blob: 65cf90e063d5adcc98f15b044e35b20e1668352a [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);
Brian Haleyab32ea52006-09-22 14:15:41 -070084int sysctl_tcp_dsack __read_mostly = 1;
85int sysctl_tcp_app_win __read_mostly = 31;
Eric Dumazetb49960a2012-05-02 02:28:41 +000086int sysctl_tcp_adv_win_scale __read_mostly = 1;
Eric Dumazet4bc2f182010-07-09 21:22:10 +000087EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
Linus Torvalds1da177e2005-04-16 15:20:36 -070088
Eric Dumazet282f23c2012-07-17 10:13:05 +020089/* rfc5961 challenge ack rate limiting */
90int sysctl_tcp_challenge_ack_limit = 100;
91
Brian Haleyab32ea52006-09-22 14:15:41 -070092int sysctl_tcp_stdurg __read_mostly;
93int sysctl_tcp_rfc1337 __read_mostly;
94int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
Ilpo Järvinenc96fd3d2007-09-20 11:36:37 -070095int sysctl_tcp_frto __read_mostly = 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -070096
Andreas Petlund7e380172010-02-18 04:48:19 +000097int sysctl_tcp_thin_dupack __read_mostly;
98
Brian Haleyab32ea52006-09-22 14:15:41 -070099int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +0000100int sysctl_tcp_early_retrans __read_mostly = 3;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700101
Linus Torvalds1da177e2005-04-16 15:20:36 -0700102#define FLAG_DATA 0x01 /* Incoming frame contained data. */
103#define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
104#define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
105#define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
106#define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
107#define FLAG_DATA_SACKED 0x20 /* New SACK. */
108#define FLAG_ECE 0x40 /* ECE in this ACK */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700109#define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
Yuchung Chenge33099f2013-03-20 13:33:00 +0000110#define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
Ilpo Järvinen2e605292007-08-02 19:46:58 -0700111#define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
Ryousei Takano564262c12007-10-25 23:03:52 -0700112#define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
Ilpo Järvinencadbd032007-12-31 04:49:21 -0800113#define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
Eric Dumazet12fb3dd2013-04-19 07:19:48 +0000114#define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700115
116#define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
117#define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
118#define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
119#define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
120
Linus Torvalds1da177e2005-04-16 15:20:36 -0700121#define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700122#define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700123
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900124/* Adapt the MSS value used to make delayed ack decision to the
Linus Torvalds1da177e2005-04-16 15:20:36 -0700125 * real world.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900126 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800127static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700128{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700129 struct inet_connection_sock *icsk = inet_csk(sk);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900130 const unsigned int lss = icsk->icsk_ack.last_seg_size;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700131 unsigned int len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700132
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900133 icsk->icsk_ack.last_seg_size = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700134
135 /* skb->len may jitter because of SACKs, even if peer
136 * sends good full-sized frames.
137 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800138 len = skb_shinfo(skb)->gso_size ? : skb->len;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700139 if (len >= icsk->icsk_ack.rcv_mss) {
140 icsk->icsk_ack.rcv_mss = len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700141 } else {
142 /* Otherwise, we make more careful check taking into account,
143 * that SACKs block is variable.
144 *
145 * "len" is invariant segment length, including TCP header.
146 */
Arnaldo Carvalho de Melo9c702202007-04-25 18:04:18 -0700147 len += skb->data - skb_transport_header(skb);
William Allen Simpsonbee7ca92009-11-10 09:51:18 +0000148 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -0700149 /* If PSH is not set, packet should be
150 * full sized, provided peer TCP is not badly broken.
151 * This observation (if it is correct 8)) allows
152 * to handle super-low mtu links fairly.
153 */
154 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -0700155 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700156 /* Subtract also invariant (if peer is RFC compliant),
157 * tcp header plus fixed timestamp option length.
158 * Resulting "len" is MSS free of SACK jitter.
159 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700160 len -= tcp_sk(sk)->tcp_header_len;
161 icsk->icsk_ack.last_seg_size = len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700162 if (len == lss) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700163 icsk->icsk_ack.rcv_mss = len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700164 return;
165 }
166 }
Alexey Kuznetsov1ef96962006-09-19 12:52:50 -0700167 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
168 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700169 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700170 }
171}
172
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700173static void tcp_incr_quickack(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700174{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700175 struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazet95c96172012-04-15 05:58:06 +0000176 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700177
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800178 if (quickacks == 0)
179 quickacks = 2;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700180 if (quickacks > icsk->icsk_ack.quick)
181 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700182}
183
stephen hemminger1b9f4092010-09-28 19:30:14 +0000184static void tcp_enter_quickack_mode(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700185{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700186 struct inet_connection_sock *icsk = inet_csk(sk);
187 tcp_incr_quickack(sk);
188 icsk->icsk_ack.pingpong = 0;
189 icsk->icsk_ack.ato = TCP_ATO_MIN;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700190}
191
192/* Send ACKs quickly, if "quick" count is not exhausted
193 * and the session is not interactive.
194 */
195
Eric Dumazeta2a385d2012-05-16 23:15:34 +0000196static inline bool tcp_in_quickack_mode(const struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700197{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700198 const struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazeta2a385d2012-05-16 23:15:34 +0000199
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700200 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700201}
202
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700203static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
204{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800205 if (tp->ecn_flags & TCP_ECN_OK)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700206 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
207}
208
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400209static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700210{
211 if (tcp_hdr(skb)->cwr)
212 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
213}
214
215static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
216{
217 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
218}
219
Eric Dumazet7a269ff2011-09-22 20:02:19 +0000220static inline void TCP_ECN_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700221{
Eric Dumazet7a269ff2011-09-22 20:02:19 +0000222 if (!(tp->ecn_flags & TCP_ECN_OK))
223 return;
224
Eric Dumazetb82d1bb2011-09-27 02:20:08 -0400225 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
Eric Dumazet7a269ff2011-09-22 20:02:19 +0000226 case INET_ECN_NOT_ECT:
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700227 /* Funny extension: if ECT is not set on a segment,
Eric Dumazet7a269ff2011-09-22 20:02:19 +0000228 * and we already seen ECT on a previous segment,
229 * it is probably a retransmit.
230 */
231 if (tp->ecn_flags & TCP_ECN_SEEN)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700232 tcp_enter_quickack_mode((struct sock *)tp);
Eric Dumazet7a269ff2011-09-22 20:02:19 +0000233 break;
234 case INET_ECN_CE:
Eric Dumazetaae06bf2012-08-06 11:04:43 +0000235 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
236 /* Better not delay acks, sender can have a very low cwnd */
237 tcp_enter_quickack_mode((struct sock *)tp);
238 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
239 }
Eric Dumazet7a269ff2011-09-22 20:02:19 +0000240 /* fallinto */
241 default:
242 tp->ecn_flags |= TCP_ECN_SEEN;
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700243 }
244}
245
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400246static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700247{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800248 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700249 tp->ecn_flags &= ~TCP_ECN_OK;
250}
251
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400252static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700253{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800254 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700255 tp->ecn_flags &= ~TCP_ECN_OK;
256}
257
Eric Dumazeta2a385d2012-05-16 23:15:34 +0000258static bool TCP_ECN_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700259{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800260 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
Eric Dumazeta2a385d2012-05-16 23:15:34 +0000261 return true;
262 return false;
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700263}
264
Linus Torvalds1da177e2005-04-16 15:20:36 -0700265/* Buffer size and advertised window tuning.
266 *
267 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
268 */
269
Eric Dumazet6ae70532013-10-01 10:23:44 -0700270static void tcp_sndbuf_expand(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700271{
Eric Dumazet6ae70532013-10-01 10:23:44 -0700272 const struct tcp_sock *tp = tcp_sk(sk);
273 int sndmem, per_mss;
274 u32 nr_segs;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700275
Eric Dumazet6ae70532013-10-01 10:23:44 -0700276 /* Worst case is non GSO/TSO : each frame consumes one skb
277 * and skb->head is kmalloced using power of two area of memory
278 */
279 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
280 MAX_TCP_HEADER +
281 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
282
283 per_mss = roundup_pow_of_two(per_mss) +
284 SKB_DATA_ALIGN(sizeof(struct sk_buff));
285
286 nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
287 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
288
289 /* Fast Recovery (RFC 5681 3.2) :
290 * Cubic needs 1.7 factor, rounded to 2 to include
291 * extra cushion (application might react slowly to POLLOUT)
292 */
293 sndmem = 2 * nr_segs * per_mss;
294
Eric Dumazet06a59ec2011-10-13 18:24:42 +0000295 if (sk->sk_sndbuf < sndmem)
296 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700297}
298
299/* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
300 *
301 * All tcp_full_space() is split to two parts: "network" buffer, allocated
302 * forward and advertised in receiver window (tp->rcv_wnd) and
303 * "application buffer", required to isolate scheduling/application
304 * latencies from network.
305 * window_clamp is maximal advertised window. It can be less than
306 * tcp_full_space(), in this case tcp_full_space() - window_clamp
307 * is reserved for "application" buffer. The less window_clamp is
308 * the smoother our behaviour from viewpoint of network, but the lower
309 * throughput and the higher sensitivity of the connection to losses. 8)
310 *
311 * rcv_ssthresh is more strict window_clamp used at "slow start"
312 * phase to predict further behaviour of this connection.
313 * It is used for two goals:
314 * - to enforce header prediction at sender, even when application
315 * requires some significant "application buffer". It is check #1.
316 * - to prevent pruning of receive queue because of misprediction
317 * of receiver window. Check #2.
318 *
319 * The scheme does not work when sender sends good segments opening
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800320 * window and then starts to feed us spaghetti. But it should work
Linus Torvalds1da177e2005-04-16 15:20:36 -0700321 * in common situations. Otherwise, we have to rely on queue collapsing.
322 */
323
324/* Slow part of check#2. */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700325static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700326{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700327 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700328 /* Optimize this! */
Eric Dumazetdfd4f0a2007-12-21 06:07:53 -0800329 int truesize = tcp_win_from_space(skb->truesize) >> 1;
330 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700331
332 while (tp->rcv_ssthresh <= window) {
333 if (truesize <= skb->len)
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700334 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700335
336 truesize >>= 1;
337 window >>= 1;
338 }
339 return 0;
340}
341
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400342static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700343{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700344 struct tcp_sock *tp = tcp_sk(sk);
345
Linus Torvalds1da177e2005-04-16 15:20:36 -0700346 /* Check #1 */
347 if (tp->rcv_ssthresh < tp->window_clamp &&
348 (int)tp->rcv_ssthresh < tcp_space(sk) &&
Glauber Costa180d8cd2011-12-11 21:47:02 +0000349 !sk_under_memory_pressure(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700350 int incr;
351
352 /* Check #2. Increase window, if skb with such overhead
353 * will fit to rcvbuf in future.
354 */
355 if (tcp_win_from_space(skb->truesize) <= skb->len)
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800356 incr = 2 * tp->advmss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700357 else
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700358 incr = __tcp_grow_window(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700359
360 if (incr) {
Eric Dumazet4d846f02012-04-16 23:28:07 +0000361 incr = max_t(int, incr, 2 * skb->len);
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800362 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
363 tp->window_clamp);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700364 inet_csk(sk)->icsk_ack.quick |= 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700365 }
366 }
367}
368
369/* 3. Tuning rcvbuf, when connection enters established state. */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700370static void tcp_fixup_rcvbuf(struct sock *sk)
371{
Eric Dumazete9266a02011-10-20 16:53:56 -0400372 u32 mss = tcp_sk(sk)->advmss;
Eric Dumazete9266a02011-10-20 16:53:56 -0400373 int rcvmem;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700374
Yuchung Cheng85f16522013-06-11 15:35:32 -0700375 rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) *
376 tcp_default_init_rwnd(mss);
Eric Dumazete9266a02011-10-20 16:53:56 -0400377
Eric Dumazetb0983d32013-09-20 13:56:58 -0700378 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
379 * Allow enough cushion so that sender is not limited by our window
380 */
381 if (sysctl_tcp_moderate_rcvbuf)
382 rcvmem <<= 2;
383
Eric Dumazete9266a02011-10-20 16:53:56 -0400384 if (sk->sk_rcvbuf < rcvmem)
385 sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700386}
387
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800388/* 4. Try to fixup all. It is made immediately after connection enters
Linus Torvalds1da177e2005-04-16 15:20:36 -0700389 * established state.
390 */
Jerry Chu10467162012-08-31 12:29:11 +0000391void tcp_init_buffer_space(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700392{
393 struct tcp_sock *tp = tcp_sk(sk);
394 int maxwin;
395
396 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
397 tcp_fixup_rcvbuf(sk);
398 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
Eric Dumazet6ae70532013-10-01 10:23:44 -0700399 tcp_sndbuf_expand(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700400
401 tp->rcvq_space.space = tp->rcv_wnd;
Eric Dumazetb0983d32013-09-20 13:56:58 -0700402 tp->rcvq_space.time = tcp_time_stamp;
403 tp->rcvq_space.seq = tp->copied_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700404
405 maxwin = tcp_full_space(sk);
406
407 if (tp->window_clamp >= maxwin) {
408 tp->window_clamp = maxwin;
409
410 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
411 tp->window_clamp = max(maxwin -
412 (maxwin >> sysctl_tcp_app_win),
413 4 * tp->advmss);
414 }
415
416 /* Force reservation of one segment. */
417 if (sysctl_tcp_app_win &&
418 tp->window_clamp > 2 * tp->advmss &&
419 tp->window_clamp + tp->advmss > maxwin)
420 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
421
422 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
423 tp->snd_cwnd_stamp = tcp_time_stamp;
424}
425
Linus Torvalds1da177e2005-04-16 15:20:36 -0700426/* 5. Recalculate window clamp after socket hit its memory bounds. */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700427static void tcp_clamp_window(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700428{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700429 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300430 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700431
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300432 icsk->icsk_ack.quick = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700433
John Heffner326f36e2005-11-10 17:11:48 -0800434 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
435 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
Glauber Costa180d8cd2011-12-11 21:47:02 +0000436 !sk_under_memory_pressure(sk) &&
437 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
John Heffner326f36e2005-11-10 17:11:48 -0800438 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
439 sysctl_tcp_rmem[2]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700440 }
John Heffner326f36e2005-11-10 17:11:48 -0800441 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800442 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700443}
444
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800445/* Initialize RCV_MSS value.
446 * RCV_MSS is an our guess about MSS used by the peer.
447 * We haven't any direct information about the MSS.
448 * It's better to underestimate the RCV_MSS rather than overestimate.
449 * Overestimations make us ACKing less frequently than needed.
450 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
451 */
452void tcp_initialize_rcv_mss(struct sock *sk)
453{
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400454 const struct tcp_sock *tp = tcp_sk(sk);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800455 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
456
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800457 hint = min(hint, tp->rcv_wnd / 2);
William Allen Simpsonbee7ca92009-11-10 09:51:18 +0000458 hint = min(hint, TCP_MSS_DEFAULT);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800459 hint = max(hint, TCP_MIN_MSS);
460
461 inet_csk(sk)->icsk_ack.rcv_mss = hint;
462}
Eric Dumazet4bc2f182010-07-09 21:22:10 +0000463EXPORT_SYMBOL(tcp_initialize_rcv_mss);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800464
Linus Torvalds1da177e2005-04-16 15:20:36 -0700465/* Receiver "autotuning" code.
466 *
467 * The algorithm for RTT estimation w/o timestamps is based on
468 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
Justin P. Mattock631dd1a2010-10-18 11:03:14 +0200469 * <http://public.lanl.gov/radiant/pubs.html#DRS>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700470 *
471 * More detail on this code can be found at
Justin P. Mattock631dd1a2010-10-18 11:03:14 +0200472 * <http://staff.psc.edu/jheffner/>,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700473 * though this reference is out of date. A new paper
474 * is pending.
475 */
476static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
477{
478 u32 new_sample = tp->rcv_rtt_est.rtt;
479 long m = sample;
480
481 if (m == 0)
482 m = 1;
483
484 if (new_sample != 0) {
485 /* If we sample in larger samples in the non-timestamp
486 * case, we could grossly overestimate the RTT especially
487 * with chatty applications or bulk transfer apps which
488 * are stalled on filesystem I/O.
489 *
490 * Also, since we are only going for a minimum in the
Stephen Hemminger31f34262005-11-15 15:17:10 -0800491 * non-timestamp case, we do not smooth things out
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800492 * else with timestamps disabled convergence takes too
Linus Torvalds1da177e2005-04-16 15:20:36 -0700493 * long.
494 */
495 if (!win_dep) {
496 m -= (new_sample >> 3);
497 new_sample += m;
Neal Cardwell18a223e2012-04-10 07:59:20 +0000498 } else {
499 m <<= 3;
500 if (m < new_sample)
501 new_sample = m;
502 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700503 } else {
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800504 /* No previous measure. */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700505 new_sample = m << 3;
506 }
507
508 if (tp->rcv_rtt_est.rtt != new_sample)
509 tp->rcv_rtt_est.rtt = new_sample;
510}
511
512static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
513{
514 if (tp->rcv_rtt_est.time == 0)
515 goto new_measure;
516 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
517 return;
Neal Cardwell651913c2012-04-27 11:29:37 -0400518 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700519
520new_measure:
521 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
522 tp->rcv_rtt_est.time = tcp_time_stamp;
523}
524
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800525static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
526 const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700527{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700528 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700529 if (tp->rx_opt.rcv_tsecr &&
530 (TCP_SKB_CB(skb)->end_seq -
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700531 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700532 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
533}
534
535/*
536 * This function should be called every time data is copied to user space.
537 * It calculates the appropriate TCP receive buffer space.
538 */
539void tcp_rcv_space_adjust(struct sock *sk)
540{
541 struct tcp_sock *tp = tcp_sk(sk);
542 int time;
Eric Dumazetb0983d32013-09-20 13:56:58 -0700543 int copied;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900544
Linus Torvalds1da177e2005-04-16 15:20:36 -0700545 time = tcp_time_stamp - tp->rcvq_space.time;
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800546 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700547 return;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900548
Eric Dumazetb0983d32013-09-20 13:56:58 -0700549 /* Number of bytes copied to user in last RTT */
550 copied = tp->copied_seq - tp->rcvq_space.seq;
551 if (copied <= tp->rcvq_space.space)
552 goto new_measure;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700553
Eric Dumazetb0983d32013-09-20 13:56:58 -0700554 /* A bit of theory :
555 * copied = bytes received in previous RTT, our base window
556 * To cope with packet losses, we need a 2x factor
557 * To cope with slow start, and sender growing its cwin by 100 %
558 * every RTT, we need a 4x factor, because the ACK we are sending
559 * now is for the next RTT, not the current one :
560 * <prev RTT . ><current RTT .. ><next RTT .... >
561 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700562
Eric Dumazetb0983d32013-09-20 13:56:58 -0700563 if (sysctl_tcp_moderate_rcvbuf &&
564 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
565 int rcvwin, rcvmem, rcvbuf;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700566
Eric Dumazetb0983d32013-09-20 13:56:58 -0700567 /* minimal window to cope with packet losses, assuming
568 * steady state. Add some cushion because of small variations.
569 */
570 rcvwin = (copied << 1) + 16 * tp->advmss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700571
Eric Dumazetb0983d32013-09-20 13:56:58 -0700572 /* If rate increased by 25%,
573 * assume slow start, rcvwin = 3 * copied
574 * If rate increased by 50%,
575 * assume sender can use 2x growth, rcvwin = 4 * copied
576 */
577 if (copied >=
578 tp->rcvq_space.space + (tp->rcvq_space.space >> 2)) {
579 if (copied >=
580 tp->rcvq_space.space + (tp->rcvq_space.space >> 1))
581 rcvwin <<= 1;
582 else
583 rcvwin += (rcvwin >> 1);
584 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700585
Eric Dumazetb0983d32013-09-20 13:56:58 -0700586 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
587 while (tcp_win_from_space(rcvmem) < tp->advmss)
588 rcvmem += 128;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700589
Eric Dumazetb0983d32013-09-20 13:56:58 -0700590 rcvbuf = min(rcvwin / tp->advmss * rcvmem, sysctl_tcp_rmem[2]);
591 if (rcvbuf > sk->sk_rcvbuf) {
592 sk->sk_rcvbuf = rcvbuf;
593
594 /* Make the window clamp follow along. */
595 tp->window_clamp = rcvwin;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700596 }
597 }
Eric Dumazetb0983d32013-09-20 13:56:58 -0700598 tp->rcvq_space.space = copied;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900599
Linus Torvalds1da177e2005-04-16 15:20:36 -0700600new_measure:
601 tp->rcvq_space.seq = tp->copied_seq;
602 tp->rcvq_space.time = tcp_time_stamp;
603}
604
605/* There is something which you must keep in mind when you analyze the
606 * behavior of the tp->ato delayed ack timeout interval. When a
607 * connection starts up, we want to ack as quickly as possible. The
608 * problem is that "good" TCP's do slow start at the beginning of data
609 * transmission. The means that until we send the first few ACK's the
610 * sender will sit on his end and only queue most of his data, because
611 * he can only send snd_cwnd unacked packets at any given time. For
612 * each ACK we send, he increments snd_cwnd and transmits more of his
613 * queue. -DaveM
614 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700615static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700616{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700617 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700618 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700619 u32 now;
620
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700621 inet_csk_schedule_ack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700622
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700623 tcp_measure_rcv_mss(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700624
625 tcp_rcv_rtt_measure(tp);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900626
Linus Torvalds1da177e2005-04-16 15:20:36 -0700627 now = tcp_time_stamp;
628
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700629 if (!icsk->icsk_ack.ato) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700630 /* The _first_ data packet received, initialize
631 * delayed ACK engine.
632 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700633 tcp_incr_quickack(sk);
634 icsk->icsk_ack.ato = TCP_ATO_MIN;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700635 } else {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700636 int m = now - icsk->icsk_ack.lrcvtime;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700637
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800638 if (m <= TCP_ATO_MIN / 2) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700639 /* The fastest case is the first. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700640 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
641 } else if (m < icsk->icsk_ack.ato) {
642 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
643 if (icsk->icsk_ack.ato > icsk->icsk_rto)
644 icsk->icsk_ack.ato = icsk->icsk_rto;
645 } else if (m > icsk->icsk_rto) {
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800646 /* Too long gap. Apparently sender failed to
Linus Torvalds1da177e2005-04-16 15:20:36 -0700647 * restart window, so that we send ACKs quickly.
648 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700649 tcp_incr_quickack(sk);
Hideo Aoki3ab224b2007-12-31 00:11:19 -0800650 sk_mem_reclaim(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700651 }
652 }
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700653 icsk->icsk_ack.lrcvtime = now;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700654
655 TCP_ECN_check_ce(tp, skb);
656
657 if (skb->len >= 128)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700658 tcp_grow_window(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700659}
660
Linus Torvalds1da177e2005-04-16 15:20:36 -0700661/* Called to compute a smoothed rtt estimate. The data fed to this
662 * routine either comes from timestamps, or from segments that were
663 * known _not_ to have been retransmitted [see Karn/Partridge
664 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
665 * piece by Van Jacobson.
666 * NOTE: the next three routines used to be one big routine.
667 * To save cycles in the RFC 1323 implementation it was better to break
668 * it up into three procedures. -- erics
669 */
Stephen Hemminger2d2abba2005-11-10 16:56:12 -0800670static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700671{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300672 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700673 long m = mrtt; /* RTT */
674
Linus Torvalds1da177e2005-04-16 15:20:36 -0700675 /* The following amusing code comes from Jacobson's
676 * article in SIGCOMM '88. Note that rtt and mdev
677 * are scaled versions of rtt and mean deviation.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900678 * This is designed to be as fast as possible
Linus Torvalds1da177e2005-04-16 15:20:36 -0700679 * m stands for "measurement".
680 *
681 * On a 1990 paper the rto value is changed to:
682 * RTO = rtt + 4 * mdev
683 *
684 * Funny. This algorithm seems to be very broken.
685 * These formulae increase RTO, when it should be decreased, increase
Stephen Hemminger31f34262005-11-15 15:17:10 -0800686 * too slowly, when it should be increased quickly, decrease too quickly
Linus Torvalds1da177e2005-04-16 15:20:36 -0700687 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
688 * does not matter how to _calculate_ it. Seems, it was trap
689 * that VJ failed to avoid. 8)
690 */
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800691 if (m == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700692 m = 1;
693 if (tp->srtt != 0) {
694 m -= (tp->srtt >> 3); /* m is now error in rtt est */
695 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
696 if (m < 0) {
697 m = -m; /* m is now abs(error) */
698 m -= (tp->mdev >> 2); /* similar update on mdev */
699 /* This is similar to one of Eifel findings.
700 * Eifel blocks mdev updates when rtt decreases.
701 * This solution is a bit different: we use finer gain
702 * for mdev in this case (alpha*beta).
703 * Like Eifel it also prevents growth of rto,
704 * but also it limits too fast rto decreases,
705 * happening in pure Eifel.
706 */
707 if (m > 0)
708 m >>= 3;
709 } else {
710 m -= (tp->mdev >> 2); /* similar update on mdev */
711 }
712 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
713 if (tp->mdev > tp->mdev_max) {
714 tp->mdev_max = tp->mdev;
715 if (tp->mdev_max > tp->rttvar)
716 tp->rttvar = tp->mdev_max;
717 }
718 if (after(tp->snd_una, tp->rtt_seq)) {
719 if (tp->mdev_max < tp->rttvar)
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800720 tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700721 tp->rtt_seq = tp->snd_nxt;
David S. Miller05bb1fa2007-08-30 22:10:28 -0700722 tp->mdev_max = tcp_rto_min(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700723 }
724 } else {
725 /* no previous measure. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800726 tp->srtt = m << 3; /* take the measured time to be rtt */
727 tp->mdev = m << 1; /* make sure rto = 3*rtt */
David S. Miller05bb1fa2007-08-30 22:10:28 -0700728 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700729 tp->rtt_seq = tp->snd_nxt;
730 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700731}
732
Eric Dumazet95bd09e2013-08-27 05:46:32 -0700733/* Set the sk_pacing_rate to allow proper sizing of TSO packets.
734 * Note: TCP stack does not yet implement pacing.
735 * FQ packet scheduler can be used to implement cheap but effective
736 * TCP pacing, to smooth the burst on large writes when packets
737 * in flight is significantly lower than cwnd (or rwin)
738 */
739static void tcp_update_pacing_rate(struct sock *sk)
740{
741 const struct tcp_sock *tp = tcp_sk(sk);
742 u64 rate;
743
744 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
745 rate = (u64)tp->mss_cache * 2 * (HZ << 3);
746
747 rate *= max(tp->snd_cwnd, tp->packets_out);
748
749 /* Correction for small srtt : minimum srtt being 8 (1 jiffy << 3),
750 * be conservative and assume srtt = 1 (125 us instead of 1.25 ms)
751 * We probably need usec resolution in the future.
752 * Note: This also takes care of possible srtt=0 case,
753 * when tcp_rtt_estimator() was not yet called.
754 */
755 if (tp->srtt > 8 + 2)
756 do_div(rate, tp->srtt);
757
Eric Dumazetba537422013-10-09 17:14:52 -0700758 /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate
759 * without any lock. We want to make sure compiler wont store
760 * intermediate values in this location.
761 */
762 ACCESS_ONCE(sk->sk_pacing_rate) = min_t(u64, rate,
763 sk->sk_max_pacing_rate);
Eric Dumazet95bd09e2013-08-27 05:46:32 -0700764}
765
Linus Torvalds1da177e2005-04-16 15:20:36 -0700766/* Calculate rto without backoff. This is the second half of Van Jacobson's
767 * routine referred to above.
768 */
stephen hemmingerf7e56a72013-12-29 11:39:51 -0800769static void tcp_set_rto(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700770{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700771 const struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700772 /* Old crap is replaced with new one. 8)
773 *
774 * More seriously:
775 * 1. If rtt variance happened to be less 50msec, it is hallucination.
776 * It cannot be less due to utterly erratic ACK generation made
777 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
778 * to do with delayed acks, because at cwnd>2 true delack timeout
779 * is invisible. Actually, Linux-2.4 also generates erratic
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800780 * ACKs in some circumstances.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700781 */
Damian Lukowskif1ecd5d2009-08-26 00:16:31 +0000782 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700783
784 /* 2. Fixups made earlier cannot be right.
785 * If we do not estimate RTO correctly without them,
786 * all the algo is pure shit and should be replaced
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800787 * with correct one. It is exactly, which we pretend to do.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700788 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700789
Ilpo Järvinenee6aac52008-12-05 22:43:08 -0800790 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
791 * guarantees that rto is higher.
792 */
Damian Lukowskif1ecd5d2009-08-26 00:16:31 +0000793 tcp_bound_rto(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700794}
795
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400796__u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700797{
798 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
799
Gerrit Renker22b71c82010-08-29 19:23:12 +0000800 if (!cwnd)
David S. Miller442b9632011-02-02 17:05:11 -0800801 cwnd = TCP_INIT_CWND;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700802 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
803}
804
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300805/*
806 * Packet counting of FACK is based on in-order assumptions, therefore TCP
807 * disables it when reordering is detected
808 */
David S. Miller4aabd8e2012-07-09 16:07:30 -0700809void tcp_disable_fack(struct tcp_sock *tp)
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300810{
Ilpo Järvinen85cc3912007-11-15 19:39:31 -0800811 /* RFC3517 uses different metric in lost marker => reset on change */
812 if (tcp_is_fack(tp))
813 tp->lost_skb_hint = NULL;
Vijay Subramanianab562222011-12-20 13:23:24 +0000814 tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300815}
816
Ryousei Takano564262c12007-10-25 23:03:52 -0700817/* Take a notice that peer is sending D-SACKs */
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300818static void tcp_dsack_seen(struct tcp_sock *tp)
819{
Vijay Subramanianab562222011-12-20 13:23:24 +0000820 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300821}
822
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300823static void tcp_update_reordering(struct sock *sk, const int metric,
824 const int ts)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700825{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300826 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700827 if (metric > tp->reordering) {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -0700828 int mib_idx;
829
Linus Torvalds1da177e2005-04-16 15:20:36 -0700830 tp->reordering = min(TCP_MAX_REORDERING, metric);
831
832 /* This exciting event is worth to be remembered. 8) */
833 if (ts)
Pavel Emelyanov40b215e2008-07-03 01:05:41 -0700834 mib_idx = LINUX_MIB_TCPTSREORDER;
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300835 else if (tcp_is_reno(tp))
Pavel Emelyanov40b215e2008-07-03 01:05:41 -0700836 mib_idx = LINUX_MIB_TCPRENOREORDER;
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300837 else if (tcp_is_fack(tp))
Pavel Emelyanov40b215e2008-07-03 01:05:41 -0700838 mib_idx = LINUX_MIB_TCPFACKREORDER;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700839 else
Pavel Emelyanov40b215e2008-07-03 01:05:41 -0700840 mib_idx = LINUX_MIB_TCPSACKREORDER;
841
Pavel Emelyanovde0744a2008-07-16 20:31:16 -0700842 NET_INC_STATS_BH(sock_net(sk), mib_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700843#if FASTRETRANS_DEBUG > 1
Joe Perches91df42b2012-05-15 14:11:54 +0000844 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
845 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
846 tp->reordering,
847 tp->fackets_out,
848 tp->sacked_out,
849 tp->undo_marker ? tp->undo_retrans : 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700850#endif
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300851 tcp_disable_fack(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700852 }
Yuchung Chengeed530b2012-05-02 13:30:03 +0000853
854 if (metric > 0)
855 tcp_disable_early_retrans(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700856}
857
Ilpo Järvinen006f5822008-09-20 21:20:20 -0700858/* This must be called before lost_out is incremented */
Ilpo Järvinenc8c213f2008-09-20 21:18:55 -0700859static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
860{
Ilpo Järvinen006f5822008-09-20 21:20:20 -0700861 if ((tp->retransmit_skb_hint == NULL) ||
Ilpo Järvinenc8c213f2008-09-20 21:18:55 -0700862 before(TCP_SKB_CB(skb)->seq,
863 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
Ilpo Järvinen006f5822008-09-20 21:20:20 -0700864 tp->retransmit_skb_hint = skb;
865
866 if (!tp->lost_out ||
867 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
868 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
Ilpo Järvinenc8c213f2008-09-20 21:18:55 -0700869}
870
Ilpo Järvinen41ea36e2008-09-20 21:19:22 -0700871static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
872{
873 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
874 tcp_verify_retransmit_hint(tp, skb);
875
876 tp->lost_out += tcp_skb_pcount(skb);
877 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
878 }
879}
880
Ilpo Järvinene1aa6802008-11-24 21:11:55 -0800881static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
882 struct sk_buff *skb)
Ilpo Järvinen006f5822008-09-20 21:20:20 -0700883{
884 tcp_verify_retransmit_hint(tp, skb);
885
886 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
887 tp->lost_out += tcp_skb_pcount(skb);
888 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
889 }
890}
891
Linus Torvalds1da177e2005-04-16 15:20:36 -0700892/* This procedure tags the retransmission queue when SACKs arrive.
893 *
894 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
895 * Packets in queue with these bits set are counted in variables
896 * sacked_out, retrans_out and lost_out, correspondingly.
897 *
898 * Valid combinations are:
899 * Tag InFlight Description
900 * 0 1 - orig segment is in flight.
901 * S 0 - nothing flies, orig reached receiver.
902 * L 0 - nothing flies, orig lost by net.
903 * R 2 - both orig and retransmit are in flight.
904 * L|R 1 - orig is lost, retransmit is in flight.
905 * S|R 1 - orig reached receiver, retrans is still in flight.
906 * (L|S|R is logically valid, it could occur when L|R is sacked,
907 * but it is equivalent to plain S and code short-curcuits it to S.
908 * L|S is logically invalid, it would mean -1 packet in flight 8))
909 *
910 * These 6 states form finite state machine, controlled by the following events:
911 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
912 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
Yuchung Cheng974c1232012-01-19 14:42:21 +0000913 * 3. Loss detection event of two flavors:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700914 * A. Scoreboard estimator decided the packet is lost.
915 * A'. Reno "three dupacks" marks head of queue lost.
Yuchung Cheng974c1232012-01-19 14:42:21 +0000916 * A''. Its FACK modification, head until snd.fack is lost.
917 * B. SACK arrives sacking SND.NXT at the moment, when the
Linus Torvalds1da177e2005-04-16 15:20:36 -0700918 * segment was retransmitted.
919 * 4. D-SACK added new rule: D-SACK changes any tag to S.
920 *
921 * It is pleasant to note, that state diagram turns out to be commutative,
922 * so that we are allowed not to be bothered by order of our actions,
923 * when multiple events arrive simultaneously. (see the function below).
924 *
925 * Reordering detection.
926 * --------------------
927 * Reordering metric is maximal distance, which a packet can be displaced
928 * in packet stream. With SACKs we can estimate it:
929 *
930 * 1. SACK fills old hole and the corresponding segment was not
931 * ever retransmitted -> reordering. Alas, we cannot use it
932 * when segment was retransmitted.
933 * 2. The last flaw is solved with D-SACK. D-SACK arrives
934 * for retransmitted and already SACKed segment -> reordering..
935 * Both of these heuristics are not used in Loss state, when we cannot
936 * account for retransmits accurately.
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -0700937 *
938 * SACK block validation.
939 * ----------------------
940 *
941 * SACK block range validation checks that the received SACK block fits to
942 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
943 * Note that SND.UNA is not included to the range though being valid because
Ilpo Järvinen0e835332007-10-01 15:28:17 -0700944 * it means that the receiver is rather inconsistent with itself reporting
945 * SACK reneging when it should advance SND.UNA. Such SACK block this is
946 * perfectly valid, however, in light of RFC2018 which explicitly states
947 * that "SACK block MUST reflect the newest segment. Even if the newest
948 * segment is going to be discarded ...", not that it looks very clever
949 * in case of head skb. Due to potentional receiver driven attacks, we
950 * choose to avoid immediate execution of a walk in write queue due to
951 * reneging and defer head skb's loss recovery to standard loss recovery
952 * procedure that will eventually trigger (nothing forbids us doing this).
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -0700953 *
954 * Implements also blockage to start_seq wrap-around. Problem lies in the
955 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
956 * there's no guarantee that it will be before snd_nxt (n). The problem
957 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
958 * wrap (s_w):
959 *
960 * <- outs wnd -> <- wrapzone ->
961 * u e n u_w e_w s n_w
962 * | | | | | | |
963 * |<------------+------+----- TCP seqno space --------------+---------->|
964 * ...-- <2^31 ->| |<--------...
965 * ...---- >2^31 ------>| |<--------...
966 *
967 * Current code wouldn't be vulnerable but it's better still to discard such
968 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
969 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
970 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
971 * equal to the ideal case (infinite seqno space without wrap caused issues).
972 *
973 * With D-SACK the lower bound is extended to cover sequence space below
974 * SND.UNA down to undo_marker, which is the last point of interest. Yet
Ryousei Takano564262c12007-10-25 23:03:52 -0700975 * again, D-SACK block must not to go across snd_una (for the same reason as
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -0700976 * for the normal SACK blocks, explained above). But there all simplicity
977 * ends, TCP might receive valid D-SACKs below that. As long as they reside
978 * fully below undo_marker they do not affect behavior in anyway and can
979 * therefore be safely ignored. In rare cases (which are more or less
980 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
981 * fragmentation and packet reordering past skb's retransmission. To consider
982 * them correctly, the acceptable range must be extended even more though
983 * the exact amount is rather hard to quantify. However, tp->max_window can
984 * be used as an exaggerated estimate.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700985 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +0000986static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
987 u32 start_seq, u32 end_seq)
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -0700988{
989 /* Too far in future, or reversed (interpretation is ambiguous) */
990 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
Eric Dumazeta2a385d2012-05-16 23:15:34 +0000991 return false;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -0700992
993 /* Nasty start_seq wrap-around check (see comments above) */
994 if (!before(start_seq, tp->snd_nxt))
Eric Dumazeta2a385d2012-05-16 23:15:34 +0000995 return false;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -0700996
Ryousei Takano564262c12007-10-25 23:03:52 -0700997 /* In outstanding window? ...This is valid exit for D-SACKs too.
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -0700998 * start_seq == snd_una is non-sensical (see comments above)
999 */
1000 if (after(start_seq, tp->snd_una))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001001 return true;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001002
1003 if (!is_dsack || !tp->undo_marker)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001004 return false;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001005
1006 /* ...Then it's D-SACK, and must reside below snd_una completely */
Zheng Yanf779b2d2011-09-18 22:37:34 -04001007 if (after(end_seq, tp->snd_una))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001008 return false;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001009
1010 if (!before(start_seq, tp->undo_marker))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001011 return true;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001012
1013 /* Too old */
1014 if (!after(end_seq, tp->undo_marker))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001015 return false;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001016
1017 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1018 * start_seq < undo_marker and end_seq >= undo_marker.
1019 */
1020 return !before(start_seq, end_seq - tp->max_window);
1021}
1022
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001023/* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
Yuchung Cheng974c1232012-01-19 14:42:21 +00001024 * Event "B". Later note: FACK people cheated me again 8), we have to account
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001025 * for reordering! Ugly, but should help.
Ilpo Järvinenf785a8e2007-10-11 17:35:41 -07001026 *
1027 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1028 * less than what is now known to be received by the other end (derived from
Ilpo Järvinen9f58f3b2007-11-15 19:42:54 -08001029 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1030 * retransmitted skbs to avoid some costly processing per ACKs.
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001031 */
Ilpo Järvinen407ef1d2007-12-02 00:47:57 +02001032static void tcp_mark_lost_retrans(struct sock *sk)
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001033{
Ilpo Järvinen9f58f3b2007-11-15 19:42:54 -08001034 const struct inet_connection_sock *icsk = inet_csk(sk);
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001035 struct tcp_sock *tp = tcp_sk(sk);
1036 struct sk_buff *skb;
Ilpo Järvinenf785a8e2007-10-11 17:35:41 -07001037 int cnt = 0;
Ilpo Järvinendf2e0142007-10-18 05:07:57 -07001038 u32 new_low_seq = tp->snd_nxt;
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001039 u32 received_upto = tcp_highest_sack_seq(tp);
Ilpo Järvinen9f58f3b2007-11-15 19:42:54 -08001040
1041 if (!tcp_is_fack(tp) || !tp->retrans_out ||
1042 !after(received_upto, tp->lost_retrans_low) ||
1043 icsk->icsk_ca_state != TCP_CA_Recovery)
Ilpo Järvinen407ef1d2007-12-02 00:47:57 +02001044 return;
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001045
1046 tcp_for_write_queue(skb, sk) {
1047 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1048
1049 if (skb == tcp_send_head(sk))
1050 break;
Ilpo Järvinenf785a8e2007-10-11 17:35:41 -07001051 if (cnt == tp->retrans_out)
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001052 break;
1053 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1054 continue;
1055
Ilpo Järvinenf785a8e2007-10-11 17:35:41 -07001056 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1057 continue;
1058
Ilpo Järvinend0af4162009-02-28 04:44:32 +00001059 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1060 * constraint here (see above) but figuring out that at
1061 * least tp->reordering SACK blocks reside between ack_seq
1062 * and received_upto is not easy task to do cheaply with
1063 * the available datastructures.
1064 *
1065 * Whether FACK should check here for tp->reordering segs
1066 * in-between one could argue for either way (it would be
1067 * rather simple to implement as we could count fack_count
1068 * during the walk and do tp->fackets_out - fack_count).
1069 */
1070 if (after(received_upto, ack_seq)) {
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001071 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1072 tp->retrans_out -= tcp_skb_pcount(skb);
1073
Ilpo Järvinen006f5822008-09-20 21:20:20 -07001074 tcp_skb_mark_lost_uncond_verify(tp, skb);
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07001075 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
Ilpo Järvinenf785a8e2007-10-11 17:35:41 -07001076 } else {
Ilpo Järvinendf2e0142007-10-18 05:07:57 -07001077 if (before(ack_seq, new_low_seq))
Ilpo Järvinenb08d6cb2007-10-11 17:36:13 -07001078 new_low_seq = ack_seq;
Ilpo Järvinenf785a8e2007-10-11 17:35:41 -07001079 cnt += tcp_skb_pcount(skb);
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001080 }
1081 }
Ilpo Järvinenb08d6cb2007-10-11 17:36:13 -07001082
1083 if (tp->retrans_out)
1084 tp->lost_retrans_low = new_low_seq;
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001085}
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001086
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001087static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1088 struct tcp_sack_block_wire *sp, int num_sacks,
1089 u32 prior_snd_una)
David S. Millerd06e0212007-06-18 22:43:06 -07001090{
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07001091 struct tcp_sock *tp = tcp_sk(sk);
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07001092 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1093 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001094 bool dup_sack = false;
David S. Millerd06e0212007-06-18 22:43:06 -07001095
1096 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001097 dup_sack = true;
Ilpo Järvinene60402d2007-08-09 15:14:46 +03001098 tcp_dsack_seen(tp);
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07001099 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
David S. Millerd06e0212007-06-18 22:43:06 -07001100 } else if (num_sacks > 1) {
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07001101 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1102 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
David S. Millerd06e0212007-06-18 22:43:06 -07001103
1104 if (!after(end_seq_0, end_seq_1) &&
1105 !before(start_seq_0, start_seq_1)) {
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001106 dup_sack = true;
Ilpo Järvinene60402d2007-08-09 15:14:46 +03001107 tcp_dsack_seen(tp);
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07001108 NET_INC_STATS_BH(sock_net(sk),
1109 LINUX_MIB_TCPDSACKOFORECV);
David S. Millerd06e0212007-06-18 22:43:06 -07001110 }
1111 }
1112
1113 /* D-SACK for already forgotten data... Do dumb counting. */
Yuchung Chengc24f6912011-02-07 12:57:04 +00001114 if (dup_sack && tp->undo_marker && tp->undo_retrans &&
David S. Millerd06e0212007-06-18 22:43:06 -07001115 !after(end_seq_0, prior_snd_una) &&
1116 after(end_seq_0, tp->undo_marker))
1117 tp->undo_retrans--;
1118
1119 return dup_sack;
1120}
1121
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001122struct tcp_sacktag_state {
1123 int reord;
1124 int fack_count;
1125 int flag;
Yuchung Cheng59c9af42013-07-22 16:20:47 -07001126 s32 rtt; /* RTT measured by SACKing never-retransmitted data */
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001127};
1128
Ilpo Järvinend1935942007-10-11 17:34:25 -07001129/* Check if skb is fully within the SACK block. In presence of GSO skbs,
1130 * the incoming SACK may not exactly match but we can find smaller MSS
1131 * aligned portion of it that matches. Therefore we might need to fragment
1132 * which may fail and creates some hassle (caller must handle error case
1133 * returns).
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001134 *
1135 * FIXME: this could be merged to shift decision code
Ilpo Järvinend1935942007-10-11 17:34:25 -07001136 */
Adrian Bunk0f79efd2007-10-26 03:57:36 -07001137static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001138 u32 start_seq, u32 end_seq)
Ilpo Järvinend1935942007-10-11 17:34:25 -07001139{
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001140 int err;
1141 bool in_sack;
Ilpo Järvinend1935942007-10-11 17:34:25 -07001142 unsigned int pkt_len;
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001143 unsigned int mss;
Ilpo Järvinend1935942007-10-11 17:34:25 -07001144
1145 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1146 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1147
1148 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1149 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001150 mss = tcp_skb_mss(skb);
Ilpo Järvinend1935942007-10-11 17:34:25 -07001151 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1152
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001153 if (!in_sack) {
Ilpo Järvinend1935942007-10-11 17:34:25 -07001154 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001155 if (pkt_len < mss)
1156 pkt_len = mss;
1157 } else {
Ilpo Järvinend1935942007-10-11 17:34:25 -07001158 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001159 if (pkt_len < mss)
1160 return -EINVAL;
1161 }
1162
1163 /* Round if necessary so that SACKs cover only full MSSes
1164 * and/or the remaining small portion (if present)
1165 */
1166 if (pkt_len > mss) {
1167 unsigned int new_len = (pkt_len / mss) * mss;
1168 if (!in_sack && new_len < pkt_len) {
1169 new_len += mss;
1170 if (new_len > skb->len)
1171 return 0;
1172 }
1173 pkt_len = new_len;
1174 }
1175 err = tcp_fragment(sk, skb, pkt_len, mss);
Ilpo Järvinend1935942007-10-11 17:34:25 -07001176 if (err < 0)
1177 return err;
1178 }
1179
1180 return in_sack;
1181}
1182
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001183/* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1184static u8 tcp_sacktag_one(struct sock *sk,
1185 struct tcp_sacktag_state *state, u8 sacked,
1186 u32 start_seq, u32 end_seq,
Yuchung Cheng59c9af42013-07-22 16:20:47 -07001187 int dup_sack, int pcount, u32 xmit_time)
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001188{
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001189 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001190 int fack_count = state->fack_count;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001191
1192 /* Account D-SACK for retransmitted packet. */
1193 if (dup_sack && (sacked & TCPCB_RETRANS)) {
Yuchung Chengc24f6912011-02-07 12:57:04 +00001194 if (tp->undo_marker && tp->undo_retrans &&
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001195 after(end_seq, tp->undo_marker))
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001196 tp->undo_retrans--;
Ilpo Järvinenede9f3b2007-12-02 00:47:58 +02001197 if (sacked & TCPCB_SACKED_ACKED)
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001198 state->reord = min(fack_count, state->reord);
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001199 }
1200
1201 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001202 if (!after(end_seq, tp->snd_una))
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001203 return sacked;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001204
1205 if (!(sacked & TCPCB_SACKED_ACKED)) {
1206 if (sacked & TCPCB_SACKED_RETRANS) {
1207 /* If the segment is not tagged as lost,
1208 * we do not clear RETRANS, believing
1209 * that retransmission is still in flight.
1210 */
1211 if (sacked & TCPCB_LOST) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001212 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001213 tp->lost_out -= pcount;
1214 tp->retrans_out -= pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001215 }
1216 } else {
1217 if (!(sacked & TCPCB_RETRANS)) {
1218 /* New sack for not retransmitted frame,
1219 * which was in hole. It is reordering.
1220 */
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001221 if (before(start_seq,
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001222 tcp_highest_sack_seq(tp)))
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001223 state->reord = min(fack_count,
1224 state->reord);
Yuchung Chenge33099f2013-03-20 13:33:00 +00001225 if (!after(end_seq, tp->high_seq))
1226 state->flag |= FLAG_ORIG_SACK_ACKED;
Yuchung Cheng59c9af42013-07-22 16:20:47 -07001227 /* Pick the earliest sequence sacked for RTT */
1228 if (state->rtt < 0)
1229 state->rtt = tcp_time_stamp - xmit_time;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001230 }
1231
1232 if (sacked & TCPCB_LOST) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001233 sacked &= ~TCPCB_LOST;
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001234 tp->lost_out -= pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001235 }
1236 }
1237
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001238 sacked |= TCPCB_SACKED_ACKED;
1239 state->flag |= FLAG_DATA_SACKED;
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001240 tp->sacked_out += pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001241
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001242 fack_count += pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001243
1244 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1245 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001246 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001247 tp->lost_cnt_hint += pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001248
1249 if (fack_count > tp->fackets_out)
1250 tp->fackets_out = fack_count;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001251 }
1252
1253 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1254 * frames and clear it. undo_retrans is decreased above, L|R frames
1255 * are accounted above as well.
1256 */
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001257 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1258 sacked &= ~TCPCB_SACKED_RETRANS;
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001259 tp->retrans_out -= pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001260 }
1261
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001262 return sacked;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001263}
1264
Neal Cardwelldaef52b2012-02-12 18:37:10 +00001265/* Shift newly-SACKed bytes from this skb to the immediately previous
1266 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1267 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001268static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1269 struct tcp_sacktag_state *state,
1270 unsigned int pcount, int shifted, int mss,
1271 bool dup_sack)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001272{
1273 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen50133162008-12-05 22:42:41 -08001274 struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
Neal Cardwelldaef52b2012-02-12 18:37:10 +00001275 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1276 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001277
1278 BUG_ON(!pcount);
1279
Neal Cardwell4c90d3b2012-02-26 10:06:19 +00001280 /* Adjust counters and hints for the newly sacked sequence
1281 * range but discard the return value since prev is already
1282 * marked. We must tag the range first because the seq
1283 * advancement below implicitly advances
1284 * tcp_highest_sack_seq() when skb is highest_sack.
1285 */
1286 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
Yuchung Cheng59c9af42013-07-22 16:20:47 -07001287 start_seq, end_seq, dup_sack, pcount,
1288 TCP_SKB_CB(skb)->when);
Neal Cardwell4c90d3b2012-02-26 10:06:19 +00001289
1290 if (skb == tp->lost_skb_hint)
Neal Cardwell0af2a0d2012-02-13 20:22:08 +00001291 tp->lost_cnt_hint += pcount;
1292
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001293 TCP_SKB_CB(prev)->end_seq += shifted;
1294 TCP_SKB_CB(skb)->seq += shifted;
1295
1296 skb_shinfo(prev)->gso_segs += pcount;
1297 BUG_ON(skb_shinfo(skb)->gso_segs < pcount);
1298 skb_shinfo(skb)->gso_segs -= pcount;
1299
1300 /* When we're adding to gso_segs == 1, gso_size will be zero,
1301 * in theory this shouldn't be necessary but as long as DSACK
1302 * code can come after this skb later on it's better to keep
1303 * setting gso_size to something.
1304 */
1305 if (!skb_shinfo(prev)->gso_size) {
1306 skb_shinfo(prev)->gso_size = mss;
Pravin B Shelarc9af6db2013-02-11 09:27:41 +00001307 skb_shinfo(prev)->gso_type = sk->sk_gso_type;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001308 }
1309
1310 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1311 if (skb_shinfo(skb)->gso_segs <= 1) {
1312 skb_shinfo(skb)->gso_size = 0;
Pravin B Shelarc9af6db2013-02-11 09:27:41 +00001313 skb_shinfo(skb)->gso_type = 0;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001314 }
1315
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001316 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1317 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1318
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001319 if (skb->len > 0) {
1320 BUG_ON(!tcp_skb_pcount(skb));
Ilpo Järvinen111cc8b2008-11-24 21:27:22 -08001321 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001322 return false;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001323 }
1324
1325 /* Whole SKB was eaten :-) */
1326
Ilpo Järvinen92ee76b2008-11-24 21:26:56 -08001327 if (skb == tp->retransmit_skb_hint)
1328 tp->retransmit_skb_hint = prev;
Ilpo Järvinen92ee76b2008-11-24 21:26:56 -08001329 if (skb == tp->lost_skb_hint) {
1330 tp->lost_skb_hint = prev;
1331 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1332 }
1333
Eric Dumazet5e8a402f2013-10-04 10:31:41 -07001334 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1335 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1336 TCP_SKB_CB(prev)->end_seq++;
1337
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001338 if (skb == tcp_highest_sack(sk))
1339 tcp_advance_highest_sack(sk, skb);
1340
1341 tcp_unlink_write_queue(skb, sk);
1342 sk_wmem_free_skb(sk, skb);
1343
Ilpo Järvinen111cc8b2008-11-24 21:27:22 -08001344 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1345
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001346 return true;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001347}
1348
1349/* I wish gso_size would have a bit more sane initialization than
1350 * something-or-zero which complicates things
1351 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001352static int tcp_skb_seglen(const struct sk_buff *skb)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001353{
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001354 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001355}
1356
1357/* Shifting pages past head area doesn't work */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001358static int skb_can_shift(const struct sk_buff *skb)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001359{
1360 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1361}
1362
1363/* Try collapsing SACK blocks spanning across multiple skbs to a single
1364 * skb.
1365 */
1366static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001367 struct tcp_sacktag_state *state,
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001368 u32 start_seq, u32 end_seq,
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001369 bool dup_sack)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001370{
1371 struct tcp_sock *tp = tcp_sk(sk);
1372 struct sk_buff *prev;
1373 int mss;
1374 int pcount = 0;
1375 int len;
1376 int in_sack;
1377
1378 if (!sk_can_gso(sk))
1379 goto fallback;
1380
1381 /* Normally R but no L won't result in plain S */
1382 if (!dup_sack &&
Ilpo Järvinen9969ca52008-12-05 22:41:06 -08001383 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001384 goto fallback;
1385 if (!skb_can_shift(skb))
1386 goto fallback;
1387 /* This frame is about to be dropped (was ACKed). */
1388 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1389 goto fallback;
1390
1391 /* Can only happen with delayed DSACK + discard craziness */
1392 if (unlikely(skb == tcp_write_queue_head(sk)))
1393 goto fallback;
1394 prev = tcp_write_queue_prev(sk, skb);
1395
1396 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1397 goto fallback;
1398
1399 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1400 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1401
1402 if (in_sack) {
1403 len = skb->len;
1404 pcount = tcp_skb_pcount(skb);
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001405 mss = tcp_skb_seglen(skb);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001406
1407 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1408 * drop this restriction as unnecessary
1409 */
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001410 if (mss != tcp_skb_seglen(prev))
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001411 goto fallback;
1412 } else {
1413 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1414 goto noop;
1415 /* CHECKME: This is non-MSS split case only?, this will
1416 * cause skipped skbs due to advancing loop btw, original
1417 * has that feature too
1418 */
1419 if (tcp_skb_pcount(skb) <= 1)
1420 goto noop;
1421
1422 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1423 if (!in_sack) {
1424 /* TODO: head merge to next could be attempted here
1425 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1426 * though it might not be worth of the additional hassle
1427 *
1428 * ...we can probably just fallback to what was done
1429 * previously. We could try merging non-SACKed ones
1430 * as well but it probably isn't going to buy off
1431 * because later SACKs might again split them, and
1432 * it would make skb timestamp tracking considerably
1433 * harder problem.
1434 */
1435 goto fallback;
1436 }
1437
1438 len = end_seq - TCP_SKB_CB(skb)->seq;
1439 BUG_ON(len < 0);
1440 BUG_ON(len > skb->len);
1441
1442 /* MSS boundaries should be honoured or else pcount will
1443 * severely break even though it makes things bit trickier.
1444 * Optimize common case to avoid most of the divides
1445 */
1446 mss = tcp_skb_mss(skb);
1447
1448 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1449 * drop this restriction as unnecessary
1450 */
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001451 if (mss != tcp_skb_seglen(prev))
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001452 goto fallback;
1453
1454 if (len == mss) {
1455 pcount = 1;
1456 } else if (len < mss) {
1457 goto noop;
1458 } else {
1459 pcount = len / mss;
1460 len = pcount * mss;
1461 }
1462 }
1463
Neal Cardwell4648dc92012-03-05 19:35:04 +00001464 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1465 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1466 goto fallback;
1467
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001468 if (!skb_shift(prev, skb, len))
1469 goto fallback;
Ilpo Järvinen9ec06ff2009-03-01 00:21:36 -08001470 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001471 goto out;
1472
1473 /* Hole filled allows collapsing with the next as well, this is very
1474 * useful when hole on every nth skb pattern happens
1475 */
1476 if (prev == tcp_write_queue_tail(sk))
1477 goto out;
1478 skb = tcp_write_queue_next(sk, prev);
1479
Ilpo Järvinenf0bc52f2008-12-05 22:40:47 -08001480 if (!skb_can_shift(skb) ||
1481 (skb == tcp_send_head(sk)) ||
1482 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001483 (mss != tcp_skb_seglen(skb)))
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001484 goto out;
1485
1486 len = skb->len;
1487 if (skb_shift(prev, skb, len)) {
1488 pcount += tcp_skb_pcount(skb);
Ilpo Järvinen9ec06ff2009-03-01 00:21:36 -08001489 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001490 }
1491
1492out:
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001493 state->fack_count += pcount;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001494 return prev;
1495
1496noop:
1497 return skb;
1498
1499fallback:
Ilpo Järvinen111cc8b2008-11-24 21:27:22 -08001500 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001501 return NULL;
1502}
1503
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001504static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1505 struct tcp_sack_block *next_dup,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001506 struct tcp_sacktag_state *state,
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001507 u32 start_seq, u32 end_seq,
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001508 bool dup_sack_in)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001509{
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001510 struct tcp_sock *tp = tcp_sk(sk);
1511 struct sk_buff *tmp;
1512
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001513 tcp_for_write_queue_from(skb, sk) {
1514 int in_sack = 0;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001515 bool dup_sack = dup_sack_in;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001516
1517 if (skb == tcp_send_head(sk))
1518 break;
1519
1520 /* queue is in-order => we can short-circuit the walk early */
1521 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1522 break;
1523
1524 if ((next_dup != NULL) &&
1525 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1526 in_sack = tcp_match_skb_to_sack(sk, skb,
1527 next_dup->start_seq,
1528 next_dup->end_seq);
1529 if (in_sack > 0)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001530 dup_sack = true;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001531 }
1532
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001533 /* skb reference here is a bit tricky to get right, since
1534 * shifting can eat and free both this skb and the next,
1535 * so not even _safe variant of the loop is enough.
1536 */
1537 if (in_sack <= 0) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001538 tmp = tcp_shift_skb_data(sk, skb, state,
1539 start_seq, end_seq, dup_sack);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001540 if (tmp != NULL) {
1541 if (tmp != skb) {
1542 skb = tmp;
1543 continue;
1544 }
1545
1546 in_sack = 0;
1547 } else {
1548 in_sack = tcp_match_skb_to_sack(sk, skb,
1549 start_seq,
1550 end_seq);
1551 }
1552 }
1553
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001554 if (unlikely(in_sack < 0))
1555 break;
1556
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001557 if (in_sack) {
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001558 TCP_SKB_CB(skb)->sacked =
1559 tcp_sacktag_one(sk,
1560 state,
1561 TCP_SKB_CB(skb)->sacked,
1562 TCP_SKB_CB(skb)->seq,
1563 TCP_SKB_CB(skb)->end_seq,
1564 dup_sack,
Yuchung Cheng59c9af42013-07-22 16:20:47 -07001565 tcp_skb_pcount(skb),
1566 TCP_SKB_CB(skb)->when);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001567
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001568 if (!before(TCP_SKB_CB(skb)->seq,
1569 tcp_highest_sack_seq(tp)))
1570 tcp_advance_highest_sack(sk, skb);
1571 }
1572
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001573 state->fack_count += tcp_skb_pcount(skb);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001574 }
1575 return skb;
1576}
1577
1578/* Avoid all extra work that is being done by sacktag while walking in
1579 * a normal way
1580 */
1581static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001582 struct tcp_sacktag_state *state,
1583 u32 skip_to_seq)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001584{
1585 tcp_for_write_queue_from(skb, sk) {
1586 if (skb == tcp_send_head(sk))
1587 break;
1588
Ilpo Järvinene8bae272008-11-24 21:12:28 -08001589 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001590 break;
Ilpo Järvinend152a7d2008-03-03 12:10:16 -08001591
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001592 state->fack_count += tcp_skb_pcount(skb);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001593 }
1594 return skb;
1595}
1596
1597static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1598 struct sock *sk,
1599 struct tcp_sack_block *next_dup,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001600 struct tcp_sacktag_state *state,
1601 u32 skip_to_seq)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001602{
1603 if (next_dup == NULL)
1604 return skb;
1605
1606 if (before(next_dup->start_seq, skip_to_seq)) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001607 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1608 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1609 next_dup->start_seq, next_dup->end_seq,
1610 1);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001611 }
1612
1613 return skb;
1614}
1615
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001616static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001617{
1618 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1619}
1620
Linus Torvalds1da177e2005-04-16 15:20:36 -07001621static int
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001622tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
Yuchung Cheng59c9af42013-07-22 16:20:47 -07001623 u32 prior_snd_una, s32 *sack_rtt)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001624{
1625 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001626 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1627 TCP_SKB_CB(ack_skb)->sacked);
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001628 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
Adam Langley4389dde2008-07-19 00:07:02 -07001629 struct tcp_sack_block sp[TCP_NUM_SACKS];
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001630 struct tcp_sack_block *cache;
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001631 struct tcp_sacktag_state state;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001632 struct sk_buff *skb;
Adam Langley4389dde2008-07-19 00:07:02 -07001633 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001634 int used_sacks;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001635 bool found_dup_sack = false;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001636 int i, j;
Baruch Evenfda03fb2007-02-04 23:35:57 -08001637 int first_sack_index;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001638
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001639 state.flag = 0;
1640 state.reord = tp->packets_out;
Yuchung Cheng59c9af42013-07-22 16:20:47 -07001641 state.rtt = -1;
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001642
Ilpo Järvinend738cd82007-03-24 21:03:23 -07001643 if (!tp->sacked_out) {
Ilpo Järvinende83c052007-10-07 23:37:55 -07001644 if (WARN_ON(tp->fackets_out))
1645 tp->fackets_out = 0;
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001646 tcp_highest_sack_reset(sk);
Ilpo Järvinend738cd82007-03-24 21:03:23 -07001647 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001648
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07001649 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
David S. Millerd06e0212007-06-18 22:43:06 -07001650 num_sacks, prior_snd_una);
1651 if (found_dup_sack)
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001652 state.flag |= FLAG_DSACKING_ACK;
Baruch Even6f746512007-02-04 23:36:42 -08001653
1654 /* Eliminate too old ACKs, but take into
1655 * account more or less fresh ones, they can
1656 * contain valid SACK info.
1657 */
1658 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1659 return 0;
1660
Ilpo Järvinen96a2d412007-11-14 15:47:18 -08001661 if (!tp->packets_out)
1662 goto out;
1663
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001664 used_sacks = 0;
1665 first_sack_index = 0;
1666 for (i = 0; i < num_sacks; i++) {
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001667 bool dup_sack = !i && found_dup_sack;
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001668
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07001669 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1670 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001671
1672 if (!tcp_is_sackblock_valid(tp, dup_sack,
1673 sp[used_sacks].start_seq,
1674 sp[used_sacks].end_seq)) {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001675 int mib_idx;
1676
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001677 if (dup_sack) {
1678 if (!tp->undo_marker)
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001679 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001680 else
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001681 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001682 } else {
1683 /* Don't count olds caused by ACK reordering */
1684 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1685 !after(sp[used_sacks].end_seq, tp->snd_una))
1686 continue;
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001687 mib_idx = LINUX_MIB_TCPSACKDISCARD;
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001688 }
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001689
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07001690 NET_INC_STATS_BH(sock_net(sk), mib_idx);
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001691 if (i == 0)
1692 first_sack_index = -1;
1693 continue;
1694 }
1695
1696 /* Ignore very old stuff early */
1697 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1698 continue;
1699
1700 used_sacks++;
1701 }
1702
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001703 /* order SACK blocks to allow in order walk of the retrans queue */
1704 for (i = used_sacks - 1; i > 0; i--) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001705 for (j = 0; j < i; j++) {
1706 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
Ilpo Järvinena0bffff2009-03-21 13:36:17 -07001707 swap(sp[j], sp[j + 1]);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001708
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001709 /* Track where the first SACK block goes to */
1710 if (j == first_sack_index)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001711 first_sack_index = j + 1;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001712 }
1713 }
1714 }
1715
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001716 skb = tcp_write_queue_head(sk);
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001717 state.fack_count = 0;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001718 i = 0;
1719
1720 if (!tp->sacked_out) {
1721 /* It's already past, so skip checking against it */
1722 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1723 } else {
1724 cache = tp->recv_sack_cache;
1725 /* Skip empty blocks in at head of the cache */
1726 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1727 !cache->end_seq)
1728 cache++;
Baruch Evenfda03fb2007-02-04 23:35:57 -08001729 }
1730
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001731 while (i < used_sacks) {
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001732 u32 start_seq = sp[i].start_seq;
1733 u32 end_seq = sp[i].end_seq;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001734 bool dup_sack = (found_dup_sack && (i == first_sack_index));
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001735 struct tcp_sack_block *next_dup = NULL;
Ilpo Järvinene56d6cd2007-11-01 00:09:37 -07001736
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001737 if (found_dup_sack && ((i + 1) == first_sack_index))
1738 next_dup = &sp[i + 1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001739
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001740 /* Skip too early cached blocks */
1741 while (tcp_sack_cache_ok(tp, cache) &&
1742 !before(start_seq, cache->end_seq))
1743 cache++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001744
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001745 /* Can skip some work by looking recv_sack_cache? */
1746 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1747 after(end_seq, cache->start_seq)) {
David S. Millerfe067e82007-03-07 12:12:44 -08001748
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001749 /* Head todo? */
1750 if (before(start_seq, cache->start_seq)) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001751 skb = tcp_sacktag_skip(skb, sk, &state,
1752 start_seq);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001753 skb = tcp_sacktag_walk(skb, sk, next_dup,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001754 &state,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001755 start_seq,
1756 cache->start_seq,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001757 dup_sack);
Baruch Evenfda03fb2007-02-04 23:35:57 -08001758 }
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001759
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001760 /* Rest of the block already fully processed? */
Ilpo Järvinen20de20b2007-11-16 16:17:05 -08001761 if (!after(end_seq, cache->end_seq))
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001762 goto advance_sp;
Ilpo Järvinen20de20b2007-11-16 16:17:05 -08001763
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001764 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001765 &state,
1766 cache->end_seq);
Ilpo Järvinene56d6cd2007-11-01 00:09:37 -07001767
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001768 /* ...tail remains todo... */
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001769 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
Ilpo Järvinen20de20b2007-11-16 16:17:05 -08001770 /* ...but better entrypoint exists! */
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001771 skb = tcp_highest_sack(sk);
1772 if (skb == NULL)
1773 break;
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001774 state.fack_count = tp->fackets_out;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001775 cache++;
1776 goto walk;
1777 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001778
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001779 skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001780 /* Check overlap against next cached too (past this one already) */
1781 cache++;
1782 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001783 }
Ilpo Järvinenfbd52eb2007-11-10 21:24:19 -08001784
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001785 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1786 skb = tcp_highest_sack(sk);
1787 if (skb == NULL)
1788 break;
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001789 state.fack_count = tp->fackets_out;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001790 }
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001791 skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001792
1793walk:
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001794 skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1795 start_seq, end_seq, dup_sack);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001796
1797advance_sp:
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001798 i++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001799 }
1800
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001801 /* Clear the head of the cache sack blocks so we can skip it next time */
1802 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1803 tp->recv_sack_cache[i].start_seq = 0;
1804 tp->recv_sack_cache[i].end_seq = 0;
1805 }
1806 for (j = 0; j < used_sacks; j++)
1807 tp->recv_sack_cache[i++] = sp[j];
1808
Ilpo Järvinen407ef1d2007-12-02 00:47:57 +02001809 tcp_mark_lost_retrans(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001810
Ilpo Järvinen86426c22007-08-09 14:45:17 +03001811 tcp_verify_left_out(tp);
1812
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001813 if ((state.reord < tp->fackets_out) &&
Yuchung Cheng9b441902013-03-20 13:32:58 +00001814 ((inet_csk(sk)->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker))
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001815 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001816
Ilpo Järvinen96a2d412007-11-14 15:47:18 -08001817out:
1818
Linus Torvalds1da177e2005-04-16 15:20:36 -07001819#if FASTRETRANS_DEBUG > 0
Ilpo Järvinen547b7922008-07-25 21:43:18 -07001820 WARN_ON((int)tp->sacked_out < 0);
1821 WARN_ON((int)tp->lost_out < 0);
1822 WARN_ON((int)tp->retrans_out < 0);
1823 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001824#endif
Yuchung Cheng59c9af42013-07-22 16:20:47 -07001825 *sack_rtt = state.rtt;
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001826 return state.flag;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001827}
1828
Ilpo Järvinen882beba2008-04-07 22:33:07 -07001829/* Limits sacked_out so that sum with lost_out isn't ever larger than
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001830 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
Ilpo Järvinen30935cf2007-02-21 23:01:36 -08001831 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001832static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001833{
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001834 u32 holes;
1835
1836 holes = max(tp->lost_out, 1U);
1837 holes = min(holes, tp->packets_out);
1838
1839 if ((tp->sacked_out + holes) > tp->packets_out) {
1840 tp->sacked_out = tp->packets_out - holes;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001841 return true;
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001842 }
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001843 return false;
Ilpo Järvinen882beba2008-04-07 22:33:07 -07001844}
1845
1846/* If we receive more dupacks than we expected counting segments
1847 * in assumption of absent reordering, interpret this as reordering.
1848 * The only another reason could be bug in receiver TCP.
1849 */
1850static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1851{
1852 struct tcp_sock *tp = tcp_sk(sk);
1853 if (tcp_limit_reno_sacked(tp))
1854 tcp_update_reordering(sk, tp->packets_out + addend, 0);
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001855}
1856
1857/* Emulate SACKs for SACKless connection: account for a new dupack. */
1858
1859static void tcp_add_reno_sack(struct sock *sk)
1860{
1861 struct tcp_sock *tp = tcp_sk(sk);
1862 tp->sacked_out++;
1863 tcp_check_reno_reordering(sk, 0);
Ilpo Järvinen005903b2007-08-09 14:44:16 +03001864 tcp_verify_left_out(tp);
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001865}
1866
1867/* Account for ACK, ACKing some data in Reno Recovery phase. */
1868
1869static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1870{
1871 struct tcp_sock *tp = tcp_sk(sk);
1872
1873 if (acked > 0) {
1874 /* One ACK acked hole. The rest eat duplicate ACKs. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001875 if (acked - 1 >= tp->sacked_out)
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001876 tp->sacked_out = 0;
1877 else
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001878 tp->sacked_out -= acked - 1;
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001879 }
1880 tcp_check_reno_reordering(sk, acked);
Ilpo Järvinen005903b2007-08-09 14:44:16 +03001881 tcp_verify_left_out(tp);
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001882}
1883
1884static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1885{
1886 tp->sacked_out = 0;
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001887}
1888
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07001889static void tcp_clear_retrans_partial(struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001890{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001891 tp->retrans_out = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001892 tp->lost_out = 0;
1893
1894 tp->undo_marker = 0;
1895 tp->undo_retrans = 0;
1896}
1897
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07001898void tcp_clear_retrans(struct tcp_sock *tp)
1899{
1900 tcp_clear_retrans_partial(tp);
1901
1902 tp->fackets_out = 0;
1903 tp->sacked_out = 0;
1904}
1905
Linus Torvalds1da177e2005-04-16 15:20:36 -07001906/* Enter Loss state. If "how" is not zero, forget all SACK information
1907 * and reset tags completely, otherwise preserve SACKs. If receiver
1908 * dropped its ofo queue, we will know this due to reneging detection.
1909 */
1910void tcp_enter_loss(struct sock *sk, int how)
1911{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001912 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001913 struct tcp_sock *tp = tcp_sk(sk);
1914 struct sk_buff *skb;
Yuchung Chenge33099f2013-03-20 13:33:00 +00001915 bool new_recovery = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001916
1917 /* Reduce ssthresh if it has not yet been made inside this window. */
Yuchung Chenge33099f2013-03-20 13:33:00 +00001918 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1919 !after(tp->high_seq, tp->snd_una) ||
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001920 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
Yuchung Chenge33099f2013-03-20 13:33:00 +00001921 new_recovery = true;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001922 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1923 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1924 tcp_ca_event(sk, CA_EVENT_LOSS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001925 }
1926 tp->snd_cwnd = 1;
1927 tp->snd_cwnd_cnt = 0;
1928 tp->snd_cwnd_stamp = tcp_time_stamp;
1929
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07001930 tcp_clear_retrans_partial(tp);
1931
1932 if (tcp_is_reno(tp))
1933 tcp_reset_reno_sack(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001934
Yuchung Cheng7ebe1832013-03-24 10:42:25 +00001935 tp->undo_marker = tp->snd_una;
1936 if (how) {
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07001937 tp->sacked_out = 0;
1938 tp->fackets_out = 0;
Ilpo Järvinenb7689202007-09-20 11:37:19 -07001939 }
Ilpo Järvinen64edc272008-09-20 21:18:32 -07001940 tcp_clear_all_retrans_hints(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001941
David S. Millerfe067e82007-03-07 12:12:44 -08001942 tcp_for_write_queue(skb, sk) {
1943 if (skb == tcp_send_head(sk))
1944 break;
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07001945
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001946 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001947 tp->undo_marker = 0;
1948 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1949 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1950 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1951 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1952 tp->lost_out += tcp_skb_pcount(skb);
Ilpo Järvinen006f5822008-09-20 21:20:20 -07001953 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001954 }
1955 }
Ilpo Järvinen005903b2007-08-09 14:44:16 +03001956 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001957
Yuchung Cheng74c181d2013-08-12 16:41:25 -07001958 /* Timeout in disordered state after receiving substantial DUPACKs
1959 * suggests that the degree of reordering is over-estimated.
1960 */
1961 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
1962 tp->sacked_out >= sysctl_tcp_reordering)
1963 tp->reordering = min_t(unsigned int, tp->reordering,
1964 sysctl_tcp_reordering);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001965 tcp_set_ca_state(sk, TCP_CA_Loss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001966 tp->high_seq = tp->snd_nxt;
1967 TCP_ECN_queue_cwr(tp);
Yuchung Chenge33099f2013-03-20 13:33:00 +00001968
1969 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1970 * loss recovery is underway except recurring timeout(s) on
1971 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1972 */
1973 tp->frto = sysctl_tcp_frto &&
1974 (new_recovery || icsk->icsk_retransmits) &&
1975 !inet_csk(sk)->icsk_mtup.probe_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001976}
1977
Ilpo Järvinencadbd032007-12-31 04:49:21 -08001978/* If ACK arrived pointing to a remembered SACK, it means that our
1979 * remembered SACKs do not reflect real state of receiver i.e.
1980 * receiver _host_ is heavily congested (or buggy).
1981 *
1982 * Do processing similar to RTO timeout.
1983 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001984static bool tcp_check_sack_reneging(struct sock *sk, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001985{
Ilpo Järvinencadbd032007-12-31 04:49:21 -08001986 if (flag & FLAG_SACK_RENEGING) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001987 struct inet_connection_sock *icsk = inet_csk(sk);
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07001988 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001989
1990 tcp_enter_loss(sk, 1);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001991 icsk->icsk_retransmits++;
David S. Millerfe067e82007-03-07 12:12:44 -08001992 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001993 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001994 icsk->icsk_rto, TCP_RTO_MAX);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001995 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001996 }
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001997 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001998}
1999
Eric Dumazetcf533ea2011-10-21 05:22:42 -04002000static inline int tcp_fackets_out(const struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002001{
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002002 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002003}
2004
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002005/* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2006 * counter when SACK is enabled (without SACK, sacked_out is used for
2007 * that purpose).
2008 *
2009 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2010 * segments up to the highest received SACK block so far and holes in
2011 * between them.
2012 *
2013 * With reordering, holes may still be in flight, so RFC3517 recovery
2014 * uses pure sacked_out (total number of SACKed segments) even though
2015 * it violates the RFC that uses duplicate ACKs, often these are equal
2016 * but when e.g. out-of-window ACKs or packet duplication occurs,
2017 * they differ. Since neither occurs due to loss, TCP should really
2018 * ignore them.
2019 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04002020static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002021{
2022 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2023}
2024
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00002025static bool tcp_pause_early_retransmit(struct sock *sk, int flag)
2026{
2027 struct tcp_sock *tp = tcp_sk(sk);
2028 unsigned long delay;
2029
2030 /* Delay early retransmit and entering fast recovery for
2031 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
2032 * available, or RTO is scheduled to fire first.
2033 */
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002034 if (sysctl_tcp_early_retrans < 2 || sysctl_tcp_early_retrans > 3 ||
2035 (flag & FLAG_ECE) || !tp->srtt)
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00002036 return false;
2037
2038 delay = max_t(unsigned long, (tp->srtt >> 5), msecs_to_jiffies(2));
2039 if (!time_after(inet_csk(sk)->icsk_timeout, (jiffies + delay)))
2040 return false;
2041
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002042 inet_csk_reset_xmit_timer(sk, ICSK_TIME_EARLY_RETRANS, delay,
2043 TCP_RTO_MAX);
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00002044 return true;
2045}
2046
Linus Torvalds1da177e2005-04-16 15:20:36 -07002047/* Linux NewReno/SACK/FACK/ECN state machine.
2048 * --------------------------------------
2049 *
2050 * "Open" Normal state, no dubious events, fast path.
2051 * "Disorder" In all the respects it is "Open",
2052 * but requires a bit more attention. It is entered when
2053 * we see some SACKs or dupacks. It is split of "Open"
2054 * mainly to move some processing from fast path to slow one.
2055 * "CWR" CWND was reduced due to some Congestion Notification event.
2056 * It can be ECN, ICMP source quench, local device congestion.
2057 * "Recovery" CWND was reduced, we are fast-retransmitting.
2058 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2059 *
2060 * tcp_fastretrans_alert() is entered:
2061 * - each incoming ACK, if state is not "Open"
2062 * - when arrived ACK is unusual, namely:
2063 * * SACK
2064 * * Duplicate ACK.
2065 * * ECN ECE.
2066 *
2067 * Counting packets in flight is pretty simple.
2068 *
2069 * in_flight = packets_out - left_out + retrans_out
2070 *
2071 * packets_out is SND.NXT-SND.UNA counted in packets.
2072 *
2073 * retrans_out is number of retransmitted segments.
2074 *
2075 * left_out is number of segments left network, but not ACKed yet.
2076 *
2077 * left_out = sacked_out + lost_out
2078 *
2079 * sacked_out: Packets, which arrived to receiver out of order
2080 * and hence not ACKed. With SACKs this number is simply
2081 * amount of SACKed data. Even without SACKs
2082 * it is easy to give pretty reliable estimate of this number,
2083 * counting duplicate ACKs.
2084 *
2085 * lost_out: Packets lost by network. TCP has no explicit
2086 * "loss notification" feedback from network (for now).
2087 * It means that this number can be only _guessed_.
2088 * Actually, it is the heuristics to predict lossage that
2089 * distinguishes different algorithms.
2090 *
2091 * F.e. after RTO, when all the queue is considered as lost,
2092 * lost_out = packets_out and in_flight = retrans_out.
2093 *
2094 * Essentially, we have now two algorithms counting
2095 * lost packets.
2096 *
2097 * FACK: It is the simplest heuristics. As soon as we decided
2098 * that something is lost, we decide that _all_ not SACKed
2099 * packets until the most forward SACK are lost. I.e.
2100 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2101 * It is absolutely correct estimate, if network does not reorder
2102 * packets. And it loses any connection to reality when reordering
2103 * takes place. We use FACK by default until reordering
2104 * is suspected on the path to this destination.
2105 *
2106 * NewReno: when Recovery is entered, we assume that one segment
2107 * is lost (classic Reno). While we are in Recovery and
2108 * a partial ACK arrives, we assume that one more packet
2109 * is lost (NewReno). This heuristics are the same in NewReno
2110 * and SACK.
2111 *
2112 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2113 * deflation etc. CWND is real congestion window, never inflated, changes
2114 * only according to classic VJ rules.
2115 *
2116 * Really tricky (and requiring careful tuning) part of algorithm
2117 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2118 * The first determines the moment _when_ we should reduce CWND and,
2119 * hence, slow down forward transmission. In fact, it determines the moment
2120 * when we decide that hole is caused by loss, rather than by a reorder.
2121 *
2122 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2123 * holes, caused by lost packets.
2124 *
2125 * And the most logically complicated part of algorithm is undo
2126 * heuristics. We detect false retransmits due to both too early
2127 * fast retransmit (reordering) and underestimated RTO, analyzing
2128 * timestamps and D-SACKs. When we detect that some segments were
2129 * retransmitted by mistake and CWND reduction was wrong, we undo
2130 * window reduction and abort recovery phase. This logic is hidden
2131 * inside several functions named tcp_try_undo_<something>.
2132 */
2133
2134/* This function decides, when we should leave Disordered state
2135 * and enter Recovery phase, reducing congestion window.
2136 *
2137 * Main question: may we further continue forward transmission
2138 * with the same cwnd?
2139 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002140static bool tcp_time_to_recover(struct sock *sk, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002141{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002142 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002143 __u32 packets_out;
2144
2145 /* Trick#1: The loss is proven. */
2146 if (tp->lost_out)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002147 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002148
2149 /* Not-A-Trick#2 : Classic rule... */
Andreas Petlundea84e552009-10-27 03:27:21 +00002150 if (tcp_dupack_heuristics(tp) > tp->reordering)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002151 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002152
Linus Torvalds1da177e2005-04-16 15:20:36 -07002153 /* Trick#4: It is still not OK... But will it be useful to delay
2154 * recovery more?
2155 */
2156 packets_out = tp->packets_out;
2157 if (packets_out <= tp->reordering &&
2158 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002159 !tcp_may_send_now(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002160 /* We have nothing to send. This connection is limited
2161 * either by receiver window or by application.
2162 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002163 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002164 }
2165
Andreas Petlund7e380172010-02-18 04:48:19 +00002166 /* If a thin stream is detected, retransmit after first
2167 * received dupack. Employ only if SACK is supported in order
2168 * to avoid possible corner-case series of spurious retransmissions
2169 * Use only if there are no unsent data.
2170 */
2171 if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2172 tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2173 tcp_is_sack(tp) && !tcp_send_head(sk))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002174 return true;
Andreas Petlund7e380172010-02-18 04:48:19 +00002175
Yuchung Chengeed530b2012-05-02 13:30:03 +00002176 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious
2177 * retransmissions due to small network reorderings, we implement
2178 * Mitigation A.3 in the RFC and delay the retransmission for a short
2179 * interval if appropriate.
2180 */
2181 if (tp->do_early_retrans && !tp->retrans_out && tp->sacked_out &&
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002182 (tp->packets_out >= (tp->sacked_out + 1) && tp->packets_out < 4) &&
Yuchung Chengeed530b2012-05-02 13:30:03 +00002183 !tcp_may_send_now(sk))
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00002184 return !tcp_pause_early_retransmit(sk, flag);
Yuchung Chengeed530b2012-05-02 13:30:03 +00002185
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002186 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002187}
2188
Yuchung Cheng974c1232012-01-19 14:42:21 +00002189/* Detect loss in event "A" above by marking head of queue up as lost.
2190 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2191 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2192 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2193 * the maximum SACKed segments to pass before reaching this limit.
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002194 */
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002195static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002196{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002197 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002198 struct sk_buff *skb;
Ilpo Järvinenc137f3d2008-04-07 22:32:38 -07002199 int cnt, oldcnt;
2200 int err;
2201 unsigned int mss;
Yuchung Cheng974c1232012-01-19 14:42:21 +00002202 /* Use SACK to deduce losses of new sequences sent during recovery */
2203 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002204
Ilpo Järvinen547b7922008-07-25 21:43:18 -07002205 WARN_ON(packets > tp->packets_out);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002206 if (tp->lost_skb_hint) {
2207 skb = tp->lost_skb_hint;
2208 cnt = tp->lost_cnt_hint;
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002209 /* Head already handled? */
2210 if (mark_head && skb != tcp_write_queue_head(sk))
2211 return;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002212 } else {
David S. Millerfe067e82007-03-07 12:12:44 -08002213 skb = tcp_write_queue_head(sk);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002214 cnt = 0;
2215 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002216
David S. Millerfe067e82007-03-07 12:12:44 -08002217 tcp_for_write_queue_from(skb, sk) {
2218 if (skb == tcp_send_head(sk))
2219 break;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002220 /* TODO: do this better */
2221 /* this is not the most efficient way to do this... */
2222 tp->lost_skb_hint = skb;
2223 tp->lost_cnt_hint = cnt;
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002224
Yuchung Cheng974c1232012-01-19 14:42:21 +00002225 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
Ilpo Järvinenc137f3d2008-04-07 22:32:38 -07002226 break;
2227
2228 oldcnt = cnt;
Ilpo Järvinenad1984e2008-01-30 20:06:02 -08002229 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002230 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2231 cnt += tcp_skb_pcount(skb);
2232
Ilpo Järvinenc137f3d2008-04-07 22:32:38 -07002233 if (cnt > packets) {
Yuchung Chengb3de7552010-09-24 13:22:06 +00002234 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
Neal Cardwellc0638c22012-03-02 21:36:51 +00002235 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
Yuchung Chengb3de7552010-09-24 13:22:06 +00002236 (oldcnt >= packets))
Ilpo Järvinenc137f3d2008-04-07 22:32:38 -07002237 break;
2238
2239 mss = skb_shinfo(skb)->gso_size;
2240 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2241 if (err < 0)
2242 break;
2243 cnt = packets;
2244 }
2245
Ilpo Järvinen41ea36e2008-09-20 21:19:22 -07002246 tcp_skb_mark_lost(tp, skb);
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002247
2248 if (mark_head)
2249 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002250 }
Ilpo Järvinen005903b2007-08-09 14:44:16 +03002251 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002252}
2253
2254/* Account newly detected lost packet(s) */
2255
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002256static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002257{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002258 struct tcp_sock *tp = tcp_sk(sk);
2259
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002260 if (tcp_is_reno(tp)) {
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002261 tcp_mark_head_lost(sk, 1, 1);
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002262 } else if (tcp_is_fack(tp)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002263 int lost = tp->fackets_out - tp->reordering;
2264 if (lost <= 0)
2265 lost = 1;
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002266 tcp_mark_head_lost(sk, lost, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002267 } else {
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002268 int sacked_upto = tp->sacked_out - tp->reordering;
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002269 if (sacked_upto >= 0)
2270 tcp_mark_head_lost(sk, sacked_upto, 0);
2271 else if (fast_rexmit)
2272 tcp_mark_head_lost(sk, 1, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002273 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002274}
2275
2276/* CWND moderation, preventing bursts due to too big ACKs
2277 * in dubious situations.
2278 */
2279static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2280{
2281 tp->snd_cwnd = min(tp->snd_cwnd,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002282 tcp_packets_in_flight(tp) + tcp_max_burst(tp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002283 tp->snd_cwnd_stamp = tcp_time_stamp;
2284}
2285
Linus Torvalds1da177e2005-04-16 15:20:36 -07002286/* Nothing was retransmitted or returned timestamp is less
2287 * than timestamp of the first retransmission.
2288 */
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00002289static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002290{
2291 return !tp->retrans_stamp ||
2292 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
Arnd Hannemannd7ee1472008-04-21 14:46:22 -07002293 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002294}
2295
2296/* Undo procedures. */
2297
2298#if FASTRETRANS_DEBUG > 1
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002299static void DBGUNDO(struct sock *sk, const char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002300{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002301 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002302 struct inet_sock *inet = inet_sk(sk);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002303
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002304 if (sk->sk_family == AF_INET) {
Joe Perches91df42b2012-05-15 14:11:54 +00002305 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2306 msg,
2307 &inet->inet_daddr, ntohs(inet->inet_dport),
2308 tp->snd_cwnd, tcp_left_out(tp),
2309 tp->snd_ssthresh, tp->prior_ssthresh,
2310 tp->packets_out);
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002311 }
Eric Dumazetdfd56b82011-12-10 09:48:31 +00002312#if IS_ENABLED(CONFIG_IPV6)
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002313 else if (sk->sk_family == AF_INET6) {
2314 struct ipv6_pinfo *np = inet6_sk(sk);
Joe Perches91df42b2012-05-15 14:11:54 +00002315 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2316 msg,
2317 &np->daddr, ntohs(inet->inet_dport),
2318 tp->snd_cwnd, tcp_left_out(tp),
2319 tp->snd_ssthresh, tp->prior_ssthresh,
2320 tp->packets_out);
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002321 }
2322#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002323}
2324#else
2325#define DBGUNDO(x...) do { } while (0)
2326#endif
2327
Yuchung Cheng7026b912013-05-29 14:20:13 +00002328static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002329{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002330 struct tcp_sock *tp = tcp_sk(sk);
2331
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002332 if (unmark_loss) {
2333 struct sk_buff *skb;
2334
2335 tcp_for_write_queue(skb, sk) {
2336 if (skb == tcp_send_head(sk))
2337 break;
2338 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2339 }
2340 tp->lost_out = 0;
2341 tcp_clear_all_retrans_hints(tp);
2342 }
2343
Linus Torvalds1da177e2005-04-16 15:20:36 -07002344 if (tp->prior_ssthresh) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002345 const struct inet_connection_sock *icsk = inet_csk(sk);
2346
2347 if (icsk->icsk_ca_ops->undo_cwnd)
2348 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002349 else
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002350 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002351
Yuchung Cheng7026b912013-05-29 14:20:13 +00002352 if (tp->prior_ssthresh > tp->snd_ssthresh) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002353 tp->snd_ssthresh = tp->prior_ssthresh;
2354 TCP_ECN_withdraw_cwr(tp);
2355 }
2356 } else {
2357 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2358 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002359 tp->snd_cwnd_stamp = tcp_time_stamp;
Yuchung Cheng7026b912013-05-29 14:20:13 +00002360 tp->undo_marker = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002361}
2362
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00002363static inline bool tcp_may_undo(const struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002364{
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002365 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002366}
2367
2368/* People celebrate: "We love our President!" */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002369static bool tcp_try_undo_recovery(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002370{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002371 struct tcp_sock *tp = tcp_sk(sk);
2372
Linus Torvalds1da177e2005-04-16 15:20:36 -07002373 if (tcp_may_undo(tp)) {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07002374 int mib_idx;
2375
Linus Torvalds1da177e2005-04-16 15:20:36 -07002376 /* Happy end! We did not retransmit anything
2377 * or our original transmission succeeded.
2378 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002379 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
Yuchung Cheng7026b912013-05-29 14:20:13 +00002380 tcp_undo_cwnd_reduction(sk, false);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002381 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07002382 mib_idx = LINUX_MIB_TCPLOSSUNDO;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002383 else
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07002384 mib_idx = LINUX_MIB_TCPFULLUNDO;
2385
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07002386 NET_INC_STATS_BH(sock_net(sk), mib_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002387 }
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002388 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002389 /* Hold old state until something *above* high_seq
2390 * is ACKed. For Reno it is MUST to prevent false
2391 * fast retransmits (RFC2582). SACK TCP is safe. */
2392 tcp_moderate_cwnd(tp);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002393 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002394 }
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002395 tcp_set_ca_state(sk, TCP_CA_Open);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002396 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002397}
2398
2399/* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
Yuchung Chengc7d9d6a2013-05-29 14:20:14 +00002400static bool tcp_try_undo_dsack(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002401{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002402 struct tcp_sock *tp = tcp_sk(sk);
2403
Linus Torvalds1da177e2005-04-16 15:20:36 -07002404 if (tp->undo_marker && !tp->undo_retrans) {
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002405 DBGUNDO(sk, "D-SACK");
Yuchung Cheng7026b912013-05-29 14:20:13 +00002406 tcp_undo_cwnd_reduction(sk, false);
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07002407 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
Yuchung Chengc7d9d6a2013-05-29 14:20:14 +00002408 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002409 }
Yuchung Chengc7d9d6a2013-05-29 14:20:14 +00002410 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002411}
2412
Ilpo Järvinen77722b12009-12-08 20:54:11 -08002413/* We can clear retrans_stamp when there are no retransmissions in the
2414 * window. It would seem that it is trivially available for us in
2415 * tp->retrans_out, however, that kind of assumptions doesn't consider
2416 * what will happen if errors occur when sending retransmission for the
2417 * second time. ...It could the that such segment has only
2418 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2419 * the head skb is enough except for some reneging corner cases that
2420 * are not worth the effort.
2421 *
2422 * Main reason for all this complexity is the fact that connection dying
2423 * time now depends on the validity of the retrans_stamp, in particular,
2424 * that successive retransmissions of a segment must not advance
2425 * retrans_stamp under any conditions.
2426 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002427static bool tcp_any_retrans_done(const struct sock *sk)
Ilpo Järvinen77722b12009-12-08 20:54:11 -08002428{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04002429 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen77722b12009-12-08 20:54:11 -08002430 struct sk_buff *skb;
2431
2432 if (tp->retrans_out)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002433 return true;
Ilpo Järvinen77722b12009-12-08 20:54:11 -08002434
2435 skb = tcp_write_queue_head(sk);
2436 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002437 return true;
Ilpo Järvinen77722b12009-12-08 20:54:11 -08002438
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002439 return false;
Ilpo Järvinen77722b12009-12-08 20:54:11 -08002440}
2441
Yuchung Chenge33099f2013-03-20 13:33:00 +00002442/* Undo during loss recovery after partial ACK or using F-RTO. */
2443static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002444{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002445 struct tcp_sock *tp = tcp_sk(sk);
2446
Yuchung Chenge33099f2013-03-20 13:33:00 +00002447 if (frto_undo || tcp_may_undo(tp)) {
Yuchung Cheng7026b912013-05-29 14:20:13 +00002448 tcp_undo_cwnd_reduction(sk, true);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002449
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002450 DBGUNDO(sk, "partial loss");
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07002451 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
Yuchung Chenge33099f2013-03-20 13:33:00 +00002452 if (frto_undo)
2453 NET_INC_STATS_BH(sock_net(sk),
2454 LINUX_MIB_TCPSPURIOUSRTOS);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002455 inet_csk(sk)->icsk_retransmits = 0;
Yuchung Chenge33099f2013-03-20 13:33:00 +00002456 if (frto_undo || tcp_is_sack(tp))
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002457 tcp_set_ca_state(sk, TCP_CA_Open);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002458 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002459 }
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002460 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002461}
2462
Yuchung Cheng684bad12012-09-02 17:38:04 +00002463/* The cwnd reduction in CWR and Recovery use the PRR algorithm
2464 * https://datatracker.ietf.org/doc/draft-ietf-tcpm-proportional-rate-reduction/
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002465 * It computes the number of packets to send (sndcnt) based on packets newly
2466 * delivered:
2467 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2468 * cwnd reductions across a full RTT.
2469 * 2) If packets in flight is lower than ssthresh (such as due to excess
2470 * losses and/or application stalls), do not perform any further cwnd
2471 * reductions, but instead slow start up to ssthresh.
2472 */
Yuchung Cheng684bad12012-09-02 17:38:04 +00002473static void tcp_init_cwnd_reduction(struct sock *sk, const bool set_ssthresh)
2474{
2475 struct tcp_sock *tp = tcp_sk(sk);
2476
2477 tp->high_seq = tp->snd_nxt;
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00002478 tp->tlp_high_seq = 0;
Yuchung Cheng684bad12012-09-02 17:38:04 +00002479 tp->snd_cwnd_cnt = 0;
2480 tp->prior_cwnd = tp->snd_cwnd;
2481 tp->prr_delivered = 0;
2482 tp->prr_out = 0;
2483 if (set_ssthresh)
2484 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2485 TCP_ECN_queue_cwr(tp);
2486}
2487
Yuchung Cheng68049732013-05-29 14:20:11 +00002488static void tcp_cwnd_reduction(struct sock *sk, const int prior_unsacked,
Yuchung Cheng684bad12012-09-02 17:38:04 +00002489 int fast_rexmit)
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002490{
2491 struct tcp_sock *tp = tcp_sk(sk);
2492 int sndcnt = 0;
2493 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
Yuchung Cheng68049732013-05-29 14:20:11 +00002494 int newly_acked_sacked = prior_unsacked -
2495 (tp->packets_out - tp->sacked_out);
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002496
Yuchung Cheng684bad12012-09-02 17:38:04 +00002497 tp->prr_delivered += newly_acked_sacked;
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002498 if (tcp_packets_in_flight(tp) > tp->snd_ssthresh) {
2499 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2500 tp->prior_cwnd - 1;
2501 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2502 } else {
2503 sndcnt = min_t(int, delta,
2504 max_t(int, tp->prr_delivered - tp->prr_out,
2505 newly_acked_sacked) + 1);
2506 }
2507
2508 sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0));
2509 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2510}
2511
Yuchung Cheng684bad12012-09-02 17:38:04 +00002512static inline void tcp_end_cwnd_reduction(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002513{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002514 struct tcp_sock *tp = tcp_sk(sk);
Nandita Dukkipatia262f0c2011-08-21 20:21:57 +00002515
Yuchung Cheng684bad12012-09-02 17:38:04 +00002516 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2517 if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR ||
2518 (tp->undo_marker && tp->snd_ssthresh < TCP_INFINITE_SSTHRESH)) {
2519 tp->snd_cwnd = tp->snd_ssthresh;
2520 tp->snd_cwnd_stamp = tcp_time_stamp;
Yuchung Cheng67d41202011-03-14 10:57:03 +00002521 }
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002522 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002523}
2524
Yuchung Cheng684bad12012-09-02 17:38:04 +00002525/* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
Yuchung Cheng09484d12012-09-02 17:38:02 +00002526void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
2527{
2528 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng09484d12012-09-02 17:38:02 +00002529
2530 tp->prior_ssthresh = 0;
Yuchung Cheng684bad12012-09-02 17:38:04 +00002531 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
Yuchung Cheng09484d12012-09-02 17:38:02 +00002532 tp->undo_marker = 0;
Yuchung Cheng684bad12012-09-02 17:38:04 +00002533 tcp_init_cwnd_reduction(sk, set_ssthresh);
Yuchung Cheng09484d12012-09-02 17:38:02 +00002534 tcp_set_ca_state(sk, TCP_CA_CWR);
2535 }
2536}
2537
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07002538static void tcp_try_keep_open(struct sock *sk)
2539{
2540 struct tcp_sock *tp = tcp_sk(sk);
2541 int state = TCP_CA_Open;
2542
Neal Cardwellf6982042011-11-16 08:58:04 +00002543 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07002544 state = TCP_CA_Disorder;
2545
2546 if (inet_csk(sk)->icsk_ca_state != state) {
2547 tcp_set_ca_state(sk, state);
2548 tp->high_seq = tp->snd_nxt;
2549 }
2550}
2551
Yuchung Cheng68049732013-05-29 14:20:11 +00002552static void tcp_try_to_open(struct sock *sk, int flag, const int prior_unsacked)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002553{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002554 struct tcp_sock *tp = tcp_sk(sk);
2555
Ilpo Järvinen86426c22007-08-09 14:45:17 +03002556 tcp_verify_left_out(tp);
2557
Yuchung Cheng9b441902013-03-20 13:32:58 +00002558 if (!tcp_any_retrans_done(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002559 tp->retrans_stamp = 0;
2560
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002561 if (flag & FLAG_ECE)
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -08002562 tcp_enter_cwr(sk, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002563
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002564 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07002565 tcp_try_keep_open(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002566 } else {
Yuchung Cheng68049732013-05-29 14:20:11 +00002567 tcp_cwnd_reduction(sk, prior_unsacked, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002568 }
2569}
2570
John Heffner5d424d52006-03-20 17:53:41 -08002571static void tcp_mtup_probe_failed(struct sock *sk)
2572{
2573 struct inet_connection_sock *icsk = inet_csk(sk);
2574
2575 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2576 icsk->icsk_mtup.probe_size = 0;
2577}
2578
Ilpo Järvinen72211e92009-03-14 14:23:04 +00002579static void tcp_mtup_probe_success(struct sock *sk)
John Heffner5d424d52006-03-20 17:53:41 -08002580{
2581 struct tcp_sock *tp = tcp_sk(sk);
2582 struct inet_connection_sock *icsk = inet_csk(sk);
2583
2584 /* FIXME: breaks with very large cwnd */
2585 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2586 tp->snd_cwnd = tp->snd_cwnd *
2587 tcp_mss_to_mtu(sk, tp->mss_cache) /
2588 icsk->icsk_mtup.probe_size;
2589 tp->snd_cwnd_cnt = 0;
2590 tp->snd_cwnd_stamp = tcp_time_stamp;
John Heffner9c6d5e52010-10-06 21:18:02 -07002591 tp->snd_ssthresh = tcp_current_ssthresh(sk);
John Heffner5d424d52006-03-20 17:53:41 -08002592
2593 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2594 icsk->icsk_mtup.probe_size = 0;
2595 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2596}
2597
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08002598/* Do a simple retransmit without using the backoff mechanisms in
2599 * tcp_timer. This is used for path mtu discovery.
2600 * The socket is already locked here.
2601 */
2602void tcp_simple_retransmit(struct sock *sk)
2603{
2604 const struct inet_connection_sock *icsk = inet_csk(sk);
2605 struct tcp_sock *tp = tcp_sk(sk);
2606 struct sk_buff *skb;
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00002607 unsigned int mss = tcp_current_mss(sk);
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08002608 u32 prior_lost = tp->lost_out;
2609
2610 tcp_for_write_queue(skb, sk) {
2611 if (skb == tcp_send_head(sk))
2612 break;
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08002613 if (tcp_skb_seglen(skb) > mss &&
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08002614 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2615 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2616 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2617 tp->retrans_out -= tcp_skb_pcount(skb);
2618 }
2619 tcp_skb_mark_lost_uncond_verify(tp, skb);
2620 }
2621 }
2622
2623 tcp_clear_retrans_hints_partial(tp);
2624
2625 if (prior_lost == tp->lost_out)
2626 return;
2627
2628 if (tcp_is_reno(tp))
2629 tcp_limit_reno_sacked(tp);
2630
2631 tcp_verify_left_out(tp);
2632
2633 /* Don't muck with the congestion window here.
2634 * Reason is that we do not increase amount of _data_
2635 * in network, but units changed and effective
2636 * cwnd/ssthresh really reduced now.
2637 */
2638 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2639 tp->high_seq = tp->snd_nxt;
2640 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2641 tp->prior_ssthresh = 0;
2642 tp->undo_marker = 0;
2643 tcp_set_ca_state(sk, TCP_CA_Loss);
2644 }
2645 tcp_xmit_retransmit_queue(sk);
2646}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00002647EXPORT_SYMBOL(tcp_simple_retransmit);
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08002648
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002649static void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2650{
2651 struct tcp_sock *tp = tcp_sk(sk);
2652 int mib_idx;
2653
2654 if (tcp_is_reno(tp))
2655 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2656 else
2657 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2658
2659 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2660
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002661 tp->prior_ssthresh = 0;
2662 tp->undo_marker = tp->snd_una;
2663 tp->undo_retrans = tp->retrans_out;
2664
2665 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2666 if (!ece_ack)
2667 tp->prior_ssthresh = tcp_current_ssthresh(sk);
Yuchung Cheng684bad12012-09-02 17:38:04 +00002668 tcp_init_cwnd_reduction(sk, true);
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002669 }
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002670 tcp_set_ca_state(sk, TCP_CA_Recovery);
2671}
2672
Yuchung Chengab42d9e2013-03-20 13:32:59 +00002673/* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2674 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2675 */
Yuchung Chenge33099f2013-03-20 13:33:00 +00002676static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack)
Yuchung Chengab42d9e2013-03-20 13:32:59 +00002677{
2678 struct inet_connection_sock *icsk = inet_csk(sk);
2679 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Chenge33099f2013-03-20 13:33:00 +00002680 bool recovered = !before(tp->snd_una, tp->high_seq);
Yuchung Chengab42d9e2013-03-20 13:32:59 +00002681
Yuchung Chenge33099f2013-03-20 13:33:00 +00002682 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2683 if (flag & FLAG_ORIG_SACK_ACKED) {
2684 /* Step 3.b. A timeout is spurious if not all data are
2685 * lost, i.e., never-retransmitted data are (s)acked.
2686 */
2687 tcp_try_undo_loss(sk, true);
2688 return;
2689 }
2690 if (after(tp->snd_nxt, tp->high_seq) &&
2691 (flag & FLAG_DATA_SACKED || is_dupack)) {
2692 tp->frto = 0; /* Loss was real: 2nd part of step 3.a */
2693 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2694 tp->high_seq = tp->snd_nxt;
2695 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
2696 TCP_NAGLE_OFF);
2697 if (after(tp->snd_nxt, tp->high_seq))
2698 return; /* Step 2.b */
2699 tp->frto = 0;
2700 }
2701 }
2702
2703 if (recovered) {
2704 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
Yuchung Chengab42d9e2013-03-20 13:32:59 +00002705 icsk->icsk_retransmits = 0;
2706 tcp_try_undo_recovery(sk);
2707 return;
2708 }
Yuchung Chengab42d9e2013-03-20 13:32:59 +00002709 if (flag & FLAG_DATA_ACKED)
2710 icsk->icsk_retransmits = 0;
Yuchung Chenge33099f2013-03-20 13:33:00 +00002711 if (tcp_is_reno(tp)) {
2712 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2713 * delivered. Lower inflight to clock out (re)tranmissions.
2714 */
2715 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2716 tcp_add_reno_sack(sk);
2717 else if (flag & FLAG_SND_UNA_ADVANCED)
2718 tcp_reset_reno_sack(tp);
2719 }
2720 if (tcp_try_undo_loss(sk, false))
Yuchung Chengab42d9e2013-03-20 13:32:59 +00002721 return;
Yuchung Chengab42d9e2013-03-20 13:32:59 +00002722 tcp_xmit_retransmit_queue(sk);
2723}
2724
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002725/* Undo during fast recovery after partial ACK. */
Yuchung Cheng7026b912013-05-29 14:20:13 +00002726static bool tcp_try_undo_partial(struct sock *sk, const int acked,
2727 const int prior_unsacked)
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002728{
2729 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002730
Yuchung Cheng7026b912013-05-29 14:20:13 +00002731 if (tp->undo_marker && tcp_packet_delayed(tp)) {
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002732 /* Plain luck! Hole if filled with delayed
2733 * packet, rather than with a retransmit.
2734 */
Yuchung Cheng7026b912013-05-29 14:20:13 +00002735 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2736
2737 /* We are getting evidence that the reordering degree is higher
2738 * than we realized. If there are no retransmits out then we
2739 * can undo. Otherwise we clock out new packets but do not
2740 * mark more packets lost or retransmit more.
2741 */
2742 if (tp->retrans_out) {
2743 tcp_cwnd_reduction(sk, prior_unsacked, 0);
2744 return true;
2745 }
2746
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002747 if (!tcp_any_retrans_done(sk))
2748 tp->retrans_stamp = 0;
2749
Yuchung Cheng7026b912013-05-29 14:20:13 +00002750 DBGUNDO(sk, "partial recovery");
2751 tcp_undo_cwnd_reduction(sk, true);
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002752 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
Yuchung Cheng7026b912013-05-29 14:20:13 +00002753 tcp_try_keep_open(sk);
2754 return true;
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002755 }
Yuchung Cheng7026b912013-05-29 14:20:13 +00002756 return false;
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002757}
2758
Linus Torvalds1da177e2005-04-16 15:20:36 -07002759/* Process an event, which can update packets-in-flight not trivially.
2760 * Main goal of this function is to calculate new estimate for left_out,
2761 * taking into account both packets sitting in receiver's buffer and
2762 * packets lost by network.
2763 *
2764 * Besides that it does CWND reduction, when packet loss is detected
2765 * and changes state of machine.
2766 *
2767 * It does _not_ decide what to send, it is made in function
2768 * tcp_xmit_retransmit_queue().
2769 */
Yuchung Cheng68049732013-05-29 14:20:11 +00002770static void tcp_fastretrans_alert(struct sock *sk, const int acked,
2771 const int prior_unsacked,
Nandita Dukkipati35f079e2013-05-21 15:12:07 +00002772 bool is_dupack, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002773{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002774 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002775 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002776 bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002777 (tcp_fackets_out(tp) > tp->reordering));
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002778 int fast_rexmit = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002779
Ilpo Järvinen3ccd3132007-12-31 04:43:32 -08002780 if (WARN_ON(!tp->packets_out && tp->sacked_out))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002781 tp->sacked_out = 0;
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07002782 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002783 tp->fackets_out = 0;
2784
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002785 /* Now state machine starts.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002786 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002787 if (flag & FLAG_ECE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002788 tp->prior_ssthresh = 0;
2789
2790 /* B. In all the states check for reneging SACKs. */
Ilpo Järvinencadbd032007-12-31 04:49:21 -08002791 if (tcp_check_sack_reneging(sk, flag))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002792 return;
2793
Yuchung Cheng974c1232012-01-19 14:42:21 +00002794 /* C. Check consistency of the current state. */
Ilpo Järvinen005903b2007-08-09 14:44:16 +03002795 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002796
Yuchung Cheng974c1232012-01-19 14:42:21 +00002797 /* D. Check state exit conditions. State can be terminated
Linus Torvalds1da177e2005-04-16 15:20:36 -07002798 * when high_seq is ACKed. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002799 if (icsk->icsk_ca_state == TCP_CA_Open) {
Ilpo Järvinen547b7922008-07-25 21:43:18 -07002800 WARN_ON(tp->retrans_out != 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002801 tp->retrans_stamp = 0;
2802 } else if (!before(tp->snd_una, tp->high_seq)) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002803 switch (icsk->icsk_ca_state) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002804 case TCP_CA_CWR:
2805 /* CWR is to be held something *above* high_seq
2806 * is ACKed for CWR bit to reach receiver. */
2807 if (tp->snd_una != tp->high_seq) {
Yuchung Cheng684bad12012-09-02 17:38:04 +00002808 tcp_end_cwnd_reduction(sk);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002809 tcp_set_ca_state(sk, TCP_CA_Open);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002810 }
2811 break;
2812
Linus Torvalds1da177e2005-04-16 15:20:36 -07002813 case TCP_CA_Recovery:
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002814 if (tcp_is_reno(tp))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002815 tcp_reset_reno_sack(tp);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002816 if (tcp_try_undo_recovery(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002817 return;
Yuchung Cheng684bad12012-09-02 17:38:04 +00002818 tcp_end_cwnd_reduction(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002819 break;
2820 }
2821 }
2822
Yuchung Cheng974c1232012-01-19 14:42:21 +00002823 /* E. Process state. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002824 switch (icsk->icsk_ca_state) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002825 case TCP_CA_Recovery:
Ilpo Järvinen2e605292007-08-02 19:46:58 -07002826 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002827 if (tcp_is_reno(tp) && is_dupack)
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002828 tcp_add_reno_sack(sk);
Yuchung Cheng7026b912013-05-29 14:20:13 +00002829 } else {
2830 if (tcp_try_undo_partial(sk, acked, prior_unsacked))
2831 return;
2832 /* Partial ACK arrived. Force fast retransmit. */
2833 do_lost = tcp_is_reno(tp) ||
2834 tcp_fackets_out(tp) > tp->reordering;
2835 }
Yuchung Chengc7d9d6a2013-05-29 14:20:14 +00002836 if (tcp_try_undo_dsack(sk)) {
2837 tcp_try_keep_open(sk);
2838 return;
2839 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002840 break;
2841 case TCP_CA_Loss:
Yuchung Chenge33099f2013-03-20 13:33:00 +00002842 tcp_process_loss(sk, flag, is_dupack);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002843 if (icsk->icsk_ca_state != TCP_CA_Open)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002844 return;
Yuchung Chengab42d9e2013-03-20 13:32:59 +00002845 /* Fall through to processing in Open state. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002846 default:
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002847 if (tcp_is_reno(tp)) {
Ilpo Järvinen2e605292007-08-02 19:46:58 -07002848 if (flag & FLAG_SND_UNA_ADVANCED)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002849 tcp_reset_reno_sack(tp);
2850 if (is_dupack)
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002851 tcp_add_reno_sack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002852 }
2853
Neal Cardwellf6982042011-11-16 08:58:04 +00002854 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002855 tcp_try_undo_dsack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002856
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00002857 if (!tcp_time_to_recover(sk, flag)) {
Yuchung Cheng68049732013-05-29 14:20:11 +00002858 tcp_try_to_open(sk, flag, prior_unsacked);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002859 return;
2860 }
2861
John Heffner5d424d52006-03-20 17:53:41 -08002862 /* MTU probe failure: don't reduce cwnd */
2863 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2864 icsk->icsk_mtup.probe_size &&
John Heffner0e7b1362006-03-20 21:32:58 -08002865 tp->snd_una == tp->mtu_probe.probe_seq_start) {
John Heffner5d424d52006-03-20 17:53:41 -08002866 tcp_mtup_probe_failed(sk);
2867 /* Restores the reduction we did in tcp_mtup_probe() */
2868 tp->snd_cwnd++;
2869 tcp_simple_retransmit(sk);
2870 return;
2871 }
2872
Linus Torvalds1da177e2005-04-16 15:20:36 -07002873 /* Otherwise enter Recovery state */
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002874 tcp_enter_recovery(sk, (flag & FLAG_ECE));
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002875 fast_rexmit = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002876 }
2877
Yuchung Cheng3e59cb02013-05-17 13:45:05 +00002878 if (do_lost)
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002879 tcp_update_scoreboard(sk, fast_rexmit);
Yuchung Cheng68049732013-05-29 14:20:11 +00002880 tcp_cwnd_reduction(sk, prior_unsacked, fast_rexmit);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002881 tcp_xmit_retransmit_queue(sk);
2882}
2883
Yuchung Chenged084952013-07-22 16:20:48 -07002884static inline bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2885 s32 seq_rtt, s32 sack_rtt)
Ilpo Järvinen41834b72008-12-05 22:43:26 -08002886{
Yuchung Cheng5b08e472013-07-22 16:20:46 -07002887 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen41834b72008-12-05 22:43:26 -08002888
Yuchung Cheng5b08e472013-07-22 16:20:46 -07002889 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2890 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2891 * Karn's algorithm forbids taking RTT if some retransmitted data
2892 * is acked (RFC6298).
2893 */
2894 if (flag & FLAG_RETRANS_DATA_ACKED)
2895 seq_rtt = -1;
2896
Yuchung Chenged084952013-07-22 16:20:48 -07002897 if (seq_rtt < 0)
2898 seq_rtt = sack_rtt;
2899
Linus Torvalds1da177e2005-04-16 15:20:36 -07002900 /* RTTM Rule: A TSecr value received in a segment is used to
2901 * update the averaged RTT measurement only if the segment
2902 * acknowledges some new data, i.e., only if it advances the
2903 * left edge of the send window.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002904 * See draft-ietf-tcplw-high-performance-00, section 3.3.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002905 */
Yuchung Cheng2909d872013-10-24 08:55:25 -07002906 if (seq_rtt < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2907 flag & FLAG_ACKED)
Yuchung Cheng5b08e472013-07-22 16:20:46 -07002908 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
Ilpo Järvinen41834b72008-12-05 22:43:26 -08002909
Yuchung Cheng5b08e472013-07-22 16:20:46 -07002910 if (seq_rtt < 0)
Yuchung Chenged084952013-07-22 16:20:48 -07002911 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002912
Yuchung Cheng5b08e472013-07-22 16:20:46 -07002913 tcp_rtt_estimator(sk, seq_rtt);
2914 tcp_set_rto(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002915
Yuchung Cheng5b08e472013-07-22 16:20:46 -07002916 /* RFC6298: only reset backoff on valid RTT measurement. */
2917 inet_csk(sk)->icsk_backoff = 0;
Yuchung Chenged084952013-07-22 16:20:48 -07002918 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002919}
2920
Yuchung Cheng375fe022013-07-22 16:20:45 -07002921/* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
Yuchung Chengbc15afa2013-10-24 08:44:25 -07002922static void tcp_synack_rtt_meas(struct sock *sk, const u32 synack_stamp)
Yuchung Cheng375fe022013-07-22 16:20:45 -07002923{
2924 struct tcp_sock *tp = tcp_sk(sk);
2925 s32 seq_rtt = -1;
2926
Yuchung Chengbc15afa2013-10-24 08:44:25 -07002927 if (synack_stamp && !tp->total_retrans)
2928 seq_rtt = tcp_time_stamp - synack_stamp;
2929
2930 /* If the ACK acks both the SYNACK and the (Fast Open'd) data packets
2931 * sent in SYN_RECV, SYNACK RTT is the smooth RTT computed in tcp_ack()
2932 */
2933 if (!tp->srtt)
2934 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, seq_rtt, -1);
Yuchung Cheng375fe022013-07-22 16:20:45 -07002935}
2936
Yuchung Cheng9f9843a72013-10-31 11:07:31 -07002937static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked, u32 in_flight)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002938{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002939 const struct inet_connection_sock *icsk = inet_csk(sk);
Yuchung Cheng9f9843a72013-10-31 11:07:31 -07002940 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked, in_flight);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002941 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002942}
2943
Linus Torvalds1da177e2005-04-16 15:20:36 -07002944/* Restart timer after forward progress on connection.
2945 * RFC2988 recommends to restart timer to now+rto.
2946 */
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00002947void tcp_rearm_rto(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002948{
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002949 const struct inet_connection_sock *icsk = inet_csk(sk);
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00002950 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002951
Jerry Chu168a8f52012-08-31 12:29:13 +00002952 /* If the retrans timer is currently being used by Fast Open
2953 * for SYN-ACK retrans purpose, stay put.
2954 */
2955 if (tp->fastopen_rsk)
2956 return;
2957
Linus Torvalds1da177e2005-04-16 15:20:36 -07002958 if (!tp->packets_out) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002959 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002960 } else {
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00002961 u32 rto = inet_csk(sk)->icsk_rto;
2962 /* Offset the time elapsed after installing regular RTO */
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002963 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
2964 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00002965 struct sk_buff *skb = tcp_write_queue_head(sk);
2966 const u32 rto_time_stamp = TCP_SKB_CB(skb)->when + rto;
2967 s32 delta = (s32)(rto_time_stamp - tcp_time_stamp);
2968 /* delta may not be positive if the socket is locked
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002969 * when the retrans timer fires and is rescheduled.
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00002970 */
2971 if (delta > 0)
2972 rto = delta;
2973 }
2974 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
2975 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002976 }
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00002977}
2978
2979/* This function is called when the delayed ER timer fires. TCP enters
2980 * fast recovery and performs fast-retransmit.
2981 */
2982void tcp_resume_early_retransmit(struct sock *sk)
2983{
2984 struct tcp_sock *tp = tcp_sk(sk);
2985
2986 tcp_rearm_rto(sk);
2987
2988 /* Stop if ER is disabled after the delayed ER timer is scheduled */
2989 if (!tp->do_early_retrans)
2990 return;
2991
2992 tcp_enter_recovery(sk, false);
2993 tcp_update_scoreboard(sk, 1);
2994 tcp_xmit_retransmit_queue(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002995}
2996
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07002997/* If we get here, the whole TSO packet has not been acked. */
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07002998static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002999{
3000 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003001 u32 packets_acked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003002
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003003 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003004
3005 packets_acked = tcp_skb_pcount(skb);
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003006 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003007 return 0;
3008 packets_acked -= tcp_skb_pcount(skb);
3009
3010 if (packets_acked) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003011 BUG_ON(tcp_skb_pcount(skb) == 0);
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003012 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003013 }
3014
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003015 return packets_acked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003016}
3017
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003018/* Remove acknowledged frames from the retransmission queue. If our packet
3019 * is before the ack sequence we can discard it as it's confirmed to have
3020 * arrived at the other end.
3021 */
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07003022static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
Yuchung Chenged084952013-07-22 16:20:48 -07003023 u32 prior_snd_una, s32 sack_rtt)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003024{
3025 struct tcp_sock *tp = tcp_sk(sk);
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08003026 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003027 struct sk_buff *skb;
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003028 u32 now = tcp_time_stamp;
Peter Senna Tschudin34a6eda2013-10-02 14:19:51 +02003029 bool fully_acked = true;
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003030 int flag = 0;
Ilpo Järvinen72018832007-12-30 04:37:55 -08003031 u32 pkts_acked = 0;
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003032 u32 reord = tp->packets_out;
Ilpo Järvinen90638a02008-09-20 21:25:52 -07003033 u32 prior_sacked = tp->sacked_out;
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003034 s32 seq_rtt = -1;
Gavin McCullagh2072c222007-12-29 19:11:21 -08003035 s32 ca_seq_rtt = -1;
Ilpo Järvinenb9ce2042007-06-15 15:08:43 -07003036 ktime_t last_ackt = net_invalid_timestamp();
Yuchung Cheng2f715c12013-10-24 08:59:27 -07003037 bool rtt_update;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003038
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003039 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09003040 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
Ilpo Järvinen72018832007-12-30 04:37:55 -08003041 u32 acked_pcount;
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003042 u8 sacked = scb->sacked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003043
Gavin McCullagh2072c222007-12-29 19:11:21 -08003044 /* Determine how many packets and what bytes were acked, tso and else */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003045 if (after(scb->end_seq, tp->snd_una)) {
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003046 if (tcp_skb_pcount(skb) == 1 ||
3047 !after(tp->snd_una, scb->seq))
3048 break;
3049
Ilpo Järvinen72018832007-12-30 04:37:55 -08003050 acked_pcount = tcp_tso_acked(sk, skb);
3051 if (!acked_pcount)
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003052 break;
3053
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003054 fully_acked = false;
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003055 } else {
Ilpo Järvinen72018832007-12-30 04:37:55 -08003056 acked_pcount = tcp_skb_pcount(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003057 }
3058
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003059 if (sacked & TCPCB_RETRANS) {
3060 if (sacked & TCPCB_SACKED_RETRANS)
Ilpo Järvinen72018832007-12-30 04:37:55 -08003061 tp->retrans_out -= acked_pcount;
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003062 flag |= FLAG_RETRANS_DATA_ACKED;
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003063 } else {
Gavin McCullagh2072c222007-12-29 19:11:21 -08003064 ca_seq_rtt = now - scb->when;
3065 last_ackt = skb->tstamp;
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003066 if (seq_rtt < 0) {
Gavin McCullagh2072c222007-12-29 19:11:21 -08003067 seq_rtt = ca_seq_rtt;
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003068 }
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003069 if (!(sacked & TCPCB_SACKED_ACKED))
Ilpo Järvinen72018832007-12-30 04:37:55 -08003070 reord = min(pkts_acked, reord);
Yuchung Chenge33099f2013-03-20 13:33:00 +00003071 if (!after(scb->end_seq, tp->high_seq))
3072 flag |= FLAG_ORIG_SACK_ACKED;
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08003073 }
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003074
3075 if (sacked & TCPCB_SACKED_ACKED)
Ilpo Järvinen72018832007-12-30 04:37:55 -08003076 tp->sacked_out -= acked_pcount;
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003077 if (sacked & TCPCB_LOST)
Ilpo Järvinen72018832007-12-30 04:37:55 -08003078 tp->lost_out -= acked_pcount;
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003079
Ilpo Järvinen72018832007-12-30 04:37:55 -08003080 tp->packets_out -= acked_pcount;
3081 pkts_acked += acked_pcount;
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003082
Ilpo Järvinen009a2e32007-09-20 11:34:38 -07003083 /* Initial outgoing SYN's get put onto the write_queue
3084 * just like anything else we transmit. It is not
3085 * true data, and if we misinform our callers that
3086 * this ACK acks real data, we will erroneously exit
3087 * connection startup slow start one packet too
3088 * quickly. This is severely frowned upon behavior.
3089 */
Eric Dumazet4de075e2011-09-27 13:25:05 -04003090 if (!(scb->tcp_flags & TCPHDR_SYN)) {
Ilpo Järvinen009a2e32007-09-20 11:34:38 -07003091 flag |= FLAG_DATA_ACKED;
3092 } else {
3093 flag |= FLAG_SYN_ACKED;
3094 tp->retrans_stamp = 0;
3095 }
3096
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003097 if (!fully_acked)
3098 break;
3099
David S. Millerfe067e82007-03-07 12:12:44 -08003100 tcp_unlink_write_queue(skb, sk);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08003101 sk_wmem_free_skb(sk, skb);
Ilpo Järvinenef9da472008-09-20 21:25:15 -07003102 if (skb == tp->retransmit_skb_hint)
3103 tp->retransmit_skb_hint = NULL;
Ilpo Järvinen90638a02008-09-20 21:25:52 -07003104 if (skb == tp->lost_skb_hint)
3105 tp->lost_skb_hint = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003106 }
3107
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07003108 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3109 tp->snd_up = tp->snd_una;
3110
Ilpo Järvinencadbd032007-12-31 04:49:21 -08003111 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3112 flag |= FLAG_SACK_RENEGING;
3113
Yuchung Cheng2f715c12013-10-24 08:59:27 -07003114 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt, sack_rtt);
Yuchung Chenged084952013-07-22 16:20:48 -07003115
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003116 if (flag & FLAG_ACKED) {
Stephen Hemminger164891a2007-04-23 22:26:16 -07003117 const struct tcp_congestion_ops *ca_ops
3118 = inet_csk(sk)->icsk_ca_ops;
3119
Yuchung Cheng2f715c12013-10-24 08:59:27 -07003120 tcp_rearm_rto(sk);
Ilpo Järvinen72211e92009-03-14 14:23:04 +00003121 if (unlikely(icsk->icsk_mtup.probe_size &&
3122 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3123 tcp_mtup_probe_success(sk);
3124 }
3125
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003126 if (tcp_is_reno(tp)) {
3127 tcp_remove_reno_sacks(sk, pkts_acked);
3128 } else {
Ilpo Järvinen59a08cb2009-02-28 04:44:28 +00003129 int delta;
3130
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003131 /* Non-retransmitted hole got filled? That's reordering */
3132 if (reord < prior_fackets)
3133 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
Ilpo Järvinen90638a02008-09-20 21:25:52 -07003134
Ilpo Järvinen59a08cb2009-02-28 04:44:28 +00003135 delta = tcp_is_fack(tp) ? pkts_acked :
3136 prior_sacked - tp->sacked_out;
3137 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003138 }
3139
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07003140 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08003141
Stephen Hemminger30cfd0b2007-07-25 23:49:34 -07003142 if (ca_ops->pkts_acked) {
3143 s32 rtt_us = -1;
Ilpo Järvinenb9ce2042007-06-15 15:08:43 -07003144
Stephen Hemminger30cfd0b2007-07-25 23:49:34 -07003145 /* Is the ACK triggering packet unambiguous? */
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003146 if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
Stephen Hemminger30cfd0b2007-07-25 23:49:34 -07003147 /* High resolution needed and available? */
3148 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
3149 !ktime_equal(last_ackt,
3150 net_invalid_timestamp()))
3151 rtt_us = ktime_us_delta(ktime_get_real(),
3152 last_ackt);
stephen hemmingerfebf0812011-03-14 07:52:12 +00003153 else if (ca_seq_rtt >= 0)
Gavin McCullagh2072c222007-12-29 19:11:21 -08003154 rtt_us = jiffies_to_usecs(ca_seq_rtt);
Stephen Hemminger30cfd0b2007-07-25 23:49:34 -07003155 }
3156
3157 ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
3158 }
Yuchung Cheng2f715c12013-10-24 08:59:27 -07003159 } else if (skb && rtt_update && sack_rtt >= 0 &&
3160 sack_rtt > (s32)(now - TCP_SKB_CB(skb)->when)) {
3161 /* Do not re-arm RTO if the sack RTT is measured from data sent
3162 * after when the head was last (re)transmitted. Otherwise the
3163 * timeout may continue to extend in loss recovery.
3164 */
3165 tcp_rearm_rto(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003166 }
3167
3168#if FASTRETRANS_DEBUG > 0
Ilpo Järvinen547b7922008-07-25 21:43:18 -07003169 WARN_ON((int)tp->sacked_out < 0);
3170 WARN_ON((int)tp->lost_out < 0);
3171 WARN_ON((int)tp->retrans_out < 0);
Ilpo Järvinene60402d2007-08-09 15:14:46 +03003172 if (!tp->packets_out && tcp_is_sack(tp)) {
Stephen Hemmingercfcabdc2007-10-09 01:59:42 -07003173 icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003174 if (tp->lost_out) {
Joe Perches91df42b2012-05-15 14:11:54 +00003175 pr_debug("Leak l=%u %d\n",
3176 tp->lost_out, icsk->icsk_ca_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003177 tp->lost_out = 0;
3178 }
3179 if (tp->sacked_out) {
Joe Perches91df42b2012-05-15 14:11:54 +00003180 pr_debug("Leak s=%u %d\n",
3181 tp->sacked_out, icsk->icsk_ca_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003182 tp->sacked_out = 0;
3183 }
3184 if (tp->retrans_out) {
Joe Perches91df42b2012-05-15 14:11:54 +00003185 pr_debug("Leak r=%u %d\n",
3186 tp->retrans_out, icsk->icsk_ca_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003187 tp->retrans_out = 0;
3188 }
3189 }
3190#endif
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003191 return flag;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003192}
3193
3194static void tcp_ack_probe(struct sock *sk)
3195{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003196 const struct tcp_sock *tp = tcp_sk(sk);
3197 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003198
3199 /* Was it a usable window open? */
3200
Ilpo Järvinen90840de2007-12-31 04:48:41 -08003201 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003202 icsk->icsk_backoff = 0;
3203 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003204 /* Socket must be waked up by subsequent tcp_data_snd_check().
3205 * This function is not for random using!
3206 */
3207 } else {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003208 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07003209 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3210 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003211 }
3212}
3213
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00003214static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003215{
Eric Dumazeta02cec22010-09-22 20:43:57 +00003216 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3217 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003218}
3219
Yuchung Cheng0f7cc9a2013-08-21 17:29:23 -07003220/* Decide wheather to run the increase function of congestion control. */
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00003221static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003222{
Yuchung Cheng0f7cc9a2013-08-21 17:29:23 -07003223 if (tcp_in_cwnd_reduction(sk))
3224 return false;
3225
3226 /* If reordering is high then always grow cwnd whenever data is
3227 * delivered regardless of its ordering. Otherwise stay conservative
Yuchung Cheng16edfe72013-09-05 15:36:09 -07003228 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
Yuchung Cheng0f7cc9a2013-08-21 17:29:23 -07003229 * new SACK or ECE mark may first advance cwnd here and later reduce
3230 * cwnd in tcp_fastretrans_alert() based on more states.
3231 */
3232 if (tcp_sk(sk)->reordering > sysctl_tcp_reordering)
3233 return flag & FLAG_FORWARD_PROGRESS;
3234
Yuchung Cheng16edfe72013-09-05 15:36:09 -07003235 return flag & FLAG_DATA_ACKED;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003236}
3237
3238/* Check that window update is acceptable.
3239 * The function assumes that snd_una<=ack<=snd_next.
3240 */
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00003241static inline bool tcp_may_update_window(const struct tcp_sock *tp,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003242 const u32 ack, const u32 ack_seq,
3243 const u32 nwin)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003244{
Eric Dumazeta02cec22010-09-22 20:43:57 +00003245 return after(ack, tp->snd_una) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07003246 after(ack_seq, tp->snd_wl1) ||
Eric Dumazeta02cec22010-09-22 20:43:57 +00003247 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003248}
3249
3250/* Update our send window.
3251 *
3252 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3253 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3254 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003255static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003256 u32 ack_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003257{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003258 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003259 int flag = 0;
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07003260 u32 nwin = ntohs(tcp_hdr(skb)->window);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003261
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07003262 if (likely(!tcp_hdr(skb)->syn))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003263 nwin <<= tp->rx_opt.snd_wscale;
3264
3265 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3266 flag |= FLAG_WIN_UPDATE;
Hantzis Fotisee7537b2009-03-02 22:42:02 -08003267 tcp_update_wl(tp, ack_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003268
3269 if (tp->snd_wnd != nwin) {
3270 tp->snd_wnd = nwin;
3271
3272 /* Note, it is the only place, where
3273 * fast path is recovered for sending TCP.
3274 */
Herbert Xu2ad41062005-10-27 18:47:46 +10003275 tp->pred_flags = 0;
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003276 tcp_fast_path_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003277
3278 if (nwin > tp->max_window) {
3279 tp->max_window = nwin;
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08003280 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003281 }
3282 }
3283 }
3284
3285 tp->snd_una = ack;
3286
3287 return flag;
3288}
3289
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003290/* RFC 5961 7 [ACK Throttling] */
3291static void tcp_send_challenge_ack(struct sock *sk)
3292{
3293 /* unprotected vars, we dont care of overwrites */
3294 static u32 challenge_timestamp;
3295 static unsigned int challenge_count;
3296 u32 now = jiffies / HZ;
3297
3298 if (now != challenge_timestamp) {
3299 challenge_timestamp = now;
3300 challenge_count = 0;
3301 }
3302 if (++challenge_count <= sysctl_tcp_challenge_ack_limit) {
3303 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK);
3304 tcp_send_ack(sk);
3305 }
3306}
3307
Eric Dumazet12fb3dd2013-04-19 07:19:48 +00003308static void tcp_store_ts_recent(struct tcp_sock *tp)
3309{
3310 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3311 tp->rx_opt.ts_recent_stamp = get_seconds();
3312}
3313
3314static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3315{
3316 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3317 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3318 * extra check below makes sure this can only happen
3319 * for pure ACK frames. -DaveM
3320 *
3321 * Not only, also it occurs for expired timestamps.
3322 */
3323
3324 if (tcp_paws_check(&tp->rx_opt, 0))
3325 tcp_store_ts_recent(tp);
3326 }
3327}
3328
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003329/* This routine deals with acks during a TLP episode.
3330 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3331 */
3332static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3333{
3334 struct tcp_sock *tp = tcp_sk(sk);
3335 bool is_tlp_dupack = (ack == tp->tlp_high_seq) &&
3336 !(flag & (FLAG_SND_UNA_ADVANCED |
3337 FLAG_NOT_DUP | FLAG_DATA_SACKED));
3338
3339 /* Mark the end of TLP episode on receiving TLP dupack or when
3340 * ack is after tlp_high_seq.
3341 */
3342 if (is_tlp_dupack) {
3343 tp->tlp_high_seq = 0;
3344 return;
3345 }
3346
3347 if (after(ack, tp->tlp_high_seq)) {
3348 tp->tlp_high_seq = 0;
3349 /* Don't reduce cwnd if DSACK arrives for TLP retrans. */
3350 if (!(flag & FLAG_DSACKING_ACK)) {
3351 tcp_init_cwnd_reduction(sk, true);
3352 tcp_set_ca_state(sk, TCP_CA_CWR);
3353 tcp_end_cwnd_reduction(sk);
Yuchung Cheng031afe42013-10-12 10:16:27 -07003354 tcp_try_keep_open(sk);
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003355 NET_INC_STATS_BH(sock_net(sk),
3356 LINUX_MIB_TCPLOSSPROBERECOVERY);
3357 }
3358 }
3359}
3360
Linus Torvalds1da177e2005-04-16 15:20:36 -07003361/* This routine deals with incoming acks, but not outgoing ones. */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003362static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003363{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003364 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003365 struct tcp_sock *tp = tcp_sk(sk);
3366 u32 prior_snd_una = tp->snd_una;
3367 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3368 u32 ack = TCP_SKB_CB(skb)->ack_seq;
Neal Cardwell7d2b55f2011-11-16 08:58:01 +00003369 bool is_dupack = false;
Eric Dumazet95bd09e2013-08-27 05:46:32 -07003370 u32 prior_in_flight, prior_cwnd = tp->snd_cwnd, prior_rtt = tp->srtt;
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003371 u32 prior_fackets;
Nandita Dukkipati35f079e2013-05-21 15:12:07 +00003372 int prior_packets = tp->packets_out;
Yuchung Cheng68049732013-05-29 14:20:11 +00003373 const int prior_unsacked = tp->packets_out - tp->sacked_out;
3374 int acked = 0; /* Number of packets newly acked */
Yuchung Cheng59c9af42013-07-22 16:20:47 -07003375 s32 sack_rtt = -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003376
John Dykstra96e0bf42009-03-22 21:49:57 -07003377 /* If the ack is older than previous acks
Linus Torvalds1da177e2005-04-16 15:20:36 -07003378 * then we can probably ignore it.
3379 */
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003380 if (before(ack, prior_snd_una)) {
3381 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3382 if (before(ack, prior_snd_una - tp->max_window)) {
3383 tcp_send_challenge_ack(sk);
3384 return -1;
3385 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003386 goto old_ack;
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003387 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003388
John Dykstra96e0bf42009-03-22 21:49:57 -07003389 /* If the ack includes data we haven't sent yet, discard
3390 * this segment (RFC793 Section 3.9).
3391 */
3392 if (after(ack, tp->snd_nxt))
3393 goto invalid_ack;
3394
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00003395 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
3396 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00003397 tcp_rearm_rto(sk);
3398
Ilpo Järvinen2e605292007-08-02 19:46:58 -07003399 if (after(ack, prior_snd_una))
3400 flag |= FLAG_SND_UNA_ADVANCED;
3401
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003402 prior_fackets = tp->fackets_out;
Ilpo Järvinen52d34082007-12-05 02:21:35 -08003403 prior_in_flight = tcp_packets_in_flight(tp);
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003404
Eric Dumazet12fb3dd2013-04-19 07:19:48 +00003405 /* ts_recent update must be made after we are sure that the packet
3406 * is in window.
3407 */
3408 if (flag & FLAG_UPDATE_TS_RECENT)
3409 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3410
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003411 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003412 /* Window is constant, pure forward advance.
3413 * No more checks are required.
3414 * Note, we use the fact that SND.UNA>=SND.WL2.
3415 */
Hantzis Fotisee7537b2009-03-02 22:42:02 -08003416 tcp_update_wl(tp, ack_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003417 tp->snd_una = ack;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003418 flag |= FLAG_WIN_UPDATE;
3419
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003420 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
Stephen Hemminger317a76f2005-06-23 12:19:55 -07003421
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07003422 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003423 } else {
3424 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3425 flag |= FLAG_DATA;
3426 else
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07003427 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003428
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003429 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003430
3431 if (TCP_SKB_CB(skb)->sacked)
Yuchung Cheng59c9af42013-07-22 16:20:47 -07003432 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3433 &sack_rtt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003434
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07003435 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003436 flag |= FLAG_ECE;
3437
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003438 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003439 }
3440
3441 /* We passed data and got it acked, remove any soft error
3442 * log. Something worked...
3443 */
3444 sk->sk_err_soft = 0;
David S. Miller4b53fb62008-07-23 16:38:45 -07003445 icsk->icsk_probes_out = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003446 tp->rcv_tstamp = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003447 if (!prior_packets)
3448 goto no_queue;
3449
Linus Torvalds1da177e2005-04-16 15:20:36 -07003450 /* See if we can take anything off of the retransmit queue. */
Yuchung Cheng68049732013-05-29 14:20:11 +00003451 acked = tp->packets_out;
Yuchung Chenged084952013-07-22 16:20:48 -07003452 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una, sack_rtt);
Yuchung Cheng68049732013-05-29 14:20:11 +00003453 acked -= tp->packets_out;
Nandita Dukkipatia262f0c2011-08-21 20:21:57 +00003454
Yuchung Cheng0f7cc9a2013-08-21 17:29:23 -07003455 /* Advance cwnd if state allows */
3456 if (tcp_may_raise_cwnd(sk, flag))
Yuchung Cheng9f9843a72013-10-31 11:07:31 -07003457 tcp_cong_avoid(sk, ack, acked, prior_in_flight);
Yuchung Cheng0f7cc9a2013-08-21 17:29:23 -07003458
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003459 if (tcp_ack_is_dubious(sk, flag)) {
Neal Cardwell7d2b55f2011-11-16 08:58:01 +00003460 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
Yuchung Cheng68049732013-05-29 14:20:11 +00003461 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3462 is_dupack, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003463 }
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003464 if (tp->tlp_high_seq)
3465 tcp_process_tlp_ack(sk, ack, flag);
3466
David S. Miller5110effe2012-07-02 02:21:03 -07003467 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) {
3468 struct dst_entry *dst = __sk_dst_get(sk);
3469 if (dst)
3470 dst_confirm(dst);
3471 }
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00003472
3473 if (icsk->icsk_pending == ICSK_TIME_RETRANS)
3474 tcp_schedule_loss_probe(sk);
Eric Dumazet95bd09e2013-08-27 05:46:32 -07003475 if (tp->srtt != prior_rtt || tp->snd_cwnd != prior_cwnd)
3476 tcp_update_pacing_rate(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003477 return 1;
3478
3479no_queue:
Neal Cardwell5628adf2011-11-16 08:58:02 +00003480 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3481 if (flag & FLAG_DSACKING_ACK)
Yuchung Cheng68049732013-05-29 14:20:11 +00003482 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3483 is_dupack, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003484 /* If this ack opens up a zero window, clear backoff. It was
3485 * being used to time the probes, and is probably far higher than
3486 * it needs to be for normal retransmission.
3487 */
David S. Millerfe067e82007-03-07 12:12:44 -08003488 if (tcp_send_head(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003489 tcp_ack_probe(sk);
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003490
3491 if (tp->tlp_high_seq)
3492 tcp_process_tlp_ack(sk, ack, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003493 return 1;
3494
John Dykstra96e0bf42009-03-22 21:49:57 -07003495invalid_ack:
3496 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3497 return -1;
3498
Linus Torvalds1da177e2005-04-16 15:20:36 -07003499old_ack:
Neal Cardwelle95ae2f2011-11-16 08:58:03 +00003500 /* If data was SACKed, tag it and see if we should send more data.
3501 * If data was DSACKed, see if we can undo a cwnd reduction.
3502 */
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07003503 if (TCP_SKB_CB(skb)->sacked) {
Yuchung Cheng59c9af42013-07-22 16:20:47 -07003504 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3505 &sack_rtt);
Yuchung Cheng68049732013-05-29 14:20:11 +00003506 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3507 is_dupack, flag);
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07003508 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003509
John Dykstra96e0bf42009-03-22 21:49:57 -07003510 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003511 return 0;
3512}
3513
Linus Torvalds1da177e2005-04-16 15:20:36 -07003514/* Look for tcp options. Normally only called on SYN and SYNACK packets.
3515 * But, this can also be called on packets in the established flow when
3516 * the fast version below fails.
3517 */
Christoph Paasch1a2c6182013-03-17 08:23:34 +00003518void tcp_parse_options(const struct sk_buff *skb,
3519 struct tcp_options_received *opt_rx, int estab,
Yuchung Cheng2100c8d2012-07-19 06:43:05 +00003520 struct tcp_fastopen_cookie *foc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003521{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003522 const unsigned char *ptr;
3523 const struct tcphdr *th = tcp_hdr(skb);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003524 int length = (th->doff * 4) - sizeof(struct tcphdr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003525
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003526 ptr = (const unsigned char *)(th + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003527 opt_rx->saw_tstamp = 0;
3528
Stephen Hemminger2de979b2007-03-08 20:45:19 -08003529 while (length > 0) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003530 int opcode = *ptr++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003531 int opsize;
3532
3533 switch (opcode) {
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003534 case TCPOPT_EOL:
3535 return;
3536 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3537 length--;
3538 continue;
3539 default:
3540 opsize = *ptr++;
3541 if (opsize < 2) /* "silly options" */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003542 return;
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003543 if (opsize > length)
3544 return; /* don't parse partial options */
3545 switch (opcode) {
3546 case TCPOPT_MSS:
3547 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07003548 u16 in_mss = get_unaligned_be16(ptr);
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003549 if (in_mss) {
3550 if (opt_rx->user_mss &&
3551 opt_rx->user_mss < in_mss)
3552 in_mss = opt_rx->user_mss;
3553 opt_rx->mss_clamp = in_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003554 }
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07003555 }
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003556 break;
3557 case TCPOPT_WINDOW:
3558 if (opsize == TCPOLEN_WINDOW && th->syn &&
David S. Millerbb5b7c12009-12-15 20:56:42 -08003559 !estab && sysctl_tcp_window_scaling) {
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003560 __u8 snd_wscale = *(__u8 *)ptr;
3561 opt_rx->wscale_ok = 1;
3562 if (snd_wscale > 14) {
Joe Perchese87cc472012-05-13 21:56:26 +00003563 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3564 __func__,
3565 snd_wscale);
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003566 snd_wscale = 14;
3567 }
3568 opt_rx->snd_wscale = snd_wscale;
3569 }
3570 break;
3571 case TCPOPT_TIMESTAMP:
3572 if ((opsize == TCPOLEN_TIMESTAMP) &&
3573 ((estab && opt_rx->tstamp_ok) ||
David S. Millerbb5b7c12009-12-15 20:56:42 -08003574 (!estab && sysctl_tcp_timestamps))) {
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003575 opt_rx->saw_tstamp = 1;
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07003576 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3577 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003578 }
3579 break;
3580 case TCPOPT_SACK_PERM:
3581 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
David S. Millerbb5b7c12009-12-15 20:56:42 -08003582 !estab && sysctl_tcp_sack) {
Vijay Subramanianab562222011-12-20 13:23:24 +00003583 opt_rx->sack_ok = TCP_SACK_SEEN;
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003584 tcp_sack_reset(opt_rx);
3585 }
3586 break;
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07003587
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003588 case TCPOPT_SACK:
3589 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3590 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3591 opt_rx->sack_ok) {
3592 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3593 }
3594 break;
3595#ifdef CONFIG_TCP_MD5SIG
3596 case TCPOPT_MD5SIG:
3597 /*
3598 * The MD5 Hash has already been
3599 * checked (see tcp_v{4,6}_do_rcv()).
3600 */
3601 break;
3602#endif
Yuchung Cheng2100c8d2012-07-19 06:43:05 +00003603 case TCPOPT_EXP:
3604 /* Fast Open option shares code 254 using a
3605 * 16 bits magic number. It's valid only in
3606 * SYN or SYN-ACK with an even size.
3607 */
3608 if (opsize < TCPOLEN_EXP_FASTOPEN_BASE ||
3609 get_unaligned_be16(ptr) != TCPOPT_FASTOPEN_MAGIC ||
3610 foc == NULL || !th->syn || (opsize & 1))
3611 break;
3612 foc->len = opsize - TCPOLEN_EXP_FASTOPEN_BASE;
3613 if (foc->len >= TCP_FASTOPEN_COOKIE_MIN &&
3614 foc->len <= TCP_FASTOPEN_COOKIE_MAX)
3615 memcpy(foc->val, ptr + 2, foc->len);
3616 else if (foc->len != 0)
3617 foc->len = -1;
3618 break;
3619
3620 }
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003621 ptr += opsize-2;
3622 length -= opsize;
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07003623 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003624 }
3625}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00003626EXPORT_SYMBOL(tcp_parse_options);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003627
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003628static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
Ilpo Järvinena4356b22008-08-23 05:12:29 -07003629{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003630 const __be32 *ptr = (const __be32 *)(th + 1);
Ilpo Järvinena4356b22008-08-23 05:12:29 -07003631
3632 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3633 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3634 tp->rx_opt.saw_tstamp = 1;
3635 ++ptr;
3636 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3637 ++ptr;
Andrew Vagine3e12022013-08-27 12:21:55 +04003638 if (*ptr)
3639 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3640 else
3641 tp->rx_opt.rcv_tsecr = 0;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003642 return true;
Ilpo Järvinena4356b22008-08-23 05:12:29 -07003643 }
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003644 return false;
Ilpo Järvinena4356b22008-08-23 05:12:29 -07003645}
3646
Linus Torvalds1da177e2005-04-16 15:20:36 -07003647/* Fast parse options. This hopes to only see timestamps.
3648 * If it is wrong it falls back on tcp_parse_options().
3649 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003650static bool tcp_fast_parse_options(const struct sk_buff *skb,
Christoph Paasch1a2c6182013-03-17 08:23:34 +00003651 const struct tcphdr *th, struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003652{
William Allen Simpson4957faade2009-12-02 18:25:27 +00003653 /* In the spirit of fast parsing, compare doff directly to constant
3654 * values. Because equality is used, short doff can be ignored here.
3655 */
3656 if (th->doff == (sizeof(*th) / 4)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003657 tp->rx_opt.saw_tstamp = 0;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003658 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003659 } else if (tp->rx_opt.tstamp_ok &&
William Allen Simpson4957faade2009-12-02 18:25:27 +00003660 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
Ilpo Järvinena4356b22008-08-23 05:12:29 -07003661 if (tcp_parse_aligned_timestamp(tp, th))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003662 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003663 }
Andrey Vaginee684b62013-02-11 05:50:19 +00003664
Christoph Paasch1a2c6182013-03-17 08:23:34 +00003665 tcp_parse_options(skb, &tp->rx_opt, 1, NULL);
Andrew Vagine3e12022013-08-27 12:21:55 +04003666 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
Andrey Vaginee684b62013-02-11 05:50:19 +00003667 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3668
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003669 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003670}
3671
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003672#ifdef CONFIG_TCP_MD5SIG
3673/*
3674 * Parse MD5 Signature option
3675 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003676const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003677{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003678 int length = (th->doff << 2) - sizeof(*th);
3679 const u8 *ptr = (const u8 *)(th + 1);
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003680
3681 /* If the TCP option is too short, we can short cut */
3682 if (length < TCPOLEN_MD5SIG)
3683 return NULL;
3684
3685 while (length > 0) {
3686 int opcode = *ptr++;
3687 int opsize;
3688
Weilong Chena22318e2013-12-23 14:37:26 +08003689 switch (opcode) {
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003690 case TCPOPT_EOL:
3691 return NULL;
3692 case TCPOPT_NOP:
3693 length--;
3694 continue;
3695 default:
3696 opsize = *ptr++;
3697 if (opsize < 2 || opsize > length)
3698 return NULL;
3699 if (opcode == TCPOPT_MD5SIG)
Dmitry Popovba78e2d2010-08-07 20:24:28 -07003700 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003701 }
3702 ptr += opsize - 2;
3703 length -= opsize;
3704 }
3705 return NULL;
3706}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00003707EXPORT_SYMBOL(tcp_parse_md5sig_option);
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003708#endif
3709
Linus Torvalds1da177e2005-04-16 15:20:36 -07003710/* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3711 *
3712 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3713 * it can pass through stack. So, the following predicate verifies that
3714 * this segment is not used for anything but congestion avoidance or
3715 * fast retransmit. Moreover, we even are able to eliminate most of such
3716 * second order effects, if we apply some small "replay" window (~RTO)
3717 * to timestamp space.
3718 *
3719 * All these measures still do not guarantee that we reject wrapped ACKs
3720 * on networks with high bandwidth, when sequence space is recycled fastly,
3721 * but it guarantees that such events will be very rare and do not affect
3722 * connection seriously. This doesn't look nice, but alas, PAWS is really
3723 * buggy extension.
3724 *
3725 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3726 * states that events when retransmit arrives after original data are rare.
3727 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3728 * the biggest problem on large power networks even with minor reordering.
3729 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3730 * up to bandwidth of 18Gigabit/sec. 8) ]
3731 */
3732
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003733static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003734{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003735 const struct tcp_sock *tp = tcp_sk(sk);
3736 const struct tcphdr *th = tcp_hdr(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003737 u32 seq = TCP_SKB_CB(skb)->seq;
3738 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3739
3740 return (/* 1. Pure ACK with correct sequence number. */
3741 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3742
3743 /* 2. ... and duplicate ACK. */
3744 ack == tp->snd_una &&
3745
3746 /* 3. ... and does not update window. */
3747 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3748
3749 /* 4. ... and sits in replay window. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003750 (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 -07003751}
3752
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00003753static inline bool tcp_paws_discard(const struct sock *sk,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003754 const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003755{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003756 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinenc887e6d2009-03-14 14:23:03 +00003757
3758 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
3759 !tcp_disordered_ack(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003760}
3761
3762/* Check segment sequence number for validity.
3763 *
3764 * Segment controls are considered valid, if the segment
3765 * fits to the window after truncation to the window. Acceptability
3766 * of data (and SYN, FIN, of course) is checked separately.
3767 * See tcp_data_queue(), for example.
3768 *
3769 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3770 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3771 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3772 * (borrowed from freebsd)
3773 */
3774
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00003775static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003776{
3777 return !before(end_seq, tp->rcv_wup) &&
3778 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3779}
3780
3781/* When we get a reset we do this. */
Jerry Chu10467162012-08-31 12:29:11 +00003782void tcp_reset(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003783{
3784 /* We want the right error as BSD sees it (and indeed as we do). */
3785 switch (sk->sk_state) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003786 case TCP_SYN_SENT:
3787 sk->sk_err = ECONNREFUSED;
3788 break;
3789 case TCP_CLOSE_WAIT:
3790 sk->sk_err = EPIPE;
3791 break;
3792 case TCP_CLOSE:
3793 return;
3794 default:
3795 sk->sk_err = ECONNRESET;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003796 }
Tom Marshalla4d25802010-09-20 15:42:05 -07003797 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3798 smp_wmb();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003799
3800 if (!sock_flag(sk, SOCK_DEAD))
3801 sk->sk_error_report(sk);
3802
3803 tcp_done(sk);
3804}
3805
3806/*
3807 * Process the FIN bit. This now behaves as it is supposed to work
3808 * and the FIN takes effect when it is validly part of sequence
3809 * space. Not before when we get holes.
3810 *
3811 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3812 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3813 * TIME-WAIT)
3814 *
3815 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3816 * close and we go into CLOSING (and later onto TIME-WAIT)
3817 *
3818 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3819 */
Eric Dumazet20c4cb72011-10-20 17:44:03 -04003820static void tcp_fin(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003821{
3822 struct tcp_sock *tp = tcp_sk(sk);
Cong Wangbcefe172013-06-15 09:39:18 +08003823 const struct dst_entry *dst;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003824
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003825 inet_csk_schedule_ack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003826
3827 sk->sk_shutdown |= RCV_SHUTDOWN;
3828 sock_set_flag(sk, SOCK_DONE);
3829
3830 switch (sk->sk_state) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003831 case TCP_SYN_RECV:
3832 case TCP_ESTABLISHED:
3833 /* Move to CLOSE_WAIT */
3834 tcp_set_state(sk, TCP_CLOSE_WAIT);
Cong Wangbcefe172013-06-15 09:39:18 +08003835 dst = __sk_dst_get(sk);
3836 if (!dst || !dst_metric(dst, RTAX_QUICKACK))
3837 inet_csk(sk)->icsk_ack.pingpong = 1;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003838 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003839
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003840 case TCP_CLOSE_WAIT:
3841 case TCP_CLOSING:
3842 /* Received a retransmission of the FIN, do
3843 * nothing.
3844 */
3845 break;
3846 case TCP_LAST_ACK:
3847 /* RFC793: Remain in the LAST-ACK state. */
3848 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003849
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003850 case TCP_FIN_WAIT1:
3851 /* This case occurs when a simultaneous close
3852 * happens, we must ack the received FIN and
3853 * enter the CLOSING state.
3854 */
3855 tcp_send_ack(sk);
3856 tcp_set_state(sk, TCP_CLOSING);
3857 break;
3858 case TCP_FIN_WAIT2:
3859 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3860 tcp_send_ack(sk);
3861 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3862 break;
3863 default:
3864 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3865 * cases we should never reach this piece of code.
3866 */
Joe Perches058bd4d2012-03-11 18:36:11 +00003867 pr_err("%s: Impossible, sk->sk_state=%d\n",
Harvey Harrison0dc47872008-03-05 20:47:47 -08003868 __func__, sk->sk_state);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003869 break;
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07003870 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003871
3872 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3873 * Probably, we should reset in this case. For now drop them.
3874 */
3875 __skb_queue_purge(&tp->out_of_order_queue);
Ilpo Järvinene60402d2007-08-09 15:14:46 +03003876 if (tcp_is_sack(tp))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003877 tcp_sack_reset(&tp->rx_opt);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08003878 sk_mem_reclaim(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003879
3880 if (!sock_flag(sk, SOCK_DEAD)) {
3881 sk->sk_state_change(sk);
3882
3883 /* Do not send POLL_HUP for half duplex close. */
3884 if (sk->sk_shutdown == SHUTDOWN_MASK ||
3885 sk->sk_state == TCP_CLOSE)
Pavel Emelyanov8d8ad9d2007-11-26 20:10:50 +08003886 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003887 else
Pavel Emelyanov8d8ad9d2007-11-26 20:10:50 +08003888 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003889 }
3890}
3891
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003892static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003893 u32 end_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003894{
3895 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3896 if (before(seq, sp->start_seq))
3897 sp->start_seq = seq;
3898 if (after(end_seq, sp->end_seq))
3899 sp->end_seq = end_seq;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003900 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003901 }
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003902 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003903}
3904
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07003905static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003906{
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07003907 struct tcp_sock *tp = tcp_sk(sk);
3908
David S. Millerbb5b7c12009-12-15 20:56:42 -08003909 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07003910 int mib_idx;
3911
Linus Torvalds1da177e2005-04-16 15:20:36 -07003912 if (before(seq, tp->rcv_nxt))
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07003913 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003914 else
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07003915 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
3916
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07003917 NET_INC_STATS_BH(sock_net(sk), mib_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003918
3919 tp->rx_opt.dsack = 1;
3920 tp->duplicate_sack[0].start_seq = seq;
3921 tp->duplicate_sack[0].end_seq = end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003922 }
3923}
3924
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07003925static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003926{
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07003927 struct tcp_sock *tp = tcp_sk(sk);
3928
Linus Torvalds1da177e2005-04-16 15:20:36 -07003929 if (!tp->rx_opt.dsack)
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07003930 tcp_dsack_set(sk, seq, end_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003931 else
3932 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3933}
3934
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003935static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003936{
3937 struct tcp_sock *tp = tcp_sk(sk);
3938
3939 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3940 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07003941 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003942 tcp_enter_quickack_mode(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003943
David S. Millerbb5b7c12009-12-15 20:56:42 -08003944 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003945 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3946
3947 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3948 end_seq = tp->rcv_nxt;
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07003949 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003950 }
3951 }
3952
3953 tcp_send_ack(sk);
3954}
3955
3956/* These routines update the SACK block as out-of-order packets arrive or
3957 * in-order packets close up the sequence space.
3958 */
3959static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
3960{
3961 int this_sack;
3962 struct tcp_sack_block *sp = &tp->selective_acks[0];
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003963 struct tcp_sack_block *swalk = sp + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003964
3965 /* See if the recent change to the first SACK eats into
3966 * or hits the sequence space of other SACK blocks, if so coalesce.
3967 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003968 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003969 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
3970 int i;
3971
3972 /* Zap SWALK, by moving every further SACK up by one slot.
3973 * Decrease num_sacks.
3974 */
3975 tp->rx_opt.num_sacks--;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003976 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
3977 sp[i] = sp[i + 1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003978 continue;
3979 }
3980 this_sack++, swalk++;
3981 }
3982}
3983
Linus Torvalds1da177e2005-04-16 15:20:36 -07003984static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
3985{
3986 struct tcp_sock *tp = tcp_sk(sk);
3987 struct tcp_sack_block *sp = &tp->selective_acks[0];
3988 int cur_sacks = tp->rx_opt.num_sacks;
3989 int this_sack;
3990
3991 if (!cur_sacks)
3992 goto new_sack;
3993
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003994 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003995 if (tcp_sack_extend(sp, seq, end_seq)) {
3996 /* Rotate this_sack to the first one. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003997 for (; this_sack > 0; this_sack--, sp--)
Ilpo Järvinena0bffff2009-03-21 13:36:17 -07003998 swap(*sp, *(sp - 1));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003999 if (cur_sacks > 1)
4000 tcp_sack_maybe_coalesce(tp);
4001 return;
4002 }
4003 }
4004
4005 /* Could not find an adjacent existing SACK, build a new one,
4006 * put it at the front, and shift everyone else down. We
4007 * always know there is at least one SACK present already here.
4008 *
4009 * If the sack array is full, forget about the last one.
4010 */
Adam Langley4389dde2008-07-19 00:07:02 -07004011 if (this_sack >= TCP_NUM_SACKS) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004012 this_sack--;
4013 tp->rx_opt.num_sacks--;
4014 sp--;
4015 }
Stephen Hemminger2de979b2007-03-08 20:45:19 -08004016 for (; this_sack > 0; this_sack--, sp--)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004017 *sp = *(sp - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004018
4019new_sack:
4020 /* Build the new head SACK, and we're done. */
4021 sp->start_seq = seq;
4022 sp->end_seq = end_seq;
4023 tp->rx_opt.num_sacks++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004024}
4025
4026/* RCV.NXT advances, some SACKs should be eaten. */
4027
4028static void tcp_sack_remove(struct tcp_sock *tp)
4029{
4030 struct tcp_sack_block *sp = &tp->selective_acks[0];
4031 int num_sacks = tp->rx_opt.num_sacks;
4032 int this_sack;
4033
4034 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
David S. Millerb03efcf2005-07-08 14:57:23 -07004035 if (skb_queue_empty(&tp->out_of_order_queue)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004036 tp->rx_opt.num_sacks = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004037 return;
4038 }
4039
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004040 for (this_sack = 0; this_sack < num_sacks;) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004041 /* Check if the start of the sack is covered by RCV.NXT. */
4042 if (!before(tp->rcv_nxt, sp->start_seq)) {
4043 int i;
4044
4045 /* RCV.NXT must cover all the block! */
Ilpo Järvinen547b7922008-07-25 21:43:18 -07004046 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004047
4048 /* Zap this SACK, by moving forward any other SACKS. */
Weilong Chena22318e2013-12-23 14:37:26 +08004049 for (i = this_sack+1; i < num_sacks; i++)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004050 tp->selective_acks[i-1] = tp->selective_acks[i];
4051 num_sacks--;
4052 continue;
4053 }
4054 this_sack++;
4055 sp++;
4056 }
Ilpo Järvinen5861f8e2009-03-14 14:23:01 +00004057 tp->rx_opt.num_sacks = num_sacks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004058}
4059
4060/* This one checks to see if we can put data from the
4061 * out_of_order queue into the receive_queue.
4062 */
4063static void tcp_ofo_queue(struct sock *sk)
4064{
4065 struct tcp_sock *tp = tcp_sk(sk);
4066 __u32 dsack_high = tp->rcv_nxt;
4067 struct sk_buff *skb;
4068
4069 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4070 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4071 break;
4072
4073 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4074 __u32 dsack = dsack_high;
4075 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4076 dsack_high = TCP_SKB_CB(skb)->end_seq;
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004077 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004078 }
4079
4080 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
Frans Popb1383382010-03-24 07:57:28 +00004081 SOCK_DEBUG(sk, "ofo packet was already received\n");
David S. Miller8728b832005-08-09 19:25:21 -07004082 __skb_unlink(skb, &tp->out_of_order_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004083 __kfree_skb(skb);
4084 continue;
4085 }
4086 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4087 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4088 TCP_SKB_CB(skb)->end_seq);
4089
David S. Miller8728b832005-08-09 19:25:21 -07004090 __skb_unlink(skb, &tp->out_of_order_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004091 __skb_queue_tail(&sk->sk_receive_queue, skb);
4092 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07004093 if (tcp_hdr(skb)->fin)
Eric Dumazet20c4cb72011-10-20 17:44:03 -04004094 tcp_fin(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004095 }
4096}
4097
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004098static bool tcp_prune_ofo_queue(struct sock *sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004099static int tcp_prune_queue(struct sock *sk);
4100
Mel Gormanc76562b2012-07-31 16:44:41 -07004101static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4102 unsigned int size)
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004103{
4104 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
Mel Gormanc76562b2012-07-31 16:44:41 -07004105 !sk_rmem_schedule(sk, skb, size)) {
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004106
4107 if (tcp_prune_queue(sk) < 0)
4108 return -1;
4109
Mel Gormanc76562b2012-07-31 16:44:41 -07004110 if (!sk_rmem_schedule(sk, skb, size)) {
Vitaliy Gusev56f367b2008-04-15 20:26:34 -07004111 if (!tcp_prune_ofo_queue(sk))
4112 return -1;
4113
Mel Gormanc76562b2012-07-31 16:44:41 -07004114 if (!sk_rmem_schedule(sk, skb, size))
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004115 return -1;
4116 }
4117 }
4118 return 0;
4119}
4120
Eric Dumazet1402d362012-04-23 07:11:42 +00004121/**
4122 * tcp_try_coalesce - try to merge skb to prior one
4123 * @sk: socket
4124 * @to: prior buffer
4125 * @from: buffer to add in queue
Eric Dumazet923dd342012-05-02 07:55:58 +00004126 * @fragstolen: pointer to boolean
Eric Dumazet1402d362012-04-23 07:11:42 +00004127 *
4128 * Before queueing skb @from after @to, try to merge them
4129 * to reduce overall memory use and queue lengths, if cost is small.
4130 * Packets in ofo or receive queues can stay a long time.
4131 * Better try to coalesce them right now to avoid future collapses.
Eric Dumazet783c1752012-04-23 17:34:36 +00004132 * Returns true if caller should free @from instead of queueing it
Eric Dumazet1402d362012-04-23 07:11:42 +00004133 */
Eric Dumazet783c1752012-04-23 17:34:36 +00004134static bool tcp_try_coalesce(struct sock *sk,
4135 struct sk_buff *to,
Eric Dumazet329033f2012-04-27 00:38:33 +00004136 struct sk_buff *from,
4137 bool *fragstolen)
Eric Dumazet1402d362012-04-23 07:11:42 +00004138{
Eric Dumazetbad43ca2012-05-19 03:02:02 +00004139 int delta;
Eric Dumazet1402d362012-04-23 07:11:42 +00004140
Eric Dumazet329033f2012-04-27 00:38:33 +00004141 *fragstolen = false;
Alexander Duyck34a802a2012-05-02 21:19:09 +00004142
Eric Dumazetbad43ca2012-05-19 03:02:02 +00004143 if (tcp_hdr(from)->fin)
4144 return false;
Eric Dumazet1ca7ee32012-05-23 17:51:37 +00004145
4146 /* Its possible this segment overlaps with prior segment in queue */
4147 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4148 return false;
4149
Eric Dumazetbad43ca2012-05-19 03:02:02 +00004150 if (!skb_try_coalesce(to, from, fragstolen, &delta))
Eric Dumazet783c1752012-04-23 17:34:36 +00004151 return false;
Alexander Duyck34a802a2012-05-02 21:19:09 +00004152
Alexander Duyck34a802a2012-05-02 21:19:09 +00004153 atomic_add(delta, &sk->sk_rmem_alloc);
4154 sk_mem_charge(sk, delta);
Alexander Duyck34a802a2012-05-02 21:19:09 +00004155 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4156 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4157 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4158 return true;
Eric Dumazet1402d362012-04-23 07:11:42 +00004159}
4160
Eric Dumazete86b29192012-03-18 11:06:44 +00004161static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4162{
4163 struct tcp_sock *tp = tcp_sk(sk);
4164 struct sk_buff *skb1;
4165 u32 seq, end_seq;
4166
4167 TCP_ECN_check_ce(tp, skb);
4168
Mel Gormanc76562b2012-07-31 16:44:41 -07004169 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
Eric Dumazeta6df1ae2012-07-16 01:41:36 +00004170 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFODROP);
Eric Dumazete86b29192012-03-18 11:06:44 +00004171 __kfree_skb(skb);
4172 return;
4173 }
4174
4175 /* Disable header prediction. */
4176 tp->pred_flags = 0;
4177 inet_csk_schedule_ack(sk);
4178
Eric Dumazeta6df1ae2012-07-16 01:41:36 +00004179 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
Eric Dumazete86b29192012-03-18 11:06:44 +00004180 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4181 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4182
4183 skb1 = skb_peek_tail(&tp->out_of_order_queue);
4184 if (!skb1) {
4185 /* Initial out of order segment, build 1 SACK. */
4186 if (tcp_is_sack(tp)) {
4187 tp->rx_opt.num_sacks = 1;
4188 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4189 tp->selective_acks[0].end_seq =
4190 TCP_SKB_CB(skb)->end_seq;
4191 }
4192 __skb_queue_head(&tp->out_of_order_queue, skb);
4193 goto end;
4194 }
4195
4196 seq = TCP_SKB_CB(skb)->seq;
4197 end_seq = TCP_SKB_CB(skb)->end_seq;
4198
4199 if (seq == TCP_SKB_CB(skb1)->end_seq) {
Eric Dumazet329033f2012-04-27 00:38:33 +00004200 bool fragstolen;
4201
4202 if (!tcp_try_coalesce(sk, skb1, skb, &fragstolen)) {
Eric Dumazet1402d362012-04-23 07:11:42 +00004203 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4204 } else {
Eric Dumazet4e4f1fc2013-09-06 10:35:58 -07004205 tcp_grow_window(sk, skb);
Eric Dumazet923dd342012-05-02 07:55:58 +00004206 kfree_skb_partial(skb, fragstolen);
Eric Dumazetc8628152012-03-18 11:07:47 +00004207 skb = NULL;
Eric Dumazetc8628152012-03-18 11:07:47 +00004208 }
Eric Dumazete86b29192012-03-18 11:06:44 +00004209
4210 if (!tp->rx_opt.num_sacks ||
4211 tp->selective_acks[0].end_seq != seq)
4212 goto add_sack;
4213
4214 /* Common case: data arrive in order after hole. */
4215 tp->selective_acks[0].end_seq = end_seq;
4216 goto end;
4217 }
4218
4219 /* Find place to insert this segment. */
4220 while (1) {
4221 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4222 break;
4223 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4224 skb1 = NULL;
4225 break;
4226 }
4227 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4228 }
4229
4230 /* Do skb overlap to previous one? */
4231 if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4232 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4233 /* All the bits are present. Drop. */
Eric Dumazeta6df1ae2012-07-16 01:41:36 +00004234 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
Eric Dumazete86b29192012-03-18 11:06:44 +00004235 __kfree_skb(skb);
4236 skb = NULL;
4237 tcp_dsack_set(sk, seq, end_seq);
4238 goto add_sack;
4239 }
4240 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4241 /* Partial overlap. */
4242 tcp_dsack_set(sk, seq,
4243 TCP_SKB_CB(skb1)->end_seq);
4244 } else {
4245 if (skb_queue_is_first(&tp->out_of_order_queue,
4246 skb1))
4247 skb1 = NULL;
4248 else
4249 skb1 = skb_queue_prev(
4250 &tp->out_of_order_queue,
4251 skb1);
4252 }
4253 }
4254 if (!skb1)
4255 __skb_queue_head(&tp->out_of_order_queue, skb);
4256 else
4257 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4258
4259 /* And clean segments covered by new one as whole. */
4260 while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4261 skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4262
4263 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4264 break;
4265 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4266 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4267 end_seq);
4268 break;
4269 }
4270 __skb_unlink(skb1, &tp->out_of_order_queue);
4271 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4272 TCP_SKB_CB(skb1)->end_seq);
Eric Dumazeta6df1ae2012-07-16 01:41:36 +00004273 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
Eric Dumazete86b29192012-03-18 11:06:44 +00004274 __kfree_skb(skb1);
4275 }
4276
4277add_sack:
4278 if (tcp_is_sack(tp))
4279 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4280end:
Eric Dumazet4e4f1fc2013-09-06 10:35:58 -07004281 if (skb) {
4282 tcp_grow_window(sk, skb);
Eric Dumazete86b29192012-03-18 11:06:44 +00004283 skb_set_owner_r(skb, sk);
Eric Dumazet4e4f1fc2013-09-06 10:35:58 -07004284 }
Eric Dumazete86b29192012-03-18 11:06:44 +00004285}
4286
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004287static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
Eric Dumazetb081f852012-05-02 09:58:29 +00004288 bool *fragstolen)
4289{
4290 int eaten;
4291 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4292
4293 __skb_pull(skb, hdrlen);
4294 eaten = (tail &&
4295 tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0;
4296 tcp_sk(sk)->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4297 if (!eaten) {
4298 __skb_queue_tail(&sk->sk_receive_queue, skb);
4299 skb_set_owner_r(skb, sk);
4300 }
4301 return eaten;
4302}
Eric Dumazete86b29192012-03-18 11:06:44 +00004303
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004304int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4305{
Mel Gormanc76562b2012-07-31 16:44:41 -07004306 struct sk_buff *skb = NULL;
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004307 struct tcphdr *th;
4308 bool fragstolen;
4309
Pavel Emelyanovc454e612012-10-29 05:05:33 +00004310 if (size == 0)
4311 return 0;
4312
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004313 skb = alloc_skb(size + sizeof(*th), sk->sk_allocation);
4314 if (!skb)
4315 goto err;
4316
Mel Gormanc76562b2012-07-31 16:44:41 -07004317 if (tcp_try_rmem_schedule(sk, skb, size + sizeof(*th)))
4318 goto err_free;
4319
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004320 th = (struct tcphdr *)skb_put(skb, sizeof(*th));
4321 skb_reset_transport_header(skb);
4322 memset(th, 0, sizeof(*th));
4323
4324 if (memcpy_fromiovec(skb_put(skb, size), msg->msg_iov, size))
4325 goto err_free;
4326
4327 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4328 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4329 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4330
4331 if (tcp_queue_rcv(sk, skb, sizeof(*th), &fragstolen)) {
4332 WARN_ON_ONCE(fragstolen); /* should not happen */
4333 __kfree_skb(skb);
4334 }
4335 return size;
4336
4337err_free:
4338 kfree_skb(skb);
4339err:
4340 return -ENOMEM;
4341}
4342
Linus Torvalds1da177e2005-04-16 15:20:36 -07004343static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4344{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04004345 const struct tcphdr *th = tcp_hdr(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004346 struct tcp_sock *tp = tcp_sk(sk);
4347 int eaten = -1;
Eric Dumazet329033f2012-04-27 00:38:33 +00004348 bool fragstolen = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004349
4350 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4351 goto drop;
4352
Eric Dumazetf84af322010-04-28 15:31:51 -07004353 skb_dst_drop(skb);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004354 __skb_pull(skb, th->doff * 4);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004355
4356 TCP_ECN_accept_cwr(tp, skb);
4357
Ilpo Järvinen5861f8e2009-03-14 14:23:01 +00004358 tp->rx_opt.dsack = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004359
4360 /* Queue data for delivery to the user.
4361 * Packets in sequence go to the receive queue.
4362 * Out of sequence packets to the out_of_order_queue.
4363 */
4364 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4365 if (tcp_receive_window(tp) == 0)
4366 goto out_of_window;
4367
4368 /* Ok. In sequence. In window. */
4369 if (tp->ucopy.task == current &&
4370 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4371 sock_owned_by_user(sk) && !tp->urg_data) {
4372 int chunk = min_t(unsigned int, skb->len,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004373 tp->ucopy.len);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004374
4375 __set_current_state(TASK_RUNNING);
4376
4377 local_bh_enable();
4378 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
4379 tp->ucopy.len -= chunk;
4380 tp->copied_seq += chunk;
Jerry Chu44f53242011-01-25 13:46:30 -08004381 eaten = (chunk == skb->len);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004382 tcp_rcv_space_adjust(sk);
4383 }
4384 local_bh_disable();
4385 }
4386
4387 if (eaten <= 0) {
4388queue_and_out:
4389 if (eaten < 0 &&
Mel Gormanc76562b2012-07-31 16:44:41 -07004390 tcp_try_rmem_schedule(sk, skb, skb->truesize))
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004391 goto drop;
4392
Eric Dumazetb081f852012-05-02 09:58:29 +00004393 eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004394 }
4395 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
Stephen Hemminger2de979b2007-03-08 20:45:19 -08004396 if (skb->len)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004397 tcp_event_data_recv(sk, skb);
Stephen Hemminger2de979b2007-03-08 20:45:19 -08004398 if (th->fin)
Eric Dumazet20c4cb72011-10-20 17:44:03 -04004399 tcp_fin(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004400
David S. Millerb03efcf2005-07-08 14:57:23 -07004401 if (!skb_queue_empty(&tp->out_of_order_queue)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004402 tcp_ofo_queue(sk);
4403
4404 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4405 * gap in queue is filled.
4406 */
David S. Millerb03efcf2005-07-08 14:57:23 -07004407 if (skb_queue_empty(&tp->out_of_order_queue))
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004408 inet_csk(sk)->icsk_ack.pingpong = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004409 }
4410
4411 if (tp->rx_opt.num_sacks)
4412 tcp_sack_remove(tp);
4413
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004414 tcp_fast_path_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004415
Eric Dumazet923dd342012-05-02 07:55:58 +00004416 if (eaten > 0)
4417 kfree_skb_partial(skb, fragstolen);
Eric Dumazet1d57f192012-09-17 12:51:39 +00004418 if (!sock_flag(sk, SOCK_DEAD))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004419 sk->sk_data_ready(sk, 0);
4420 return;
4421 }
4422
4423 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4424 /* A retransmit, 2nd most common case. Force an immediate ack. */
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07004425 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004426 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004427
4428out_of_window:
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004429 tcp_enter_quickack_mode(sk);
4430 inet_csk_schedule_ack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004431drop:
4432 __kfree_skb(skb);
4433 return;
4434 }
4435
4436 /* Out of window. F.e. zero window probe. */
4437 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4438 goto out_of_window;
4439
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004440 tcp_enter_quickack_mode(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004441
4442 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4443 /* Partial packet, seq < rcv_next < end_seq */
4444 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4445 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4446 TCP_SKB_CB(skb)->end_seq);
4447
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004448 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004449
Linus Torvalds1da177e2005-04-16 15:20:36 -07004450 /* If window is closed, drop tail of packet. But after
4451 * remembering D-SACK for its head made in previous line.
4452 */
4453 if (!tcp_receive_window(tp))
4454 goto out_of_window;
4455 goto queue_and_out;
4456 }
4457
Eric Dumazete86b29192012-03-18 11:06:44 +00004458 tcp_data_queue_ofo(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004459}
4460
Ilpo Järvinen2cf46632008-08-23 05:11:41 -07004461static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4462 struct sk_buff_head *list)
4463{
David S. Miller91521942009-05-28 21:35:47 -07004464 struct sk_buff *next = NULL;
4465
4466 if (!skb_queue_is_last(list, skb))
4467 next = skb_queue_next(list, skb);
Ilpo Järvinen2cf46632008-08-23 05:11:41 -07004468
4469 __skb_unlink(skb, list);
4470 __kfree_skb(skb);
4471 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4472
4473 return next;
4474}
4475
Linus Torvalds1da177e2005-04-16 15:20:36 -07004476/* Collapse contiguous sequence of skbs head..tail with
4477 * sequence numbers start..end.
David S. Miller91521942009-05-28 21:35:47 -07004478 *
4479 * If tail is NULL, this means until the end of the list.
4480 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07004481 * Segments with FIN/SYN are not collapsed (only because this
4482 * simplifies code)
4483 */
4484static void
David S. Miller8728b832005-08-09 19:25:21 -07004485tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4486 struct sk_buff *head, struct sk_buff *tail,
4487 u32 start, u32 end)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004488{
David S. Miller91521942009-05-28 21:35:47 -07004489 struct sk_buff *skb, *n;
4490 bool end_of_skbs;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004491
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08004492 /* First, check that queue is collapsible and find
Linus Torvalds1da177e2005-04-16 15:20:36 -07004493 * the point where collapsing can be useful. */
David S. Miller91521942009-05-28 21:35:47 -07004494 skb = head;
4495restart:
4496 end_of_skbs = true;
4497 skb_queue_walk_from_safe(list, skb, n) {
4498 if (skb == tail)
4499 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004500 /* No new bits? It is possible on ofo queue. */
4501 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
Ilpo Järvinen2cf46632008-08-23 05:11:41 -07004502 skb = tcp_collapse_one(sk, skb, list);
David S. Miller91521942009-05-28 21:35:47 -07004503 if (!skb)
4504 break;
4505 goto restart;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004506 }
4507
4508 /* The first skb to collapse is:
4509 * - not SYN/FIN and
4510 * - bloated or contains data before "start" or
4511 * overlaps to the next one.
4512 */
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07004513 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07004514 (tcp_win_from_space(skb->truesize) > skb->len ||
David S. Miller91521942009-05-28 21:35:47 -07004515 before(TCP_SKB_CB(skb)->seq, start))) {
4516 end_of_skbs = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004517 break;
David S. Miller91521942009-05-28 21:35:47 -07004518 }
4519
4520 if (!skb_queue_is_last(list, skb)) {
4521 struct sk_buff *next = skb_queue_next(list, skb);
4522 if (next != tail &&
4523 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4524 end_of_skbs = false;
4525 break;
4526 }
4527 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004528
4529 /* Decided to skip this, advance start seq. */
4530 start = TCP_SKB_CB(skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004531 }
David S. Miller91521942009-05-28 21:35:47 -07004532 if (end_of_skbs || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004533 return;
4534
4535 while (before(start, end)) {
4536 struct sk_buff *nskb;
Chuck Leverc2636b42007-10-23 21:07:32 -07004537 unsigned int header = skb_headroom(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004538 int copy = SKB_MAX_ORDER(header, 0);
4539
4540 /* Too big header? This can happen with IPv6. */
4541 if (copy < 0)
4542 return;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004543 if (end - start < copy)
4544 copy = end - start;
4545 nskb = alloc_skb(copy + header, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004546 if (!nskb)
4547 return;
Arnaldo Carvalho de Meloc51957d2007-03-10 12:47:22 -03004548
Arnaldo Carvalho de Melo98e399f2007-03-19 15:33:04 -07004549 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
Arnaldo Carvalho de Melo9c702202007-04-25 18:04:18 -07004550 skb_set_network_header(nskb, (skb_network_header(skb) -
4551 skb->head));
4552 skb_set_transport_header(nskb, (skb_transport_header(skb) -
4553 skb->head));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004554 skb_reserve(nskb, header);
4555 memcpy(nskb->head, skb->head, header);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004556 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4557 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
David S. Miller43f59c82008-09-21 21:28:51 -07004558 __skb_queue_before(list, skb, nskb);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08004559 skb_set_owner_r(nskb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004560
4561 /* Copy data, releasing collapsed skbs. */
4562 while (copy > 0) {
4563 int offset = start - TCP_SKB_CB(skb)->seq;
4564 int size = TCP_SKB_CB(skb)->end_seq - start;
4565
Kris Katterjohn09a62662006-01-08 22:24:28 -08004566 BUG_ON(offset < 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004567 if (size > 0) {
4568 size = min(copy, size);
4569 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4570 BUG();
4571 TCP_SKB_CB(nskb)->end_seq += size;
4572 copy -= size;
4573 start += size;
4574 }
4575 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
Ilpo Järvinen2cf46632008-08-23 05:11:41 -07004576 skb = tcp_collapse_one(sk, skb, list);
David S. Miller91521942009-05-28 21:35:47 -07004577 if (!skb ||
4578 skb == tail ||
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07004579 tcp_hdr(skb)->syn ||
4580 tcp_hdr(skb)->fin)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004581 return;
4582 }
4583 }
4584 }
4585}
4586
4587/* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4588 * and tcp_collapse() them until all the queue is collapsed.
4589 */
4590static void tcp_collapse_ofo_queue(struct sock *sk)
4591{
4592 struct tcp_sock *tp = tcp_sk(sk);
4593 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4594 struct sk_buff *head;
4595 u32 start, end;
4596
4597 if (skb == NULL)
4598 return;
4599
4600 start = TCP_SKB_CB(skb)->seq;
4601 end = TCP_SKB_CB(skb)->end_seq;
4602 head = skb;
4603
4604 for (;;) {
David S. Miller91521942009-05-28 21:35:47 -07004605 struct sk_buff *next = NULL;
4606
4607 if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4608 next = skb_queue_next(&tp->out_of_order_queue, skb);
4609 skb = next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004610
4611 /* Segment is terminated when we see gap or when
4612 * we are at the end of all the queue. */
David S. Miller91521942009-05-28 21:35:47 -07004613 if (!skb ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07004614 after(TCP_SKB_CB(skb)->seq, end) ||
4615 before(TCP_SKB_CB(skb)->end_seq, start)) {
David S. Miller8728b832005-08-09 19:25:21 -07004616 tcp_collapse(sk, &tp->out_of_order_queue,
4617 head, skb, start, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004618 head = skb;
David S. Miller91521942009-05-28 21:35:47 -07004619 if (!skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004620 break;
4621 /* Start new segment */
4622 start = TCP_SKB_CB(skb)->seq;
4623 end = TCP_SKB_CB(skb)->end_seq;
4624 } else {
4625 if (before(TCP_SKB_CB(skb)->seq, start))
4626 start = TCP_SKB_CB(skb)->seq;
4627 if (after(TCP_SKB_CB(skb)->end_seq, end))
4628 end = TCP_SKB_CB(skb)->end_seq;
4629 }
4630 }
4631}
4632
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004633/*
4634 * Purge the out-of-order queue.
Vitaliy Gusev56f367b2008-04-15 20:26:34 -07004635 * Return true if queue was pruned.
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004636 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004637static bool tcp_prune_ofo_queue(struct sock *sk)
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004638{
4639 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004640 bool res = false;
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004641
4642 if (!skb_queue_empty(&tp->out_of_order_queue)) {
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07004643 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004644 __skb_queue_purge(&tp->out_of_order_queue);
4645
4646 /* Reset SACK state. A conforming SACK implementation will
4647 * do the same at a timeout based retransmit. When a connection
4648 * is in a sad state like this, we care only about integrity
4649 * of the connection not performance.
4650 */
4651 if (tp->rx_opt.sack_ok)
4652 tcp_sack_reset(&tp->rx_opt);
4653 sk_mem_reclaim(sk);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004654 res = true;
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004655 }
Vitaliy Gusev56f367b2008-04-15 20:26:34 -07004656 return res;
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004657}
4658
Linus Torvalds1da177e2005-04-16 15:20:36 -07004659/* Reduce allocated memory if we can, trying to get
4660 * the socket within its memory limits again.
4661 *
4662 * Return less than zero if we should start dropping frames
4663 * until the socket owning process reads some of the data
4664 * to stabilize the situation.
4665 */
4666static int tcp_prune_queue(struct sock *sk)
4667{
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004668 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004669
4670 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4671
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07004672 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004673
4674 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004675 tcp_clamp_window(sk);
Glauber Costa180d8cd2011-12-11 21:47:02 +00004676 else if (sk_under_memory_pressure(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004677 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4678
4679 tcp_collapse_ofo_queue(sk);
David S. Miller91521942009-05-28 21:35:47 -07004680 if (!skb_queue_empty(&sk->sk_receive_queue))
4681 tcp_collapse(sk, &sk->sk_receive_queue,
4682 skb_peek(&sk->sk_receive_queue),
4683 NULL,
4684 tp->copied_seq, tp->rcv_nxt);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08004685 sk_mem_reclaim(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004686
4687 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4688 return 0;
4689
4690 /* Collapsing did not help, destructive actions follow.
4691 * This must not ever occur. */
4692
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004693 tcp_prune_ofo_queue(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004694
4695 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4696 return 0;
4697
4698 /* If we are really being abused, tell the caller to silently
4699 * drop receive data on the floor. It will get retransmitted
4700 * and hopefully then we'll have sufficient space.
4701 */
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07004702 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004703
4704 /* Massive buffer overcommit. */
4705 tp->pred_flags = 0;
4706 return -1;
4707}
4708
Linus Torvalds1da177e2005-04-16 15:20:36 -07004709/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4710 * As additional protections, we do not touch cwnd in retransmission phases,
4711 * and if application hit its sndbuf limit recently.
4712 */
4713void tcp_cwnd_application_limited(struct sock *sk)
4714{
4715 struct tcp_sock *tp = tcp_sk(sk);
4716
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03004717 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07004718 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4719 /* Limited by application or receiver window. */
Ilpo Järvinend254bcd2006-08-04 16:57:42 -07004720 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4721 u32 win_used = max(tp->snd_cwnd_used, init_win);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004722 if (win_used < tp->snd_cwnd) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03004723 tp->snd_ssthresh = tcp_current_ssthresh(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004724 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4725 }
4726 tp->snd_cwnd_used = 0;
4727 }
4728 tp->snd_cwnd_stamp = tcp_time_stamp;
4729}
4730
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004731static bool tcp_should_expand_sndbuf(const struct sock *sk)
David S. Miller0d9901d2005-07-05 15:21:10 -07004732{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04004733 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004734
David S. Miller0d9901d2005-07-05 15:21:10 -07004735 /* If the user specified a specific send buffer setting, do
4736 * not modify it.
4737 */
4738 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004739 return false;
David S. Miller0d9901d2005-07-05 15:21:10 -07004740
4741 /* If we are under global TCP memory pressure, do not expand. */
Glauber Costa180d8cd2011-12-11 21:47:02 +00004742 if (sk_under_memory_pressure(sk))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004743 return false;
David S. Miller0d9901d2005-07-05 15:21:10 -07004744
4745 /* If we are under soft global TCP memory pressure, do not expand. */
Glauber Costa180d8cd2011-12-11 21:47:02 +00004746 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004747 return false;
David S. Miller0d9901d2005-07-05 15:21:10 -07004748
4749 /* If we filled the congestion window, do not expand. */
4750 if (tp->packets_out >= tp->snd_cwnd)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004751 return false;
David S. Miller0d9901d2005-07-05 15:21:10 -07004752
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004753 return true;
David S. Miller0d9901d2005-07-05 15:21:10 -07004754}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004755
4756/* When incoming ACK allowed to free some skb from write_queue,
4757 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4758 * on the exit from tcp input handler.
4759 *
4760 * PROBLEM: sndbuf expansion does not work well with largesend.
4761 */
4762static void tcp_new_space(struct sock *sk)
4763{
4764 struct tcp_sock *tp = tcp_sk(sk);
4765
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004766 if (tcp_should_expand_sndbuf(sk)) {
Eric Dumazet6ae70532013-10-01 10:23:44 -07004767 tcp_sndbuf_expand(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004768 tp->snd_cwnd_stamp = tcp_time_stamp;
4769 }
4770
4771 sk->sk_write_space(sk);
4772}
4773
Stephen Hemminger40efc6f2006-01-03 16:03:49 -08004774static void tcp_check_space(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004775{
4776 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4777 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4778 if (sk->sk_socket &&
4779 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4780 tcp_new_space(sk);
4781 }
4782}
4783
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004784static inline void tcp_data_snd_check(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004785{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004786 tcp_push_pending_frames(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004787 tcp_check_space(sk);
4788}
4789
4790/*
4791 * Check if sending an ack is needed.
4792 */
4793static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4794{
4795 struct tcp_sock *tp = tcp_sk(sk);
4796
4797 /* More than one full frame received... */
Joe Perches9d4fb272009-11-23 10:41:23 -08004798 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07004799 /* ... and right edge of window advances far enough.
4800 * (tcp_recvmsg() will send ACK otherwise). Or...
4801 */
Joe Perches9d4fb272009-11-23 10:41:23 -08004802 __tcp_select_window(sk) >= tp->rcv_wnd) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07004803 /* We ACK each frame or... */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004804 tcp_in_quickack_mode(sk) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07004805 /* We have out of order data. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004806 (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004807 /* Then ack it now */
4808 tcp_send_ack(sk);
4809 } else {
4810 /* Else, send delayed ack. */
4811 tcp_send_delayed_ack(sk);
4812 }
4813}
4814
Stephen Hemminger40efc6f2006-01-03 16:03:49 -08004815static inline void tcp_ack_snd_check(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004816{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004817 if (!inet_csk_ack_scheduled(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004818 /* We sent a data segment already. */
4819 return;
4820 }
4821 __tcp_ack_snd_check(sk, 1);
4822}
4823
4824/*
4825 * This routine is only called when we have urgent data
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08004826 * signaled. Its the 'slow' part of tcp_urg. It could be
Linus Torvalds1da177e2005-04-16 15:20:36 -07004827 * moved inline now as tcp_urg is only called from one
4828 * place. We handle URGent data wrong. We have to - as
4829 * BSD still doesn't use the correction from RFC961.
4830 * For 1003.1g we should support a new option TCP_STDURG to permit
4831 * either form (or just set the sysctl tcp_stdurg).
4832 */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004833
Eric Dumazetcf533ea2011-10-21 05:22:42 -04004834static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004835{
4836 struct tcp_sock *tp = tcp_sk(sk);
4837 u32 ptr = ntohs(th->urg_ptr);
4838
4839 if (ptr && !sysctl_tcp_stdurg)
4840 ptr--;
4841 ptr += ntohl(th->seq);
4842
4843 /* Ignore urgent data that we've already seen and read. */
4844 if (after(tp->copied_seq, ptr))
4845 return;
4846
4847 /* Do not replay urg ptr.
4848 *
4849 * NOTE: interesting situation not covered by specs.
4850 * Misbehaving sender may send urg ptr, pointing to segment,
4851 * which we already have in ofo queue. We are not able to fetch
4852 * such data and will stay in TCP_URG_NOTYET until will be eaten
4853 * by recvmsg(). Seems, we are not obliged to handle such wicked
4854 * situations. But it is worth to think about possibility of some
4855 * DoSes using some hypothetical application level deadlock.
4856 */
4857 if (before(ptr, tp->rcv_nxt))
4858 return;
4859
4860 /* Do we already have a newer (or duplicate) urgent pointer? */
4861 if (tp->urg_data && !after(ptr, tp->urg_seq))
4862 return;
4863
4864 /* Tell the world about our new urgent pointer. */
4865 sk_send_sigurg(sk);
4866
4867 /* We may be adding urgent data when the last byte read was
4868 * urgent. To do this requires some care. We cannot just ignore
4869 * tp->copied_seq since we would read the last urgent byte again
4870 * as data, nor can we alter copied_seq until this data arrives
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08004871 * or we break the semantics of SIOCATMARK (and thus sockatmark())
Linus Torvalds1da177e2005-04-16 15:20:36 -07004872 *
4873 * NOTE. Double Dutch. Rendering to plain English: author of comment
4874 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4875 * and expect that both A and B disappear from stream. This is _wrong_.
4876 * Though this happens in BSD with high probability, this is occasional.
4877 * Any application relying on this is buggy. Note also, that fix "works"
4878 * only in this artificial test. Insert some normal data between A and B and we will
4879 * decline of BSD again. Verdict: it is better to remove to trap
4880 * buggy users.
4881 */
4882 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004883 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004884 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4885 tp->copied_seq++;
4886 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
David S. Miller8728b832005-08-09 19:25:21 -07004887 __skb_unlink(skb, &sk->sk_receive_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004888 __kfree_skb(skb);
4889 }
4890 }
4891
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004892 tp->urg_data = TCP_URG_NOTYET;
4893 tp->urg_seq = ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004894
4895 /* Disable header prediction. */
4896 tp->pred_flags = 0;
4897}
4898
4899/* This is the 'fast' part of urgent handling. */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04004900static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004901{
4902 struct tcp_sock *tp = tcp_sk(sk);
4903
4904 /* Check if we get a new urgent pointer - normally not. */
4905 if (th->urg)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004906 tcp_check_urg(sk, th);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004907
4908 /* Do we wait for any urgent data? - normally not... */
4909 if (tp->urg_data == TCP_URG_NOTYET) {
4910 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4911 th->syn;
4912
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004913 /* Is the urgent pointer pointing into this packet? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004914 if (ptr < skb->len) {
4915 u8 tmp;
4916 if (skb_copy_bits(skb, ptr, &tmp, 1))
4917 BUG();
4918 tp->urg_data = TCP_URG_VALID | tmp;
4919 if (!sock_flag(sk, SOCK_DEAD))
4920 sk->sk_data_ready(sk, 0);
4921 }
4922 }
4923}
4924
4925static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4926{
4927 struct tcp_sock *tp = tcp_sk(sk);
4928 int chunk = skb->len - hlen;
4929 int err;
4930
4931 local_bh_enable();
Herbert Xu60476372007-04-09 11:59:39 -07004932 if (skb_csum_unnecessary(skb))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004933 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4934 else
4935 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4936 tp->ucopy.iov);
4937
4938 if (!err) {
4939 tp->ucopy.len -= chunk;
4940 tp->copied_seq += chunk;
4941 tcp_rcv_space_adjust(sk);
4942 }
4943
4944 local_bh_disable();
4945 return err;
4946}
4947
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004948static __sum16 __tcp_checksum_complete_user(struct sock *sk,
4949 struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004950{
Al Virob51655b2006-11-14 21:40:42 -08004951 __sum16 result;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004952
4953 if (sock_owned_by_user(sk)) {
4954 local_bh_enable();
4955 result = __tcp_checksum_complete(skb);
4956 local_bh_disable();
4957 } else {
4958 result = __tcp_checksum_complete(skb);
4959 }
4960 return result;
4961}
4962
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00004963static inline bool tcp_checksum_complete_user(struct sock *sk,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004964 struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004965{
Herbert Xu60476372007-04-09 11:59:39 -07004966 return !skb_csum_unnecessary(skb) &&
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004967 __tcp_checksum_complete_user(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004968}
4969
Chris Leech1a2449a2006-05-23 18:05:53 -07004970#ifdef CONFIG_NET_DMA
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004971static bool tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004972 int hlen)
Chris Leech1a2449a2006-05-23 18:05:53 -07004973{
4974 struct tcp_sock *tp = tcp_sk(sk);
4975 int chunk = skb->len - hlen;
4976 int dma_cookie;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004977 bool copied_early = false;
Chris Leech1a2449a2006-05-23 18:05:53 -07004978
4979 if (tp->ucopy.wakeup)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004980 return false;
Chris Leech1a2449a2006-05-23 18:05:53 -07004981
4982 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
Dave Jianga2bd1142012-04-04 16:10:46 -07004983 tp->ucopy.dma_chan = net_dma_find_channel();
Chris Leech1a2449a2006-05-23 18:05:53 -07004984
Herbert Xu60476372007-04-09 11:59:39 -07004985 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
Chris Leech1a2449a2006-05-23 18:05:53 -07004986
4987 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004988 skb, hlen,
4989 tp->ucopy.iov, chunk,
4990 tp->ucopy.pinned_list);
Chris Leech1a2449a2006-05-23 18:05:53 -07004991
4992 if (dma_cookie < 0)
4993 goto out;
4994
4995 tp->ucopy.dma_cookie = dma_cookie;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004996 copied_early = true;
Chris Leech1a2449a2006-05-23 18:05:53 -07004997
4998 tp->ucopy.len -= chunk;
4999 tp->copied_seq += chunk;
5000 tcp_rcv_space_adjust(sk);
5001
5002 if ((tp->ucopy.len == 0) ||
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07005003 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
Chris Leech1a2449a2006-05-23 18:05:53 -07005004 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
5005 tp->ucopy.wakeup = 1;
5006 sk->sk_data_ready(sk, 0);
5007 }
5008 } else if (chunk > 0) {
5009 tp->ucopy.wakeup = 1;
5010 sk->sk_data_ready(sk, 0);
5011 }
5012out:
5013 return copied_early;
5014}
5015#endif /* CONFIG_NET_DMA */
5016
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005017/* Does PAWS and seqno based validation of an incoming segment, flags will
5018 * play significant role here.
5019 */
Eric Dumazet0c246042012-07-17 01:41:30 +00005020static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5021 const struct tcphdr *th, int syn_inerr)
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005022{
5023 struct tcp_sock *tp = tcp_sk(sk);
5024
5025 /* RFC1323: H1. Apply PAWS check first. */
Christoph Paasch1a2c6182013-03-17 08:23:34 +00005026 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005027 tcp_paws_discard(sk, skb)) {
5028 if (!th->rst) {
5029 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5030 tcp_send_dupack(sk, skb);
5031 goto discard;
5032 }
5033 /* Reset is accepted even if it did not pass PAWS. */
5034 }
5035
5036 /* Step 1: check sequence number */
5037 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5038 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5039 * (RST) segments are validated by checking their SEQ-fields."
5040 * And page 69: "If an incoming segment is not acceptable,
5041 * an acknowledgment should be sent in reply (unless the RST
5042 * bit is set, if so drop the segment and return)".
5043 */
Eric Dumazete3715892012-07-17 12:29:30 +00005044 if (!th->rst) {
5045 if (th->syn)
5046 goto syn_challenge;
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005047 tcp_send_dupack(sk, skb);
Eric Dumazete3715892012-07-17 12:29:30 +00005048 }
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005049 goto discard;
5050 }
5051
5052 /* Step 2: check RST bit */
5053 if (th->rst) {
Eric Dumazet282f23c2012-07-17 10:13:05 +02005054 /* RFC 5961 3.2 :
5055 * If sequence number exactly matches RCV.NXT, then
5056 * RESET the connection
5057 * else
5058 * Send a challenge ACK
5059 */
5060 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt)
5061 tcp_reset(sk);
5062 else
5063 tcp_send_challenge_ack(sk);
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005064 goto discard;
5065 }
5066
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005067 /* step 3: check security and precedence [ignored] */
5068
Eric Dumazet0c246042012-07-17 01:41:30 +00005069 /* step 4: Check for a SYN
5070 * RFC 5691 4.2 : Send a challenge ack
5071 */
5072 if (th->syn) {
Eric Dumazete3715892012-07-17 12:29:30 +00005073syn_challenge:
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005074 if (syn_inerr)
5075 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
Eric Dumazet0c246042012-07-17 01:41:30 +00005076 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5077 tcp_send_challenge_ack(sk);
5078 goto discard;
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005079 }
5080
Eric Dumazet0c246042012-07-17 01:41:30 +00005081 return true;
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005082
5083discard:
5084 __kfree_skb(skb);
Eric Dumazet0c246042012-07-17 01:41:30 +00005085 return false;
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005086}
5087
Linus Torvalds1da177e2005-04-16 15:20:36 -07005088/*
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005089 * TCP receive function for the ESTABLISHED state.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005090 *
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005091 * It is split into a fast path and a slow path. The fast path is
Linus Torvalds1da177e2005-04-16 15:20:36 -07005092 * disabled when:
5093 * - A zero window was announced from us - zero window probing
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005094 * is only handled properly in the slow path.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005095 * - Out of order segments arrived.
5096 * - Urgent data is expected.
5097 * - There is no buffer space left
5098 * - Unexpected TCP flags/window values/header lengths are received
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005099 * (detected by checking the TCP header against pred_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005100 * - Data is sent in both directions. Fast path only supports pure senders
5101 * or pure receivers (this means either the sequence number or the ack
5102 * value must stay constant)
5103 * - Unexpected TCP option.
5104 *
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005105 * When these conditions are not satisfied it drops into a standard
Linus Torvalds1da177e2005-04-16 15:20:36 -07005106 * receive procedure patterned after RFC793 to handle all cases.
5107 * The first three cases are guaranteed by proper pred_flags setting,
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005108 * the rest is checked inline. Fast processing is turned on in
Linus Torvalds1da177e2005-04-16 15:20:36 -07005109 * tcp_data_queue when everything is OK.
5110 */
Vijay Subramanianc995ae22013-09-03 12:23:22 -07005111void tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5112 const struct tcphdr *th, unsigned int len)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005113{
5114 struct tcp_sock *tp = tcp_sk(sk);
5115
Eric Dumazet5d299f32012-08-06 05:09:33 +00005116 if (unlikely(sk->sk_rx_dst == NULL))
5117 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005118 /*
5119 * Header prediction.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005120 * The code loosely follows the one in the famous
Linus Torvalds1da177e2005-04-16 15:20:36 -07005121 * "30 instruction TCP receive" Van Jacobson mail.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005122 *
5123 * Van's trick is to deposit buffers into socket queue
Linus Torvalds1da177e2005-04-16 15:20:36 -07005124 * on a device interrupt, to call tcp_recv function
5125 * on the receive process context and checksum and copy
5126 * the buffer to user space. smart...
5127 *
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005128 * Our current scheme is not silly either but we take the
Linus Torvalds1da177e2005-04-16 15:20:36 -07005129 * extra cost of the net_bh soft interrupt processing...
5130 * We do checksum and copy also but from device to kernel.
5131 */
5132
5133 tp->rx_opt.saw_tstamp = 0;
5134
5135 /* pred_flags is 0xS?10 << 16 + snd_wnd
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08005136 * if header_prediction is to be made
Linus Torvalds1da177e2005-04-16 15:20:36 -07005137 * 'S' will always be tp->tcp_header_len >> 2
5138 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005139 * turn it off (when there are holes in the receive
Linus Torvalds1da177e2005-04-16 15:20:36 -07005140 * space for instance)
5141 * PSH flag is ignored.
5142 */
5143
5144 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
John Dykstra96e0bf42009-03-22 21:49:57 -07005145 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5146 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005147 int tcp_header_len = tp->tcp_header_len;
5148
5149 /* Timestamp header prediction: tcp_header_len
5150 * is automatically equal to th->doff*4 due to pred_flags
5151 * match.
5152 */
5153
5154 /* Check timestamp */
5155 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005156 /* No? Slow path! */
Ilpo Järvinena4356b22008-08-23 05:12:29 -07005157 if (!tcp_parse_aligned_timestamp(tp, th))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005158 goto slow_path;
5159
Linus Torvalds1da177e2005-04-16 15:20:36 -07005160 /* If PAWS failed, check it more carefully in slow path */
5161 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5162 goto slow_path;
5163
5164 /* DO NOT update ts_recent here, if checksum fails
5165 * and timestamp was corrupted part, it will result
5166 * in a hung connection since we will drop all
5167 * future packets due to the PAWS test.
5168 */
5169 }
5170
5171 if (len <= tcp_header_len) {
5172 /* Bulk data transfer: sender */
5173 if (len == tcp_header_len) {
5174 /* Predicted packet is in window by definition.
5175 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5176 * Hence, check seq<=rcv_wup reduces to:
5177 */
5178 if (tcp_header_len ==
5179 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5180 tp->rcv_nxt == tp->rcv_wup)
5181 tcp_store_ts_recent(tp);
5182
Linus Torvalds1da177e2005-04-16 15:20:36 -07005183 /* We know that such packets are checksummed
5184 * on entry.
5185 */
5186 tcp_ack(sk, skb, 0);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005187 __kfree_skb(skb);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005188 tcp_data_snd_check(sk);
Vijay Subramanianc995ae22013-09-03 12:23:22 -07005189 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005190 } else { /* Header too small */
Pavel Emelyanov63231bd2008-07-16 20:22:25 -07005191 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005192 goto discard;
5193 }
5194 } else {
5195 int eaten = 0;
Chris Leech1a2449a2006-05-23 18:05:53 -07005196 int copied_early = 0;
Eric Dumazetb081f852012-05-02 09:58:29 +00005197 bool fragstolen = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005198
Chris Leech1a2449a2006-05-23 18:05:53 -07005199 if (tp->copied_seq == tp->rcv_nxt &&
5200 len - tcp_header_len <= tp->ucopy.len) {
5201#ifdef CONFIG_NET_DMA
Jiri Kosina59ea33a2012-07-27 10:38:50 +00005202 if (tp->ucopy.task == current &&
5203 sock_owned_by_user(sk) &&
5204 tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
Chris Leech1a2449a2006-05-23 18:05:53 -07005205 copied_early = 1;
5206 eaten = 1;
5207 }
5208#endif
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005209 if (tp->ucopy.task == current &&
5210 sock_owned_by_user(sk) && !copied_early) {
Chris Leech1a2449a2006-05-23 18:05:53 -07005211 __set_current_state(TASK_RUNNING);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005212
Chris Leech1a2449a2006-05-23 18:05:53 -07005213 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
5214 eaten = 1;
5215 }
5216 if (eaten) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005217 /* Predicted packet is in window by definition.
5218 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5219 * Hence, check seq<=rcv_wup reduces to:
5220 */
5221 if (tcp_header_len ==
5222 (sizeof(struct tcphdr) +
5223 TCPOLEN_TSTAMP_ALIGNED) &&
5224 tp->rcv_nxt == tp->rcv_wup)
5225 tcp_store_ts_recent(tp);
5226
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005227 tcp_rcv_rtt_measure_ts(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005228
5229 __skb_pull(skb, tcp_header_len);
5230 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07005231 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005232 }
Chris Leech1a2449a2006-05-23 18:05:53 -07005233 if (copied_early)
5234 tcp_cleanup_rbuf(sk, skb->len);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005235 }
5236 if (!eaten) {
5237 if (tcp_checksum_complete_user(sk, skb))
5238 goto csum_error;
5239
Neal Cardwellaab2b4b2013-03-04 06:23:05 +00005240 if ((int)skb->truesize > sk->sk_forward_alloc)
5241 goto step5;
5242
Linus Torvalds1da177e2005-04-16 15:20:36 -07005243 /* Predicted packet is in window by definition.
5244 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5245 * Hence, check seq<=rcv_wup reduces to:
5246 */
5247 if (tcp_header_len ==
5248 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5249 tp->rcv_nxt == tp->rcv_wup)
5250 tcp_store_ts_recent(tp);
5251
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005252 tcp_rcv_rtt_measure_ts(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005253
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07005254 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005255
5256 /* Bulk data transfer: receiver */
Eric Dumazetb081f852012-05-02 09:58:29 +00005257 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5258 &fragstolen);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005259 }
5260
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005261 tcp_event_data_recv(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005262
5263 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5264 /* Well, only one small jumplet in fast path... */
5265 tcp_ack(sk, skb, FLAG_DATA);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005266 tcp_data_snd_check(sk);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005267 if (!inet_csk_ack_scheduled(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005268 goto no_ack;
5269 }
5270
Ali Saidi53240c22008-10-07 15:31:19 -07005271 if (!copied_early || tp->rcv_nxt != tp->rcv_wup)
5272 __tcp_ack_snd_check(sk, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005273no_ack:
Chris Leech1a2449a2006-05-23 18:05:53 -07005274#ifdef CONFIG_NET_DMA
5275 if (copied_early)
5276 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
5277 else
5278#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07005279 if (eaten)
Eric Dumazetb081f852012-05-02 09:58:29 +00005280 kfree_skb_partial(skb, fragstolen);
Eric Dumazet1d57f192012-09-17 12:51:39 +00005281 sk->sk_data_ready(sk, 0);
Vijay Subramanianc995ae22013-09-03 12:23:22 -07005282 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005283 }
5284 }
5285
5286slow_path:
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005287 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005288 goto csum_error;
5289
Eric Dumazet7b514a82013-01-10 16:18:47 +00005290 if (!th->ack && !th->rst)
Eric Dumazetc3ae62af2012-12-26 12:44:34 +00005291 goto discard;
5292
Linus Torvalds1da177e2005-04-16 15:20:36 -07005293 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07005294 * Standard slow path.
5295 */
5296
Eric Dumazet0c246042012-07-17 01:41:30 +00005297 if (!tcp_validate_incoming(sk, skb, th, 1))
Vijay Subramanianc995ae22013-09-03 12:23:22 -07005298 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005299
5300step5:
Eric Dumazet12fb3dd2013-04-19 07:19:48 +00005301 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
John Dykstra96e0bf42009-03-22 21:49:57 -07005302 goto discard;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005303
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005304 tcp_rcv_rtt_measure_ts(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005305
5306 /* Process urgent data. */
5307 tcp_urg(sk, skb, th);
5308
5309 /* step 7: process the segment text */
5310 tcp_data_queue(sk, skb);
5311
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005312 tcp_data_snd_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005313 tcp_ack_snd_check(sk);
Vijay Subramanianc995ae22013-09-03 12:23:22 -07005314 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005315
5316csum_error:
Eric Dumazet6a5dc9e2013-04-29 08:39:56 +00005317 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_CSUMERRORS);
Pavel Emelyanov63231bd2008-07-16 20:22:25 -07005318 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005319
5320discard:
5321 __kfree_skb(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005322}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00005323EXPORT_SYMBOL(tcp_rcv_established);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005324
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005325void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5326{
5327 struct tcp_sock *tp = tcp_sk(sk);
5328 struct inet_connection_sock *icsk = inet_csk(sk);
5329
5330 tcp_set_state(sk, TCP_ESTABLISHED);
5331
David S. Miller41063e92012-06-19 21:22:05 -07005332 if (skb != NULL) {
Eric Dumazet5d299f32012-08-06 05:09:33 +00005333 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005334 security_inet_conn_established(sk, skb);
David S. Miller41063e92012-06-19 21:22:05 -07005335 }
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005336
5337 /* Make sure socket is routed, for correct metrics. */
5338 icsk->icsk_af_ops->rebuild_header(sk);
5339
5340 tcp_init_metrics(sk);
5341
5342 tcp_init_congestion_control(sk);
5343
5344 /* Prevent spurious tcp_cwnd_restart() on first data
5345 * packet.
5346 */
5347 tp->lsndtime = tcp_time_stamp;
5348
5349 tcp_init_buffer_space(sk);
5350
5351 if (sock_flag(sk, SOCK_KEEPOPEN))
5352 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5353
5354 if (!tp->rx_opt.snd_wscale)
5355 __tcp_fast_path_on(tp, tp->snd_wnd);
5356 else
5357 tp->pred_flags = 0;
5358
5359 if (!sock_flag(sk, SOCK_DEAD)) {
5360 sk->sk_state_change(sk);
5361 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5362 }
5363}
5364
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005365static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5366 struct tcp_fastopen_cookie *cookie)
5367{
5368 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng67da22d2012-07-19 06:43:11 +00005369 struct sk_buff *data = tp->syn_data ? tcp_write_queue_head(sk) : NULL;
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005370 u16 mss = tp->rx_opt.mss_clamp;
Yuchung Chengaab48742012-07-19 06:43:10 +00005371 bool syn_drop;
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005372
5373 if (mss == tp->rx_opt.user_mss) {
5374 struct tcp_options_received opt;
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005375
5376 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5377 tcp_clear_options(&opt);
5378 opt.user_mss = opt.mss_clamp = 0;
Christoph Paasch1a2c6182013-03-17 08:23:34 +00005379 tcp_parse_options(synack, &opt, 0, NULL);
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005380 mss = opt.mss_clamp;
5381 }
5382
Yuchung Cheng67da22d2012-07-19 06:43:11 +00005383 if (!tp->syn_fastopen) /* Ignore an unsolicited cookie */
5384 cookie->len = -1;
5385
Yuchung Chengaab48742012-07-19 06:43:10 +00005386 /* The SYN-ACK neither has cookie nor acknowledges the data. Presumably
5387 * the remote receives only the retransmitted (regular) SYNs: either
5388 * the original SYN-data or the corresponding SYN-ACK is lost.
5389 */
Yuchung Cheng66555e92013-01-31 11:16:46 -08005390 syn_drop = (cookie->len <= 0 && data && tp->total_retrans);
Yuchung Chengaab48742012-07-19 06:43:10 +00005391
5392 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop);
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005393
5394 if (data) { /* Retransmit unacked data in SYN */
Yuchung Cheng93b174a2012-12-06 08:45:32 +00005395 tcp_for_write_queue_from(data, sk) {
5396 if (data == tcp_send_head(sk) ||
5397 __tcp_retransmit_skb(sk, data))
5398 break;
5399 }
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005400 tcp_rearm_rto(sk);
5401 return true;
5402 }
Yuchung Cheng6f736012012-10-19 15:14:44 +00005403 tp->syn_data_acked = tp->syn_data;
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005404 return false;
5405}
5406
Linus Torvalds1da177e2005-04-16 15:20:36 -07005407static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
Eric Dumazetcf533ea2011-10-21 05:22:42 -04005408 const struct tcphdr *th, unsigned int len)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005409{
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08005410 struct inet_connection_sock *icsk = inet_csk(sk);
William Allen Simpson4957faade2009-12-02 18:25:27 +00005411 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005412 struct tcp_fastopen_cookie foc = { .len = -1 };
William Allen Simpson4957faade2009-12-02 18:25:27 +00005413 int saved_clamp = tp->rx_opt.mss_clamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005414
Christoph Paasch1a2c6182013-03-17 08:23:34 +00005415 tcp_parse_options(skb, &tp->rx_opt, 0, &foc);
Andrew Vagine3e12022013-08-27 12:21:55 +04005416 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
Andrey Vaginee684b62013-02-11 05:50:19 +00005417 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005418
5419 if (th->ack) {
5420 /* rfc793:
5421 * "If the state is SYN-SENT then
5422 * first check the ACK bit
5423 * If the ACK bit is set
5424 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5425 * a reset (unless the RST bit is set, if so drop
5426 * the segment and return)"
Linus Torvalds1da177e2005-04-16 15:20:36 -07005427 */
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005428 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5429 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005430 goto reset_and_undo;
5431
5432 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5433 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5434 tcp_time_stamp)) {
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07005435 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005436 goto reset_and_undo;
5437 }
5438
5439 /* Now ACK is acceptable.
5440 *
5441 * "If the RST bit is set
5442 * If the ACK was acceptable then signal the user "error:
5443 * connection reset", drop the segment, enter CLOSED state,
5444 * delete TCB, and return."
5445 */
5446
5447 if (th->rst) {
5448 tcp_reset(sk);
5449 goto discard;
5450 }
5451
5452 /* rfc793:
5453 * "fifth, if neither of the SYN or RST bits is set then
5454 * drop the segment and return."
5455 *
5456 * See note below!
5457 * --ANK(990513)
5458 */
5459 if (!th->syn)
5460 goto discard_and_undo;
5461
5462 /* rfc793:
5463 * "If the SYN bit is on ...
5464 * are acceptable then ...
5465 * (our SYN has been ACKed), change the connection
5466 * state to ESTABLISHED..."
5467 */
5468
5469 TCP_ECN_rcv_synack(tp, th);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005470
Razvan Ghitulete1b3a6922012-08-14 16:30:20 +03005471 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005472 tcp_ack(sk, skb, FLAG_SLOWPATH);
5473
5474 /* Ok.. it's good. Set up sequence numbers and
5475 * move to established.
5476 */
5477 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5478 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5479
5480 /* RFC1323: The window in SYN & SYN/ACK segments is
5481 * never scaled.
5482 */
5483 tp->snd_wnd = ntohs(th->window);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005484
5485 if (!tp->rx_opt.wscale_ok) {
5486 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5487 tp->window_clamp = min(tp->window_clamp, 65535U);
5488 }
5489
5490 if (tp->rx_opt.saw_tstamp) {
5491 tp->rx_opt.tstamp_ok = 1;
5492 tp->tcp_header_len =
5493 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5494 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5495 tcp_store_ts_recent(tp);
5496 } else {
5497 tp->tcp_header_len = sizeof(struct tcphdr);
5498 }
5499
Ilpo Järvinene60402d2007-08-09 15:14:46 +03005500 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5501 tcp_enable_fack(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005502
John Heffner5d424d52006-03-20 17:53:41 -08005503 tcp_mtup_init(sk);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08005504 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005505 tcp_initialize_rcv_mss(sk);
5506
5507 /* Remember, tcp_poll() does not lock socket!
5508 * Change state from SYN-SENT only after copied_seq
5509 * is initialized. */
5510 tp->copied_seq = tp->rcv_nxt;
William Allen Simpson4957faade2009-12-02 18:25:27 +00005511
Ralf Baechlee16aa202006-12-07 00:11:33 -08005512 smp_mb();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005513
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005514 tcp_finish_connect(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005515
Yuchung Cheng67da22d2012-07-19 06:43:11 +00005516 if ((tp->syn_fastopen || tp->syn_data) &&
5517 tcp_rcv_fastopen_synack(sk, skb, &foc))
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00005518 return -1;
5519
Arnaldo Carvalho de Melo295f7322005-08-09 20:11:56 -07005520 if (sk->sk_write_pending ||
5521 icsk->icsk_accept_queue.rskq_defer_accept ||
5522 icsk->icsk_ack.pingpong) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005523 /* Save one ACK. Data will be ready after
5524 * several ticks, if write_pending is set.
5525 *
5526 * It may be deleted, but with this feature tcpdumps
5527 * look so _wonderfully_ clever, that I was not able
5528 * to stand against the temptation 8) --ANK
5529 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005530 inet_csk_schedule_ack(sk);
Arnaldo Carvalho de Melo295f7322005-08-09 20:11:56 -07005531 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005532 tcp_enter_quickack_mode(sk);
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07005533 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5534 TCP_DELACK_MAX, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005535
5536discard:
5537 __kfree_skb(skb);
5538 return 0;
5539 } else {
5540 tcp_send_ack(sk);
5541 }
5542 return -1;
5543 }
5544
5545 /* No ACK in the segment */
5546
5547 if (th->rst) {
5548 /* rfc793:
5549 * "If the RST bit is set
5550 *
5551 * Otherwise (no ACK) drop the segment and return."
5552 */
5553
5554 goto discard_and_undo;
5555 }
5556
5557 /* PAWS check. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005558 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
Ilpo Järvinenc887e6d2009-03-14 14:23:03 +00005559 tcp_paws_reject(&tp->rx_opt, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005560 goto discard_and_undo;
5561
5562 if (th->syn) {
5563 /* We see SYN without ACK. It is attempt of
5564 * simultaneous connect with crossed SYNs.
5565 * Particularly, it can be connect to self.
5566 */
5567 tcp_set_state(sk, TCP_SYN_RECV);
5568
5569 if (tp->rx_opt.saw_tstamp) {
5570 tp->rx_opt.tstamp_ok = 1;
5571 tcp_store_ts_recent(tp);
5572 tp->tcp_header_len =
5573 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5574 } else {
5575 tp->tcp_header_len = sizeof(struct tcphdr);
5576 }
5577
5578 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5579 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5580
5581 /* RFC1323: The window in SYN & SYN/ACK segments is
5582 * never scaled.
5583 */
5584 tp->snd_wnd = ntohs(th->window);
5585 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5586 tp->max_window = tp->snd_wnd;
5587
5588 TCP_ECN_rcv_syn(tp, th);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005589
John Heffner5d424d52006-03-20 17:53:41 -08005590 tcp_mtup_init(sk);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08005591 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005592 tcp_initialize_rcv_mss(sk);
5593
Linus Torvalds1da177e2005-04-16 15:20:36 -07005594 tcp_send_synack(sk);
5595#if 0
5596 /* Note, we could accept data and URG from this segment.
Jerry Chu168a8f52012-08-31 12:29:13 +00005597 * There are no obstacles to make this (except that we must
5598 * either change tcp_recvmsg() to prevent it from returning data
5599 * before 3WHS completes per RFC793, or employ TCP Fast Open).
Linus Torvalds1da177e2005-04-16 15:20:36 -07005600 *
5601 * However, if we ignore data in ACKless segments sometimes,
5602 * we have no reasons to accept it sometimes.
5603 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5604 * is not flawless. So, discard packet for sanity.
5605 * Uncomment this return to process the data.
5606 */
5607 return -1;
5608#else
5609 goto discard;
5610#endif
5611 }
5612 /* "fifth, if neither of the SYN or RST bits is set then
5613 * drop the segment and return."
5614 */
5615
5616discard_and_undo:
5617 tcp_clear_options(&tp->rx_opt);
5618 tp->rx_opt.mss_clamp = saved_clamp;
5619 goto discard;
5620
5621reset_and_undo:
5622 tcp_clear_options(&tp->rx_opt);
5623 tp->rx_opt.mss_clamp = saved_clamp;
5624 return 1;
5625}
5626
Linus Torvalds1da177e2005-04-16 15:20:36 -07005627/*
5628 * This function implements the receiving procedure of RFC 793 for
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005629 * all states except ESTABLISHED and TIME_WAIT.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005630 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5631 * address independent.
5632 */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005633
Linus Torvalds1da177e2005-04-16 15:20:36 -07005634int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
Eric Dumazetcf533ea2011-10-21 05:22:42 -04005635 const struct tcphdr *th, unsigned int len)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005636{
5637 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -08005638 struct inet_connection_sock *icsk = inet_csk(sk);
Jerry Chu168a8f52012-08-31 12:29:13 +00005639 struct request_sock *req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005640 int queued = 0;
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005641 bool acceptable;
Yuchung Chengbc15afa2013-10-24 08:44:25 -07005642 u32 synack_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005643
5644 tp->rx_opt.saw_tstamp = 0;
5645
5646 switch (sk->sk_state) {
5647 case TCP_CLOSE:
5648 goto discard;
5649
5650 case TCP_LISTEN:
Stephen Hemminger2de979b2007-03-08 20:45:19 -08005651 if (th->ack)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005652 return 1;
5653
Stephen Hemminger2de979b2007-03-08 20:45:19 -08005654 if (th->rst)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005655 goto discard;
5656
Stephen Hemminger2de979b2007-03-08 20:45:19 -08005657 if (th->syn) {
Eric Dumazetfdf5af02011-12-02 23:41:42 +00005658 if (th->fin)
5659 goto discard;
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -08005660 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005661 return 1;
5662
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005663 /* Now we have several options: In theory there is
5664 * nothing else in the frame. KA9Q has an option to
Linus Torvalds1da177e2005-04-16 15:20:36 -07005665 * send data with the syn, BSD accepts data with the
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005666 * syn up to the [to be] advertised window and
5667 * Solaris 2.1 gives you a protocol error. For now
5668 * we just ignore it, that fits the spec precisely
Linus Torvalds1da177e2005-04-16 15:20:36 -07005669 * and avoids incompatibilities. It would be nice in
5670 * future to drop through and process the data.
5671 *
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005672 * Now that TTCP is starting to be used we ought to
Linus Torvalds1da177e2005-04-16 15:20:36 -07005673 * queue this data.
5674 * But, this leaves one open to an easy denial of
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005675 * service attack, and SYN cookies can't defend
Linus Torvalds1da177e2005-04-16 15:20:36 -07005676 * against this problem. So, we drop the data
Masayuki Nakagawafb7e2392007-01-23 20:15:06 -08005677 * in the interest of security over speed unless
5678 * it's still in use.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005679 */
Masayuki Nakagawafb7e2392007-01-23 20:15:06 -08005680 kfree_skb(skb);
5681 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005682 }
5683 goto discard;
5684
5685 case TCP_SYN_SENT:
Linus Torvalds1da177e2005-04-16 15:20:36 -07005686 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5687 if (queued >= 0)
5688 return queued;
5689
5690 /* Do step6 onward by hand. */
5691 tcp_urg(sk, skb, th);
5692 __kfree_skb(skb);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005693 tcp_data_snd_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005694 return 0;
5695 }
5696
Jerry Chu168a8f52012-08-31 12:29:13 +00005697 req = tp->fastopen_rsk;
5698 if (req != NULL) {
Jerry Chu37561f62012-10-22 11:26:36 +00005699 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
Jerry Chu168a8f52012-08-31 12:29:13 +00005700 sk->sk_state != TCP_FIN_WAIT1);
5701
5702 if (tcp_check_req(sk, skb, req, NULL, true) == NULL)
5703 goto discard;
Neal Cardwelle69bebd2012-09-22 04:18:57 +00005704 }
Eric Dumazetc3ae62af2012-12-26 12:44:34 +00005705
Eric Dumazet7b514a82013-01-10 16:18:47 +00005706 if (!th->ack && !th->rst)
Eric Dumazetc3ae62af2012-12-26 12:44:34 +00005707 goto discard;
5708
Neal Cardwelle69bebd2012-09-22 04:18:57 +00005709 if (!tcp_validate_incoming(sk, skb, th, 0))
Eric Dumazet0c246042012-07-17 01:41:30 +00005710 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005711
5712 /* step 5: check the ACK field */
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005713 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
5714 FLAG_UPDATE_TS_RECENT) > 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005715
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005716 switch (sk->sk_state) {
5717 case TCP_SYN_RECV:
Joe Perches61eb90032013-05-24 18:36:13 +00005718 if (!acceptable)
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005719 return 1;
Joe Perches61eb90032013-05-24 18:36:13 +00005720
5721 /* Once we leave TCP_SYN_RECV, we no longer need req
5722 * so release it.
5723 */
5724 if (req) {
Yuchung Chengbc15afa2013-10-24 08:44:25 -07005725 synack_stamp = tcp_rsk(req)->snt_synack;
Joe Perches61eb90032013-05-24 18:36:13 +00005726 tp->total_retrans = req->num_retrans;
Joe Perches61eb90032013-05-24 18:36:13 +00005727 reqsk_fastopen_remove(sk, req, false);
5728 } else {
Yuchung Chengbc15afa2013-10-24 08:44:25 -07005729 synack_stamp = tp->lsndtime;
Joe Perches61eb90032013-05-24 18:36:13 +00005730 /* Make sure socket is routed, for correct metrics. */
5731 icsk->icsk_af_ops->rebuild_header(sk);
5732 tcp_init_congestion_control(sk);
5733
5734 tcp_mtup_init(sk);
Joe Perches61eb90032013-05-24 18:36:13 +00005735 tp->copied_seq = tp->rcv_nxt;
Eric Dumazetb0983d32013-09-20 13:56:58 -07005736 tcp_init_buffer_space(sk);
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005737 }
Joe Perches61eb90032013-05-24 18:36:13 +00005738 smp_mb();
5739 tcp_set_state(sk, TCP_ESTABLISHED);
5740 sk->sk_state_change(sk);
5741
5742 /* Note, that this wakeup is only for marginal crossed SYN case.
5743 * Passively open sockets are not waked up, because
5744 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5745 */
5746 if (sk->sk_socket)
5747 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5748
5749 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5750 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
5751 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
Yuchung Chengbc15afa2013-10-24 08:44:25 -07005752 tcp_synack_rtt_meas(sk, synack_stamp);
Joe Perches61eb90032013-05-24 18:36:13 +00005753
5754 if (tp->rx_opt.tstamp_ok)
5755 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5756
5757 if (req) {
5758 /* Re-arm the timer because data may have been sent out.
5759 * This is similar to the regular data transmission case
5760 * when new data has just been ack'ed.
5761 *
5762 * (TFO) - we could try to be more aggressive and
5763 * retransmitting any data sooner based on when they
5764 * are sent out.
5765 */
5766 tcp_rearm_rto(sk);
5767 } else
5768 tcp_init_metrics(sk);
5769
Neal Cardwell02cf4eb2013-10-21 15:40:19 -04005770 tcp_update_pacing_rate(sk);
5771
Joe Perches61eb90032013-05-24 18:36:13 +00005772 /* Prevent spurious tcp_cwnd_restart() on first data packet */
5773 tp->lsndtime = tcp_time_stamp;
5774
5775 tcp_initialize_rcv_mss(sk);
5776 tcp_fast_path_on(tp);
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005777 break;
5778
Joe Perchesc48b22d2013-05-24 18:06:58 +00005779 case TCP_FIN_WAIT1: {
5780 struct dst_entry *dst;
5781 int tmo;
5782
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005783 /* If we enter the TCP_FIN_WAIT1 state and we are a
5784 * Fast Open socket and this is the first acceptable
5785 * ACK we have received, this would have acknowledged
5786 * our SYNACK so stop the SYNACK timer.
5787 */
5788 if (req != NULL) {
5789 /* Return RST if ack_seq is invalid.
5790 * Note that RFC793 only says to generate a
5791 * DUPACK for it but for TCP Fast Open it seems
5792 * better to treat this case like TCP_SYN_RECV
5793 * above.
5794 */
5795 if (!acceptable)
5796 return 1;
5797 /* We no longer need the request sock. */
5798 reqsk_fastopen_remove(sk, req, false);
5799 tcp_rearm_rto(sk);
5800 }
Joe Perchesc48b22d2013-05-24 18:06:58 +00005801 if (tp->snd_una != tp->write_seq)
5802 break;
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005803
Joe Perchesc48b22d2013-05-24 18:06:58 +00005804 tcp_set_state(sk, TCP_FIN_WAIT2);
5805 sk->sk_shutdown |= SEND_SHUTDOWN;
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005806
Joe Perchesc48b22d2013-05-24 18:06:58 +00005807 dst = __sk_dst_get(sk);
5808 if (dst)
5809 dst_confirm(dst);
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005810
Joe Perchesc48b22d2013-05-24 18:06:58 +00005811 if (!sock_flag(sk, SOCK_DEAD)) {
5812 /* Wake up lingering close() */
5813 sk->sk_state_change(sk);
5814 break;
5815 }
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005816
Joe Perchesc48b22d2013-05-24 18:06:58 +00005817 if (tp->linger2 < 0 ||
5818 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5819 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5820 tcp_done(sk);
5821 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5822 return 1;
5823 }
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005824
Joe Perchesc48b22d2013-05-24 18:06:58 +00005825 tmo = tcp_fin_time(sk);
5826 if (tmo > TCP_TIMEWAIT_LEN) {
5827 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5828 } else if (th->fin || sock_owned_by_user(sk)) {
5829 /* Bad case. We could lose such FIN otherwise.
5830 * It is not a big problem, but it looks confusing
5831 * and not so rare event. We still can lose it now,
5832 * if it spins in bh_lock_sock(), but it is really
5833 * marginal case.
5834 */
5835 inet_csk_reset_keepalive_timer(sk, tmo);
5836 } else {
5837 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5838 goto discard;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005839 }
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005840 break;
Joe Perchesc48b22d2013-05-24 18:06:58 +00005841 }
Eric Dumazet1f6afc82013-05-24 15:03:54 +00005842
5843 case TCP_CLOSING:
5844 if (tp->snd_una == tp->write_seq) {
5845 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5846 goto discard;
5847 }
5848 break;
5849
5850 case TCP_LAST_ACK:
5851 if (tp->snd_una == tp->write_seq) {
5852 tcp_update_metrics(sk);
5853 tcp_done(sk);
5854 goto discard;
5855 }
5856 break;
Eric Dumazetc3ae62af2012-12-26 12:44:34 +00005857 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07005858
5859 /* step 6: check the URG bit */
5860 tcp_urg(sk, skb, th);
5861
5862 /* step 7: process the segment text */
5863 switch (sk->sk_state) {
5864 case TCP_CLOSE_WAIT:
5865 case TCP_CLOSING:
5866 case TCP_LAST_ACK:
5867 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5868 break;
5869 case TCP_FIN_WAIT1:
5870 case TCP_FIN_WAIT2:
5871 /* RFC 793 says to queue data in these states,
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005872 * RFC 1122 says we MUST send a reset.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005873 * BSD 4.4 also does reset.
5874 */
5875 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5876 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5877 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07005878 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005879 tcp_reset(sk);
5880 return 1;
5881 }
5882 }
5883 /* Fall through */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005884 case TCP_ESTABLISHED:
Linus Torvalds1da177e2005-04-16 15:20:36 -07005885 tcp_data_queue(sk, skb);
5886 queued = 1;
5887 break;
5888 }
5889
5890 /* tcp_data could move socket to TIME-WAIT */
5891 if (sk->sk_state != TCP_CLOSE) {
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005892 tcp_data_snd_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005893 tcp_ack_snd_check(sk);
5894 }
5895
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005896 if (!queued) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005897discard:
5898 __kfree_skb(skb);
5899 }
5900 return 0;
5901}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005902EXPORT_SYMBOL(tcp_rcv_state_process);