005483 tcp_rcv_rtt_measure_ts(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005484
5485 __skb_pull(skb, tcp_header_len);
5486 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07005487 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005488 }
Chris Leech1a2449a2006-05-23 18:05:53 -07005489 if (copied_early)
5490 tcp_cleanup_rbuf(sk, skb->len);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005491 }
5492 if (!eaten) {
5493 if (tcp_checksum_complete_user(sk, skb))
5494 goto csum_error;
5495
5496 /* Predicted packet is in window by definition.
5497 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5498 * Hence, check seq<=rcv_wup reduces to:
5499 */
5500 if (tcp_header_len ==
5501 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5502 tp->rcv_nxt == tp->rcv_wup)
5503 tcp_store_ts_recent(tp);
5504
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005505 tcp_rcv_rtt_measure_ts(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005506
5507 if ((int)skb->truesize > sk->sk_forward_alloc)
5508 goto step5;
5509
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07005510 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005511
5512 /* Bulk data transfer: receiver */
Eric Dumazetb081f852012-05-02 09:58:29 +00005513 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5514 &fragstolen);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005515 }
5516
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005517 tcp_event_data_recv(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005518
5519 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5520 /* Well, only one small jumplet in fast path... */
5521 tcp_ack(sk, skb, FLAG_DATA);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005522 tcp_data_snd_check(sk);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005523 if (!inet_csk_ack_scheduled(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005524 goto no_ack;
5525 }
5526
Ali Saidi53240c22008-10-07 15:31:19 -07005527 if (!copied_early || tp->rcv_nxt != tp->rcv_wup)
5528 __tcp_ack_snd_check(sk, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005529no_ack:
Chris Leech1a2449a2006-05-23 18:05:53 -07005530#ifdef CONFIG_NET_DMA
5531 if (copied_early)
5532 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
5533 else
5534#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07005535 if (eaten)
Eric Dumazetb081f852012-05-02 09:58:29 +00005536 kfree_skb_partial(skb, fragstolen);
Eric Dumazet1d57f192012-09-17 12:51:39 +00005537 sk->sk_data_ready(sk, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005538 return 0;
5539 }
5540 }
5541
5542slow_path:
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005543 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005544 goto csum_error;
5545
5546 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07005547 * Standard slow path.
5548 */
5549
Eric Dumazet0c246042012-07-17 01:41:30 +00005550 if (!tcp_validate_incoming(sk, skb, th, 1))
5551 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005552
5553step5:
John Dykstra96e0bf42009-03-22 21:49:57 -07005554 if (th->ack && tcp_ack(sk, skb, FLAG_SLOWPATH) < 0)
5555 goto discard;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005556
Eric Dumazetbd090df2012-11-13 05:37:18 +00005557 /* ts_recent update must be made after we are sure that the packet
5558 * is in window.
5559 */
5560 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
5561
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005562 tcp_rcv_rtt_measure_ts(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005563
5564 /* Process urgent data. */
5565 tcp_urg(sk, skb, th);
5566
5567 /* step 7: process the segment text */
5568 tcp_data_queue(sk, skb);
5569
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005570 tcp_data_snd_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005571 tcp_ack_snd_check(sk);
5572 return 0;
5573
5574csum_error:
Pavel Emelyanov63231bd2008-07-16 20:22:25 -07005575 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005576
5577discard:
5578 __kfree_skb(skb);
5579 return 0;
5580}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00005581EXPORT_SYMBOL(tcp_rcv_established);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005582
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005583void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5584{
5585 struct tcp_sock *tp = tcp_sk(sk);
5586 struct inet_connection_sock *icsk = inet_csk(sk);
5587
5588 tcp_set_state(sk, TCP_ESTABLISHED);
5589
David S. Miller41063e92012-06-19 21:22:05 -07005590 if (skb != NULL) {
Eric Dumazet5d299f32012-08-06 05:09:33 +00005591 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005592 security_inet_conn_established(sk, skb);
David S. Miller41063e92012-06-19 21:22:05 -07005593 }
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005594
5595 /* Make sure socket is routed, for correct metrics. */
5596 icsk->icsk_af_ops->rebuild_header(sk);
5597
5598 tcp_init_metrics(sk);
5599
5600 tcp_init_congestion_control(sk);
5601
5602 /* Prevent spurious tcp_cwnd_restart() on first data
5603 * packet.
5604 */
5605 tp->lsndtime = tcp_time_stamp;
5606
5607 tcp_init_buffer_space(sk);
5608
5609 if (sock_flag(sk, SOCK_KEEPOPEN))
5610 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5611
5612 if (!tp->rx_opt.snd_wscale)
5613 __tcp_fast_path_on(tp, tp->snd_wnd);
5614 else
5615 tp->pred_flags = 0;
5616
5617 if (!sock_flag(sk, SOCK_DEAD)) {
5618 sk->sk_state_change(sk);
5619 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5620 }
5621}
5622
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005623static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5624 struct tcp_fastopen_cookie *cookie)
5625{
5626 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng67da22d2012-07-19 06:43:11 +00005627 struct sk_buff *data = tp->syn_data ? tcp_write_queue_head(sk) : NULL;
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005628 u16 mss = tp->rx_opt.mss_clamp;
Yuchung Chengaab48742012-07-19 06:43:10 +00005629 bool syn_drop;
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005630
5631 if (mss == tp->rx_opt.user_mss) {
5632 struct tcp_options_received opt;
5633 const u8 *hash_location;
5634
5635 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5636 tcp_clear_options(&opt);
5637 opt.user_mss = opt.mss_clamp = 0;
5638 tcp_parse_options(synack, &opt, &hash_location, 0, NULL);
5639 mss = opt.mss_clamp;
5640 }
5641
Yuchung Cheng67da22d2012-07-19 06:43:11 +00005642 if (!tp->syn_fastopen) /* Ignore an unsolicited cookie */
5643 cookie->len = -1;
5644
Yuchung Chengaab48742012-07-19 06:43:10 +00005645 /* The SYN-ACK neither has cookie nor acknowledges the data. Presumably
5646 * the remote receives only the retransmitted (regular) SYNs: either
5647 * the original SYN-data or the corresponding SYN-ACK is lost.
5648 */
5649 syn_drop = (cookie->len <= 0 && data &&
5650 inet_csk(sk)->icsk_retransmits);
5651
5652 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop);
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005653
5654 if (data) { /* Retransmit unacked data in SYN */
Yuchung Cheng93b174a2012-12-06 08:45:32 +00005655 tcp_for_write_queue_from(data, sk) {
5656 if (data == tcp_send_head(sk) ||
5657 __tcp_retransmit_skb(sk, data))
5658 break;
5659 }
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005660 tcp_rearm_rto(sk);
5661 return true;
5662 }
Yuchung Cheng6f736012012-10-19 15:14:44 +00005663 tp->syn_data_acked = tp->syn_data;
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005664 return false;
5665}
5666
Linus Torvalds1da177e2005-04-16 15:20:36 -07005667static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
Eric Dumazetcf533ea2011-10-21 05:22:42 -04005668 const struct tcphdr *th, unsigned int len)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005669{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04005670 const u8 *hash_location;
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08005671 struct inet_connection_sock *icsk = inet_csk(sk);
William Allen Simpson4957faade2009-12-02 18:25:27 +00005672 struct tcp_sock *tp = tcp_sk(sk);
William Allen Simpson4957faade2009-12-02 18:25:27 +00005673 struct tcp_cookie_values *cvp = tp->cookie_values;
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005674 struct tcp_fastopen_cookie foc = { .len = -1 };
William Allen Simpson4957faade2009-12-02 18:25:27 +00005675 int saved_clamp = tp->rx_opt.mss_clamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005676
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005677 tcp_parse_options(skb, &tp->rx_opt, &hash_location, 0, &foc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005678
5679 if (th->ack) {
5680 /* rfc793:
5681 * "If the state is SYN-SENT then
5682 * first check the ACK bit
5683 * If the ACK bit is set
5684 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5685 * a reset (unless the RST bit is set, if so drop
5686 * the segment and return)"
Linus Torvalds1da177e2005-04-16 15:20:36 -07005687 */
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005688 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5689 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005690 goto reset_and_undo;
5691
5692 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5693 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5694 tcp_time_stamp)) {
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07005695 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005696 goto reset_and_undo;
5697 }
5698
5699 /* Now ACK is acceptable.
5700 *
5701 * "If the RST bit is set
5702 * If the ACK was acceptable then signal the user "error:
5703 * connection reset", drop the segment, enter CLOSED state,
5704 * delete TCB, and return."
5705 */
5706
5707 if (th->rst) {
5708 tcp_reset(sk);
5709 goto discard;
5710 }
5711
5712 /* rfc793:
5713 * "fifth, if neither of the SYN or RST bits is set then
5714 * drop the segment and return."
5715 *
5716 * See note below!
5717 * --ANK(990513)
5718 */
5719 if (!th->syn)
5720 goto discard_and_undo;
5721
5722 /* rfc793:
5723 * "If the SYN bit is on ...
5724 * are acceptable then ...
5725 * (our SYN has been ACKed), change the connection
5726 * state to ESTABLISHED..."
5727 */
5728
5729 TCP_ECN_rcv_synack(tp, th);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005730
Razvan Ghitulete1b3a6922012-08-14 16:30:20 +03005731 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005732 tcp_ack(sk, skb, FLAG_SLOWPATH);
5733
5734 /* Ok.. it's good. Set up sequence numbers and
5735 * move to established.
5736 */
5737 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5738 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5739
5740 /* RFC1323: The window in SYN & SYN/ACK segments is
5741 * never scaled.
5742 */
5743 tp->snd_wnd = ntohs(th->window);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005744
5745 if (!tp->rx_opt.wscale_ok) {
5746 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5747 tp->window_clamp = min(tp->window_clamp, 65535U);
5748 }
5749
5750 if (tp->rx_opt.saw_tstamp) {
5751 tp->rx_opt.tstamp_ok = 1;
5752 tp->tcp_header_len =
5753 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5754 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5755 tcp_store_ts_recent(tp);
5756 } else {
5757 tp->tcp_header_len = sizeof(struct tcphdr);
5758 }
5759
Ilpo Järvinene60402d2007-08-09 15:14:46 +03005760 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5761 tcp_enable_fack(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005762
John Heffner5d424d52006-03-20 17:53:41 -08005763 tcp_mtup_init(sk);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08005764 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005765 tcp_initialize_rcv_mss(sk);
5766
5767 /* Remember, tcp_poll() does not lock socket!
5768 * Change state from SYN-SENT only after copied_seq
5769 * is initialized. */
5770 tp->copied_seq = tp->rcv_nxt;
William Allen Simpson4957faade2009-12-02 18:25:27 +00005771
5772 if (cvp != NULL &&
5773 cvp->cookie_pair_size > 0 &&
5774 tp->rx_opt.cookie_plus > 0) {
5775 int cookie_size = tp->rx_opt.cookie_plus
5776 - TCPOLEN_COOKIE_BASE;
5777 int cookie_pair_size = cookie_size
5778 + cvp->cookie_desired;
5779
5780 /* A cookie extension option was sent and returned.
5781 * Note that each incoming SYNACK replaces the
5782 * Responder cookie. The initial exchange is most
5783 * fragile, as protection against spoofing relies
5784 * entirely upon the sequence and timestamp (above).
5785 * This replacement strategy allows the correct pair to
5786 * pass through, while any others will be filtered via
5787 * Responder verification later.
5788 */
5789 if (sizeof(cvp->cookie_pair) >= cookie_pair_size) {
5790 memcpy(&cvp->cookie_pair[cvp->cookie_desired],
5791 hash_location, cookie_size);
5792 cvp->cookie_pair_size = cookie_pair_size;
5793 }
5794 }
5795
Ralf Baechlee16aa202006-12-07 00:11:33 -08005796 smp_mb();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005797
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005798 tcp_finish_connect(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005799
Yuchung Cheng67da22d2012-07-19 06:43:11 +00005800 if ((tp->syn_fastopen || tp->syn_data) &&
5801 tcp_rcv_fastopen_synack(sk, skb, &foc))
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005802 return -1;
5803
Arnaldo Carvalho de Melo295f7322005-08-09 20:11:56 -07005804 if (sk->sk_write_pending ||
5805 icsk->icsk_accept_queue.rskq_defer_accept ||
5806 icsk->icsk_ack.pingpong) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005807 /* Save one ACK. Data will be ready after
5808 * several ticks, if write_pending is set.
5809 *
5810 * It may be deleted, but with this feature tcpdumps
5811 * look so _wonderfully_ clever, that I was not able
5812 * to stand against the temptation 8) --ANK
5813 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005814 inet_csk_schedule_ack(sk);
Arnaldo Carvalho de Melo295f7322005-08-09 20:11:56 -07005815 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005816 tcp_enter_quickack_mode(sk);
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07005817 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5818 TCP_DELACK_MAX, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005819
5820discard:
5821 __kfree_skb(skb);
5822 return 0;
5823 } else {
5824 tcp_send_ack(sk);
5825 }
5826 return -1;
5827 }
5828
5829 /* No ACK in the segment */
5830
5831 if (th->rst) {
5832 /* rfc793:
5833 * "If the RST bit is set
5834 *
5835 * Otherwise (no ACK) drop the segment and return."
5836 */
5837
5838 goto discard_and_undo;
5839 }
5840
5841 /* PAWS check. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005842 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
Ilpo Järvinenc887e6d2009-03-14 14:23:03 +00005843 tcp_paws_reject(&tp->rx_opt, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005844 goto discard_and_undo;
5845
5846 if (th->syn) {
5847 /* We see SYN without ACK. It is attempt of
5848 * simultaneous connect with crossed SYNs.
5849 * Particularly, it can be connect to self.
5850 */
5851 tcp_set_state(sk, TCP_SYN_RECV);
5852
5853 if (tp->rx_opt.saw_tstamp) {
5854 tp->rx_opt.tstamp_ok = 1;
5855 tcp_store_ts_recent(tp);
5856 tp->tcp_header_len =
5857 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5858 } else {
5859 tp->tcp_header_len = sizeof(struct tcphdr);
5860 }
5861
5862 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5863 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5864
5865 /* RFC1323: The window in SYN & SYN/ACK segments is
5866 * never scaled.
5867 */
5868 tp->snd_wnd = ntohs(th->window);
5869 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5870 tp->max_window = tp->snd_wnd;
5871
5872 TCP_ECN_rcv_syn(tp, th);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005873
John Heffner5d424d52006-03-20 17:53:41 -08005874 tcp_mtup_init(sk);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08005875 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005876 tcp_initialize_rcv_mss(sk);
5877
Linus Torvalds1da177e2005-04-16 15:20:36 -07005878 tcp_send_synack(sk);
5879#if 0
5880 /* Note, we could accept data and URG from this segment.
Jerry Chu168a8f52012-08-31 12:29:13 +00005881 * There are no obstacles to make this (except that we must
5882 * either change tcp_recvmsg() to prevent it from returning data
5883 * before 3WHS completes per RFC793, or employ TCP Fast Open).
Linus Torvalds1da177e2005-04-16 15:20:36 -07005884 *
5885 * However, if we ignore data in ACKless segments sometimes,
5886 * we have no reasons to accept it sometimes.
5887 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5888 * is not flawless. So, discard packet for sanity.
5889 * Uncomment this return to process the data.
5890 */
5891 return -1;
5892#else
5893 goto discard;
5894#endif
5895 }
5896 /* "fifth, if neither of the SYN or RST bits is set then
5897 * drop the segment and return."
5898 */
5899
5900discard_and_undo:
5901 tcp_clear_options(&tp->rx_opt);
5902 tp->rx_opt.mss_clamp = saved_clamp;
5903 goto discard;
5904
5905reset_and_undo:
5906 tcp_clear_options(&tp->rx_opt);
5907 tp->rx_opt.mss_clamp = saved_clamp;
5908 return 1;
5909}
5910
Linus Torvalds1da177e2005-04-16 15:20:36 -07005911/*
5912 * This function implements the receiving procedure of RFC 793 for
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005913 * all states except ESTABLISHED and TIME_WAIT.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005914 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5915 * address independent.
5916 */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005917
Linus Torvalds1da177e2005-04-16 15:20:36 -07005918int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
Eric Dumazetcf533ea2011-10-21 05:22:42 -04005919 const struct tcphdr *th, unsigned int len)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005920{
5921 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -08005922 struct inet_connection_sock *icsk = inet_csk(sk);
Jerry Chu168a8f52012-08-31 12:29:13 +00005923 struct request_sock *req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005924 int queued = 0;
5925
5926 tp->rx_opt.saw_tstamp = 0;
5927
5928 switch (sk->sk_state) {
5929 case TCP_CLOSE:
5930 goto discard;
5931
5932 case TCP_LISTEN:
Stephen Hemminger2de979b2007-03-08 20:45:19 -08005933 if (th->ack)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005934 return 1;
5935
Stephen Hemminger2de979b2007-03-08 20:45:19 -08005936 if (th->rst)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005937 goto discard;
5938
Stephen Hemminger2de979b2007-03-08 20:45:19 -08005939 if (th->syn) {
Eric Dumazetfdf5af02011-12-02 23:41:42 +00005940 if (th->fin)
5941 goto discard;
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -08005942 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005943 return 1;
5944
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005945 /* Now we have several options: In theory there is
5946 * nothing else in the frame. KA9Q has an option to
Linus Torvalds1da177e2005-04-16 15:20:36 -07005947 * send data with the syn, BSD accepts data with the
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005948 * syn up to the [to be] advertised window and
5949 * Solaris 2.1 gives you a protocol error. For now
5950 * we just ignore it, that fits the spec precisely
Linus Torvalds1da177e2005-04-16 15:20:36 -07005951 * and avoids incompatibilities. It would be nice in
5952 * future to drop through and process the data.
5953 *
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005954 * Now that TTCP is starting to be used we ought to
Linus Torvalds1da177e2005-04-16 15:20:36 -07005955 * queue this data.
5956 * But, this leaves one open to an easy denial of
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005957 * service attack, and SYN cookies can't defend
Linus Torvalds1da177e2005-04-16 15:20:36 -07005958 * against this problem. So, we drop the data
Masayuki Nakagawafb7e2392007-01-23 20:15:06 -08005959 * in the interest of security over speed unless
5960 * it's still in use.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005961 */
Masayuki Nakagawafb7e2392007-01-23 20:15:06 -08005962 kfree_skb(skb);
5963 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005964 }
5965 goto discard;
5966
5967 case TCP_SYN_SENT:
Linus Torvalds1da177e2005-04-16 15:20:36 -07005968 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5969 if (queued >= 0)
5970 return queued;
5971
5972 /* Do step6 onward by hand. */
5973 tcp_urg(sk, skb, th);
5974 __kfree_skb(skb);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005975 tcp_data_snd_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005976 return 0;
5977 }
5978
Jerry Chu168a8f52012-08-31 12:29:13 +00005979 req = tp->fastopen_rsk;
5980 if (req != NULL) {
Jerry Chu37561f62012-10-22 11:26:36 +00005981 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
Jerry Chu168a8f52012-08-31 12:29:13 +00005982 sk->sk_state != TCP_FIN_WAIT1);
5983
5984 if (tcp_check_req(sk, skb, req, NULL, true) == NULL)
5985 goto discard;
Neal Cardwelle69bebd2012-09-22 04:18:57 +00005986 }
5987 if (!tcp_validate_incoming(sk, skb, th, 0))
Eric Dumazet0c246042012-07-17 01:41:30 +00005988 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005989
5990 /* step 5: check the ACK field */
5991 if (th->ack) {
John Dykstra96e0bf42009-03-22 21:49:57 -07005992 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH) > 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005993
Stephen Hemminger2de979b2007-03-08 20:45:19 -08005994 switch (sk->sk_state) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005995 case TCP_SYN_RECV:
5996 if (acceptable) {
Jerry Chu168a8f52012-08-31 12:29:13 +00005997 /* Once we leave TCP_SYN_RECV, we no longer
5998 * need req so release it.
5999 */
6000 if (req) {
Neal Cardwell016818d2012-09-22 04:18:55 +00006001 tcp_synack_rtt_meas(sk, req);
Eric Dumazete6c022a2012-10-27 23:16:46 +00006002 tp->total_retrans = req->num_retrans;
Neal Cardwell30099b22012-09-22 17:03:47 +00006003
Jerry Chu168a8f52012-08-31 12:29:13 +00006004 reqsk_fastopen_remove(sk, req, false);
6005 } else {
6006 /* Make sure socket is routed, for
6007 * correct metrics.
6008 */
6009 icsk->icsk_af_ops->rebuild_header(sk);
6010 tcp_init_congestion_control(sk);
6011
6012 tcp_mtup_init(sk);
6013 tcp_init_buffer_space(sk);
6014 tp->copied_seq = tp->rcv_nxt;
6015 }
Ralf Baechlee16aa202006-12-07 00:11:33 -08006016 smp_mb();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006017 tcp_set_state(sk, TCP_ESTABLISHED);
6018 sk->sk_state_change(sk);
6019
6020 /* Note, that this wakeup is only for marginal
6021 * crossed SYN case. Passively open sockets
6022 * are not waked up, because sk->sk_sleep ==
6023 * NULL and sk->sk_socket == NULL.
6024 */
Pavel Emelyanov8d8ad9d2007-11-26 20:10:50 +08006025 if (sk->sk_socket)
6026 sk_wake_async(sk,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08006027 SOCK_WAKE_IO, POLL_OUT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006028
6029 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
6030 tp->snd_wnd = ntohs(th->window) <<
6031 tp->rx_opt.snd_wscale;
Hantzis Fotisee7537b2009-03-02 22:42:02 -08006032 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006033
Linus Torvalds1da177e2005-04-16 15:20:36 -07006034 if (tp->rx_opt.tstamp_ok)
6035 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6036
Jerry Chu168a8f52012-08-31 12:29:13 +00006037 if (req) {
6038 /* Re-arm the timer because data may
6039 * have been sent out. This is similar
6040 * to the regular data transmission case
6041 * when new data has just been ack'ed.
6042 *
6043 * (TFO) - we could try to be more
6044 * aggressive and retranmitting any data
6045 * sooner based on when they were sent
6046 * out.
6047 */
6048 tcp_rearm_rto(sk);
6049 } else
6050 tcp_init_metrics(sk);
Stephen Hemminger317a76f2005-06-23 12:19:55 -07006051
Linus Torvalds1da177e2005-04-16 15:20:36 -07006052 /* Prevent spurious tcp_cwnd_restart() on
6053 * first data packet.
6054 */
6055 tp->lsndtime = tcp_time_stamp;
6056
6057 tcp_initialize_rcv_mss(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006058 tcp_fast_path_on(tp);
6059 } else {
6060 return 1;
6061 }
6062 break;
6063
6064 case TCP_FIN_WAIT1:
Jerry Chu168a8f52012-08-31 12:29:13 +00006065 /* If we enter the TCP_FIN_WAIT1 state and we are a
6066 * Fast Open socket and this is the first acceptable
6067 * ACK we have received, this would have acknowledged
6068 * our SYNACK so stop the SYNACK timer.
6069 */
Jerry Chu37561f62012-10-22 11:26:36 +00006070 if (req != NULL) {
6071 /* Return RST if ack_seq is invalid.
6072 * Note that RFC793 only says to generate a
6073 * DUPACK for it but for TCP Fast Open it seems
6074 * better to treat this case like TCP_SYN_RECV
6075 * above.
6076 */
6077 if (!acceptable)
6078 return 1;
Jerry Chu168a8f52012-08-31 12:29:13 +00006079 /* We no longer need the request sock. */
6080 reqsk_fastopen_remove(sk, req, false);
6081 tcp_rearm_rto(sk);
6082 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07006083 if (tp->snd_una == tp->write_seq) {
David S. Miller5110effe2012-07-02 02:21:03 -07006084 struct dst_entry *dst;
6085
Linus Torvalds1da177e2005-04-16 15:20:36 -07006086 tcp_set_state(sk, TCP_FIN_WAIT2);
6087 sk->sk_shutdown |= SEND_SHUTDOWN;
David S. Miller5110effe2012-07-02 02:21:03 -07006088
6089 dst = __sk_dst_get(sk);
6090 if (dst)
6091 dst_confirm(dst);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006092
6093 if (!sock_flag(sk, SOCK_DEAD))
6094 /* Wake up lingering close() */
6095 sk->sk_state_change(sk);
6096 else {
6097 int tmo;
6098
6099 if (tp->linger2 < 0 ||
6100 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6101 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
6102 tcp_done(sk);
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07006103 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006104 return 1;
6105 }
6106
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07006107 tmo = tcp_fin_time(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006108 if (tmo > TCP_TIMEWAIT_LEN) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07006109 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006110 } else if (th->fin || sock_owned_by_user(sk)) {
6111 /* Bad case. We could lose such FIN otherwise.
6112 * It is not a big problem, but it looks confusing
6113 * and not so rare event. We still can lose it now,
6114 * if it spins in bh_lock_sock(), but it is really
6115 * marginal case.
6116 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07006117 inet_csk_reset_keepalive_timer(sk, tmo);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006118 } else {
6119 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6120 goto discard;
6121 }
6122 }
6123 }
6124 break;
6125
6126 case TCP_CLOSING:
6127 if (tp->snd_una == tp->write_seq) {
6128 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6129 goto discard;
6130 }
6131 break;
6132
6133 case TCP_LAST_ACK:
6134 if (tp->snd_una == tp->write_seq) {
6135 tcp_update_metrics(sk);
6136 tcp_done(sk);
6137 goto discard;
6138 }
6139 break;
6140 }
6141 } else
6142 goto discard;
6143
Eric Dumazetbd090df2012-11-13 05:37:18 +00006144 /* ts_recent update must be made after we are sure that the packet
6145 * is in window.
6146 */
6147 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
6148
Linus Torvalds1da177e2005-04-16 15:20:36 -07006149 /* step 6: check the URG bit */
6150 tcp_urg(sk, skb, th);
6151
6152 /* step 7: process the segment text */
6153 switch (sk->sk_state) {
6154 case TCP_CLOSE_WAIT:
6155 case TCP_CLOSING:
6156 case TCP_LAST_ACK:
6157 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
6158 break;
6159 case TCP_FIN_WAIT1:
6160 case TCP_FIN_WAIT2:
6161 /* RFC 793 says to queue data in these states,
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09006162 * RFC 1122 says we MUST send a reset.
Linus Torvalds1da177e2005-04-16 15:20:36 -07006163 * BSD 4.4 also does reset.
6164 */
6165 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6166 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6167 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07006168 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006169 tcp_reset(sk);
6170 return 1;
6171 }
6172 }
6173 /* Fall through */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09006174 case TCP_ESTABLISHED:
Linus Torvalds1da177e2005-04-16 15:20:36 -07006175 tcp_data_queue(sk, skb);
6176 queued = 1;
6177 break;
6178 }
6179
6180 /* tcp_data could move socket to TIME-WAIT */
6181 if (sk->sk_state != TCP_CLOSE) {
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07006182 tcp_data_snd_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006183 tcp_ack_snd_check(sk);
6184 }
6185
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09006186 if (!queued) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07006187discard:
6188 __kfree_skb(skb);
6189 }
6190 return 0;
6191}
Linus Torvalds1da177e2005-04-16 15:20:36 -07006192EXPORT_SYMBOL(tcp_rcv_state_process);