blob: 2bc8e27a163daafed09a3bf2ee6a145414848d72 [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
Linus Torvalds1da177e2005-04-16 15:20:36 -070064#include <linux/mm.h>
65#include <linux/module.h>
66#include <linux/sysctl.h>
Ilpo Järvinena0bffff2009-03-21 13:36:17 -070067#include <linux/kernel.h>
Satoru SATOH5ffc02a2008-05-04 22:14:42 -070068#include <net/dst.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070069#include <net/tcp.h>
70#include <net/inet_common.h>
71#include <linux/ipsec.h>
72#include <asm/unaligned.h>
Chris Leech1a2449a2006-05-23 18:05:53 -070073#include <net/netdma.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070074
Brian Haleyab32ea52006-09-22 14:15:41 -070075int sysctl_tcp_timestamps __read_mostly = 1;
76int sysctl_tcp_window_scaling __read_mostly = 1;
77int sysctl_tcp_sack __read_mostly = 1;
78int sysctl_tcp_fack __read_mostly = 1;
79int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
80int sysctl_tcp_ecn __read_mostly;
81int sysctl_tcp_dsack __read_mostly = 1;
82int sysctl_tcp_app_win __read_mostly = 31;
83int sysctl_tcp_adv_win_scale __read_mostly = 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -070084
Brian Haleyab32ea52006-09-22 14:15:41 -070085int sysctl_tcp_stdurg __read_mostly;
86int sysctl_tcp_rfc1337 __read_mostly;
87int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
Ilpo Järvinenc96fd3d2007-09-20 11:36:37 -070088int sysctl_tcp_frto __read_mostly = 2;
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -080089int sysctl_tcp_frto_response __read_mostly;
Brian Haleyab32ea52006-09-22 14:15:41 -070090int sysctl_tcp_nometrics_save __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -070091
Brian Haleyab32ea52006-09-22 14:15:41 -070092int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
93int sysctl_tcp_abc __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -070094
Linus Torvalds1da177e2005-04-16 15:20:36 -070095#define FLAG_DATA 0x01 /* Incoming frame contained data. */
96#define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
97#define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
98#define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
99#define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
100#define FLAG_DATA_SACKED 0x20 /* New SACK. */
101#define FLAG_ECE 0x40 /* ECE in this ACK */
102#define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */
103#define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
Ilpo Järvinen4dc26652007-02-21 23:16:11 -0800104#define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */
Ilpo Järvinen2e605292007-08-02 19:46:58 -0700105#define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
Ryousei Takano564262c2007-10-25 23:03:52 -0700106#define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
Ilpo Järvinen009a2e32007-09-20 11:34:38 -0700107#define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */
Ilpo Järvinencadbd032007-12-31 04:49:21 -0800108#define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700109
110#define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
111#define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
112#define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
113#define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
Ilpo Järvinen2e605292007-08-02 19:46:58 -0700114#define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700115
Linus Torvalds1da177e2005-04-16 15:20:36 -0700116#define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700117#define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700118
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900119/* Adapt the MSS value used to make delayed ack decision to the
Linus Torvalds1da177e2005-04-16 15:20:36 -0700120 * real world.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900121 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800122static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700123{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700124 struct inet_connection_sock *icsk = inet_csk(sk);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900125 const unsigned int lss = icsk->icsk_ack.last_seg_size;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700126 unsigned int len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700127
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900128 icsk->icsk_ack.last_seg_size = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700129
130 /* skb->len may jitter because of SACKs, even if peer
131 * sends good full-sized frames.
132 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800133 len = skb_shinfo(skb)->gso_size ? : skb->len;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700134 if (len >= icsk->icsk_ack.rcv_mss) {
135 icsk->icsk_ack.rcv_mss = len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700136 } else {
137 /* Otherwise, we make more careful check taking into account,
138 * that SACKs block is variable.
139 *
140 * "len" is invariant segment length, including TCP header.
141 */
Arnaldo Carvalho de Melo9c702202007-04-25 18:04:18 -0700142 len += skb->data - skb_transport_header(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700143 if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) ||
144 /* If PSH is not set, packet should be
145 * full sized, provided peer TCP is not badly broken.
146 * This observation (if it is correct 8)) allows
147 * to handle super-low mtu links fairly.
148 */
149 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -0700150 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700151 /* Subtract also invariant (if peer is RFC compliant),
152 * tcp header plus fixed timestamp option length.
153 * Resulting "len" is MSS free of SACK jitter.
154 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700155 len -= tcp_sk(sk)->tcp_header_len;
156 icsk->icsk_ack.last_seg_size = len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700157 if (len == lss) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700158 icsk->icsk_ack.rcv_mss = len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700159 return;
160 }
161 }
Alexey Kuznetsov1ef96962006-09-19 12:52:50 -0700162 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
163 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700164 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700165 }
166}
167
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700168static void tcp_incr_quickack(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700169{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700170 struct inet_connection_sock *icsk = inet_csk(sk);
171 unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700172
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800173 if (quickacks == 0)
174 quickacks = 2;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700175 if (quickacks > icsk->icsk_ack.quick)
176 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700177}
178
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700179void tcp_enter_quickack_mode(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700180{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700181 struct inet_connection_sock *icsk = inet_csk(sk);
182 tcp_incr_quickack(sk);
183 icsk->icsk_ack.pingpong = 0;
184 icsk->icsk_ack.ato = TCP_ATO_MIN;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700185}
186
187/* Send ACKs quickly, if "quick" count is not exhausted
188 * and the session is not interactive.
189 */
190
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700191static inline int tcp_in_quickack_mode(const struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700192{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700193 const struct inet_connection_sock *icsk = inet_csk(sk);
194 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700195}
196
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700197static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
198{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800199 if (tp->ecn_flags & TCP_ECN_OK)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700200 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
201}
202
203static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, struct sk_buff *skb)
204{
205 if (tcp_hdr(skb)->cwr)
206 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
207}
208
209static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
210{
211 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
212}
213
214static inline void TCP_ECN_check_ce(struct tcp_sock *tp, struct sk_buff *skb)
215{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800216 if (tp->ecn_flags & TCP_ECN_OK) {
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700217 if (INET_ECN_is_ce(TCP_SKB_CB(skb)->flags))
218 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
219 /* Funny extension: if ECT is not set on a segment,
220 * it is surely retransmit. It is not in ECN RFC,
221 * but Linux follows this rule. */
222 else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb)->flags)))
223 tcp_enter_quickack_mode((struct sock *)tp);
224 }
225}
226
227static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, struct tcphdr *th)
228{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800229 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700230 tp->ecn_flags &= ~TCP_ECN_OK;
231}
232
233static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, struct tcphdr *th)
234{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800235 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700236 tp->ecn_flags &= ~TCP_ECN_OK;
237}
238
239static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock *tp, struct tcphdr *th)
240{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800241 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700242 return 1;
243 return 0;
244}
245
Linus Torvalds1da177e2005-04-16 15:20:36 -0700246/* Buffer size and advertised window tuning.
247 *
248 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
249 */
250
251static void tcp_fixup_sndbuf(struct sock *sk)
252{
253 int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 +
254 sizeof(struct sk_buff);
255
256 if (sk->sk_sndbuf < 3 * sndmem)
257 sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]);
258}
259
260/* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
261 *
262 * All tcp_full_space() is split to two parts: "network" buffer, allocated
263 * forward and advertised in receiver window (tp->rcv_wnd) and
264 * "application buffer", required to isolate scheduling/application
265 * latencies from network.
266 * window_clamp is maximal advertised window. It can be less than
267 * tcp_full_space(), in this case tcp_full_space() - window_clamp
268 * is reserved for "application" buffer. The less window_clamp is
269 * the smoother our behaviour from viewpoint of network, but the lower
270 * throughput and the higher sensitivity of the connection to losses. 8)
271 *
272 * rcv_ssthresh is more strict window_clamp used at "slow start"
273 * phase to predict further behaviour of this connection.
274 * It is used for two goals:
275 * - to enforce header prediction at sender, even when application
276 * requires some significant "application buffer". It is check #1.
277 * - to prevent pruning of receive queue because of misprediction
278 * of receiver window. Check #2.
279 *
280 * The scheme does not work when sender sends good segments opening
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800281 * window and then starts to feed us spaghetti. But it should work
Linus Torvalds1da177e2005-04-16 15:20:36 -0700282 * in common situations. Otherwise, we have to rely on queue collapsing.
283 */
284
285/* Slow part of check#2. */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700286static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700287{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700288 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700289 /* Optimize this! */
Eric Dumazetdfd4f0a2007-12-21 06:07:53 -0800290 int truesize = tcp_win_from_space(skb->truesize) >> 1;
291 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700292
293 while (tp->rcv_ssthresh <= window) {
294 if (truesize <= skb->len)
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700295 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700296
297 truesize >>= 1;
298 window >>= 1;
299 }
300 return 0;
301}
302
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800303static void tcp_grow_window(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700304{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700305 struct tcp_sock *tp = tcp_sk(sk);
306
Linus Torvalds1da177e2005-04-16 15:20:36 -0700307 /* Check #1 */
308 if (tp->rcv_ssthresh < tp->window_clamp &&
309 (int)tp->rcv_ssthresh < tcp_space(sk) &&
310 !tcp_memory_pressure) {
311 int incr;
312
313 /* Check #2. Increase window, if skb with such overhead
314 * will fit to rcvbuf in future.
315 */
316 if (tcp_win_from_space(skb->truesize) <= skb->len)
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800317 incr = 2 * tp->advmss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700318 else
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700319 incr = __tcp_grow_window(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700320
321 if (incr) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800322 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
323 tp->window_clamp);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700324 inet_csk(sk)->icsk_ack.quick |= 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700325 }
326 }
327}
328
329/* 3. Tuning rcvbuf, when connection enters established state. */
330
331static void tcp_fixup_rcvbuf(struct sock *sk)
332{
333 struct tcp_sock *tp = tcp_sk(sk);
334 int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
335
336 /* Try to select rcvbuf so that 4 mss-sized segments
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800337 * will fit to window and corresponding skbs will fit to our rcvbuf.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700338 * (was 3; 4 is minimum to allow fast retransmit to work.)
339 */
340 while (tcp_win_from_space(rcvmem) < tp->advmss)
341 rcvmem += 128;
342 if (sk->sk_rcvbuf < 4 * rcvmem)
343 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]);
344}
345
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800346/* 4. Try to fixup all. It is made immediately after connection enters
Linus Torvalds1da177e2005-04-16 15:20:36 -0700347 * established state.
348 */
349static void tcp_init_buffer_space(struct sock *sk)
350{
351 struct tcp_sock *tp = tcp_sk(sk);
352 int maxwin;
353
354 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
355 tcp_fixup_rcvbuf(sk);
356 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
357 tcp_fixup_sndbuf(sk);
358
359 tp->rcvq_space.space = tp->rcv_wnd;
360
361 maxwin = tcp_full_space(sk);
362
363 if (tp->window_clamp >= maxwin) {
364 tp->window_clamp = maxwin;
365
366 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
367 tp->window_clamp = max(maxwin -
368 (maxwin >> sysctl_tcp_app_win),
369 4 * tp->advmss);
370 }
371
372 /* Force reservation of one segment. */
373 if (sysctl_tcp_app_win &&
374 tp->window_clamp > 2 * tp->advmss &&
375 tp->window_clamp + tp->advmss > maxwin)
376 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
377
378 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
379 tp->snd_cwnd_stamp = tcp_time_stamp;
380}
381
Linus Torvalds1da177e2005-04-16 15:20:36 -0700382/* 5. Recalculate window clamp after socket hit its memory bounds. */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700383static void tcp_clamp_window(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700384{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700385 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300386 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700387
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300388 icsk->icsk_ack.quick = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700389
John Heffner326f36e2005-11-10 17:11:48 -0800390 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
391 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
392 !tcp_memory_pressure &&
393 atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
394 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
395 sysctl_tcp_rmem[2]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700396 }
John Heffner326f36e2005-11-10 17:11:48 -0800397 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800398 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700399}
400
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800401/* Initialize RCV_MSS value.
402 * RCV_MSS is an our guess about MSS used by the peer.
403 * We haven't any direct information about the MSS.
404 * It's better to underestimate the RCV_MSS rather than overestimate.
405 * Overestimations make us ACKing less frequently than needed.
406 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
407 */
408void tcp_initialize_rcv_mss(struct sock *sk)
409{
410 struct tcp_sock *tp = tcp_sk(sk);
411 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
412
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800413 hint = min(hint, tp->rcv_wnd / 2);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800414 hint = min(hint, TCP_MIN_RCVMSS);
415 hint = max(hint, TCP_MIN_MSS);
416
417 inet_csk(sk)->icsk_ack.rcv_mss = hint;
418}
419
Linus Torvalds1da177e2005-04-16 15:20:36 -0700420/* Receiver "autotuning" code.
421 *
422 * The algorithm for RTT estimation w/o timestamps is based on
423 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
424 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
425 *
426 * More detail on this code can be found at
427 * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
428 * though this reference is out of date. A new paper
429 * is pending.
430 */
431static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
432{
433 u32 new_sample = tp->rcv_rtt_est.rtt;
434 long m = sample;
435
436 if (m == 0)
437 m = 1;
438
439 if (new_sample != 0) {
440 /* If we sample in larger samples in the non-timestamp
441 * case, we could grossly overestimate the RTT especially
442 * with chatty applications or bulk transfer apps which
443 * are stalled on filesystem I/O.
444 *
445 * Also, since we are only going for a minimum in the
Stephen Hemminger31f34262005-11-15 15:17:10 -0800446 * non-timestamp case, we do not smooth things out
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800447 * else with timestamps disabled convergence takes too
Linus Torvalds1da177e2005-04-16 15:20:36 -0700448 * long.
449 */
450 if (!win_dep) {
451 m -= (new_sample >> 3);
452 new_sample += m;
453 } else if (m < new_sample)
454 new_sample = m << 3;
455 } else {
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800456 /* No previous measure. */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700457 new_sample = m << 3;
458 }
459
460 if (tp->rcv_rtt_est.rtt != new_sample)
461 tp->rcv_rtt_est.rtt = new_sample;
462}
463
464static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
465{
466 if (tp->rcv_rtt_est.time == 0)
467 goto new_measure;
468 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
469 return;
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800470 tcp_rcv_rtt_update(tp, jiffies - tp->rcv_rtt_est.time, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700471
472new_measure:
473 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
474 tp->rcv_rtt_est.time = tcp_time_stamp;
475}
476
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800477static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
478 const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700479{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700480 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700481 if (tp->rx_opt.rcv_tsecr &&
482 (TCP_SKB_CB(skb)->end_seq -
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700483 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700484 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
485}
486
487/*
488 * This function should be called every time data is copied to user space.
489 * It calculates the appropriate TCP receive buffer space.
490 */
491void tcp_rcv_space_adjust(struct sock *sk)
492{
493 struct tcp_sock *tp = tcp_sk(sk);
494 int time;
495 int space;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900496
Linus Torvalds1da177e2005-04-16 15:20:36 -0700497 if (tp->rcvq_space.time == 0)
498 goto new_measure;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900499
Linus Torvalds1da177e2005-04-16 15:20:36 -0700500 time = tcp_time_stamp - tp->rcvq_space.time;
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800501 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700502 return;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900503
Linus Torvalds1da177e2005-04-16 15:20:36 -0700504 space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
505
506 space = max(tp->rcvq_space.space, space);
507
508 if (tp->rcvq_space.space != space) {
509 int rcvmem;
510
511 tp->rcvq_space.space = space;
512
John Heffner6fcf9412006-02-09 17:06:57 -0800513 if (sysctl_tcp_moderate_rcvbuf &&
514 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700515 int new_clamp = space;
516
517 /* Receive space grows, normalize in order to
518 * take into account packet headers and sk_buff
519 * structure overhead.
520 */
521 space /= tp->advmss;
522 if (!space)
523 space = 1;
524 rcvmem = (tp->advmss + MAX_TCP_HEADER +
525 16 + sizeof(struct sk_buff));
526 while (tcp_win_from_space(rcvmem) < tp->advmss)
527 rcvmem += 128;
528 space *= rcvmem;
529 space = min(space, sysctl_tcp_rmem[2]);
530 if (space > sk->sk_rcvbuf) {
531 sk->sk_rcvbuf = space;
532
533 /* Make the window clamp follow along. */
534 tp->window_clamp = new_clamp;
535 }
536 }
537 }
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900538
Linus Torvalds1da177e2005-04-16 15:20:36 -0700539new_measure:
540 tp->rcvq_space.seq = tp->copied_seq;
541 tp->rcvq_space.time = tcp_time_stamp;
542}
543
544/* There is something which you must keep in mind when you analyze the
545 * behavior of the tp->ato delayed ack timeout interval. When a
546 * connection starts up, we want to ack as quickly as possible. The
547 * problem is that "good" TCP's do slow start at the beginning of data
548 * transmission. The means that until we send the first few ACK's the
549 * sender will sit on his end and only queue most of his data, because
550 * he can only send snd_cwnd unacked packets at any given time. For
551 * each ACK we send, he increments snd_cwnd and transmits more of his
552 * queue. -DaveM
553 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700554static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700555{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700556 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700557 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700558 u32 now;
559
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700560 inet_csk_schedule_ack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700561
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700562 tcp_measure_rcv_mss(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700563
564 tcp_rcv_rtt_measure(tp);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900565
Linus Torvalds1da177e2005-04-16 15:20:36 -0700566 now = tcp_time_stamp;
567
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700568 if (!icsk->icsk_ack.ato) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700569 /* The _first_ data packet received, initialize
570 * delayed ACK engine.
571 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700572 tcp_incr_quickack(sk);
573 icsk->icsk_ack.ato = TCP_ATO_MIN;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700574 } else {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700575 int m = now - icsk->icsk_ack.lrcvtime;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700576
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800577 if (m <= TCP_ATO_MIN / 2) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700578 /* The fastest case is the first. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700579 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
580 } else if (m < icsk->icsk_ack.ato) {
581 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
582 if (icsk->icsk_ack.ato > icsk->icsk_rto)
583 icsk->icsk_ack.ato = icsk->icsk_rto;
584 } else if (m > icsk->icsk_rto) {
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800585 /* Too long gap. Apparently sender failed to
Linus Torvalds1da177e2005-04-16 15:20:36 -0700586 * restart window, so that we send ACKs quickly.
587 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700588 tcp_incr_quickack(sk);
Hideo Aoki3ab224b2007-12-31 00:11:19 -0800589 sk_mem_reclaim(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700590 }
591 }
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700592 icsk->icsk_ack.lrcvtime = now;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700593
594 TCP_ECN_check_ce(tp, skb);
595
596 if (skb->len >= 128)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700597 tcp_grow_window(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700598}
599
David S. Miller05bb1fa2007-08-30 22:10:28 -0700600static u32 tcp_rto_min(struct sock *sk)
601{
602 struct dst_entry *dst = __sk_dst_get(sk);
603 u32 rto_min = TCP_RTO_MIN;
604
David S. Miller5c127c52007-08-31 14:39:44 -0700605 if (dst && dst_metric_locked(dst, RTAX_RTO_MIN))
Stephen Hemmingerc1e20f72008-07-18 23:02:15 -0700606 rto_min = dst_metric_rtt(dst, RTAX_RTO_MIN);
David S. Miller05bb1fa2007-08-30 22:10:28 -0700607 return rto_min;
608}
609
Linus Torvalds1da177e2005-04-16 15:20:36 -0700610/* Called to compute a smoothed rtt estimate. The data fed to this
611 * routine either comes from timestamps, or from segments that were
612 * known _not_ to have been retransmitted [see Karn/Partridge
613 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
614 * piece by Van Jacobson.
615 * NOTE: the next three routines used to be one big routine.
616 * To save cycles in the RFC 1323 implementation it was better to break
617 * it up into three procedures. -- erics
618 */
Stephen Hemminger2d2abba2005-11-10 16:56:12 -0800619static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700620{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300621 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700622 long m = mrtt; /* RTT */
623
Linus Torvalds1da177e2005-04-16 15:20:36 -0700624 /* The following amusing code comes from Jacobson's
625 * article in SIGCOMM '88. Note that rtt and mdev
626 * are scaled versions of rtt and mean deviation.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900627 * This is designed to be as fast as possible
Linus Torvalds1da177e2005-04-16 15:20:36 -0700628 * m stands for "measurement".
629 *
630 * On a 1990 paper the rto value is changed to:
631 * RTO = rtt + 4 * mdev
632 *
633 * Funny. This algorithm seems to be very broken.
634 * These formulae increase RTO, when it should be decreased, increase
Stephen Hemminger31f34262005-11-15 15:17:10 -0800635 * too slowly, when it should be increased quickly, decrease too quickly
Linus Torvalds1da177e2005-04-16 15:20:36 -0700636 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
637 * does not matter how to _calculate_ it. Seems, it was trap
638 * that VJ failed to avoid. 8)
639 */
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800640 if (m == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700641 m = 1;
642 if (tp->srtt != 0) {
643 m -= (tp->srtt >> 3); /* m is now error in rtt est */
644 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
645 if (m < 0) {
646 m = -m; /* m is now abs(error) */
647 m -= (tp->mdev >> 2); /* similar update on mdev */
648 /* This is similar to one of Eifel findings.
649 * Eifel blocks mdev updates when rtt decreases.
650 * This solution is a bit different: we use finer gain
651 * for mdev in this case (alpha*beta).
652 * Like Eifel it also prevents growth of rto,
653 * but also it limits too fast rto decreases,
654 * happening in pure Eifel.
655 */
656 if (m > 0)
657 m >>= 3;
658 } else {
659 m -= (tp->mdev >> 2); /* similar update on mdev */
660 }
661 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
662 if (tp->mdev > tp->mdev_max) {
663 tp->mdev_max = tp->mdev;
664 if (tp->mdev_max > tp->rttvar)
665 tp->rttvar = tp->mdev_max;
666 }
667 if (after(tp->snd_una, tp->rtt_seq)) {
668 if (tp->mdev_max < tp->rttvar)
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800669 tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700670 tp->rtt_seq = tp->snd_nxt;
David S. Miller05bb1fa2007-08-30 22:10:28 -0700671 tp->mdev_max = tcp_rto_min(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700672 }
673 } else {
674 /* no previous measure. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800675 tp->srtt = m << 3; /* take the measured time to be rtt */
676 tp->mdev = m << 1; /* make sure rto = 3*rtt */
David S. Miller05bb1fa2007-08-30 22:10:28 -0700677 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700678 tp->rtt_seq = tp->snd_nxt;
679 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700680}
681
682/* Calculate rto without backoff. This is the second half of Van Jacobson's
683 * routine referred to above.
684 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700685static inline void tcp_set_rto(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700686{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700687 const struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700688 /* Old crap is replaced with new one. 8)
689 *
690 * More seriously:
691 * 1. If rtt variance happened to be less 50msec, it is hallucination.
692 * It cannot be less due to utterly erratic ACK generation made
693 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
694 * to do with delayed acks, because at cwnd>2 true delack timeout
695 * is invisible. Actually, Linux-2.4 also generates erratic
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800696 * ACKs in some circumstances.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700697 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700698 inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700699
700 /* 2. Fixups made earlier cannot be right.
701 * If we do not estimate RTO correctly without them,
702 * all the algo is pure shit and should be replaced
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800703 * with correct one. It is exactly, which we pretend to do.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700704 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700705
Ilpo Järvinenee6aac52008-12-05 22:43:08 -0800706 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
707 * guarantees that rto is higher.
708 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700709 if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX)
710 inet_csk(sk)->icsk_rto = TCP_RTO_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700711}
712
713/* Save metrics learned by this TCP session.
714 This function is called only, when TCP finishes successfully
715 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
716 */
717void tcp_update_metrics(struct sock *sk)
718{
719 struct tcp_sock *tp = tcp_sk(sk);
720 struct dst_entry *dst = __sk_dst_get(sk);
721
722 if (sysctl_tcp_nometrics_save)
723 return;
724
725 dst_confirm(dst);
726
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800727 if (dst && (dst->flags & DST_HOST)) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300728 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700729 int m;
Stephen Hemmingerc1e20f72008-07-18 23:02:15 -0700730 unsigned long rtt;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700731
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300732 if (icsk->icsk_backoff || !tp->srtt) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700733 /* This session failed to estimate rtt. Why?
734 * Probably, no packets returned in time.
735 * Reset our results.
736 */
737 if (!(dst_metric_locked(dst, RTAX_RTT)))
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800738 dst->metrics[RTAX_RTT - 1] = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700739 return;
740 }
741
Stephen Hemmingerc1e20f72008-07-18 23:02:15 -0700742 rtt = dst_metric_rtt(dst, RTAX_RTT);
743 m = rtt - tp->srtt;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700744
745 /* If newly calculated rtt larger than stored one,
746 * store new one. Otherwise, use EWMA. Remember,
747 * rtt overestimation is always better than underestimation.
748 */
749 if (!(dst_metric_locked(dst, RTAX_RTT))) {
750 if (m <= 0)
Stephen Hemmingerc1e20f72008-07-18 23:02:15 -0700751 set_dst_metric_rtt(dst, RTAX_RTT, tp->srtt);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700752 else
Stephen Hemmingerc1e20f72008-07-18 23:02:15 -0700753 set_dst_metric_rtt(dst, RTAX_RTT, rtt - (m >> 3));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700754 }
755
756 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
Stephen Hemmingerc1e20f72008-07-18 23:02:15 -0700757 unsigned long var;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700758 if (m < 0)
759 m = -m;
760
761 /* Scale deviation to rttvar fixed point */
762 m >>= 1;
763 if (m < tp->mdev)
764 m = tp->mdev;
765
Stephen Hemmingerc1e20f72008-07-18 23:02:15 -0700766 var = dst_metric_rtt(dst, RTAX_RTTVAR);
767 if (m >= var)
768 var = m;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700769 else
Stephen Hemmingerc1e20f72008-07-18 23:02:15 -0700770 var -= (var - m) >> 2;
771
772 set_dst_metric_rtt(dst, RTAX_RTTVAR, var);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700773 }
774
775 if (tp->snd_ssthresh >= 0xFFFF) {
776 /* Slow start still did not finish. */
777 if (dst_metric(dst, RTAX_SSTHRESH) &&
778 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
779 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
780 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
781 if (!dst_metric_locked(dst, RTAX_CWND) &&
782 tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800783 dst->metrics[RTAX_CWND - 1] = tp->snd_cwnd;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700784 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300785 icsk->icsk_ca_state == TCP_CA_Open) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700786 /* Cong. avoidance phase, cwnd is reliable. */
787 if (!dst_metric_locked(dst, RTAX_SSTHRESH))
788 dst->metrics[RTAX_SSTHRESH-1] =
789 max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
790 if (!dst_metric_locked(dst, RTAX_CWND))
Satoru SATOH5ffc02a2008-05-04 22:14:42 -0700791 dst->metrics[RTAX_CWND-1] = (dst_metric(dst, RTAX_CWND) + tp->snd_cwnd) >> 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700792 } else {
793 /* Else slow start did not finish, cwnd is non-sense,
794 ssthresh may be also invalid.
795 */
796 if (!dst_metric_locked(dst, RTAX_CWND))
Satoru SATOH5ffc02a2008-05-04 22:14:42 -0700797 dst->metrics[RTAX_CWND-1] = (dst_metric(dst, RTAX_CWND) + tp->snd_ssthresh) >> 1;
798 if (dst_metric(dst, RTAX_SSTHRESH) &&
Linus Torvalds1da177e2005-04-16 15:20:36 -0700799 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
Satoru SATOH5ffc02a2008-05-04 22:14:42 -0700800 tp->snd_ssthresh > dst_metric(dst, RTAX_SSTHRESH))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700801 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
802 }
803
804 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
Satoru SATOH5ffc02a2008-05-04 22:14:42 -0700805 if (dst_metric(dst, RTAX_REORDERING) < tp->reordering &&
Linus Torvalds1da177e2005-04-16 15:20:36 -0700806 tp->reordering != sysctl_tcp_reordering)
807 dst->metrics[RTAX_REORDERING-1] = tp->reordering;
808 }
809 }
810}
811
Gerrit Renker410e27a2008-09-09 13:27:22 +0200812/* Numbers are taken from RFC3390.
813 *
814 * John Heffner states:
815 *
816 * The RFC specifies a window of no more than 4380 bytes
817 * unless 2*MSS > 4380. Reading the pseudocode in the RFC
818 * is a bit misleading because they use a clamp at 4380 bytes
819 * rather than use a multiplier in the relevant range.
820 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700821__u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
822{
823 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
824
Gerrit Renker410e27a2008-09-09 13:27:22 +0200825 if (!cwnd) {
826 if (tp->mss_cache > 1460)
827 cwnd = 2;
828 else
829 cwnd = (tp->mss_cache > 1095) ? 3 : 4;
830 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700831 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
832}
833
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800834/* Set slow start threshold and cwnd not falling to slow start */
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -0800835void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800836{
837 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -0800838 const struct inet_connection_sock *icsk = inet_csk(sk);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800839
840 tp->prior_ssthresh = 0;
841 tp->bytes_acked = 0;
Ilpo Järvinene01f9d72007-03-02 13:27:25 -0800842 if (icsk->icsk_ca_state < TCP_CA_CWR) {
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800843 tp->undo_marker = 0;
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -0800844 if (set_ssthresh)
845 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800846 tp->snd_cwnd = min(tp->snd_cwnd,
847 tcp_packets_in_flight(tp) + 1U);
848 tp->snd_cwnd_cnt = 0;
849 tp->high_seq = tp->snd_nxt;
850 tp->snd_cwnd_stamp = tcp_time_stamp;
851 TCP_ECN_queue_cwr(tp);
852
853 tcp_set_ca_state(sk, TCP_CA_CWR);
854 }
855}
856
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300857/*
858 * Packet counting of FACK is based on in-order assumptions, therefore TCP
859 * disables it when reordering is detected
860 */
861static void tcp_disable_fack(struct tcp_sock *tp)
862{
Ilpo Järvinen85cc3912007-11-15 19:39:31 -0800863 /* RFC3517 uses different metric in lost marker => reset on change */
864 if (tcp_is_fack(tp))
865 tp->lost_skb_hint = NULL;
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300866 tp->rx_opt.sack_ok &= ~2;
867}
868
Ryousei Takano564262c2007-10-25 23:03:52 -0700869/* Take a notice that peer is sending D-SACKs */
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300870static void tcp_dsack_seen(struct tcp_sock *tp)
871{
872 tp->rx_opt.sack_ok |= 4;
873}
874
Linus Torvalds1da177e2005-04-16 15:20:36 -0700875/* Initialize metrics on socket. */
876
877static void tcp_init_metrics(struct sock *sk)
878{
879 struct tcp_sock *tp = tcp_sk(sk);
880 struct dst_entry *dst = __sk_dst_get(sk);
881
882 if (dst == NULL)
883 goto reset;
884
885 dst_confirm(dst);
886
887 if (dst_metric_locked(dst, RTAX_CWND))
888 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
889 if (dst_metric(dst, RTAX_SSTHRESH)) {
890 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
891 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
892 tp->snd_ssthresh = tp->snd_cwnd_clamp;
893 }
894 if (dst_metric(dst, RTAX_REORDERING) &&
895 tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300896 tcp_disable_fack(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700897 tp->reordering = dst_metric(dst, RTAX_REORDERING);
898 }
899
900 if (dst_metric(dst, RTAX_RTT) == 0)
901 goto reset;
902
Stephen Hemmingerc1e20f72008-07-18 23:02:15 -0700903 if (!tp->srtt && dst_metric_rtt(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700904 goto reset;
905
906 /* Initial rtt is determined from SYN,SYN-ACK.
907 * The segment is small and rtt may appear much
908 * less than real one. Use per-dst memory
909 * to make it more realistic.
910 *
911 * A bit of theory. RTT is time passed after "normal" sized packet
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800912 * is sent until it is ACKed. In normal circumstances sending small
Linus Torvalds1da177e2005-04-16 15:20:36 -0700913 * packets force peer to delay ACKs and calculation is correct too.
914 * The algorithm is adaptive and, provided we follow specs, it
915 * NEVER underestimate RTT. BUT! If peer tries to make some clever
916 * tricks sort of "quick acks" for time long enough to decrease RTT
917 * to low value, and then abruptly stops to do it and starts to delay
918 * ACKs, wait for troubles.
919 */
Stephen Hemmingerc1e20f72008-07-18 23:02:15 -0700920 if (dst_metric_rtt(dst, RTAX_RTT) > tp->srtt) {
921 tp->srtt = dst_metric_rtt(dst, RTAX_RTT);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700922 tp->rtt_seq = tp->snd_nxt;
923 }
Stephen Hemmingerc1e20f72008-07-18 23:02:15 -0700924 if (dst_metric_rtt(dst, RTAX_RTTVAR) > tp->mdev) {
925 tp->mdev = dst_metric_rtt(dst, RTAX_RTTVAR);
Satoru SATOH488faa22007-12-16 14:00:19 -0800926 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700927 }
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700928 tcp_set_rto(sk);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700929 if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700930 goto reset;
931 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
932 tp->snd_cwnd_stamp = tcp_time_stamp;
933 return;
934
935reset:
936 /* Play conservative. If timestamps are not
937 * supported, TCP will fail to recalculate correct
938 * rtt, if initial rto is too small. FORGET ALL AND RESET!
939 */
940 if (!tp->rx_opt.saw_tstamp && tp->srtt) {
941 tp->srtt = 0;
942 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700943 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700944 }
945}
946
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300947static void tcp_update_reordering(struct sock *sk, const int metric,
948 const int ts)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700949{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300950 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700951 if (metric > tp->reordering) {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -0700952 int mib_idx;
953
Linus Torvalds1da177e2005-04-16 15:20:36 -0700954 tp->reordering = min(TCP_MAX_REORDERING, metric);
955
956 /* This exciting event is worth to be remembered. 8) */
957 if (ts)
Pavel Emelyanov40b215e2008-07-03 01:05:41 -0700958 mib_idx = LINUX_MIB_TCPTSREORDER;
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300959 else if (tcp_is_reno(tp))
Pavel Emelyanov40b215e2008-07-03 01:05:41 -0700960 mib_idx = LINUX_MIB_TCPRENOREORDER;
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300961 else if (tcp_is_fack(tp))
Pavel Emelyanov40b215e2008-07-03 01:05:41 -0700962 mib_idx = LINUX_MIB_TCPFACKREORDER;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700963 else
Pavel Emelyanov40b215e2008-07-03 01:05:41 -0700964 mib_idx = LINUX_MIB_TCPSACKREORDER;
965
Pavel Emelyanovde0744a2008-07-16 20:31:16 -0700966 NET_INC_STATS_BH(sock_net(sk), mib_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700967#if FASTRETRANS_DEBUG > 1
968 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300969 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700970 tp->reordering,
971 tp->fackets_out,
972 tp->sacked_out,
973 tp->undo_marker ? tp->undo_retrans : 0);
974#endif
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300975 tcp_disable_fack(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700976 }
977}
978
Ilpo Järvinen006f5822008-09-20 21:20:20 -0700979/* This must be called before lost_out is incremented */
Ilpo Järvinenc8c213f2008-09-20 21:18:55 -0700980static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
981{
Ilpo Järvinen006f5822008-09-20 21:20:20 -0700982 if ((tp->retransmit_skb_hint == NULL) ||
Ilpo Järvinenc8c213f2008-09-20 21:18:55 -0700983 before(TCP_SKB_CB(skb)->seq,
984 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
Ilpo Järvinen006f5822008-09-20 21:20:20 -0700985 tp->retransmit_skb_hint = skb;
986
987 if (!tp->lost_out ||
988 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
989 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
Ilpo Järvinenc8c213f2008-09-20 21:18:55 -0700990}
991
Ilpo Järvinen41ea36e2008-09-20 21:19:22 -0700992static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
993{
994 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
995 tcp_verify_retransmit_hint(tp, skb);
996
997 tp->lost_out += tcp_skb_pcount(skb);
998 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
999 }
1000}
1001
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08001002static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
1003 struct sk_buff *skb)
Ilpo Järvinen006f5822008-09-20 21:20:20 -07001004{
1005 tcp_verify_retransmit_hint(tp, skb);
1006
1007 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1008 tp->lost_out += tcp_skb_pcount(skb);
1009 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1010 }
1011}
1012
Linus Torvalds1da177e2005-04-16 15:20:36 -07001013/* This procedure tags the retransmission queue when SACKs arrive.
1014 *
1015 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1016 * Packets in queue with these bits set are counted in variables
1017 * sacked_out, retrans_out and lost_out, correspondingly.
1018 *
1019 * Valid combinations are:
1020 * Tag InFlight Description
1021 * 0 1 - orig segment is in flight.
1022 * S 0 - nothing flies, orig reached receiver.
1023 * L 0 - nothing flies, orig lost by net.
1024 * R 2 - both orig and retransmit are in flight.
1025 * L|R 1 - orig is lost, retransmit is in flight.
1026 * S|R 1 - orig reached receiver, retrans is still in flight.
1027 * (L|S|R is logically valid, it could occur when L|R is sacked,
1028 * but it is equivalent to plain S and code short-curcuits it to S.
1029 * L|S is logically invalid, it would mean -1 packet in flight 8))
1030 *
1031 * These 6 states form finite state machine, controlled by the following events:
1032 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1033 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1034 * 3. Loss detection event of one of three flavors:
1035 * A. Scoreboard estimator decided the packet is lost.
1036 * A'. Reno "three dupacks" marks head of queue lost.
1037 * A''. Its FACK modfication, head until snd.fack is lost.
1038 * B. SACK arrives sacking data transmitted after never retransmitted
1039 * hole was sent out.
1040 * C. SACK arrives sacking SND.NXT at the moment, when the
1041 * segment was retransmitted.
1042 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1043 *
1044 * It is pleasant to note, that state diagram turns out to be commutative,
1045 * so that we are allowed not to be bothered by order of our actions,
1046 * when multiple events arrive simultaneously. (see the function below).
1047 *
1048 * Reordering detection.
1049 * --------------------
1050 * Reordering metric is maximal distance, which a packet can be displaced
1051 * in packet stream. With SACKs we can estimate it:
1052 *
1053 * 1. SACK fills old hole and the corresponding segment was not
1054 * ever retransmitted -> reordering. Alas, we cannot use it
1055 * when segment was retransmitted.
1056 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1057 * for retransmitted and already SACKed segment -> reordering..
1058 * Both of these heuristics are not used in Loss state, when we cannot
1059 * account for retransmits accurately.
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001060 *
1061 * SACK block validation.
1062 * ----------------------
1063 *
1064 * SACK block range validation checks that the received SACK block fits to
1065 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1066 * Note that SND.UNA is not included to the range though being valid because
Ilpo Järvinen0e835332007-10-01 15:28:17 -07001067 * it means that the receiver is rather inconsistent with itself reporting
1068 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1069 * perfectly valid, however, in light of RFC2018 which explicitly states
1070 * that "SACK block MUST reflect the newest segment. Even if the newest
1071 * segment is going to be discarded ...", not that it looks very clever
1072 * in case of head skb. Due to potentional receiver driven attacks, we
1073 * choose to avoid immediate execution of a walk in write queue due to
1074 * reneging and defer head skb's loss recovery to standard loss recovery
1075 * procedure that will eventually trigger (nothing forbids us doing this).
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001076 *
1077 * Implements also blockage to start_seq wrap-around. Problem lies in the
1078 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1079 * there's no guarantee that it will be before snd_nxt (n). The problem
1080 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1081 * wrap (s_w):
1082 *
1083 * <- outs wnd -> <- wrapzone ->
1084 * u e n u_w e_w s n_w
1085 * | | | | | | |
1086 * |<------------+------+----- TCP seqno space --------------+---------->|
1087 * ...-- <2^31 ->| |<--------...
1088 * ...---- >2^31 ------>| |<--------...
1089 *
1090 * Current code wouldn't be vulnerable but it's better still to discard such
1091 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1092 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1093 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1094 * equal to the ideal case (infinite seqno space without wrap caused issues).
1095 *
1096 * With D-SACK the lower bound is extended to cover sequence space below
1097 * SND.UNA down to undo_marker, which is the last point of interest. Yet
Ryousei Takano564262c2007-10-25 23:03:52 -07001098 * again, D-SACK block must not to go across snd_una (for the same reason as
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001099 * for the normal SACK blocks, explained above). But there all simplicity
1100 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1101 * fully below undo_marker they do not affect behavior in anyway and can
1102 * therefore be safely ignored. In rare cases (which are more or less
1103 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1104 * fragmentation and packet reordering past skb's retransmission. To consider
1105 * them correctly, the acceptable range must be extended even more though
1106 * the exact amount is rather hard to quantify. However, tp->max_window can
1107 * be used as an exaggerated estimate.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001108 */
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001109static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack,
1110 u32 start_seq, u32 end_seq)
1111{
1112 /* Too far in future, or reversed (interpretation is ambiguous) */
1113 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1114 return 0;
1115
1116 /* Nasty start_seq wrap-around check (see comments above) */
1117 if (!before(start_seq, tp->snd_nxt))
1118 return 0;
1119
Ryousei Takano564262c2007-10-25 23:03:52 -07001120 /* In outstanding window? ...This is valid exit for D-SACKs too.
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001121 * start_seq == snd_una is non-sensical (see comments above)
1122 */
1123 if (after(start_seq, tp->snd_una))
1124 return 1;
1125
1126 if (!is_dsack || !tp->undo_marker)
1127 return 0;
1128
1129 /* ...Then it's D-SACK, and must reside below snd_una completely */
1130 if (!after(end_seq, tp->snd_una))
1131 return 0;
1132
1133 if (!before(start_seq, tp->undo_marker))
1134 return 1;
1135
1136 /* Too old */
1137 if (!after(end_seq, tp->undo_marker))
1138 return 0;
1139
1140 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1141 * start_seq < undo_marker and end_seq >= undo_marker.
1142 */
1143 return !before(start_seq, end_seq - tp->max_window);
1144}
1145
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001146/* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1147 * Event "C". Later note: FACK people cheated me again 8), we have to account
1148 * for reordering! Ugly, but should help.
Ilpo Järvinenf785a8e2007-10-11 17:35:41 -07001149 *
1150 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1151 * less than what is now known to be received by the other end (derived from
Ilpo Järvinen9f58f3b2007-11-15 19:42:54 -08001152 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1153 * retransmitted skbs to avoid some costly processing per ACKs.
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001154 */
Ilpo Järvinen407ef1d2007-12-02 00:47:57 +02001155static void tcp_mark_lost_retrans(struct sock *sk)
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001156{
Ilpo Järvinen9f58f3b2007-11-15 19:42:54 -08001157 const struct inet_connection_sock *icsk = inet_csk(sk);
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001158 struct tcp_sock *tp = tcp_sk(sk);
1159 struct sk_buff *skb;
Ilpo Järvinenf785a8e2007-10-11 17:35:41 -07001160 int cnt = 0;
Ilpo Järvinendf2e0142007-10-18 05:07:57 -07001161 u32 new_low_seq = tp->snd_nxt;
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001162 u32 received_upto = tcp_highest_sack_seq(tp);
Ilpo Järvinen9f58f3b2007-11-15 19:42:54 -08001163
1164 if (!tcp_is_fack(tp) || !tp->retrans_out ||
1165 !after(received_upto, tp->lost_retrans_low) ||
1166 icsk->icsk_ca_state != TCP_CA_Recovery)
Ilpo Järvinen407ef1d2007-12-02 00:47:57 +02001167 return;
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001168
1169 tcp_for_write_queue(skb, sk) {
1170 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1171
1172 if (skb == tcp_send_head(sk))
1173 break;
Ilpo Järvinenf785a8e2007-10-11 17:35:41 -07001174 if (cnt == tp->retrans_out)
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001175 break;
1176 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1177 continue;
1178
Ilpo Järvinenf785a8e2007-10-11 17:35:41 -07001179 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1180 continue;
1181
Ilpo Järvinend0af4162009-02-28 04:44:32 +00001182 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1183 * constraint here (see above) but figuring out that at
1184 * least tp->reordering SACK blocks reside between ack_seq
1185 * and received_upto is not easy task to do cheaply with
1186 * the available datastructures.
1187 *
1188 * Whether FACK should check here for tp->reordering segs
1189 * in-between one could argue for either way (it would be
1190 * rather simple to implement as we could count fack_count
1191 * during the walk and do tp->fackets_out - fack_count).
1192 */
1193 if (after(received_upto, ack_seq)) {
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001194 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1195 tp->retrans_out -= tcp_skb_pcount(skb);
1196
Ilpo Järvinen006f5822008-09-20 21:20:20 -07001197 tcp_skb_mark_lost_uncond_verify(tp, skb);
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07001198 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
Ilpo Järvinenf785a8e2007-10-11 17:35:41 -07001199 } else {
Ilpo Järvinendf2e0142007-10-18 05:07:57 -07001200 if (before(ack_seq, new_low_seq))
Ilpo Järvinenb08d6cb2007-10-11 17:36:13 -07001201 new_low_seq = ack_seq;
Ilpo Järvinenf785a8e2007-10-11 17:35:41 -07001202 cnt += tcp_skb_pcount(skb);
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001203 }
1204 }
Ilpo Järvinenb08d6cb2007-10-11 17:36:13 -07001205
1206 if (tp->retrans_out)
1207 tp->lost_retrans_low = new_low_seq;
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001208}
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001209
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07001210static int tcp_check_dsack(struct sock *sk, struct sk_buff *ack_skb,
David S. Millerd06e0212007-06-18 22:43:06 -07001211 struct tcp_sack_block_wire *sp, int num_sacks,
1212 u32 prior_snd_una)
1213{
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07001214 struct tcp_sock *tp = tcp_sk(sk);
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07001215 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1216 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
David S. Millerd06e0212007-06-18 22:43:06 -07001217 int dup_sack = 0;
1218
1219 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1220 dup_sack = 1;
Ilpo Järvinene60402d2007-08-09 15:14:46 +03001221 tcp_dsack_seen(tp);
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07001222 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
David S. Millerd06e0212007-06-18 22:43:06 -07001223 } else if (num_sacks > 1) {
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07001224 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1225 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
David S. Millerd06e0212007-06-18 22:43:06 -07001226
1227 if (!after(end_seq_0, end_seq_1) &&
1228 !before(start_seq_0, start_seq_1)) {
1229 dup_sack = 1;
Ilpo Järvinene60402d2007-08-09 15:14:46 +03001230 tcp_dsack_seen(tp);
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07001231 NET_INC_STATS_BH(sock_net(sk),
1232 LINUX_MIB_TCPDSACKOFORECV);
David S. Millerd06e0212007-06-18 22:43:06 -07001233 }
1234 }
1235
1236 /* D-SACK for already forgotten data... Do dumb counting. */
1237 if (dup_sack &&
1238 !after(end_seq_0, prior_snd_una) &&
1239 after(end_seq_0, tp->undo_marker))
1240 tp->undo_retrans--;
1241
1242 return dup_sack;
1243}
1244
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001245struct tcp_sacktag_state {
1246 int reord;
1247 int fack_count;
1248 int flag;
1249};
1250
Ilpo Järvinend1935942007-10-11 17:34:25 -07001251/* Check if skb is fully within the SACK block. In presence of GSO skbs,
1252 * the incoming SACK may not exactly match but we can find smaller MSS
1253 * aligned portion of it that matches. Therefore we might need to fragment
1254 * which may fail and creates some hassle (caller must handle error case
1255 * returns).
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001256 *
1257 * FIXME: this could be merged to shift decision code
Ilpo Järvinend1935942007-10-11 17:34:25 -07001258 */
Adrian Bunk0f79efd2007-10-26 03:57:36 -07001259static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1260 u32 start_seq, u32 end_seq)
Ilpo Järvinend1935942007-10-11 17:34:25 -07001261{
1262 int in_sack, err;
1263 unsigned int pkt_len;
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001264 unsigned int mss;
Ilpo Järvinend1935942007-10-11 17:34:25 -07001265
1266 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1267 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1268
1269 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1270 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001271 mss = tcp_skb_mss(skb);
Ilpo Järvinend1935942007-10-11 17:34:25 -07001272 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1273
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001274 if (!in_sack) {
Ilpo Järvinend1935942007-10-11 17:34:25 -07001275 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001276 if (pkt_len < mss)
1277 pkt_len = mss;
1278 } else {
Ilpo Järvinend1935942007-10-11 17:34:25 -07001279 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001280 if (pkt_len < mss)
1281 return -EINVAL;
1282 }
1283
1284 /* Round if necessary so that SACKs cover only full MSSes
1285 * and/or the remaining small portion (if present)
1286 */
1287 if (pkt_len > mss) {
1288 unsigned int new_len = (pkt_len / mss) * mss;
1289 if (!in_sack && new_len < pkt_len) {
1290 new_len += mss;
1291 if (new_len > skb->len)
1292 return 0;
1293 }
1294 pkt_len = new_len;
1295 }
1296 err = tcp_fragment(sk, skb, pkt_len, mss);
Ilpo Järvinend1935942007-10-11 17:34:25 -07001297 if (err < 0)
1298 return err;
1299 }
1300
1301 return in_sack;
1302}
1303
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001304static u8 tcp_sacktag_one(struct sk_buff *skb, struct sock *sk,
1305 struct tcp_sacktag_state *state,
1306 int dup_sack, int pcount)
Ilpo Järvinen9e10c472007-11-15 19:44:56 -08001307{
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001308 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen9e10c472007-11-15 19:44:56 -08001309 u8 sacked = TCP_SKB_CB(skb)->sacked;
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001310 int fack_count = state->fack_count;
Ilpo Järvinen9e10c472007-11-15 19:44:56 -08001311
1312 /* Account D-SACK for retransmitted packet. */
1313 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1314 if (after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1315 tp->undo_retrans--;
Ilpo Järvinenede9f3b2007-12-02 00:47:58 +02001316 if (sacked & TCPCB_SACKED_ACKED)
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001317 state->reord = min(fack_count, state->reord);
Ilpo Järvinen9e10c472007-11-15 19:44:56 -08001318 }
1319
1320 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1321 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001322 return sacked;
Ilpo Järvinen9e10c472007-11-15 19:44:56 -08001323
1324 if (!(sacked & TCPCB_SACKED_ACKED)) {
1325 if (sacked & TCPCB_SACKED_RETRANS) {
1326 /* If the segment is not tagged as lost,
1327 * we do not clear RETRANS, believing
1328 * that retransmission is still in flight.
1329 */
1330 if (sacked & TCPCB_LOST) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001331 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001332 tp->lost_out -= pcount;
1333 tp->retrans_out -= pcount;
Ilpo Järvinen9e10c472007-11-15 19:44:56 -08001334 }
1335 } else {
1336 if (!(sacked & TCPCB_RETRANS)) {
1337 /* New sack for not retransmitted frame,
1338 * which was in hole. It is reordering.
1339 */
1340 if (before(TCP_SKB_CB(skb)->seq,
1341 tcp_highest_sack_seq(tp)))
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001342 state->reord = min(fack_count,
1343 state->reord);
Ilpo Järvinen9e10c472007-11-15 19:44:56 -08001344
1345 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1346 if (!after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark))
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001347 state->flag |= FLAG_ONLY_ORIG_SACKED;
Ilpo Järvinen9e10c472007-11-15 19:44:56 -08001348 }
1349
1350 if (sacked & TCPCB_LOST) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001351 sacked &= ~TCPCB_LOST;
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001352 tp->lost_out -= pcount;
Ilpo Järvinen9e10c472007-11-15 19:44:56 -08001353 }
1354 }
1355
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001356 sacked |= TCPCB_SACKED_ACKED;
1357 state->flag |= FLAG_DATA_SACKED;
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001358 tp->sacked_out += pcount;
Ilpo Järvinen9e10c472007-11-15 19:44:56 -08001359
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001360 fack_count += pcount;
Ilpo Järvinen9e10c472007-11-15 19:44:56 -08001361
1362 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1363 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1364 before(TCP_SKB_CB(skb)->seq,
1365 TCP_SKB_CB(tp->lost_skb_hint)->seq))
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001366 tp->lost_cnt_hint += pcount;
Ilpo Järvinen9e10c472007-11-15 19:44:56 -08001367
1368 if (fack_count > tp->fackets_out)
1369 tp->fackets_out = fack_count;
Ilpo Järvinen9e10c472007-11-15 19:44:56 -08001370 }
1371
1372 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1373 * frames and clear it. undo_retrans is decreased above, L|R frames
1374 * are accounted above as well.
1375 */
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001376 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1377 sacked &= ~TCPCB_SACKED_RETRANS;
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001378 tp->retrans_out -= pcount;
Ilpo Järvinen9e10c472007-11-15 19:44:56 -08001379 }
1380
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001381 return sacked;
Ilpo Järvinen9e10c472007-11-15 19:44:56 -08001382}
1383
Ilpo Järvinen50133162008-12-05 22:42:41 -08001384static int tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001385 struct tcp_sacktag_state *state,
Ilpo Järvinen9ec06ff2009-03-01 00:21:36 -08001386 unsigned int pcount, int shifted, int mss,
1387 int dup_sack)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001388{
1389 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen50133162008-12-05 22:42:41 -08001390 struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001391
1392 BUG_ON(!pcount);
1393
Ilpo Järvinen92ee76b2008-11-24 21:26:56 -08001394 /* Tweak before seqno plays */
1395 if (!tcp_is_fack(tp) && tcp_is_sack(tp) && tp->lost_skb_hint &&
1396 !before(TCP_SKB_CB(tp->lost_skb_hint)->seq, TCP_SKB_CB(skb)->seq))
1397 tp->lost_cnt_hint += pcount;
1398
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001399 TCP_SKB_CB(prev)->end_seq += shifted;
1400 TCP_SKB_CB(skb)->seq += shifted;
1401
1402 skb_shinfo(prev)->gso_segs += pcount;
1403 BUG_ON(skb_shinfo(skb)->gso_segs < pcount);
1404 skb_shinfo(skb)->gso_segs -= pcount;
1405
1406 /* When we're adding to gso_segs == 1, gso_size will be zero,
1407 * in theory this shouldn't be necessary but as long as DSACK
1408 * code can come after this skb later on it's better to keep
1409 * setting gso_size to something.
1410 */
1411 if (!skb_shinfo(prev)->gso_size) {
1412 skb_shinfo(prev)->gso_size = mss;
1413 skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1414 }
1415
1416 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1417 if (skb_shinfo(skb)->gso_segs <= 1) {
1418 skb_shinfo(skb)->gso_size = 0;
1419 skb_shinfo(skb)->gso_type = 0;
1420 }
1421
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001422 /* We discard results */
Ilpo Järvinen9ec06ff2009-03-01 00:21:36 -08001423 tcp_sacktag_one(skb, sk, state, dup_sack, pcount);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001424
1425 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1426 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1427
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001428 if (skb->len > 0) {
1429 BUG_ON(!tcp_skb_pcount(skb));
Ilpo Järvinen111cc8b2008-11-24 21:27:22 -08001430 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001431 return 0;
1432 }
1433
1434 /* Whole SKB was eaten :-) */
1435
Ilpo Järvinen92ee76b2008-11-24 21:26:56 -08001436 if (skb == tp->retransmit_skb_hint)
1437 tp->retransmit_skb_hint = prev;
1438 if (skb == tp->scoreboard_skb_hint)
1439 tp->scoreboard_skb_hint = prev;
1440 if (skb == tp->lost_skb_hint) {
1441 tp->lost_skb_hint = prev;
1442 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1443 }
1444
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001445 TCP_SKB_CB(skb)->flags |= TCP_SKB_CB(prev)->flags;
1446 if (skb == tcp_highest_sack(sk))
1447 tcp_advance_highest_sack(sk, skb);
1448
1449 tcp_unlink_write_queue(skb, sk);
1450 sk_wmem_free_skb(sk, skb);
1451
Ilpo Järvinen111cc8b2008-11-24 21:27:22 -08001452 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1453
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001454 return 1;
1455}
1456
1457/* I wish gso_size would have a bit more sane initialization than
1458 * something-or-zero which complicates things
1459 */
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001460static int tcp_skb_seglen(struct sk_buff *skb)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001461{
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001462 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001463}
1464
1465/* Shifting pages past head area doesn't work */
1466static int skb_can_shift(struct sk_buff *skb)
1467{
1468 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1469}
1470
1471/* Try collapsing SACK blocks spanning across multiple skbs to a single
1472 * skb.
1473 */
1474static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001475 struct tcp_sacktag_state *state,
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001476 u32 start_seq, u32 end_seq,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001477 int dup_sack)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001478{
1479 struct tcp_sock *tp = tcp_sk(sk);
1480 struct sk_buff *prev;
1481 int mss;
1482 int pcount = 0;
1483 int len;
1484 int in_sack;
1485
1486 if (!sk_can_gso(sk))
1487 goto fallback;
1488
1489 /* Normally R but no L won't result in plain S */
1490 if (!dup_sack &&
Ilpo Järvinen9969ca52008-12-05 22:41:06 -08001491 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001492 goto fallback;
1493 if (!skb_can_shift(skb))
1494 goto fallback;
1495 /* This frame is about to be dropped (was ACKed). */
1496 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1497 goto fallback;
1498
1499 /* Can only happen with delayed DSACK + discard craziness */
1500 if (unlikely(skb == tcp_write_queue_head(sk)))
1501 goto fallback;
1502 prev = tcp_write_queue_prev(sk, skb);
1503
1504 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1505 goto fallback;
1506
1507 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1508 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1509
1510 if (in_sack) {
1511 len = skb->len;
1512 pcount = tcp_skb_pcount(skb);
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001513 mss = tcp_skb_seglen(skb);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001514
1515 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1516 * drop this restriction as unnecessary
1517 */
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001518 if (mss != tcp_skb_seglen(prev))
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001519 goto fallback;
1520 } else {
1521 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1522 goto noop;
1523 /* CHECKME: This is non-MSS split case only?, this will
1524 * cause skipped skbs due to advancing loop btw, original
1525 * has that feature too
1526 */
1527 if (tcp_skb_pcount(skb) <= 1)
1528 goto noop;
1529
1530 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1531 if (!in_sack) {
1532 /* TODO: head merge to next could be attempted here
1533 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1534 * though it might not be worth of the additional hassle
1535 *
1536 * ...we can probably just fallback to what was done
1537 * previously. We could try merging non-SACKed ones
1538 * as well but it probably isn't going to buy off
1539 * because later SACKs might again split them, and
1540 * it would make skb timestamp tracking considerably
1541 * harder problem.
1542 */
1543 goto fallback;
1544 }
1545
1546 len = end_seq - TCP_SKB_CB(skb)->seq;
1547 BUG_ON(len < 0);
1548 BUG_ON(len > skb->len);
1549
1550 /* MSS boundaries should be honoured or else pcount will
1551 * severely break even though it makes things bit trickier.
1552 * Optimize common case to avoid most of the divides
1553 */
1554 mss = tcp_skb_mss(skb);
1555
1556 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1557 * drop this restriction as unnecessary
1558 */
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001559 if (mss != tcp_skb_seglen(prev))
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001560 goto fallback;
1561
1562 if (len == mss) {
1563 pcount = 1;
1564 } else if (len < mss) {
1565 goto noop;
1566 } else {
1567 pcount = len / mss;
1568 len = pcount * mss;
1569 }
1570 }
1571
1572 if (!skb_shift(prev, skb, len))
1573 goto fallback;
Ilpo Järvinen9ec06ff2009-03-01 00:21:36 -08001574 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001575 goto out;
1576
1577 /* Hole filled allows collapsing with the next as well, this is very
1578 * useful when hole on every nth skb pattern happens
1579 */
1580 if (prev == tcp_write_queue_tail(sk))
1581 goto out;
1582 skb = tcp_write_queue_next(sk, prev);
1583
Ilpo Järvinenf0bc52f2008-12-05 22:40:47 -08001584 if (!skb_can_shift(skb) ||
1585 (skb == tcp_send_head(sk)) ||
1586 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001587 (mss != tcp_skb_seglen(skb)))
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001588 goto out;
1589
1590 len = skb->len;
1591 if (skb_shift(prev, skb, len)) {
1592 pcount += tcp_skb_pcount(skb);
Ilpo Järvinen9ec06ff2009-03-01 00:21:36 -08001593 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001594 }
1595
1596out:
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001597 state->fack_count += pcount;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001598 return prev;
1599
1600noop:
1601 return skb;
1602
1603fallback:
Ilpo Järvinen111cc8b2008-11-24 21:27:22 -08001604 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001605 return NULL;
1606}
1607
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001608static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1609 struct tcp_sack_block *next_dup,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001610 struct tcp_sacktag_state *state,
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001611 u32 start_seq, u32 end_seq,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001612 int dup_sack_in)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001613{
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001614 struct tcp_sock *tp = tcp_sk(sk);
1615 struct sk_buff *tmp;
1616
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001617 tcp_for_write_queue_from(skb, sk) {
1618 int in_sack = 0;
1619 int dup_sack = dup_sack_in;
1620
1621 if (skb == tcp_send_head(sk))
1622 break;
1623
1624 /* queue is in-order => we can short-circuit the walk early */
1625 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1626 break;
1627
1628 if ((next_dup != NULL) &&
1629 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1630 in_sack = tcp_match_skb_to_sack(sk, skb,
1631 next_dup->start_seq,
1632 next_dup->end_seq);
1633 if (in_sack > 0)
1634 dup_sack = 1;
1635 }
1636
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001637 /* skb reference here is a bit tricky to get right, since
1638 * shifting can eat and free both this skb and the next,
1639 * so not even _safe variant of the loop is enough.
1640 */
1641 if (in_sack <= 0) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001642 tmp = tcp_shift_skb_data(sk, skb, state,
1643 start_seq, end_seq, dup_sack);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001644 if (tmp != NULL) {
1645 if (tmp != skb) {
1646 skb = tmp;
1647 continue;
1648 }
1649
1650 in_sack = 0;
1651 } else {
1652 in_sack = tcp_match_skb_to_sack(sk, skb,
1653 start_seq,
1654 end_seq);
1655 }
1656 }
1657
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001658 if (unlikely(in_sack < 0))
1659 break;
1660
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001661 if (in_sack) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001662 TCP_SKB_CB(skb)->sacked = tcp_sacktag_one(skb, sk,
1663 state,
1664 dup_sack,
1665 tcp_skb_pcount(skb));
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001666
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001667 if (!before(TCP_SKB_CB(skb)->seq,
1668 tcp_highest_sack_seq(tp)))
1669 tcp_advance_highest_sack(sk, skb);
1670 }
1671
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001672 state->fack_count += tcp_skb_pcount(skb);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001673 }
1674 return skb;
1675}
1676
1677/* Avoid all extra work that is being done by sacktag while walking in
1678 * a normal way
1679 */
1680static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001681 struct tcp_sacktag_state *state,
1682 u32 skip_to_seq)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001683{
1684 tcp_for_write_queue_from(skb, sk) {
1685 if (skb == tcp_send_head(sk))
1686 break;
1687
Ilpo Järvinene8bae272008-11-24 21:12:28 -08001688 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001689 break;
Ilpo Järvinend152a7d2008-03-03 12:10:16 -08001690
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001691 state->fack_count += tcp_skb_pcount(skb);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001692 }
1693 return skb;
1694}
1695
1696static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1697 struct sock *sk,
1698 struct tcp_sack_block *next_dup,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001699 struct tcp_sacktag_state *state,
1700 u32 skip_to_seq)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001701{
1702 if (next_dup == NULL)
1703 return skb;
1704
1705 if (before(next_dup->start_seq, skip_to_seq)) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001706 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1707 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1708 next_dup->start_seq, next_dup->end_seq,
1709 1);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001710 }
1711
1712 return skb;
1713}
1714
1715static int tcp_sack_cache_ok(struct tcp_sock *tp, struct tcp_sack_block *cache)
1716{
1717 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1718}
1719
Linus Torvalds1da177e2005-04-16 15:20:36 -07001720static int
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001721tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb,
1722 u32 prior_snd_una)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001723{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001724 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001725 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo9c702202007-04-25 18:04:18 -07001726 unsigned char *ptr = (skb_transport_header(ack_skb) +
1727 TCP_SKB_CB(ack_skb)->sacked);
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001728 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
Adam Langley4389dde2008-07-19 00:07:02 -07001729 struct tcp_sack_block sp[TCP_NUM_SACKS];
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001730 struct tcp_sack_block *cache;
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001731 struct tcp_sacktag_state state;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001732 struct sk_buff *skb;
Adam Langley4389dde2008-07-19 00:07:02 -07001733 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001734 int used_sacks;
Ilpo Järvinen7769f402007-06-15 15:14:04 -07001735 int found_dup_sack = 0;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001736 int i, j;
Baruch Evenfda03fb2007-02-04 23:35:57 -08001737 int first_sack_index;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001738
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001739 state.flag = 0;
1740 state.reord = tp->packets_out;
1741
Ilpo Järvinend738cd82007-03-24 21:03:23 -07001742 if (!tp->sacked_out) {
Ilpo Järvinende83c052007-10-07 23:37:55 -07001743 if (WARN_ON(tp->fackets_out))
1744 tp->fackets_out = 0;
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001745 tcp_highest_sack_reset(sk);
Ilpo Järvinend738cd82007-03-24 21:03:23 -07001746 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001747
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07001748 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
David S. Millerd06e0212007-06-18 22:43:06 -07001749 num_sacks, prior_snd_una);
1750 if (found_dup_sack)
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001751 state.flag |= FLAG_DSACKING_ACK;
Baruch Even6f746512007-02-04 23:36:42 -08001752
1753 /* Eliminate too old ACKs, but take into
1754 * account more or less fresh ones, they can
1755 * contain valid SACK info.
1756 */
1757 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1758 return 0;
1759
Ilpo Järvinen96a2d412007-11-14 15:47:18 -08001760 if (!tp->packets_out)
1761 goto out;
1762
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001763 used_sacks = 0;
1764 first_sack_index = 0;
1765 for (i = 0; i < num_sacks; i++) {
1766 int dup_sack = !i && found_dup_sack;
1767
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07001768 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1769 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001770
1771 if (!tcp_is_sackblock_valid(tp, dup_sack,
1772 sp[used_sacks].start_seq,
1773 sp[used_sacks].end_seq)) {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001774 int mib_idx;
1775
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001776 if (dup_sack) {
1777 if (!tp->undo_marker)
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001778 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001779 else
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001780 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001781 } else {
1782 /* Don't count olds caused by ACK reordering */
1783 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1784 !after(sp[used_sacks].end_seq, tp->snd_una))
1785 continue;
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001786 mib_idx = LINUX_MIB_TCPSACKDISCARD;
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001787 }
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001788
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07001789 NET_INC_STATS_BH(sock_net(sk), mib_idx);
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001790 if (i == 0)
1791 first_sack_index = -1;
1792 continue;
1793 }
1794
1795 /* Ignore very old stuff early */
1796 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1797 continue;
1798
1799 used_sacks++;
1800 }
1801
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001802 /* order SACK blocks to allow in order walk of the retrans queue */
1803 for (i = used_sacks - 1; i > 0; i--) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001804 for (j = 0; j < i; j++) {
1805 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
Ilpo Järvinena0bffff2009-03-21 13:36:17 -07001806 swap(sp[j], sp[j + 1]);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001807
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001808 /* Track where the first SACK block goes to */
1809 if (j == first_sack_index)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001810 first_sack_index = j + 1;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001811 }
1812 }
1813 }
1814
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001815 skb = tcp_write_queue_head(sk);
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001816 state.fack_count = 0;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001817 i = 0;
1818
1819 if (!tp->sacked_out) {
1820 /* It's already past, so skip checking against it */
1821 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1822 } else {
1823 cache = tp->recv_sack_cache;
1824 /* Skip empty blocks in at head of the cache */
1825 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1826 !cache->end_seq)
1827 cache++;
Baruch Evenfda03fb2007-02-04 23:35:57 -08001828 }
1829
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001830 while (i < used_sacks) {
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001831 u32 start_seq = sp[i].start_seq;
1832 u32 end_seq = sp[i].end_seq;
Ilpo Järvinen7769f402007-06-15 15:14:04 -07001833 int dup_sack = (found_dup_sack && (i == first_sack_index));
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001834 struct tcp_sack_block *next_dup = NULL;
Ilpo Järvinene56d6cd2007-11-01 00:09:37 -07001835
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001836 if (found_dup_sack && ((i + 1) == first_sack_index))
1837 next_dup = &sp[i + 1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001838
1839 /* Event "B" in the comment above. */
1840 if (after(end_seq, tp->high_seq))
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001841 state.flag |= FLAG_DATA_LOST;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001842
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001843 /* Skip too early cached blocks */
1844 while (tcp_sack_cache_ok(tp, cache) &&
1845 !before(start_seq, cache->end_seq))
1846 cache++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001847
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001848 /* Can skip some work by looking recv_sack_cache? */
1849 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1850 after(end_seq, cache->start_seq)) {
David S. Millerfe067e82007-03-07 12:12:44 -08001851
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001852 /* Head todo? */
1853 if (before(start_seq, cache->start_seq)) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001854 skb = tcp_sacktag_skip(skb, sk, &state,
1855 start_seq);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001856 skb = tcp_sacktag_walk(skb, sk, next_dup,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001857 &state,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001858 start_seq,
1859 cache->start_seq,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001860 dup_sack);
Baruch Evenfda03fb2007-02-04 23:35:57 -08001861 }
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001862
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001863 /* Rest of the block already fully processed? */
Ilpo Järvinen20de20b2007-11-16 16:17:05 -08001864 if (!after(end_seq, cache->end_seq))
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001865 goto advance_sp;
Ilpo Järvinen20de20b2007-11-16 16:17:05 -08001866
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001867 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001868 &state,
1869 cache->end_seq);
Ilpo Järvinene56d6cd2007-11-01 00:09:37 -07001870
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001871 /* ...tail remains todo... */
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001872 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
Ilpo Järvinen20de20b2007-11-16 16:17:05 -08001873 /* ...but better entrypoint exists! */
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001874 skb = tcp_highest_sack(sk);
1875 if (skb == NULL)
1876 break;
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001877 state.fack_count = tp->fackets_out;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001878 cache++;
1879 goto walk;
1880 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001881
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001882 skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001883 /* Check overlap against next cached too (past this one already) */
1884 cache++;
1885 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001886 }
Ilpo Järvinenfbd52eb2007-11-10 21:24:19 -08001887
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001888 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1889 skb = tcp_highest_sack(sk);
1890 if (skb == NULL)
1891 break;
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001892 state.fack_count = tp->fackets_out;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001893 }
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001894 skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001895
1896walk:
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001897 skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1898 start_seq, end_seq, dup_sack);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001899
1900advance_sp:
Ilpo Järvinenfbd52eb2007-11-10 21:24:19 -08001901 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1902 * due to in-order walk
1903 */
1904 if (after(end_seq, tp->frto_highmark))
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001905 state.flag &= ~FLAG_ONLY_ORIG_SACKED;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001906
1907 i++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001908 }
1909
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001910 /* Clear the head of the cache sack blocks so we can skip it next time */
1911 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1912 tp->recv_sack_cache[i].start_seq = 0;
1913 tp->recv_sack_cache[i].end_seq = 0;
1914 }
1915 for (j = 0; j < used_sacks; j++)
1916 tp->recv_sack_cache[i++] = sp[j];
1917
Ilpo Järvinen407ef1d2007-12-02 00:47:57 +02001918 tcp_mark_lost_retrans(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001919
Ilpo Järvinen86426c22007-08-09 14:45:17 +03001920 tcp_verify_left_out(tp);
1921
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001922 if ((state.reord < tp->fackets_out) &&
Ilpo Järvinenf5771112007-11-15 19:35:11 -08001923 ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) &&
Ilpo Järvinenc5e7af02007-02-23 16:22:06 -08001924 (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001925 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001926
Ilpo Järvinen96a2d412007-11-14 15:47:18 -08001927out:
1928
Linus Torvalds1da177e2005-04-16 15:20:36 -07001929#if FASTRETRANS_DEBUG > 0
Ilpo Järvinen547b7922008-07-25 21:43:18 -07001930 WARN_ON((int)tp->sacked_out < 0);
1931 WARN_ON((int)tp->lost_out < 0);
1932 WARN_ON((int)tp->retrans_out < 0);
1933 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001934#endif
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001935 return state.flag;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001936}
1937
Ilpo Järvinen882beba2008-04-07 22:33:07 -07001938/* Limits sacked_out so that sum with lost_out isn't ever larger than
1939 * packets_out. Returns zero if sacked_out adjustement wasn't necessary.
Ilpo Järvinen30935cf2007-02-21 23:01:36 -08001940 */
Ilpo Järvinen8eecaba2008-11-25 13:45:29 -08001941static int tcp_limit_reno_sacked(struct tcp_sock *tp)
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001942{
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001943 u32 holes;
1944
1945 holes = max(tp->lost_out, 1U);
1946 holes = min(holes, tp->packets_out);
1947
1948 if ((tp->sacked_out + holes) > tp->packets_out) {
1949 tp->sacked_out = tp->packets_out - holes;
Ilpo Järvinen882beba2008-04-07 22:33:07 -07001950 return 1;
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001951 }
Ilpo Järvinen882beba2008-04-07 22:33:07 -07001952 return 0;
1953}
1954
1955/* If we receive more dupacks than we expected counting segments
1956 * in assumption of absent reordering, interpret this as reordering.
1957 * The only another reason could be bug in receiver TCP.
1958 */
1959static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1960{
1961 struct tcp_sock *tp = tcp_sk(sk);
1962 if (tcp_limit_reno_sacked(tp))
1963 tcp_update_reordering(sk, tp->packets_out + addend, 0);
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001964}
1965
1966/* Emulate SACKs for SACKless connection: account for a new dupack. */
1967
1968static void tcp_add_reno_sack(struct sock *sk)
1969{
1970 struct tcp_sock *tp = tcp_sk(sk);
1971 tp->sacked_out++;
1972 tcp_check_reno_reordering(sk, 0);
Ilpo Järvinen005903b2007-08-09 14:44:16 +03001973 tcp_verify_left_out(tp);
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001974}
1975
1976/* Account for ACK, ACKing some data in Reno Recovery phase. */
1977
1978static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1979{
1980 struct tcp_sock *tp = tcp_sk(sk);
1981
1982 if (acked > 0) {
1983 /* One ACK acked hole. The rest eat duplicate ACKs. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001984 if (acked - 1 >= tp->sacked_out)
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001985 tp->sacked_out = 0;
1986 else
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001987 tp->sacked_out -= acked - 1;
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001988 }
1989 tcp_check_reno_reordering(sk, acked);
Ilpo Järvinen005903b2007-08-09 14:44:16 +03001990 tcp_verify_left_out(tp);
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001991}
1992
1993static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1994{
1995 tp->sacked_out = 0;
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001996}
1997
Ilpo Järvinen62ab2222008-05-08 01:09:11 -07001998static int tcp_is_sackfrto(const struct tcp_sock *tp)
1999{
2000 return (sysctl_tcp_frto == 0x2) && !tcp_is_reno(tp);
2001}
2002
Ilpo Järvinen95eacd22007-10-01 15:27:42 -07002003/* F-RTO can only be used if TCP has never retransmitted anything other than
2004 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
2005 */
Ilpo Järvinen46d0de42007-02-21 23:10:39 -08002006int tcp_use_frto(struct sock *sk)
Ilpo Järvinenbdaae172007-02-21 22:59:58 -08002007{
2008 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen6adb4f72008-04-07 22:33:57 -07002009 const struct inet_connection_sock *icsk = inet_csk(sk);
Ilpo Järvinen46d0de42007-02-21 23:10:39 -08002010 struct sk_buff *skb;
Ilpo Järvinenbdaae172007-02-21 22:59:58 -08002011
Ilpo Järvinen575ee712007-04-30 00:39:55 -07002012 if (!sysctl_tcp_frto)
Ilpo Järvinen46d0de42007-02-21 23:10:39 -08002013 return 0;
2014
Ilpo Järvinen6adb4f72008-04-07 22:33:57 -07002015 /* MTU probe and F-RTO won't really play nicely along currently */
2016 if (icsk->icsk_mtup.probe_size)
2017 return 0;
2018
Ilpo Järvinen62ab2222008-05-08 01:09:11 -07002019 if (tcp_is_sackfrto(tp))
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08002020 return 1;
2021
Ilpo Järvinen46d0de42007-02-21 23:10:39 -08002022 /* Avoid expensive walking of rexmit queue if possible */
2023 if (tp->retrans_out > 1)
2024 return 0;
2025
David S. Millerfe067e82007-03-07 12:12:44 -08002026 skb = tcp_write_queue_head(sk);
David S. Miller28e34872008-09-23 02:51:41 -07002027 if (tcp_skb_is_last(sk, skb))
2028 return 1;
David S. Millerfe067e82007-03-07 12:12:44 -08002029 skb = tcp_write_queue_next(sk, skb); /* Skips head */
2030 tcp_for_write_queue_from(skb, sk) {
2031 if (skb == tcp_send_head(sk))
2032 break;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002033 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
Ilpo Järvinen46d0de42007-02-21 23:10:39 -08002034 return 0;
2035 /* Short-circuit when first non-SACKed skb has been checked */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002036 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
Ilpo Järvinen46d0de42007-02-21 23:10:39 -08002037 break;
2038 }
2039 return 1;
Ilpo Järvinenbdaae172007-02-21 22:59:58 -08002040}
2041
Ilpo Järvinen30935cf2007-02-21 23:01:36 -08002042/* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
2043 * recovery a bit and use heuristics in tcp_process_frto() to detect if
Ilpo Järvinend1a54c62007-02-21 23:11:57 -08002044 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
2045 * keep retrans_out counting accurate (with SACK F-RTO, other than head
2046 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
2047 * bits are handled if the Loss state is really to be entered (in
2048 * tcp_enter_frto_loss).
Ilpo Järvinen7487c482007-02-21 23:02:30 -08002049 *
2050 * Do like tcp_enter_loss() would; when RTO expires the second time it
2051 * does:
2052 * "Reduce ssthresh if it has not yet been made inside this window."
Linus Torvalds1da177e2005-04-16 15:20:36 -07002053 */
2054void tcp_enter_frto(struct sock *sk)
2055{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002056 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002057 struct tcp_sock *tp = tcp_sk(sk);
2058 struct sk_buff *skb;
2059
Ilpo Järvinen7487c482007-02-21 23:02:30 -08002060 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002061 tp->snd_una == tp->high_seq ||
Ilpo Järvinen7487c482007-02-21 23:02:30 -08002062 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
2063 !icsk->icsk_retransmits)) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002064 tp->prior_ssthresh = tcp_current_ssthresh(sk);
Ilpo Järvinen66e93e42007-02-21 23:13:47 -08002065 /* Our state is too optimistic in ssthresh() call because cwnd
Ryousei Takano564262c2007-10-25 23:03:52 -07002066 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
Ilpo Järvinen66e93e42007-02-21 23:13:47 -08002067 * recovery has not yet completed. Pattern would be this: RTO,
2068 * Cumulative ACK, RTO (2xRTO for the same segment does not end
2069 * up here twice).
2070 * RFC4138 should be more specific on what to do, even though
2071 * RTO is quite unlikely to occur after the first Cumulative ACK
2072 * due to back-off and complexity of triggering events ...
2073 */
2074 if (tp->frto_counter) {
2075 u32 stored_cwnd;
2076 stored_cwnd = tp->snd_cwnd;
2077 tp->snd_cwnd = 2;
2078 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2079 tp->snd_cwnd = stored_cwnd;
2080 } else {
2081 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2082 }
2083 /* ... in theory, cong.control module could do "any tricks" in
2084 * ssthresh(), which means that ca_state, lost bits and lost_out
2085 * counter would have to be faked before the call occurs. We
2086 * consider that too expensive, unlikely and hacky, so modules
2087 * using these in ssthresh() must deal these incompatibility
2088 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
2089 */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002090 tcp_ca_event(sk, CA_EVENT_FRTO);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002091 }
2092
Linus Torvalds1da177e2005-04-16 15:20:36 -07002093 tp->undo_marker = tp->snd_una;
2094 tp->undo_retrans = 0;
2095
David S. Millerfe067e82007-03-07 12:12:44 -08002096 skb = tcp_write_queue_head(sk);
Ilpo Järvinen009a2e32007-09-20 11:34:38 -07002097 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2098 tp->undo_marker = 0;
Ilpo Järvinend1a54c62007-02-21 23:11:57 -08002099 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
Ilpo Järvinen522e7542007-02-21 22:54:52 -08002100 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
Ilpo Järvinend1a54c62007-02-21 23:11:57 -08002101 tp->retrans_out -= tcp_skb_pcount(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002102 }
Ilpo Järvinen005903b2007-08-09 14:44:16 +03002103 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002104
Ilpo Järvinen746aa322007-11-13 21:01:23 -08002105 /* Too bad if TCP was application limited */
2106 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2107
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08002108 /* Earlier loss recovery underway (see RFC4138; Appendix B).
2109 * The last condition is necessary at least in tp->frto_counter case.
2110 */
Ilpo Järvinen62ab2222008-05-08 01:09:11 -07002111 if (tcp_is_sackfrto(tp) && (tp->frto_counter ||
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08002112 ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
2113 after(tp->high_seq, tp->snd_una)) {
2114 tp->frto_highmark = tp->high_seq;
2115 } else {
2116 tp->frto_highmark = tp->snd_nxt;
2117 }
Ilpo Järvinen7b0eb222007-02-21 23:03:35 -08002118 tcp_set_ca_state(sk, TCP_CA_Disorder);
2119 tp->high_seq = tp->snd_nxt;
Ilpo Järvinen7487c482007-02-21 23:02:30 -08002120 tp->frto_counter = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002121}
2122
2123/* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
2124 * which indicates that we should follow the traditional RTO recovery,
2125 * i.e. mark everything lost and do go-back-N retransmission.
2126 */
Ilpo Järvinend1a54c62007-02-21 23:11:57 -08002127static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002128{
2129 struct tcp_sock *tp = tcp_sk(sk);
2130 struct sk_buff *skb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002131
Linus Torvalds1da177e2005-04-16 15:20:36 -07002132 tp->lost_out = 0;
Ilpo Järvinend1a54c62007-02-21 23:11:57 -08002133 tp->retrans_out = 0;
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002134 if (tcp_is_reno(tp))
Ilpo Järvinen9bff40f2007-05-27 01:53:49 -07002135 tcp_reset_reno_sack(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002136
David S. Millerfe067e82007-03-07 12:12:44 -08002137 tcp_for_write_queue(skb, sk) {
2138 if (skb == tcp_send_head(sk))
2139 break;
Ilpo Järvinen23aeeec2007-11-13 21:03:13 -08002140
2141 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
Ilpo Järvinend1a54c62007-02-21 23:11:57 -08002142 /*
2143 * Count the retransmission made on RTO correctly (only when
2144 * waiting for the first ACK and did not get it)...
2145 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002146 if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) {
Ilpo Järvinen0a9f2a42007-07-15 00:19:29 -07002147 /* For some reason this R-bit might get cleared? */
2148 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
2149 tp->retrans_out += tcp_skb_pcount(skb);
Ilpo Järvinend1a54c62007-02-21 23:11:57 -08002150 /* ...enter this if branch just for the first segment */
2151 flag |= FLAG_DATA_ACKED;
2152 } else {
Ilpo Järvinen009a2e32007-09-20 11:34:38 -07002153 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2154 tp->undo_marker = 0;
Ilpo Järvinen23aeeec2007-11-13 21:03:13 -08002155 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
Ilpo Järvinend1a54c62007-02-21 23:11:57 -08002156 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002157
Ilpo Järvinen79d445162008-05-13 02:54:19 -07002158 /* Marking forward transmissions that were made after RTO lost
2159 * can cause unnecessary retransmissions in some scenarios,
2160 * SACK blocks will mitigate that in some but not in all cases.
2161 * We used to not mark them but it was causing break-ups with
2162 * receivers that do only in-order receival.
2163 *
2164 * TODO: we could detect presence of such receiver and select
2165 * different behavior per flow.
2166 */
2167 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
Ilpo Järvinen9bff40f2007-05-27 01:53:49 -07002168 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2169 tp->lost_out += tcp_skb_pcount(skb);
Ilpo Järvinen006f5822008-09-20 21:20:20 -07002170 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002171 }
2172 }
Ilpo Järvinen005903b2007-08-09 14:44:16 +03002173 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002174
Ilpo Järvinen95c49222007-02-21 23:05:18 -08002175 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002176 tp->snd_cwnd_cnt = 0;
2177 tp->snd_cwnd_stamp = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002178 tp->frto_counter = 0;
Ilpo Järvinen16e90682007-10-11 17:32:31 -07002179 tp->bytes_acked = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002180
2181 tp->reordering = min_t(unsigned int, tp->reordering,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002182 sysctl_tcp_reordering);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002183 tcp_set_ca_state(sk, TCP_CA_Loss);
Ilpo Järvinen79d445162008-05-13 02:54:19 -07002184 tp->high_seq = tp->snd_nxt;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002185 TCP_ECN_queue_cwr(tp);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002186
Ilpo Järvinen64edc272008-09-20 21:18:32 -07002187 tcp_clear_all_retrans_hints(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002188}
2189
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07002190static void tcp_clear_retrans_partial(struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002191{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002192 tp->retrans_out = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002193 tp->lost_out = 0;
2194
2195 tp->undo_marker = 0;
2196 tp->undo_retrans = 0;
2197}
2198
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07002199void tcp_clear_retrans(struct tcp_sock *tp)
2200{
2201 tcp_clear_retrans_partial(tp);
2202
2203 tp->fackets_out = 0;
2204 tp->sacked_out = 0;
2205}
2206
Linus Torvalds1da177e2005-04-16 15:20:36 -07002207/* Enter Loss state. If "how" is not zero, forget all SACK information
2208 * and reset tags completely, otherwise preserve SACKs. If receiver
2209 * dropped its ofo queue, we will know this due to reneging detection.
2210 */
2211void tcp_enter_loss(struct sock *sk, int how)
2212{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002213 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002214 struct tcp_sock *tp = tcp_sk(sk);
2215 struct sk_buff *skb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002216
2217 /* Reduce ssthresh if it has not yet been made inside this window. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002218 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
2219 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2220 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2221 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2222 tcp_ca_event(sk, CA_EVENT_LOSS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002223 }
2224 tp->snd_cwnd = 1;
2225 tp->snd_cwnd_cnt = 0;
2226 tp->snd_cwnd_stamp = tcp_time_stamp;
2227
Stephen Hemminger9772efb2005-11-10 17:09:53 -08002228 tp->bytes_acked = 0;
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07002229 tcp_clear_retrans_partial(tp);
2230
2231 if (tcp_is_reno(tp))
2232 tcp_reset_reno_sack(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002233
Ilpo Järvinenb7689202007-09-20 11:37:19 -07002234 if (!how) {
2235 /* Push undo marker, if it was plain RTO and nothing
2236 * was retransmitted. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002237 tp->undo_marker = tp->snd_una;
Ilpo Järvinenb7689202007-09-20 11:37:19 -07002238 } else {
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07002239 tp->sacked_out = 0;
2240 tp->fackets_out = 0;
Ilpo Järvinenb7689202007-09-20 11:37:19 -07002241 }
Ilpo Järvinen64edc272008-09-20 21:18:32 -07002242 tcp_clear_all_retrans_hints(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002243
David S. Millerfe067e82007-03-07 12:12:44 -08002244 tcp_for_write_queue(skb, sk) {
2245 if (skb == tcp_send_head(sk))
2246 break;
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07002247
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002248 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002249 tp->undo_marker = 0;
2250 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
2251 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
2252 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2253 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2254 tp->lost_out += tcp_skb_pcount(skb);
Ilpo Järvinen006f5822008-09-20 21:20:20 -07002255 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002256 }
2257 }
Ilpo Järvinen005903b2007-08-09 14:44:16 +03002258 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002259
2260 tp->reordering = min_t(unsigned int, tp->reordering,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002261 sysctl_tcp_reordering);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002262 tcp_set_ca_state(sk, TCP_CA_Loss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002263 tp->high_seq = tp->snd_nxt;
2264 TCP_ECN_queue_cwr(tp);
Ryousei Takano564262c2007-10-25 23:03:52 -07002265 /* Abort F-RTO algorithm if one is in progress */
Ilpo Järvinen580e5722007-05-19 13:56:57 -07002266 tp->frto_counter = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002267}
2268
Ilpo Järvinencadbd032007-12-31 04:49:21 -08002269/* If ACK arrived pointing to a remembered SACK, it means that our
2270 * remembered SACKs do not reflect real state of receiver i.e.
2271 * receiver _host_ is heavily congested (or buggy).
2272 *
2273 * Do processing similar to RTO timeout.
2274 */
2275static int tcp_check_sack_reneging(struct sock *sk, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002276{
Ilpo Järvinencadbd032007-12-31 04:49:21 -08002277 if (flag & FLAG_SACK_RENEGING) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002278 struct inet_connection_sock *icsk = inet_csk(sk);
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07002279 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002280
2281 tcp_enter_loss(sk, 1);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002282 icsk->icsk_retransmits++;
David S. Millerfe067e82007-03-07 12:12:44 -08002283 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002284 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002285 icsk->icsk_rto, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002286 return 1;
2287 }
2288 return 0;
2289}
2290
2291static inline int tcp_fackets_out(struct tcp_sock *tp)
2292{
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002293 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002294}
2295
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002296/* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2297 * counter when SACK is enabled (without SACK, sacked_out is used for
2298 * that purpose).
2299 *
2300 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2301 * segments up to the highest received SACK block so far and holes in
2302 * between them.
2303 *
2304 * With reordering, holes may still be in flight, so RFC3517 recovery
2305 * uses pure sacked_out (total number of SACKed segments) even though
2306 * it violates the RFC that uses duplicate ACKs, often these are equal
2307 * but when e.g. out-of-window ACKs or packet duplication occurs,
2308 * they differ. Since neither occurs due to loss, TCP should really
2309 * ignore them.
2310 */
2311static inline int tcp_dupack_heurestics(struct tcp_sock *tp)
2312{
2313 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2314}
2315
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002316static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002317{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002318 return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002319}
2320
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002321static inline int tcp_head_timedout(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002322{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002323 struct tcp_sock *tp = tcp_sk(sk);
2324
Linus Torvalds1da177e2005-04-16 15:20:36 -07002325 return tp->packets_out &&
David S. Millerfe067e82007-03-07 12:12:44 -08002326 tcp_skb_timedout(sk, tcp_write_queue_head(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002327}
2328
2329/* Linux NewReno/SACK/FACK/ECN state machine.
2330 * --------------------------------------
2331 *
2332 * "Open" Normal state, no dubious events, fast path.
2333 * "Disorder" In all the respects it is "Open",
2334 * but requires a bit more attention. It is entered when
2335 * we see some SACKs or dupacks. It is split of "Open"
2336 * mainly to move some processing from fast path to slow one.
2337 * "CWR" CWND was reduced due to some Congestion Notification event.
2338 * It can be ECN, ICMP source quench, local device congestion.
2339 * "Recovery" CWND was reduced, we are fast-retransmitting.
2340 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2341 *
2342 * tcp_fastretrans_alert() is entered:
2343 * - each incoming ACK, if state is not "Open"
2344 * - when arrived ACK is unusual, namely:
2345 * * SACK
2346 * * Duplicate ACK.
2347 * * ECN ECE.
2348 *
2349 * Counting packets in flight is pretty simple.
2350 *
2351 * in_flight = packets_out - left_out + retrans_out
2352 *
2353 * packets_out is SND.NXT-SND.UNA counted in packets.
2354 *
2355 * retrans_out is number of retransmitted segments.
2356 *
2357 * left_out is number of segments left network, but not ACKed yet.
2358 *
2359 * left_out = sacked_out + lost_out
2360 *
2361 * sacked_out: Packets, which arrived to receiver out of order
2362 * and hence not ACKed. With SACKs this number is simply
2363 * amount of SACKed data. Even without SACKs
2364 * it is easy to give pretty reliable estimate of this number,
2365 * counting duplicate ACKs.
2366 *
2367 * lost_out: Packets lost by network. TCP has no explicit
2368 * "loss notification" feedback from network (for now).
2369 * It means that this number can be only _guessed_.
2370 * Actually, it is the heuristics to predict lossage that
2371 * distinguishes different algorithms.
2372 *
2373 * F.e. after RTO, when all the queue is considered as lost,
2374 * lost_out = packets_out and in_flight = retrans_out.
2375 *
2376 * Essentially, we have now two algorithms counting
2377 * lost packets.
2378 *
2379 * FACK: It is the simplest heuristics. As soon as we decided
2380 * that something is lost, we decide that _all_ not SACKed
2381 * packets until the most forward SACK are lost. I.e.
2382 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2383 * It is absolutely correct estimate, if network does not reorder
2384 * packets. And it loses any connection to reality when reordering
2385 * takes place. We use FACK by default until reordering
2386 * is suspected on the path to this destination.
2387 *
2388 * NewReno: when Recovery is entered, we assume that one segment
2389 * is lost (classic Reno). While we are in Recovery and
2390 * a partial ACK arrives, we assume that one more packet
2391 * is lost (NewReno). This heuristics are the same in NewReno
2392 * and SACK.
2393 *
2394 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2395 * deflation etc. CWND is real congestion window, never inflated, changes
2396 * only according to classic VJ rules.
2397 *
2398 * Really tricky (and requiring careful tuning) part of algorithm
2399 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2400 * The first determines the moment _when_ we should reduce CWND and,
2401 * hence, slow down forward transmission. In fact, it determines the moment
2402 * when we decide that hole is caused by loss, rather than by a reorder.
2403 *
2404 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2405 * holes, caused by lost packets.
2406 *
2407 * And the most logically complicated part of algorithm is undo
2408 * heuristics. We detect false retransmits due to both too early
2409 * fast retransmit (reordering) and underestimated RTO, analyzing
2410 * timestamps and D-SACKs. When we detect that some segments were
2411 * retransmitted by mistake and CWND reduction was wrong, we undo
2412 * window reduction and abort recovery phase. This logic is hidden
2413 * inside several functions named tcp_try_undo_<something>.
2414 */
2415
2416/* This function decides, when we should leave Disordered state
2417 * and enter Recovery phase, reducing congestion window.
2418 *
2419 * Main question: may we further continue forward transmission
2420 * with the same cwnd?
2421 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002422static int tcp_time_to_recover(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002423{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002424 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002425 __u32 packets_out;
2426
Ryousei Takano564262c2007-10-25 23:03:52 -07002427 /* Do not perform any recovery during F-RTO algorithm */
Ilpo Järvinen52c63f12007-02-21 23:06:52 -08002428 if (tp->frto_counter)
2429 return 0;
2430
Linus Torvalds1da177e2005-04-16 15:20:36 -07002431 /* Trick#1: The loss is proven. */
2432 if (tp->lost_out)
2433 return 1;
2434
2435 /* Not-A-Trick#2 : Classic rule... */
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002436 if (tcp_dupack_heurestics(tp) > tp->reordering)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002437 return 1;
2438
2439 /* Trick#3 : when we use RFC2988 timer restart, fast
2440 * retransmit can be triggered by timeout of queue head.
2441 */
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002442 if (tcp_is_fack(tp) && tcp_head_timedout(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002443 return 1;
2444
2445 /* Trick#4: It is still not OK... But will it be useful to delay
2446 * recovery more?
2447 */
2448 packets_out = tp->packets_out;
2449 if (packets_out <= tp->reordering &&
2450 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002451 !tcp_may_send_now(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002452 /* We have nothing to send. This connection is limited
2453 * either by receiver window or by application.
2454 */
2455 return 1;
2456 }
2457
2458 return 0;
2459}
2460
Ilpo Järvinen7363a5b2009-02-28 04:44:33 +00002461/* New heuristics: it is possible only after we switched to restart timer
2462 * each time when something is ACKed. Hence, we can detect timed out packets
2463 * during fast retransmit without falling to slow start.
2464 *
2465 * Usefulness of this as is very questionable, since we should know which of
2466 * the segments is the next to timeout which is relatively expensive to find
2467 * in general case unless we add some data structure just for that. The
2468 * current approach certainly won't find the right one too often and when it
2469 * finally does find _something_ it usually marks large part of the window
2470 * right away (because a retransmission with a larger timestamp blocks the
2471 * loop from advancing). -ij
2472 */
2473static void tcp_timeout_skbs(struct sock *sk)
2474{
2475 struct tcp_sock *tp = tcp_sk(sk);
2476 struct sk_buff *skb;
2477
2478 if (!tcp_is_fack(tp) || !tcp_head_timedout(sk))
2479 return;
2480
2481 skb = tp->scoreboard_skb_hint;
2482 if (tp->scoreboard_skb_hint == NULL)
2483 skb = tcp_write_queue_head(sk);
2484
2485 tcp_for_write_queue_from(skb, sk) {
2486 if (skb == tcp_send_head(sk))
2487 break;
2488 if (!tcp_skb_timedout(sk, skb))
2489 break;
2490
2491 tcp_skb_mark_lost(tp, skb);
2492 }
2493
2494 tp->scoreboard_skb_hint = skb;
2495
2496 tcp_verify_left_out(tp);
2497}
2498
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002499/* Mark head of queue up as lost. With RFC3517 SACK, the packets is
2500 * is against sacked "cnt", otherwise it's against facked "cnt"
2501 */
Ilpo Järvinen1b69d742008-04-07 22:31:38 -07002502static void tcp_mark_head_lost(struct sock *sk, int packets)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002503{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002504 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002505 struct sk_buff *skb;
Ilpo Järvinenc137f3d2008-04-07 22:32:38 -07002506 int cnt, oldcnt;
2507 int err;
2508 unsigned int mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002509
Ilpo Järvinen547b7922008-07-25 21:43:18 -07002510 WARN_ON(packets > tp->packets_out);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002511 if (tp->lost_skb_hint) {
2512 skb = tp->lost_skb_hint;
2513 cnt = tp->lost_cnt_hint;
2514 } else {
David S. Millerfe067e82007-03-07 12:12:44 -08002515 skb = tcp_write_queue_head(sk);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002516 cnt = 0;
2517 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002518
David S. Millerfe067e82007-03-07 12:12:44 -08002519 tcp_for_write_queue_from(skb, sk) {
2520 if (skb == tcp_send_head(sk))
2521 break;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002522 /* TODO: do this better */
2523 /* this is not the most efficient way to do this... */
2524 tp->lost_skb_hint = skb;
2525 tp->lost_cnt_hint = cnt;
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002526
Ilpo Järvinenc137f3d2008-04-07 22:32:38 -07002527 if (after(TCP_SKB_CB(skb)->end_seq, tp->high_seq))
2528 break;
2529
2530 oldcnt = cnt;
Ilpo Järvinenad1984e2008-01-30 20:06:02 -08002531 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002532 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2533 cnt += tcp_skb_pcount(skb);
2534
Ilpo Järvinenc137f3d2008-04-07 22:32:38 -07002535 if (cnt > packets) {
2536 if (tcp_is_sack(tp) || (oldcnt >= packets))
2537 break;
2538
2539 mss = skb_shinfo(skb)->gso_size;
2540 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2541 if (err < 0)
2542 break;
2543 cnt = packets;
2544 }
2545
Ilpo Järvinen41ea36e2008-09-20 21:19:22 -07002546 tcp_skb_mark_lost(tp, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002547 }
Ilpo Järvinen005903b2007-08-09 14:44:16 +03002548 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002549}
2550
2551/* Account newly detected lost packet(s) */
2552
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002553static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002554{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002555 struct tcp_sock *tp = tcp_sk(sk);
2556
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002557 if (tcp_is_reno(tp)) {
Ilpo Järvinen1b69d742008-04-07 22:31:38 -07002558 tcp_mark_head_lost(sk, 1);
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002559 } else if (tcp_is_fack(tp)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002560 int lost = tp->fackets_out - tp->reordering;
2561 if (lost <= 0)
2562 lost = 1;
Ilpo Järvinen1b69d742008-04-07 22:31:38 -07002563 tcp_mark_head_lost(sk, lost);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002564 } else {
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002565 int sacked_upto = tp->sacked_out - tp->reordering;
Ilpo Järvinen1b69d742008-04-07 22:31:38 -07002566 if (sacked_upto < fast_rexmit)
2567 sacked_upto = fast_rexmit;
2568 tcp_mark_head_lost(sk, sacked_upto);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002569 }
2570
Ilpo Järvinen7363a5b2009-02-28 04:44:33 +00002571 tcp_timeout_skbs(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002572}
2573
2574/* CWND moderation, preventing bursts due to too big ACKs
2575 * in dubious situations.
2576 */
2577static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2578{
2579 tp->snd_cwnd = min(tp->snd_cwnd,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002580 tcp_packets_in_flight(tp) + tcp_max_burst(tp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002581 tp->snd_cwnd_stamp = tcp_time_stamp;
2582}
2583
Stephen Hemminger72dc5b92006-06-05 17:30:08 -07002584/* Lower bound on congestion window is slow start threshold
2585 * unless congestion avoidance choice decides to overide it.
2586 */
2587static inline u32 tcp_cwnd_min(const struct sock *sk)
2588{
2589 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2590
2591 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
2592}
2593
Linus Torvalds1da177e2005-04-16 15:20:36 -07002594/* Decrease cwnd each second ack. */
Ilpo Järvinen1e757f92007-07-30 19:48:37 -07002595static void tcp_cwnd_down(struct sock *sk, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002596{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002597 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002598 int decr = tp->snd_cwnd_cnt + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002599
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002600 if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK)) ||
2601 (tcp_is_reno(tp) && !(flag & FLAG_NOT_DUP))) {
2602 tp->snd_cwnd_cnt = decr & 1;
Ilpo Järvinen1e757f92007-07-30 19:48:37 -07002603 decr >>= 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002604
Ilpo Järvinen1e757f92007-07-30 19:48:37 -07002605 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
2606 tp->snd_cwnd -= decr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002607
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002608 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
Ilpo Järvinen1e757f92007-07-30 19:48:37 -07002609 tp->snd_cwnd_stamp = tcp_time_stamp;
2610 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002611}
2612
2613/* Nothing was retransmitted or returned timestamp is less
2614 * than timestamp of the first retransmission.
2615 */
2616static inline int tcp_packet_delayed(struct tcp_sock *tp)
2617{
2618 return !tp->retrans_stamp ||
2619 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
Arnd Hannemannd7ee1472008-04-21 14:46:22 -07002620 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002621}
2622
2623/* Undo procedures. */
2624
2625#if FASTRETRANS_DEBUG > 1
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002626static void DBGUNDO(struct sock *sk, const char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002627{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002628 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002629 struct inet_sock *inet = inet_sk(sk);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002630
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002631 if (sk->sk_family == AF_INET) {
Harvey Harrison673d57e2008-10-31 00:53:57 -07002632 printk(KERN_DEBUG "Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002633 msg,
Harvey Harrison673d57e2008-10-31 00:53:57 -07002634 &inet->daddr, ntohs(inet->dport),
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002635 tp->snd_cwnd, tcp_left_out(tp),
2636 tp->snd_ssthresh, tp->prior_ssthresh,
2637 tp->packets_out);
2638 }
2639#if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2640 else if (sk->sk_family == AF_INET6) {
2641 struct ipv6_pinfo *np = inet6_sk(sk);
Harvey Harrison5b095d9892008-10-29 12:52:50 -07002642 printk(KERN_DEBUG "Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002643 msg,
Harvey Harrison0c6ce782008-10-28 16:09:23 -07002644 &np->daddr, ntohs(inet->dport),
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002645 tp->snd_cwnd, tcp_left_out(tp),
2646 tp->snd_ssthresh, tp->prior_ssthresh,
2647 tp->packets_out);
2648 }
2649#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002650}
2651#else
2652#define DBGUNDO(x...) do { } while (0)
2653#endif
2654
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002655static void tcp_undo_cwr(struct sock *sk, const int undo)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002656{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002657 struct tcp_sock *tp = tcp_sk(sk);
2658
Linus Torvalds1da177e2005-04-16 15:20:36 -07002659 if (tp->prior_ssthresh) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002660 const struct inet_connection_sock *icsk = inet_csk(sk);
2661
2662 if (icsk->icsk_ca_ops->undo_cwnd)
2663 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002664 else
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002665 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002666
2667 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
2668 tp->snd_ssthresh = tp->prior_ssthresh;
2669 TCP_ECN_withdraw_cwr(tp);
2670 }
2671 } else {
2672 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2673 }
2674 tcp_moderate_cwnd(tp);
2675 tp->snd_cwnd_stamp = tcp_time_stamp;
2676}
2677
2678static inline int tcp_may_undo(struct tcp_sock *tp)
2679{
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002680 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002681}
2682
2683/* People celebrate: "We love our President!" */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002684static int tcp_try_undo_recovery(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002685{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002686 struct tcp_sock *tp = tcp_sk(sk);
2687
Linus Torvalds1da177e2005-04-16 15:20:36 -07002688 if (tcp_may_undo(tp)) {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07002689 int mib_idx;
2690
Linus Torvalds1da177e2005-04-16 15:20:36 -07002691 /* Happy end! We did not retransmit anything
2692 * or our original transmission succeeded.
2693 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002694 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002695 tcp_undo_cwr(sk, 1);
2696 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07002697 mib_idx = LINUX_MIB_TCPLOSSUNDO;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002698 else
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07002699 mib_idx = LINUX_MIB_TCPFULLUNDO;
2700
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07002701 NET_INC_STATS_BH(sock_net(sk), mib_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002702 tp->undo_marker = 0;
2703 }
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002704 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002705 /* Hold old state until something *above* high_seq
2706 * is ACKed. For Reno it is MUST to prevent false
2707 * fast retransmits (RFC2582). SACK TCP is safe. */
2708 tcp_moderate_cwnd(tp);
2709 return 1;
2710 }
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002711 tcp_set_ca_state(sk, TCP_CA_Open);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002712 return 0;
2713}
2714
2715/* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002716static void tcp_try_undo_dsack(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002717{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002718 struct tcp_sock *tp = tcp_sk(sk);
2719
Linus Torvalds1da177e2005-04-16 15:20:36 -07002720 if (tp->undo_marker && !tp->undo_retrans) {
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002721 DBGUNDO(sk, "D-SACK");
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002722 tcp_undo_cwr(sk, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002723 tp->undo_marker = 0;
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07002724 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002725 }
2726}
2727
2728/* Undo during fast recovery after partial ACK. */
2729
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002730static int tcp_try_undo_partial(struct sock *sk, int acked)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002731{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002732 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002733 /* Partial ACK arrived. Force Hoe's retransmit. */
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002734 int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002735
2736 if (tcp_may_undo(tp)) {
2737 /* Plain luck! Hole if filled with delayed
2738 * packet, rather than with a retransmit.
2739 */
2740 if (tp->retrans_out == 0)
2741 tp->retrans_stamp = 0;
2742
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002743 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002744
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002745 DBGUNDO(sk, "Hoe");
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002746 tcp_undo_cwr(sk, 0);
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07002747 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002748
2749 /* So... Do not make Hoe's retransmit yet.
2750 * If the first packet was delayed, the rest
2751 * ones are most probably delayed as well.
2752 */
2753 failed = 0;
2754 }
2755 return failed;
2756}
2757
2758/* Undo during loss recovery after partial ACK. */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002759static int tcp_try_undo_loss(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002760{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002761 struct tcp_sock *tp = tcp_sk(sk);
2762
Linus Torvalds1da177e2005-04-16 15:20:36 -07002763 if (tcp_may_undo(tp)) {
2764 struct sk_buff *skb;
David S. Millerfe067e82007-03-07 12:12:44 -08002765 tcp_for_write_queue(skb, sk) {
2766 if (skb == tcp_send_head(sk))
2767 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002768 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2769 }
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002770
Ilpo Järvinen5af4ec22007-09-20 11:30:48 -07002771 tcp_clear_all_retrans_hints(tp);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002772
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002773 DBGUNDO(sk, "partial loss");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002774 tp->lost_out = 0;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002775 tcp_undo_cwr(sk, 1);
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07002776 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002777 inet_csk(sk)->icsk_retransmits = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002778 tp->undo_marker = 0;
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002779 if (tcp_is_sack(tp))
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002780 tcp_set_ca_state(sk, TCP_CA_Open);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002781 return 1;
2782 }
2783 return 0;
2784}
2785
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002786static inline void tcp_complete_cwr(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002787{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002788 struct tcp_sock *tp = tcp_sk(sk);
Stephen Hemminger317a76f2005-06-23 12:19:55 -07002789 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002790 tp->snd_cwnd_stamp = tcp_time_stamp;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002791 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002792}
2793
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07002794static void tcp_try_keep_open(struct sock *sk)
2795{
2796 struct tcp_sock *tp = tcp_sk(sk);
2797 int state = TCP_CA_Open;
2798
2799 if (tcp_left_out(tp) || tp->retrans_out || tp->undo_marker)
2800 state = TCP_CA_Disorder;
2801
2802 if (inet_csk(sk)->icsk_ca_state != state) {
2803 tcp_set_ca_state(sk, state);
2804 tp->high_seq = tp->snd_nxt;
2805 }
2806}
2807
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002808static void tcp_try_to_open(struct sock *sk, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002809{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002810 struct tcp_sock *tp = tcp_sk(sk);
2811
Ilpo Järvinen86426c22007-08-09 14:45:17 +03002812 tcp_verify_left_out(tp);
2813
Ilpo Järvinena1c1f282008-05-13 02:53:26 -07002814 if (!tp->frto_counter && tp->retrans_out == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002815 tp->retrans_stamp = 0;
2816
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002817 if (flag & FLAG_ECE)
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -08002818 tcp_enter_cwr(sk, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002819
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002820 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07002821 tcp_try_keep_open(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002822 tcp_moderate_cwnd(tp);
2823 } else {
Ilpo Järvinen1e757f92007-07-30 19:48:37 -07002824 tcp_cwnd_down(sk, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002825 }
2826}
2827
John Heffner5d424d52006-03-20 17:53:41 -08002828static void tcp_mtup_probe_failed(struct sock *sk)
2829{
2830 struct inet_connection_sock *icsk = inet_csk(sk);
2831
2832 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2833 icsk->icsk_mtup.probe_size = 0;
2834}
2835
Ilpo Järvinen72211e92009-03-14 14:23:04 +00002836static void tcp_mtup_probe_success(struct sock *sk)
John Heffner5d424d52006-03-20 17:53:41 -08002837{
2838 struct tcp_sock *tp = tcp_sk(sk);
2839 struct inet_connection_sock *icsk = inet_csk(sk);
2840
2841 /* FIXME: breaks with very large cwnd */
2842 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2843 tp->snd_cwnd = tp->snd_cwnd *
2844 tcp_mss_to_mtu(sk, tp->mss_cache) /
2845 icsk->icsk_mtup.probe_size;
2846 tp->snd_cwnd_cnt = 0;
2847 tp->snd_cwnd_stamp = tcp_time_stamp;
2848 tp->rcv_ssthresh = tcp_current_ssthresh(sk);
2849
2850 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2851 icsk->icsk_mtup.probe_size = 0;
2852 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2853}
2854
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08002855/* Do a simple retransmit without using the backoff mechanisms in
2856 * tcp_timer. This is used for path mtu discovery.
2857 * The socket is already locked here.
2858 */
2859void tcp_simple_retransmit(struct sock *sk)
2860{
2861 const struct inet_connection_sock *icsk = inet_csk(sk);
2862 struct tcp_sock *tp = tcp_sk(sk);
2863 struct sk_buff *skb;
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00002864 unsigned int mss = tcp_current_mss(sk);
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08002865 u32 prior_lost = tp->lost_out;
2866
2867 tcp_for_write_queue(skb, sk) {
2868 if (skb == tcp_send_head(sk))
2869 break;
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08002870 if (tcp_skb_seglen(skb) > mss &&
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08002871 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2872 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2873 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2874 tp->retrans_out -= tcp_skb_pcount(skb);
2875 }
2876 tcp_skb_mark_lost_uncond_verify(tp, skb);
2877 }
2878 }
2879
2880 tcp_clear_retrans_hints_partial(tp);
2881
2882 if (prior_lost == tp->lost_out)
2883 return;
2884
2885 if (tcp_is_reno(tp))
2886 tcp_limit_reno_sacked(tp);
2887
2888 tcp_verify_left_out(tp);
2889
2890 /* Don't muck with the congestion window here.
2891 * Reason is that we do not increase amount of _data_
2892 * in network, but units changed and effective
2893 * cwnd/ssthresh really reduced now.
2894 */
2895 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2896 tp->high_seq = tp->snd_nxt;
2897 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2898 tp->prior_ssthresh = 0;
2899 tp->undo_marker = 0;
2900 tcp_set_ca_state(sk, TCP_CA_Loss);
2901 }
2902 tcp_xmit_retransmit_queue(sk);
2903}
2904
Linus Torvalds1da177e2005-04-16 15:20:36 -07002905/* Process an event, which can update packets-in-flight not trivially.
2906 * Main goal of this function is to calculate new estimate for left_out,
2907 * taking into account both packets sitting in receiver's buffer and
2908 * packets lost by network.
2909 *
2910 * Besides that it does CWND reduction, when packet loss is detected
2911 * and changes state of machine.
2912 *
2913 * It does _not_ decide what to send, it is made in function
2914 * tcp_xmit_retransmit_queue().
2915 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002916static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002917{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002918 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002919 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002920 int is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
2921 int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002922 (tcp_fackets_out(tp) > tp->reordering));
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07002923 int fast_rexmit = 0, mib_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002924
Ilpo Järvinen3ccd3132007-12-31 04:43:32 -08002925 if (WARN_ON(!tp->packets_out && tp->sacked_out))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002926 tp->sacked_out = 0;
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07002927 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002928 tp->fackets_out = 0;
2929
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002930 /* Now state machine starts.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002931 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002932 if (flag & FLAG_ECE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002933 tp->prior_ssthresh = 0;
2934
2935 /* B. In all the states check for reneging SACKs. */
Ilpo Järvinencadbd032007-12-31 04:49:21 -08002936 if (tcp_check_sack_reneging(sk, flag))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002937 return;
2938
2939 /* C. Process data loss notification, provided it is valid. */
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002940 if (tcp_is_fack(tp) && (flag & FLAG_DATA_LOST) &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07002941 before(tp->snd_una, tp->high_seq) &&
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002942 icsk->icsk_ca_state != TCP_CA_Open &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07002943 tp->fackets_out > tp->reordering) {
Ilpo Järvinen1b69d742008-04-07 22:31:38 -07002944 tcp_mark_head_lost(sk, tp->fackets_out - tp->reordering);
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07002945 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002946 }
2947
Ilpo Järvinen005903b2007-08-09 14:44:16 +03002948 /* D. Check consistency of the current state. */
2949 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002950
2951 /* E. Check state exit conditions. State can be terminated
2952 * when high_seq is ACKed. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002953 if (icsk->icsk_ca_state == TCP_CA_Open) {
Ilpo Järvinen547b7922008-07-25 21:43:18 -07002954 WARN_ON(tp->retrans_out != 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002955 tp->retrans_stamp = 0;
2956 } else if (!before(tp->snd_una, tp->high_seq)) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002957 switch (icsk->icsk_ca_state) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002958 case TCP_CA_Loss:
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002959 icsk->icsk_retransmits = 0;
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002960 if (tcp_try_undo_recovery(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002961 return;
2962 break;
2963
2964 case TCP_CA_CWR:
2965 /* CWR is to be held something *above* high_seq
2966 * is ACKed for CWR bit to reach receiver. */
2967 if (tp->snd_una != tp->high_seq) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002968 tcp_complete_cwr(sk);
2969 tcp_set_ca_state(sk, TCP_CA_Open);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002970 }
2971 break;
2972
2973 case TCP_CA_Disorder:
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002974 tcp_try_undo_dsack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002975 if (!tp->undo_marker ||
2976 /* For SACK case do not Open to allow to undo
2977 * catching for all duplicate ACKs. */
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002978 tcp_is_reno(tp) || tp->snd_una != tp->high_seq) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002979 tp->undo_marker = 0;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002980 tcp_set_ca_state(sk, TCP_CA_Open);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002981 }
2982 break;
2983
2984 case TCP_CA_Recovery:
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002985 if (tcp_is_reno(tp))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002986 tcp_reset_reno_sack(tp);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002987 if (tcp_try_undo_recovery(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002988 return;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002989 tcp_complete_cwr(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002990 break;
2991 }
2992 }
2993
2994 /* F. Process state. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002995 switch (icsk->icsk_ca_state) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002996 case TCP_CA_Recovery:
Ilpo Järvinen2e605292007-08-02 19:46:58 -07002997 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002998 if (tcp_is_reno(tp) && is_dupack)
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002999 tcp_add_reno_sack(sk);
Ilpo Järvinen1b6d4272007-08-09 14:53:36 +03003000 } else
3001 do_lost = tcp_try_undo_partial(sk, pkts_acked);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003002 break;
3003 case TCP_CA_Loss:
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003004 if (flag & FLAG_DATA_ACKED)
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003005 icsk->icsk_retransmits = 0;
Ilpo Järvinen882beba2008-04-07 22:33:07 -07003006 if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED)
3007 tcp_reset_reno_sack(tp);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003008 if (!tcp_try_undo_loss(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003009 tcp_moderate_cwnd(tp);
3010 tcp_xmit_retransmit_queue(sk);
3011 return;
3012 }
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003013 if (icsk->icsk_ca_state != TCP_CA_Open)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003014 return;
3015 /* Loss is undone; fall through to processing in Open state. */
3016 default:
Ilpo Järvinene60402d2007-08-09 15:14:46 +03003017 if (tcp_is_reno(tp)) {
Ilpo Järvinen2e605292007-08-02 19:46:58 -07003018 if (flag & FLAG_SND_UNA_ADVANCED)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003019 tcp_reset_reno_sack(tp);
3020 if (is_dupack)
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003021 tcp_add_reno_sack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003022 }
3023
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003024 if (icsk->icsk_ca_state == TCP_CA_Disorder)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003025 tcp_try_undo_dsack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003026
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003027 if (!tcp_time_to_recover(sk)) {
3028 tcp_try_to_open(sk, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003029 return;
3030 }
3031
John Heffner5d424d52006-03-20 17:53:41 -08003032 /* MTU probe failure: don't reduce cwnd */
3033 if (icsk->icsk_ca_state < TCP_CA_CWR &&
3034 icsk->icsk_mtup.probe_size &&
John Heffner0e7b1362006-03-20 21:32:58 -08003035 tp->snd_una == tp->mtu_probe.probe_seq_start) {
John Heffner5d424d52006-03-20 17:53:41 -08003036 tcp_mtup_probe_failed(sk);
3037 /* Restores the reduction we did in tcp_mtup_probe() */
3038 tp->snd_cwnd++;
3039 tcp_simple_retransmit(sk);
3040 return;
3041 }
3042
Linus Torvalds1da177e2005-04-16 15:20:36 -07003043 /* Otherwise enter Recovery state */
3044
Ilpo Järvinene60402d2007-08-09 15:14:46 +03003045 if (tcp_is_reno(tp))
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07003046 mib_idx = LINUX_MIB_TCPRENORECOVERY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003047 else
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07003048 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
3049
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07003050 NET_INC_STATS_BH(sock_net(sk), mib_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003051
3052 tp->high_seq = tp->snd_nxt;
3053 tp->prior_ssthresh = 0;
3054 tp->undo_marker = tp->snd_una;
3055 tp->undo_retrans = tp->retrans_out;
3056
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003057 if (icsk->icsk_ca_state < TCP_CA_CWR) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003058 if (!(flag & FLAG_ECE))
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003059 tp->prior_ssthresh = tcp_current_ssthresh(sk);
3060 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003061 TCP_ECN_queue_cwr(tp);
3062 }
3063
Stephen Hemminger9772efb2005-11-10 17:09:53 -08003064 tp->bytes_acked = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003065 tp->snd_cwnd_cnt = 0;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003066 tcp_set_ca_state(sk, TCP_CA_Recovery);
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08003067 fast_rexmit = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003068 }
3069
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08003070 if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk)))
3071 tcp_update_scoreboard(sk, fast_rexmit);
Ilpo Järvinen1e757f92007-07-30 19:48:37 -07003072 tcp_cwnd_down(sk, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003073 tcp_xmit_retransmit_queue(sk);
3074}
3075
Ilpo Järvinen41834b72008-12-05 22:43:26 -08003076static void tcp_valid_rtt_meas(struct sock *sk, u32 seq_rtt)
3077{
3078 tcp_rtt_estimator(sk, seq_rtt);
3079 tcp_set_rto(sk);
3080 inet_csk(sk)->icsk_backoff = 0;
3081}
3082
Linus Torvalds1da177e2005-04-16 15:20:36 -07003083/* Read draft-ietf-tcplw-high-performance before mucking
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08003084 * with this code. (Supersedes RFC1323)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003085 */
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08003086static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003087{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003088 /* RTTM Rule: A TSecr value received in a segment is used to
3089 * update the averaged RTT measurement only if the segment
3090 * acknowledges some new data, i.e., only if it advances the
3091 * left edge of the send window.
3092 *
3093 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3094 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
3095 *
3096 * Changed: reset backoff as soon as we see the first valid sample.
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08003097 * If we do not, we get strongly overestimated rto. With timestamps
Linus Torvalds1da177e2005-04-16 15:20:36 -07003098 * samples are accepted even from very old segments: f.e., when rtt=1
3099 * increases to 8, we retransmit 5 times and after 8 seconds delayed
3100 * answer arrives rto becomes 120 seconds! If at least one of segments
3101 * in window is lost... Voila. --ANK (010210)
3102 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003103 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen41834b72008-12-05 22:43:26 -08003104
3105 tcp_valid_rtt_meas(sk, tcp_time_stamp - tp->rx_opt.rcv_tsecr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003106}
3107
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08003108static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003109{
3110 /* We don't have a timestamp. Can only use
3111 * packets that are not retransmitted to determine
3112 * rtt estimates. Also, we must not reset the
3113 * backoff for rto until we get a non-retransmitted
3114 * packet. This allows us to deal with a situation
3115 * where the network delay has increased suddenly.
3116 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
3117 */
3118
3119 if (flag & FLAG_RETRANS_DATA_ACKED)
3120 return;
3121
Ilpo Järvinen41834b72008-12-05 22:43:26 -08003122 tcp_valid_rtt_meas(sk, seq_rtt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003123}
3124
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003125static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08003126 const s32 seq_rtt)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003127{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003128 const struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003129 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
3130 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08003131 tcp_ack_saw_tstamp(sk, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003132 else if (seq_rtt >= 0)
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08003133 tcp_ack_no_tstamp(sk, seq_rtt, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003134}
3135
Ilpo Järvinenc3a05c62007-12-02 00:47:59 +02003136static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003137{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003138 const struct inet_connection_sock *icsk = inet_csk(sk);
Ilpo Järvinenc3a05c62007-12-02 00:47:59 +02003139 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003140 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003141}
3142
Linus Torvalds1da177e2005-04-16 15:20:36 -07003143/* Restart timer after forward progress on connection.
3144 * RFC2988 recommends to restart timer to now+rto.
3145 */
Ilpo Järvinen6728e7d2007-08-24 22:53:26 -07003146static void tcp_rearm_rto(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003147{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003148 struct tcp_sock *tp = tcp_sk(sk);
3149
Linus Torvalds1da177e2005-04-16 15:20:36 -07003150 if (!tp->packets_out) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003151 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003152 } else {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003153 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3154 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003155 }
3156}
3157
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003158/* If we get here, the whole TSO packet has not been acked. */
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003159static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003160{
3161 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003162 u32 packets_acked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003163
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003164 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003165
3166 packets_acked = tcp_skb_pcount(skb);
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003167 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003168 return 0;
3169 packets_acked -= tcp_skb_pcount(skb);
3170
3171 if (packets_acked) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003172 BUG_ON(tcp_skb_pcount(skb) == 0);
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003173 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003174 }
3175
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003176 return packets_acked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003177}
3178
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003179/* Remove acknowledged frames from the retransmission queue. If our packet
3180 * is before the ack sequence we can discard it as it's confirmed to have
3181 * arrived at the other end.
3182 */
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07003183static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3184 u32 prior_snd_una)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003185{
3186 struct tcp_sock *tp = tcp_sk(sk);
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08003187 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003188 struct sk_buff *skb;
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003189 u32 now = tcp_time_stamp;
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003190 int fully_acked = 1;
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003191 int flag = 0;
Ilpo Järvinen72018832007-12-30 04:37:55 -08003192 u32 pkts_acked = 0;
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003193 u32 reord = tp->packets_out;
Ilpo Järvinen90638a02008-09-20 21:25:52 -07003194 u32 prior_sacked = tp->sacked_out;
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003195 s32 seq_rtt = -1;
Gavin McCullagh2072c222007-12-29 19:11:21 -08003196 s32 ca_seq_rtt = -1;
Ilpo Järvinenb9ce2042007-06-15 15:08:43 -07003197 ktime_t last_ackt = net_invalid_timestamp();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003198
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003199 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09003200 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
Ilpo Järvinen72018832007-12-30 04:37:55 -08003201 u32 acked_pcount;
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003202 u8 sacked = scb->sacked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003203
Gavin McCullagh2072c222007-12-29 19:11:21 -08003204 /* Determine how many packets and what bytes were acked, tso and else */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003205 if (after(scb->end_seq, tp->snd_una)) {
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003206 if (tcp_skb_pcount(skb) == 1 ||
3207 !after(tp->snd_una, scb->seq))
3208 break;
3209
Ilpo Järvinen72018832007-12-30 04:37:55 -08003210 acked_pcount = tcp_tso_acked(sk, skb);
3211 if (!acked_pcount)
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003212 break;
3213
3214 fully_acked = 0;
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003215 } else {
Ilpo Järvinen72018832007-12-30 04:37:55 -08003216 acked_pcount = tcp_skb_pcount(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003217 }
3218
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003219 if (sacked & TCPCB_RETRANS) {
3220 if (sacked & TCPCB_SACKED_RETRANS)
Ilpo Järvinen72018832007-12-30 04:37:55 -08003221 tp->retrans_out -= acked_pcount;
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003222 flag |= FLAG_RETRANS_DATA_ACKED;
3223 ca_seq_rtt = -1;
3224 seq_rtt = -1;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003225 if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1))
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003226 flag |= FLAG_NONHEAD_RETRANS_ACKED;
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003227 } else {
Gavin McCullagh2072c222007-12-29 19:11:21 -08003228 ca_seq_rtt = now - scb->when;
3229 last_ackt = skb->tstamp;
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003230 if (seq_rtt < 0) {
Gavin McCullagh2072c222007-12-29 19:11:21 -08003231 seq_rtt = ca_seq_rtt;
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003232 }
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003233 if (!(sacked & TCPCB_SACKED_ACKED))
Ilpo Järvinen72018832007-12-30 04:37:55 -08003234 reord = min(pkts_acked, reord);
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08003235 }
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003236
3237 if (sacked & TCPCB_SACKED_ACKED)
Ilpo Järvinen72018832007-12-30 04:37:55 -08003238 tp->sacked_out -= acked_pcount;
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003239 if (sacked & TCPCB_LOST)
Ilpo Järvinen72018832007-12-30 04:37:55 -08003240 tp->lost_out -= acked_pcount;
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003241
Ilpo Järvinen72018832007-12-30 04:37:55 -08003242 tp->packets_out -= acked_pcount;
3243 pkts_acked += acked_pcount;
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003244
Ilpo Järvinen009a2e32007-09-20 11:34:38 -07003245 /* Initial outgoing SYN's get put onto the write_queue
3246 * just like anything else we transmit. It is not
3247 * true data, and if we misinform our callers that
3248 * this ACK acks real data, we will erroneously exit
3249 * connection startup slow start one packet too
3250 * quickly. This is severely frowned upon behavior.
3251 */
3252 if (!(scb->flags & TCPCB_FLAG_SYN)) {
3253 flag |= FLAG_DATA_ACKED;
3254 } else {
3255 flag |= FLAG_SYN_ACKED;
3256 tp->retrans_stamp = 0;
3257 }
3258
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003259 if (!fully_acked)
3260 break;
3261
David S. Millerfe067e82007-03-07 12:12:44 -08003262 tcp_unlink_write_queue(skb, sk);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08003263 sk_wmem_free_skb(sk, skb);
Ilpo Järvinen90638a02008-09-20 21:25:52 -07003264 tp->scoreboard_skb_hint = NULL;
Ilpo Järvinenef9da472008-09-20 21:25:15 -07003265 if (skb == tp->retransmit_skb_hint)
3266 tp->retransmit_skb_hint = NULL;
Ilpo Järvinen90638a02008-09-20 21:25:52 -07003267 if (skb == tp->lost_skb_hint)
3268 tp->lost_skb_hint = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003269 }
3270
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07003271 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3272 tp->snd_up = tp->snd_una;
3273
Ilpo Järvinencadbd032007-12-31 04:49:21 -08003274 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3275 flag |= FLAG_SACK_RENEGING;
3276
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003277 if (flag & FLAG_ACKED) {
Stephen Hemminger164891a2007-04-23 22:26:16 -07003278 const struct tcp_congestion_ops *ca_ops
3279 = inet_csk(sk)->icsk_ca_ops;
3280
Ilpo Järvinen72211e92009-03-14 14:23:04 +00003281 if (unlikely(icsk->icsk_mtup.probe_size &&
3282 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3283 tcp_mtup_probe_success(sk);
3284 }
3285
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003286 tcp_ack_update_rtt(sk, flag, seq_rtt);
Ilpo Järvinen6728e7d2007-08-24 22:53:26 -07003287 tcp_rearm_rto(sk);
Stephen Hemminger317a76f2005-06-23 12:19:55 -07003288
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003289 if (tcp_is_reno(tp)) {
3290 tcp_remove_reno_sacks(sk, pkts_acked);
3291 } else {
Ilpo Järvinen59a08cb2009-02-28 04:44:28 +00003292 int delta;
3293
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003294 /* Non-retransmitted hole got filled? That's reordering */
3295 if (reord < prior_fackets)
3296 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
Ilpo Järvinen90638a02008-09-20 21:25:52 -07003297
Ilpo Järvinen59a08cb2009-02-28 04:44:28 +00003298 delta = tcp_is_fack(tp) ? pkts_acked :
3299 prior_sacked - tp->sacked_out;
3300 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003301 }
3302
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07003303 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08003304
Stephen Hemminger30cfd0b2007-07-25 23:49:34 -07003305 if (ca_ops->pkts_acked) {
3306 s32 rtt_us = -1;
Ilpo Järvinenb9ce2042007-06-15 15:08:43 -07003307
Stephen Hemminger30cfd0b2007-07-25 23:49:34 -07003308 /* Is the ACK triggering packet unambiguous? */
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003309 if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
Stephen Hemminger30cfd0b2007-07-25 23:49:34 -07003310 /* High resolution needed and available? */
3311 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
3312 !ktime_equal(last_ackt,
3313 net_invalid_timestamp()))
3314 rtt_us = ktime_us_delta(ktime_get_real(),
3315 last_ackt);
Gavin McCullagh2072c222007-12-29 19:11:21 -08003316 else if (ca_seq_rtt > 0)
3317 rtt_us = jiffies_to_usecs(ca_seq_rtt);
Stephen Hemminger30cfd0b2007-07-25 23:49:34 -07003318 }
3319
3320 ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
3321 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003322 }
3323
3324#if FASTRETRANS_DEBUG > 0
Ilpo Järvinen547b7922008-07-25 21:43:18 -07003325 WARN_ON((int)tp->sacked_out < 0);
3326 WARN_ON((int)tp->lost_out < 0);
3327 WARN_ON((int)tp->retrans_out < 0);
Ilpo Järvinene60402d2007-08-09 15:14:46 +03003328 if (!tp->packets_out && tcp_is_sack(tp)) {
Stephen Hemmingercfcabdc2007-10-09 01:59:42 -07003329 icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003330 if (tp->lost_out) {
3331 printk(KERN_DEBUG "Leak l=%u %d\n",
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003332 tp->lost_out, icsk->icsk_ca_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003333 tp->lost_out = 0;
3334 }
3335 if (tp->sacked_out) {
3336 printk(KERN_DEBUG "Leak s=%u %d\n",
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003337 tp->sacked_out, icsk->icsk_ca_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003338 tp->sacked_out = 0;
3339 }
3340 if (tp->retrans_out) {
3341 printk(KERN_DEBUG "Leak r=%u %d\n",
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003342 tp->retrans_out, icsk->icsk_ca_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003343 tp->retrans_out = 0;
3344 }
3345 }
3346#endif
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003347 return flag;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003348}
3349
3350static void tcp_ack_probe(struct sock *sk)
3351{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003352 const struct tcp_sock *tp = tcp_sk(sk);
3353 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003354
3355 /* Was it a usable window open? */
3356
Ilpo Järvinen90840de2007-12-31 04:48:41 -08003357 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003358 icsk->icsk_backoff = 0;
3359 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003360 /* Socket must be waked up by subsequent tcp_data_snd_check().
3361 * This function is not for random using!
3362 */
3363 } else {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003364 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07003365 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3366 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003367 }
3368}
3369
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003370static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003371{
3372 return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003373 inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003374}
3375
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003376static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003377{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003378 const struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003379 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003380 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003381}
3382
3383/* Check that window update is acceptable.
3384 * The function assumes that snd_una<=ack<=snd_next.
3385 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003386static inline int tcp_may_update_window(const struct tcp_sock *tp,
3387 const u32 ack, const u32 ack_seq,
3388 const u32 nwin)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003389{
3390 return (after(ack, tp->snd_una) ||
3391 after(ack_seq, tp->snd_wl1) ||
3392 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
3393}
3394
3395/* Update our send window.
3396 *
3397 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3398 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3399 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003400static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack,
3401 u32 ack_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003402{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003403 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003404 int flag = 0;
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07003405 u32 nwin = ntohs(tcp_hdr(skb)->window);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003406
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07003407 if (likely(!tcp_hdr(skb)->syn))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003408 nwin <<= tp->rx_opt.snd_wscale;
3409
3410 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3411 flag |= FLAG_WIN_UPDATE;
Hantzis Fotisee7537b2009-03-02 22:42:02 -08003412 tcp_update_wl(tp, ack_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003413
3414 if (tp->snd_wnd != nwin) {
3415 tp->snd_wnd = nwin;
3416
3417 /* Note, it is the only place, where
3418 * fast path is recovered for sending TCP.
3419 */
Herbert Xu2ad41062005-10-27 18:47:46 +10003420 tp->pred_flags = 0;
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003421 tcp_fast_path_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003422
3423 if (nwin > tp->max_window) {
3424 tp->max_window = nwin;
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08003425 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003426 }
3427 }
3428 }
3429
3430 tp->snd_una = ack;
3431
3432 return flag;
3433}
3434
Ilpo Järvinen9ead9a12007-02-21 22:56:19 -08003435/* A very conservative spurious RTO response algorithm: reduce cwnd and
3436 * continue in congestion avoidance.
3437 */
3438static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
3439{
3440 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
Ilpo Järvinenaa8b6a72007-02-21 23:06:03 -08003441 tp->snd_cwnd_cnt = 0;
Ilpo Järvinen16e90682007-10-11 17:32:31 -07003442 tp->bytes_acked = 0;
Ilpo Järvinen46323652007-05-19 13:56:23 -07003443 TCP_ECN_queue_cwr(tp);
Ilpo Järvinen9ead9a12007-02-21 22:56:19 -08003444 tcp_moderate_cwnd(tp);
3445}
3446
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -08003447/* A conservative spurious RTO response algorithm: reduce cwnd using
3448 * rate halving and continue in congestion avoidance.
3449 */
3450static void tcp_ratehalving_spur_to_response(struct sock *sk)
3451{
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -08003452 tcp_enter_cwr(sk, 0);
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -08003453}
3454
Ilpo Järvinene317f6f2007-03-02 13:34:19 -08003455static void tcp_undo_spur_to_response(struct sock *sk, int flag)
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -08003456{
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003457 if (flag & FLAG_ECE)
Ilpo Järvinene317f6f2007-03-02 13:34:19 -08003458 tcp_ratehalving_spur_to_response(sk);
3459 else
3460 tcp_undo_cwr(sk, 1);
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -08003461}
3462
Ilpo Järvinen30935cf2007-02-21 23:01:36 -08003463/* F-RTO spurious RTO detection algorithm (RFC4138)
3464 *
Ilpo Järvinen6408d202007-02-21 23:04:11 -08003465 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3466 * comments). State (ACK number) is kept in frto_counter. When ACK advances
3467 * window (but not to or beyond highest sequence sent before RTO):
Ilpo Järvinen30935cf2007-02-21 23:01:36 -08003468 * On First ACK, send two new segments out.
3469 * On Second ACK, RTO was likely spurious. Do spurious response (response
3470 * algorithm is not part of the F-RTO detection algorithm
3471 * given in RFC4138 but can be selected separately).
3472 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
Ilpo Järvinend551e452007-04-30 00:42:20 -07003473 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3474 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3475 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
Ilpo Järvinen30935cf2007-02-21 23:01:36 -08003476 *
3477 * Rationale: if the RTO was spurious, new ACKs should arrive from the
3478 * original window even after we transmit two new data segments.
3479 *
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08003480 * SACK version:
3481 * on first step, wait until first cumulative ACK arrives, then move to
3482 * the second step. In second step, the next ACK decides.
3483 *
Ilpo Järvinen30935cf2007-02-21 23:01:36 -08003484 * F-RTO is implemented (mainly) in four functions:
3485 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
3486 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3487 * called when tcp_use_frto() showed green light
3488 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3489 * - tcp_enter_frto_loss() is called if there is not enough evidence
3490 * to prove that the RTO is indeed spurious. It transfers the control
3491 * from F-RTO to the conventional RTO recovery
3492 */
Ilpo Järvinen2e605292007-08-02 19:46:58 -07003493static int tcp_process_frto(struct sock *sk, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003494{
3495 struct tcp_sock *tp = tcp_sk(sk);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09003496
Ilpo Järvinen005903b2007-08-09 14:44:16 +03003497 tcp_verify_left_out(tp);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09003498
Ilpo Järvinen7487c482007-02-21 23:02:30 -08003499 /* Duplicate the behavior from Loss state (fastretrans_alert) */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003500 if (flag & FLAG_DATA_ACKED)
Ilpo Järvinen7487c482007-02-21 23:02:30 -08003501 inet_csk(sk)->icsk_retransmits = 0;
3502
Ilpo Järvinen009a2e32007-09-20 11:34:38 -07003503 if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||
3504 ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
3505 tp->undo_marker = 0;
3506
Ilpo Järvinen95c49222007-02-21 23:05:18 -08003507 if (!before(tp->snd_una, tp->frto_highmark)) {
Ilpo Järvinend551e452007-04-30 00:42:20 -07003508 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
Ilpo Järvinen7c9a4a52007-02-21 23:08:34 -08003509 return 1;
Ilpo Järvinen95c49222007-02-21 23:05:18 -08003510 }
3511
Ilpo Järvinen62ab2222008-05-08 01:09:11 -07003512 if (!tcp_is_sackfrto(tp)) {
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08003513 /* RFC4138 shortcoming in step 2; should also have case c):
3514 * ACK isn't duplicate nor advances window, e.g., opposite dir
3515 * data, winupdate
3516 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003517 if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP))
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08003518 return 1;
Ilpo Järvinen6408d202007-02-21 23:04:11 -08003519
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003520 if (!(flag & FLAG_DATA_ACKED)) {
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08003521 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
3522 flag);
3523 return 1;
3524 }
3525 } else {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003526 if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08003527 /* Prevent sending of new data. */
3528 tp->snd_cwnd = min(tp->snd_cwnd,
3529 tcp_packets_in_flight(tp));
3530 return 1;
3531 }
3532
Ilpo Järvinend551e452007-04-30 00:42:20 -07003533 if ((tp->frto_counter >= 2) &&
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003534 (!(flag & FLAG_FORWARD_PROGRESS) ||
3535 ((flag & FLAG_DATA_SACKED) &&
3536 !(flag & FLAG_ONLY_ORIG_SACKED)))) {
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08003537 /* RFC4138 shortcoming (see comment above) */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003538 if (!(flag & FLAG_FORWARD_PROGRESS) &&
3539 (flag & FLAG_NOT_DUP))
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08003540 return 1;
3541
3542 tcp_enter_frto_loss(sk, 3, flag);
3543 return 1;
3544 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003545 }
3546
3547 if (tp->frto_counter == 1) {
Ilpo Järvinen3e6f0492007-12-05 02:20:21 -08003548 /* tcp_may_send_now needs to see updated state */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003549 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
Ilpo Järvinen94d0ea72007-02-21 23:07:27 -08003550 tp->frto_counter = 2;
Ilpo Järvinen3e6f0492007-12-05 02:20:21 -08003551
3552 if (!tcp_may_send_now(sk))
3553 tcp_enter_frto_loss(sk, 2, flag);
3554
Ilpo Järvinen7c9a4a52007-02-21 23:08:34 -08003555 return 1;
Ilpo Järvinend551e452007-04-30 00:42:20 -07003556 } else {
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -08003557 switch (sysctl_tcp_frto_response) {
3558 case 2:
Ilpo Järvinene317f6f2007-03-02 13:34:19 -08003559 tcp_undo_spur_to_response(sk, flag);
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -08003560 break;
3561 case 1:
3562 tcp_conservative_spur_to_response(tp);
3563 break;
3564 default:
3565 tcp_ratehalving_spur_to_response(sk);
3566 break;
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07003567 }
Ilpo Järvinen94d0ea72007-02-21 23:07:27 -08003568 tp->frto_counter = 0;
Ilpo Järvinen009a2e32007-09-20 11:34:38 -07003569 tp->undo_marker = 0;
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07003570 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSPURIOUSRTOS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003571 }
Ilpo Järvinen7c9a4a52007-02-21 23:08:34 -08003572 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003573}
3574
Linus Torvalds1da177e2005-04-16 15:20:36 -07003575/* This routine deals with incoming acks, but not outgoing ones. */
3576static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
3577{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003578 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003579 struct tcp_sock *tp = tcp_sk(sk);
3580 u32 prior_snd_una = tp->snd_una;
3581 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3582 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3583 u32 prior_in_flight;
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003584 u32 prior_fackets;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003585 int prior_packets;
Ilpo Järvinen7c9a4a52007-02-21 23:08:34 -08003586 int frto_cwnd = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003587
John Dykstra96e0bf42009-03-22 21:49:57 -07003588 /* If the ack is older than previous acks
Linus Torvalds1da177e2005-04-16 15:20:36 -07003589 * then we can probably ignore it.
3590 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003591 if (before(ack, prior_snd_una))
3592 goto old_ack;
3593
John Dykstra96e0bf42009-03-22 21:49:57 -07003594 /* If the ack includes data we haven't sent yet, discard
3595 * this segment (RFC793 Section 3.9).
3596 */
3597 if (after(ack, tp->snd_nxt))
3598 goto invalid_ack;
3599
Ilpo Järvinen2e605292007-08-02 19:46:58 -07003600 if (after(ack, prior_snd_una))
3601 flag |= FLAG_SND_UNA_ADVANCED;
3602
Daikichi Osuga3fdf3f02006-08-29 02:01:44 -07003603 if (sysctl_tcp_abc) {
3604 if (icsk->icsk_ca_state < TCP_CA_CWR)
3605 tp->bytes_acked += ack - prior_snd_una;
3606 else if (icsk->icsk_ca_state == TCP_CA_Loss)
3607 /* we assume just one segment left network */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003608 tp->bytes_acked += min(ack - prior_snd_una,
3609 tp->mss_cache);
Daikichi Osuga3fdf3f02006-08-29 02:01:44 -07003610 }
Stephen Hemminger9772efb2005-11-10 17:09:53 -08003611
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003612 prior_fackets = tp->fackets_out;
Ilpo Järvinen52d34082007-12-05 02:21:35 -08003613 prior_in_flight = tcp_packets_in_flight(tp);
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003614
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003615 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003616 /* Window is constant, pure forward advance.
3617 * No more checks are required.
3618 * Note, we use the fact that SND.UNA>=SND.WL2.
3619 */
Hantzis Fotisee7537b2009-03-02 22:42:02 -08003620 tcp_update_wl(tp, ack_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003621 tp->snd_una = ack;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003622 flag |= FLAG_WIN_UPDATE;
3623
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003624 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
Stephen Hemminger317a76f2005-06-23 12:19:55 -07003625
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07003626 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003627 } else {
3628 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3629 flag |= FLAG_DATA;
3630 else
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07003631 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003632
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003633 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003634
3635 if (TCP_SKB_CB(skb)->sacked)
3636 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3637
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07003638 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003639 flag |= FLAG_ECE;
3640
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003641 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003642 }
3643
3644 /* We passed data and got it acked, remove any soft error
3645 * log. Something worked...
3646 */
3647 sk->sk_err_soft = 0;
David S. Miller4b53fb62008-07-23 16:38:45 -07003648 icsk->icsk_probes_out = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003649 tp->rcv_tstamp = tcp_time_stamp;
3650 prior_packets = tp->packets_out;
3651 if (!prior_packets)
3652 goto no_queue;
3653
Linus Torvalds1da177e2005-04-16 15:20:36 -07003654 /* See if we can take anything off of the retransmit queue. */
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07003655 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003656
Ilpo Järvinene1cd8f72007-11-14 15:55:09 -08003657 if (tp->frto_counter)
3658 frto_cwnd = tcp_process_frto(sk, flag);
Ilpo Järvinen3de96472007-10-01 15:28:48 -07003659 /* Guarantee sacktag reordering detection against wrap-arounds */
3660 if (before(tp->frto_highmark, tp->snd_una))
3661 tp->frto_highmark = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003662
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003663 if (tcp_ack_is_dubious(sk, flag)) {
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08003664 /* Advance CWND, if state allows this. */
Ilpo Järvinen7c9a4a52007-02-21 23:08:34 -08003665 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
3666 tcp_may_raise_cwnd(sk, flag))
Ilpo Järvinenc3a05c62007-12-02 00:47:59 +02003667 tcp_cong_avoid(sk, ack, prior_in_flight);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003668 tcp_fastretrans_alert(sk, prior_packets - tp->packets_out,
3669 flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003670 } else {
Ilpo Järvinen7c9a4a52007-02-21 23:08:34 -08003671 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
Ilpo Järvinenc3a05c62007-12-02 00:47:59 +02003672 tcp_cong_avoid(sk, ack, prior_in_flight);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003673 }
3674
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003675 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003676 dst_confirm(sk->sk_dst_cache);
3677
3678 return 1;
3679
3680no_queue:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003681 /* If this ack opens up a zero window, clear backoff. It was
3682 * being used to time the probes, and is probably far higher than
3683 * it needs to be for normal retransmission.
3684 */
David S. Millerfe067e82007-03-07 12:12:44 -08003685 if (tcp_send_head(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003686 tcp_ack_probe(sk);
3687 return 1;
3688
John Dykstra96e0bf42009-03-22 21:49:57 -07003689invalid_ack:
3690 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3691 return -1;
3692
Linus Torvalds1da177e2005-04-16 15:20:36 -07003693old_ack:
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07003694 if (TCP_SKB_CB(skb)->sacked) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003695 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07003696 if (icsk->icsk_ca_state == TCP_CA_Open)
3697 tcp_try_keep_open(sk);
3698 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003699
John Dykstra96e0bf42009-03-22 21:49:57 -07003700 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003701 return 0;
3702}
3703
Linus Torvalds1da177e2005-04-16 15:20:36 -07003704/* Look for tcp options. Normally only called on SYN and SYNACK packets.
3705 * But, this can also be called on packets in the established flow when
3706 * the fast version below fails.
3707 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003708void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx,
3709 int estab)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003710{
3711 unsigned char *ptr;
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07003712 struct tcphdr *th = tcp_hdr(skb);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003713 int length = (th->doff * 4) - sizeof(struct tcphdr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003714
3715 ptr = (unsigned char *)(th + 1);
3716 opt_rx->saw_tstamp = 0;
3717
Stephen Hemminger2de979b2007-03-08 20:45:19 -08003718 while (length > 0) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003719 int opcode = *ptr++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003720 int opsize;
3721
3722 switch (opcode) {
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003723 case TCPOPT_EOL:
3724 return;
3725 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3726 length--;
3727 continue;
3728 default:
3729 opsize = *ptr++;
3730 if (opsize < 2) /* "silly options" */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003731 return;
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003732 if (opsize > length)
3733 return; /* don't parse partial options */
3734 switch (opcode) {
3735 case TCPOPT_MSS:
3736 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07003737 u16 in_mss = get_unaligned_be16(ptr);
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003738 if (in_mss) {
3739 if (opt_rx->user_mss &&
3740 opt_rx->user_mss < in_mss)
3741 in_mss = opt_rx->user_mss;
3742 opt_rx->mss_clamp = in_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003743 }
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07003744 }
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003745 break;
3746 case TCPOPT_WINDOW:
3747 if (opsize == TCPOLEN_WINDOW && th->syn &&
3748 !estab && sysctl_tcp_window_scaling) {
3749 __u8 snd_wscale = *(__u8 *)ptr;
3750 opt_rx->wscale_ok = 1;
3751 if (snd_wscale > 14) {
3752 if (net_ratelimit())
3753 printk(KERN_INFO "tcp_parse_options: Illegal window "
3754 "scaling value %d >14 received.\n",
3755 snd_wscale);
3756 snd_wscale = 14;
3757 }
3758 opt_rx->snd_wscale = snd_wscale;
3759 }
3760 break;
3761 case TCPOPT_TIMESTAMP:
3762 if ((opsize == TCPOLEN_TIMESTAMP) &&
3763 ((estab && opt_rx->tstamp_ok) ||
3764 (!estab && sysctl_tcp_timestamps))) {
3765 opt_rx->saw_tstamp = 1;
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07003766 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3767 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003768 }
3769 break;
3770 case TCPOPT_SACK_PERM:
3771 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3772 !estab && sysctl_tcp_sack) {
3773 opt_rx->sack_ok = 1;
3774 tcp_sack_reset(opt_rx);
3775 }
3776 break;
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07003777
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003778 case TCPOPT_SACK:
3779 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3780 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3781 opt_rx->sack_ok) {
3782 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3783 }
3784 break;
3785#ifdef CONFIG_TCP_MD5SIG
3786 case TCPOPT_MD5SIG:
3787 /*
3788 * The MD5 Hash has already been
3789 * checked (see tcp_v{4,6}_do_rcv()).
3790 */
3791 break;
3792#endif
3793 }
3794
3795 ptr += opsize-2;
3796 length -= opsize;
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07003797 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003798 }
3799}
3800
Ilpo Järvinena4356b22008-08-23 05:12:29 -07003801static int tcp_parse_aligned_timestamp(struct tcp_sock *tp, struct tcphdr *th)
3802{
3803 __be32 *ptr = (__be32 *)(th + 1);
3804
3805 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3806 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3807 tp->rx_opt.saw_tstamp = 1;
3808 ++ptr;
3809 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3810 ++ptr;
3811 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3812 return 1;
3813 }
3814 return 0;
3815}
3816
Linus Torvalds1da177e2005-04-16 15:20:36 -07003817/* Fast parse options. This hopes to only see timestamps.
3818 * If it is wrong it falls back on tcp_parse_options().
3819 */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -08003820static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
3821 struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003822{
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003823 if (th->doff == sizeof(struct tcphdr) >> 2) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003824 tp->rx_opt.saw_tstamp = 0;
3825 return 0;
3826 } else if (tp->rx_opt.tstamp_ok &&
3827 th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
Ilpo Järvinena4356b22008-08-23 05:12:29 -07003828 if (tcp_parse_aligned_timestamp(tp, th))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003829 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003830 }
3831 tcp_parse_options(skb, &tp->rx_opt, 1);
3832 return 1;
3833}
3834
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003835#ifdef CONFIG_TCP_MD5SIG
3836/*
3837 * Parse MD5 Signature option
3838 */
3839u8 *tcp_parse_md5sig_option(struct tcphdr *th)
3840{
3841 int length = (th->doff << 2) - sizeof (*th);
3842 u8 *ptr = (u8*)(th + 1);
3843
3844 /* If the TCP option is too short, we can short cut */
3845 if (length < TCPOLEN_MD5SIG)
3846 return NULL;
3847
3848 while (length > 0) {
3849 int opcode = *ptr++;
3850 int opsize;
3851
3852 switch(opcode) {
3853 case TCPOPT_EOL:
3854 return NULL;
3855 case TCPOPT_NOP:
3856 length--;
3857 continue;
3858 default:
3859 opsize = *ptr++;
3860 if (opsize < 2 || opsize > length)
3861 return NULL;
3862 if (opcode == TCPOPT_MD5SIG)
3863 return ptr;
3864 }
3865 ptr += opsize - 2;
3866 length -= opsize;
3867 }
3868 return NULL;
3869}
3870#endif
3871
Linus Torvalds1da177e2005-04-16 15:20:36 -07003872static inline void tcp_store_ts_recent(struct tcp_sock *tp)
3873{
3874 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
James Morris9d729f72007-03-04 16:12:44 -08003875 tp->rx_opt.ts_recent_stamp = get_seconds();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003876}
3877
3878static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3879{
3880 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3881 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3882 * extra check below makes sure this can only happen
3883 * for pure ACK frames. -DaveM
3884 *
3885 * Not only, also it occurs for expired timestamps.
3886 */
3887
Ilpo Järvinenc887e6d2009-03-14 14:23:03 +00003888 if (tcp_paws_check(&tp->rx_opt, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003889 tcp_store_ts_recent(tp);
3890 }
3891}
3892
3893/* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3894 *
3895 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3896 * it can pass through stack. So, the following predicate verifies that
3897 * this segment is not used for anything but congestion avoidance or
3898 * fast retransmit. Moreover, we even are able to eliminate most of such
3899 * second order effects, if we apply some small "replay" window (~RTO)
3900 * to timestamp space.
3901 *
3902 * All these measures still do not guarantee that we reject wrapped ACKs
3903 * on networks with high bandwidth, when sequence space is recycled fastly,
3904 * but it guarantees that such events will be very rare and do not affect
3905 * connection seriously. This doesn't look nice, but alas, PAWS is really
3906 * buggy extension.
3907 *
3908 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3909 * states that events when retransmit arrives after original data are rare.
3910 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3911 * the biggest problem on large power networks even with minor reordering.
3912 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3913 * up to bandwidth of 18Gigabit/sec. 8) ]
3914 */
3915
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003916static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003917{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003918 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07003919 struct tcphdr *th = tcp_hdr(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003920 u32 seq = TCP_SKB_CB(skb)->seq;
3921 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3922
3923 return (/* 1. Pure ACK with correct sequence number. */
3924 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3925
3926 /* 2. ... and duplicate ACK. */
3927 ack == tp->snd_una &&
3928
3929 /* 3. ... and does not update window. */
3930 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3931
3932 /* 4. ... and sits in replay window. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003933 (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 -07003934}
3935
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003936static inline int tcp_paws_discard(const struct sock *sk,
3937 const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003938{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003939 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinenc887e6d2009-03-14 14:23:03 +00003940
3941 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
3942 !tcp_disordered_ack(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003943}
3944
3945/* Check segment sequence number for validity.
3946 *
3947 * Segment controls are considered valid, if the segment
3948 * fits to the window after truncation to the window. Acceptability
3949 * of data (and SYN, FIN, of course) is checked separately.
3950 * See tcp_data_queue(), for example.
3951 *
3952 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3953 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3954 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3955 * (borrowed from freebsd)
3956 */
3957
3958static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
3959{
3960 return !before(end_seq, tp->rcv_wup) &&
3961 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3962}
3963
3964/* When we get a reset we do this. */
3965static void tcp_reset(struct sock *sk)
3966{
3967 /* We want the right error as BSD sees it (and indeed as we do). */
3968 switch (sk->sk_state) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003969 case TCP_SYN_SENT:
3970 sk->sk_err = ECONNREFUSED;
3971 break;
3972 case TCP_CLOSE_WAIT:
3973 sk->sk_err = EPIPE;
3974 break;
3975 case TCP_CLOSE:
3976 return;
3977 default:
3978 sk->sk_err = ECONNRESET;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003979 }
3980
3981 if (!sock_flag(sk, SOCK_DEAD))
3982 sk->sk_error_report(sk);
3983
3984 tcp_done(sk);
3985}
3986
3987/*
3988 * Process the FIN bit. This now behaves as it is supposed to work
3989 * and the FIN takes effect when it is validly part of sequence
3990 * space. Not before when we get holes.
3991 *
3992 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3993 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3994 * TIME-WAIT)
3995 *
3996 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3997 * close and we go into CLOSING (and later onto TIME-WAIT)
3998 *
3999 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4000 */
4001static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
4002{
4003 struct tcp_sock *tp = tcp_sk(sk);
4004
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004005 inet_csk_schedule_ack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004006
4007 sk->sk_shutdown |= RCV_SHUTDOWN;
4008 sock_set_flag(sk, SOCK_DONE);
4009
4010 switch (sk->sk_state) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004011 case TCP_SYN_RECV:
4012 case TCP_ESTABLISHED:
4013 /* Move to CLOSE_WAIT */
4014 tcp_set_state(sk, TCP_CLOSE_WAIT);
4015 inet_csk(sk)->icsk_ack.pingpong = 1;
4016 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004017
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004018 case TCP_CLOSE_WAIT:
4019 case TCP_CLOSING:
4020 /* Received a retransmission of the FIN, do
4021 * nothing.
4022 */
4023 break;
4024 case TCP_LAST_ACK:
4025 /* RFC793: Remain in the LAST-ACK state. */
4026 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004027
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004028 case TCP_FIN_WAIT1:
4029 /* This case occurs when a simultaneous close
4030 * happens, we must ack the received FIN and
4031 * enter the CLOSING state.
4032 */
4033 tcp_send_ack(sk);
4034 tcp_set_state(sk, TCP_CLOSING);
4035 break;
4036 case TCP_FIN_WAIT2:
4037 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4038 tcp_send_ack(sk);
4039 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4040 break;
4041 default:
4042 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4043 * cases we should never reach this piece of code.
4044 */
4045 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
Harvey Harrison0dc47872008-03-05 20:47:47 -08004046 __func__, sk->sk_state);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004047 break;
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07004048 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004049
4050 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4051 * Probably, we should reset in this case. For now drop them.
4052 */
4053 __skb_queue_purge(&tp->out_of_order_queue);
Ilpo Järvinene60402d2007-08-09 15:14:46 +03004054 if (tcp_is_sack(tp))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004055 tcp_sack_reset(&tp->rx_opt);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08004056 sk_mem_reclaim(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004057
4058 if (!sock_flag(sk, SOCK_DEAD)) {
4059 sk->sk_state_change(sk);
4060
4061 /* Do not send POLL_HUP for half duplex close. */
4062 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4063 sk->sk_state == TCP_CLOSE)
Pavel Emelyanov8d8ad9d2007-11-26 20:10:50 +08004064 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004065 else
Pavel Emelyanov8d8ad9d2007-11-26 20:10:50 +08004066 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004067 }
4068}
4069
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004070static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4071 u32 end_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004072{
4073 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4074 if (before(seq, sp->start_seq))
4075 sp->start_seq = seq;
4076 if (after(end_seq, sp->end_seq))
4077 sp->end_seq = end_seq;
4078 return 1;
4079 }
4080 return 0;
4081}
4082
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004083static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004084{
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004085 struct tcp_sock *tp = tcp_sk(sk);
4086
Ilpo Järvinene60402d2007-08-09 15:14:46 +03004087 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07004088 int mib_idx;
4089
Linus Torvalds1da177e2005-04-16 15:20:36 -07004090 if (before(seq, tp->rcv_nxt))
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07004091 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004092 else
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07004093 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4094
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07004095 NET_INC_STATS_BH(sock_net(sk), mib_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004096
4097 tp->rx_opt.dsack = 1;
4098 tp->duplicate_sack[0].start_seq = seq;
4099 tp->duplicate_sack[0].end_seq = end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004100 }
4101}
4102
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004103static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004104{
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004105 struct tcp_sock *tp = tcp_sk(sk);
4106
Linus Torvalds1da177e2005-04-16 15:20:36 -07004107 if (!tp->rx_opt.dsack)
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004108 tcp_dsack_set(sk, seq, end_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004109 else
4110 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4111}
4112
4113static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
4114{
4115 struct tcp_sock *tp = tcp_sk(sk);
4116
4117 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4118 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07004119 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004120 tcp_enter_quickack_mode(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004121
Ilpo Järvinene60402d2007-08-09 15:14:46 +03004122 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004123 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4124
4125 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4126 end_seq = tp->rcv_nxt;
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004127 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004128 }
4129 }
4130
4131 tcp_send_ack(sk);
4132}
4133
4134/* These routines update the SACK block as out-of-order packets arrive or
4135 * in-order packets close up the sequence space.
4136 */
4137static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4138{
4139 int this_sack;
4140 struct tcp_sack_block *sp = &tp->selective_acks[0];
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004141 struct tcp_sack_block *swalk = sp + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004142
4143 /* See if the recent change to the first SACK eats into
4144 * or hits the sequence space of other SACK blocks, if so coalesce.
4145 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004146 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004147 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4148 int i;
4149
4150 /* Zap SWALK, by moving every further SACK up by one slot.
4151 * Decrease num_sacks.
4152 */
4153 tp->rx_opt.num_sacks--;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004154 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4155 sp[i] = sp[i + 1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07004156 continue;
4157 }
4158 this_sack++, swalk++;
4159 }
4160}
4161
Linus Torvalds1da177e2005-04-16 15:20:36 -07004162static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4163{
4164 struct tcp_sock *tp = tcp_sk(sk);
4165 struct tcp_sack_block *sp = &tp->selective_acks[0];
4166 int cur_sacks = tp->rx_opt.num_sacks;
4167 int this_sack;
4168
4169 if (!cur_sacks)
4170 goto new_sack;
4171
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004172 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004173 if (tcp_sack_extend(sp, seq, end_seq)) {
4174 /* Rotate this_sack to the first one. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004175 for (; this_sack > 0; this_sack--, sp--)
Ilpo Järvinena0bffff2009-03-21 13:36:17 -07004176 swap(*sp, *(sp - 1));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004177 if (cur_sacks > 1)
4178 tcp_sack_maybe_coalesce(tp);
4179 return;
4180 }
4181 }
4182
4183 /* Could not find an adjacent existing SACK, build a new one,
4184 * put it at the front, and shift everyone else down. We
4185 * always know there is at least one SACK present already here.
4186 *
4187 * If the sack array is full, forget about the last one.
4188 */
Adam Langley4389dde2008-07-19 00:07:02 -07004189 if (this_sack >= TCP_NUM_SACKS) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004190 this_sack--;
4191 tp->rx_opt.num_sacks--;
4192 sp--;
4193 }
Stephen Hemminger2de979b2007-03-08 20:45:19 -08004194 for (; this_sack > 0; this_sack--, sp--)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004195 *sp = *(sp - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004196
4197new_sack:
4198 /* Build the new head SACK, and we're done. */
4199 sp->start_seq = seq;
4200 sp->end_seq = end_seq;
4201 tp->rx_opt.num_sacks++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004202}
4203
4204/* RCV.NXT advances, some SACKs should be eaten. */
4205
4206static void tcp_sack_remove(struct tcp_sock *tp)
4207{
4208 struct tcp_sack_block *sp = &tp->selective_acks[0];
4209 int num_sacks = tp->rx_opt.num_sacks;
4210 int this_sack;
4211
4212 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
David S. Millerb03efcf2005-07-08 14:57:23 -07004213 if (skb_queue_empty(&tp->out_of_order_queue)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004214 tp->rx_opt.num_sacks = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004215 return;
4216 }
4217
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004218 for (this_sack = 0; this_sack < num_sacks;) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004219 /* Check if the start of the sack is covered by RCV.NXT. */
4220 if (!before(tp->rcv_nxt, sp->start_seq)) {
4221 int i;
4222
4223 /* RCV.NXT must cover all the block! */
Ilpo Järvinen547b7922008-07-25 21:43:18 -07004224 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004225
4226 /* Zap this SACK, by moving forward any other SACKS. */
4227 for (i=this_sack+1; i < num_sacks; i++)
4228 tp->selective_acks[i-1] = tp->selective_acks[i];
4229 num_sacks--;
4230 continue;
4231 }
4232 this_sack++;
4233 sp++;
4234 }
Ilpo Järvinen5861f8e2009-03-14 14:23:01 +00004235 tp->rx_opt.num_sacks = num_sacks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004236}
4237
4238/* This one checks to see if we can put data from the
4239 * out_of_order queue into the receive_queue.
4240 */
4241static void tcp_ofo_queue(struct sock *sk)
4242{
4243 struct tcp_sock *tp = tcp_sk(sk);
4244 __u32 dsack_high = tp->rcv_nxt;
4245 struct sk_buff *skb;
4246
4247 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4248 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4249 break;
4250
4251 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4252 __u32 dsack = dsack_high;
4253 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4254 dsack_high = TCP_SKB_CB(skb)->end_seq;
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004255 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004256 }
4257
4258 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4259 SOCK_DEBUG(sk, "ofo packet was already received \n");
David S. Miller8728b832005-08-09 19:25:21 -07004260 __skb_unlink(skb, &tp->out_of_order_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004261 __kfree_skb(skb);
4262 continue;
4263 }
4264 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4265 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4266 TCP_SKB_CB(skb)->end_seq);
4267
David S. Miller8728b832005-08-09 19:25:21 -07004268 __skb_unlink(skb, &tp->out_of_order_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004269 __skb_queue_tail(&sk->sk_receive_queue, skb);
4270 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07004271 if (tcp_hdr(skb)->fin)
4272 tcp_fin(skb, sk, tcp_hdr(skb));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004273 }
4274}
4275
Vitaliy Gusev56f367b2008-04-15 20:26:34 -07004276static int tcp_prune_ofo_queue(struct sock *sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004277static int tcp_prune_queue(struct sock *sk);
4278
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004279static inline int tcp_try_rmem_schedule(struct sock *sk, unsigned int size)
4280{
4281 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4282 !sk_rmem_schedule(sk, size)) {
4283
4284 if (tcp_prune_queue(sk) < 0)
4285 return -1;
4286
4287 if (!sk_rmem_schedule(sk, size)) {
Vitaliy Gusev56f367b2008-04-15 20:26:34 -07004288 if (!tcp_prune_ofo_queue(sk))
4289 return -1;
4290
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004291 if (!sk_rmem_schedule(sk, size))
4292 return -1;
4293 }
4294 }
4295 return 0;
4296}
4297
Linus Torvalds1da177e2005-04-16 15:20:36 -07004298static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4299{
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07004300 struct tcphdr *th = tcp_hdr(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004301 struct tcp_sock *tp = tcp_sk(sk);
4302 int eaten = -1;
4303
4304 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4305 goto drop;
4306
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004307 __skb_pull(skb, th->doff * 4);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004308
4309 TCP_ECN_accept_cwr(tp, skb);
4310
Ilpo Järvinen5861f8e2009-03-14 14:23:01 +00004311 tp->rx_opt.dsack = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004312
4313 /* Queue data for delivery to the user.
4314 * Packets in sequence go to the receive queue.
4315 * Out of sequence packets to the out_of_order_queue.
4316 */
4317 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4318 if (tcp_receive_window(tp) == 0)
4319 goto out_of_window;
4320
4321 /* Ok. In sequence. In window. */
4322 if (tp->ucopy.task == current &&
4323 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4324 sock_owned_by_user(sk) && !tp->urg_data) {
4325 int chunk = min_t(unsigned int, skb->len,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004326 tp->ucopy.len);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004327
4328 __set_current_state(TASK_RUNNING);
4329
4330 local_bh_enable();
4331 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
4332 tp->ucopy.len -= chunk;
4333 tp->copied_seq += chunk;
4334 eaten = (chunk == skb->len && !th->fin);
4335 tcp_rcv_space_adjust(sk);
4336 }
4337 local_bh_disable();
4338 }
4339
4340 if (eaten <= 0) {
4341queue_and_out:
4342 if (eaten < 0 &&
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004343 tcp_try_rmem_schedule(sk, skb->truesize))
4344 goto drop;
4345
Hideo Aoki3ab224b2007-12-31 00:11:19 -08004346 skb_set_owner_r(skb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004347 __skb_queue_tail(&sk->sk_receive_queue, skb);
4348 }
4349 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
Stephen Hemminger2de979b2007-03-08 20:45:19 -08004350 if (skb->len)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004351 tcp_event_data_recv(sk, skb);
Stephen Hemminger2de979b2007-03-08 20:45:19 -08004352 if (th->fin)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004353 tcp_fin(skb, sk, th);
4354
David S. Millerb03efcf2005-07-08 14:57:23 -07004355 if (!skb_queue_empty(&tp->out_of_order_queue)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004356 tcp_ofo_queue(sk);
4357
4358 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4359 * gap in queue is filled.
4360 */
David S. Millerb03efcf2005-07-08 14:57:23 -07004361 if (skb_queue_empty(&tp->out_of_order_queue))
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004362 inet_csk(sk)->icsk_ack.pingpong = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004363 }
4364
4365 if (tp->rx_opt.num_sacks)
4366 tcp_sack_remove(tp);
4367
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004368 tcp_fast_path_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004369
4370 if (eaten > 0)
4371 __kfree_skb(skb);
4372 else if (!sock_flag(sk, SOCK_DEAD))
4373 sk->sk_data_ready(sk, 0);
4374 return;
4375 }
4376
4377 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4378 /* A retransmit, 2nd most common case. Force an immediate ack. */
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07004379 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004380 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004381
4382out_of_window:
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004383 tcp_enter_quickack_mode(sk);
4384 inet_csk_schedule_ack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004385drop:
4386 __kfree_skb(skb);
4387 return;
4388 }
4389
4390 /* Out of window. F.e. zero window probe. */
4391 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4392 goto out_of_window;
4393
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004394 tcp_enter_quickack_mode(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004395
4396 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4397 /* Partial packet, seq < rcv_next < end_seq */
4398 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4399 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4400 TCP_SKB_CB(skb)->end_seq);
4401
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004402 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004403
Linus Torvalds1da177e2005-04-16 15:20:36 -07004404 /* If window is closed, drop tail of packet. But after
4405 * remembering D-SACK for its head made in previous line.
4406 */
4407 if (!tcp_receive_window(tp))
4408 goto out_of_window;
4409 goto queue_and_out;
4410 }
4411
4412 TCP_ECN_check_ce(tp, skb);
4413
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004414 if (tcp_try_rmem_schedule(sk, skb->truesize))
4415 goto drop;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004416
4417 /* Disable header prediction. */
4418 tp->pred_flags = 0;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004419 inet_csk_schedule_ack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004420
4421 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4422 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4423
Hideo Aoki3ab224b2007-12-31 00:11:19 -08004424 skb_set_owner_r(skb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004425
4426 if (!skb_peek(&tp->out_of_order_queue)) {
4427 /* Initial out of order segment, build 1 SACK. */
Ilpo Järvinene60402d2007-08-09 15:14:46 +03004428 if (tcp_is_sack(tp)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004429 tp->rx_opt.num_sacks = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004430 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4431 tp->selective_acks[0].end_seq =
4432 TCP_SKB_CB(skb)->end_seq;
4433 }
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004434 __skb_queue_head(&tp->out_of_order_queue, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004435 } else {
4436 struct sk_buff *skb1 = tp->out_of_order_queue.prev;
4437 u32 seq = TCP_SKB_CB(skb)->seq;
4438 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4439
4440 if (seq == TCP_SKB_CB(skb1)->end_seq) {
Gerrit Renker7de6c032008-04-14 00:05:09 -07004441 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004442
4443 if (!tp->rx_opt.num_sacks ||
4444 tp->selective_acks[0].end_seq != seq)
4445 goto add_sack;
4446
4447 /* Common case: data arrive in order after hole. */
4448 tp->selective_acks[0].end_seq = end_seq;
4449 return;
4450 }
4451
4452 /* Find place to insert this segment. */
4453 do {
4454 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4455 break;
4456 } while ((skb1 = skb1->prev) !=
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004457 (struct sk_buff *)&tp->out_of_order_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004458
4459 /* Do skb overlap to previous one? */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004460 if (skb1 != (struct sk_buff *)&tp->out_of_order_queue &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07004461 before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4462 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4463 /* All the bits are present. Drop. */
4464 __kfree_skb(skb);
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004465 tcp_dsack_set(sk, seq, end_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004466 goto add_sack;
4467 }
4468 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4469 /* Partial overlap. */
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004470 tcp_dsack_set(sk, seq,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004471 TCP_SKB_CB(skb1)->end_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004472 } else {
4473 skb1 = skb1->prev;
4474 }
4475 }
David S. Miller43f59c82008-09-21 21:28:51 -07004476 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004477
Linus Torvalds1da177e2005-04-16 15:20:36 -07004478 /* And clean segments covered by new one as whole. */
4479 while ((skb1 = skb->next) !=
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004480 (struct sk_buff *)&tp->out_of_order_queue &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07004481 after(end_seq, TCP_SKB_CB(skb1)->seq)) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004482 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004483 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004484 end_seq);
4485 break;
4486 }
4487 __skb_unlink(skb1, &tp->out_of_order_queue);
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004488 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004489 TCP_SKB_CB(skb1)->end_seq);
4490 __kfree_skb(skb1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004491 }
4492
4493add_sack:
Ilpo Järvinene60402d2007-08-09 15:14:46 +03004494 if (tcp_is_sack(tp))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004495 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4496 }
4497}
4498
Ilpo Järvinen2cf46632008-08-23 05:11:41 -07004499static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4500 struct sk_buff_head *list)
4501{
4502 struct sk_buff *next = skb->next;
4503
4504 __skb_unlink(skb, list);
4505 __kfree_skb(skb);
4506 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4507
4508 return next;
4509}
4510
Linus Torvalds1da177e2005-04-16 15:20:36 -07004511/* Collapse contiguous sequence of skbs head..tail with
4512 * sequence numbers start..end.
4513 * Segments with FIN/SYN are not collapsed (only because this
4514 * simplifies code)
4515 */
4516static void
David S. Miller8728b832005-08-09 19:25:21 -07004517tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4518 struct sk_buff *head, struct sk_buff *tail,
4519 u32 start, u32 end)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004520{
4521 struct sk_buff *skb;
4522
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08004523 /* First, check that queue is collapsible and find
Linus Torvalds1da177e2005-04-16 15:20:36 -07004524 * the point where collapsing can be useful. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004525 for (skb = head; skb != tail;) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004526 /* No new bits? It is possible on ofo queue. */
4527 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
Ilpo Järvinen2cf46632008-08-23 05:11:41 -07004528 skb = tcp_collapse_one(sk, skb, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004529 continue;
4530 }
4531
4532 /* The first skb to collapse is:
4533 * - not SYN/FIN and
4534 * - bloated or contains data before "start" or
4535 * overlaps to the next one.
4536 */
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07004537 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07004538 (tcp_win_from_space(skb->truesize) > skb->len ||
4539 before(TCP_SKB_CB(skb)->seq, start) ||
4540 (skb->next != tail &&
4541 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
4542 break;
4543
4544 /* Decided to skip this, advance start seq. */
4545 start = TCP_SKB_CB(skb)->end_seq;
4546 skb = skb->next;
4547 }
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07004548 if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004549 return;
4550
4551 while (before(start, end)) {
4552 struct sk_buff *nskb;
Chuck Leverc2636b42007-10-23 21:07:32 -07004553 unsigned int header = skb_headroom(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004554 int copy = SKB_MAX_ORDER(header, 0);
4555
4556 /* Too big header? This can happen with IPv6. */
4557 if (copy < 0)
4558 return;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004559 if (end - start < copy)
4560 copy = end - start;
4561 nskb = alloc_skb(copy + header, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004562 if (!nskb)
4563 return;
Arnaldo Carvalho de Meloc51957d2007-03-10 12:47:22 -03004564
Arnaldo Carvalho de Melo98e399f2007-03-19 15:33:04 -07004565 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
Arnaldo Carvalho de Melo9c702202007-04-25 18:04:18 -07004566 skb_set_network_header(nskb, (skb_network_header(skb) -
4567 skb->head));
4568 skb_set_transport_header(nskb, (skb_transport_header(skb) -
4569 skb->head));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004570 skb_reserve(nskb, header);
4571 memcpy(nskb->head, skb->head, header);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004572 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4573 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
David S. Miller43f59c82008-09-21 21:28:51 -07004574 __skb_queue_before(list, skb, nskb);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08004575 skb_set_owner_r(nskb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004576
4577 /* Copy data, releasing collapsed skbs. */
4578 while (copy > 0) {
4579 int offset = start - TCP_SKB_CB(skb)->seq;
4580 int size = TCP_SKB_CB(skb)->end_seq - start;
4581
Kris Katterjohn09a62662006-01-08 22:24:28 -08004582 BUG_ON(offset < 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004583 if (size > 0) {
4584 size = min(copy, size);
4585 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4586 BUG();
4587 TCP_SKB_CB(nskb)->end_seq += size;
4588 copy -= size;
4589 start += size;
4590 }
4591 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
Ilpo Järvinen2cf46632008-08-23 05:11:41 -07004592 skb = tcp_collapse_one(sk, skb, list);
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07004593 if (skb == tail ||
4594 tcp_hdr(skb)->syn ||
4595 tcp_hdr(skb)->fin)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004596 return;
4597 }
4598 }
4599 }
4600}
4601
4602/* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4603 * and tcp_collapse() them until all the queue is collapsed.
4604 */
4605static void tcp_collapse_ofo_queue(struct sock *sk)
4606{
4607 struct tcp_sock *tp = tcp_sk(sk);
4608 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4609 struct sk_buff *head;
4610 u32 start, end;
4611
4612 if (skb == NULL)
4613 return;
4614
4615 start = TCP_SKB_CB(skb)->seq;
4616 end = TCP_SKB_CB(skb)->end_seq;
4617 head = skb;
4618
4619 for (;;) {
4620 skb = skb->next;
4621
4622 /* Segment is terminated when we see gap or when
4623 * we are at the end of all the queue. */
4624 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
4625 after(TCP_SKB_CB(skb)->seq, end) ||
4626 before(TCP_SKB_CB(skb)->end_seq, start)) {
David S. Miller8728b832005-08-09 19:25:21 -07004627 tcp_collapse(sk, &tp->out_of_order_queue,
4628 head, skb, start, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004629 head = skb;
4630 if (skb == (struct sk_buff *)&tp->out_of_order_queue)
4631 break;
4632 /* Start new segment */
4633 start = TCP_SKB_CB(skb)->seq;
4634 end = TCP_SKB_CB(skb)->end_seq;
4635 } else {
4636 if (before(TCP_SKB_CB(skb)->seq, start))
4637 start = TCP_SKB_CB(skb)->seq;
4638 if (after(TCP_SKB_CB(skb)->end_seq, end))
4639 end = TCP_SKB_CB(skb)->end_seq;
4640 }
4641 }
4642}
4643
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004644/*
4645 * Purge the out-of-order queue.
Vitaliy Gusev56f367b2008-04-15 20:26:34 -07004646 * Return true if queue was pruned.
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004647 */
Vitaliy Gusev56f367b2008-04-15 20:26:34 -07004648static int tcp_prune_ofo_queue(struct sock *sk)
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004649{
4650 struct tcp_sock *tp = tcp_sk(sk);
Vitaliy Gusev56f367b2008-04-15 20:26:34 -07004651 int res = 0;
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004652
4653 if (!skb_queue_empty(&tp->out_of_order_queue)) {
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07004654 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004655 __skb_queue_purge(&tp->out_of_order_queue);
4656
4657 /* Reset SACK state. A conforming SACK implementation will
4658 * do the same at a timeout based retransmit. When a connection
4659 * is in a sad state like this, we care only about integrity
4660 * of the connection not performance.
4661 */
4662 if (tp->rx_opt.sack_ok)
4663 tcp_sack_reset(&tp->rx_opt);
4664 sk_mem_reclaim(sk);
Vitaliy Gusev56f367b2008-04-15 20:26:34 -07004665 res = 1;
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004666 }
Vitaliy Gusev56f367b2008-04-15 20:26:34 -07004667 return res;
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004668}
4669
Linus Torvalds1da177e2005-04-16 15:20:36 -07004670/* Reduce allocated memory if we can, trying to get
4671 * the socket within its memory limits again.
4672 *
4673 * Return less than zero if we should start dropping frames
4674 * until the socket owning process reads some of the data
4675 * to stabilize the situation.
4676 */
4677static int tcp_prune_queue(struct sock *sk)
4678{
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004679 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004680
4681 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4682
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07004683 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004684
4685 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004686 tcp_clamp_window(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004687 else if (tcp_memory_pressure)
4688 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4689
4690 tcp_collapse_ofo_queue(sk);
David S. Miller8728b832005-08-09 19:25:21 -07004691 tcp_collapse(sk, &sk->sk_receive_queue,
4692 sk->sk_receive_queue.next,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004693 (struct sk_buff *)&sk->sk_receive_queue,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004694 tp->copied_seq, tp->rcv_nxt);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08004695 sk_mem_reclaim(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004696
4697 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4698 return 0;
4699
4700 /* Collapsing did not help, destructive actions follow.
4701 * This must not ever occur. */
4702
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004703 tcp_prune_ofo_queue(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004704
4705 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4706 return 0;
4707
4708 /* If we are really being abused, tell the caller to silently
4709 * drop receive data on the floor. It will get retransmitted
4710 * and hopefully then we'll have sufficient space.
4711 */
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07004712 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004713
4714 /* Massive buffer overcommit. */
4715 tp->pred_flags = 0;
4716 return -1;
4717}
4718
Linus Torvalds1da177e2005-04-16 15:20:36 -07004719/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4720 * As additional protections, we do not touch cwnd in retransmission phases,
4721 * and if application hit its sndbuf limit recently.
4722 */
4723void tcp_cwnd_application_limited(struct sock *sk)
4724{
4725 struct tcp_sock *tp = tcp_sk(sk);
4726
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03004727 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07004728 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4729 /* Limited by application or receiver window. */
Ilpo Järvinend254bcd2006-08-04 16:57:42 -07004730 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4731 u32 win_used = max(tp->snd_cwnd_used, init_win);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004732 if (win_used < tp->snd_cwnd) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03004733 tp->snd_ssthresh = tcp_current_ssthresh(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004734 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4735 }
4736 tp->snd_cwnd_used = 0;
4737 }
4738 tp->snd_cwnd_stamp = tcp_time_stamp;
4739}
4740
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004741static int tcp_should_expand_sndbuf(struct sock *sk)
David S. Miller0d9901d2005-07-05 15:21:10 -07004742{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004743 struct tcp_sock *tp = tcp_sk(sk);
4744
David S. Miller0d9901d2005-07-05 15:21:10 -07004745 /* If the user specified a specific send buffer setting, do
4746 * not modify it.
4747 */
4748 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4749 return 0;
4750
4751 /* If we are under global TCP memory pressure, do not expand. */
4752 if (tcp_memory_pressure)
4753 return 0;
4754
4755 /* If we are under soft global TCP memory pressure, do not expand. */
4756 if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
4757 return 0;
4758
4759 /* If we filled the congestion window, do not expand. */
4760 if (tp->packets_out >= tp->snd_cwnd)
4761 return 0;
4762
4763 return 1;
4764}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004765
4766/* When incoming ACK allowed to free some skb from write_queue,
4767 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4768 * on the exit from tcp input handler.
4769 *
4770 * PROBLEM: sndbuf expansion does not work well with largesend.
4771 */
4772static void tcp_new_space(struct sock *sk)
4773{
4774 struct tcp_sock *tp = tcp_sk(sk);
4775
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004776 if (tcp_should_expand_sndbuf(sk)) {
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004777 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
Ilpo Järvinen4a7e5602008-10-07 14:43:31 -07004778 MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
4779 int demanded = max_t(unsigned int, tp->snd_cwnd,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004780 tp->reordering + 1);
4781 sndmem *= 2 * demanded;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004782 if (sndmem > sk->sk_sndbuf)
4783 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
4784 tp->snd_cwnd_stamp = tcp_time_stamp;
4785 }
4786
4787 sk->sk_write_space(sk);
4788}
4789
Stephen Hemminger40efc6f2006-01-03 16:03:49 -08004790static void tcp_check_space(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004791{
4792 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4793 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4794 if (sk->sk_socket &&
4795 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4796 tcp_new_space(sk);
4797 }
4798}
4799
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004800static inline void tcp_data_snd_check(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004801{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004802 tcp_push_pending_frames(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004803 tcp_check_space(sk);
4804}
4805
4806/*
4807 * Check if sending an ack is needed.
4808 */
4809static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4810{
4811 struct tcp_sock *tp = tcp_sk(sk);
4812
4813 /* More than one full frame received... */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004814 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss
Linus Torvalds1da177e2005-04-16 15:20:36 -07004815 /* ... and right edge of window advances far enough.
4816 * (tcp_recvmsg() will send ACK otherwise). Or...
4817 */
4818 && __tcp_select_window(sk) >= tp->rcv_wnd) ||
4819 /* We ACK each frame or... */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004820 tcp_in_quickack_mode(sk) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07004821 /* We have out of order data. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004822 (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004823 /* Then ack it now */
4824 tcp_send_ack(sk);
4825 } else {
4826 /* Else, send delayed ack. */
4827 tcp_send_delayed_ack(sk);
4828 }
4829}
4830
Stephen Hemminger40efc6f2006-01-03 16:03:49 -08004831static inline void tcp_ack_snd_check(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004832{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004833 if (!inet_csk_ack_scheduled(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004834 /* We sent a data segment already. */
4835 return;
4836 }
4837 __tcp_ack_snd_check(sk, 1);
4838}
4839
4840/*
4841 * This routine is only called when we have urgent data
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08004842 * signaled. Its the 'slow' part of tcp_urg. It could be
Linus Torvalds1da177e2005-04-16 15:20:36 -07004843 * moved inline now as tcp_urg is only called from one
4844 * place. We handle URGent data wrong. We have to - as
4845 * BSD still doesn't use the correction from RFC961.
4846 * For 1003.1g we should support a new option TCP_STDURG to permit
4847 * either form (or just set the sysctl tcp_stdurg).
4848 */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004849
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004850static void tcp_check_urg(struct sock *sk, struct tcphdr *th)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004851{
4852 struct tcp_sock *tp = tcp_sk(sk);
4853 u32 ptr = ntohs(th->urg_ptr);
4854
4855 if (ptr && !sysctl_tcp_stdurg)
4856 ptr--;
4857 ptr += ntohl(th->seq);
4858
4859 /* Ignore urgent data that we've already seen and read. */
4860 if (after(tp->copied_seq, ptr))
4861 return;
4862
4863 /* Do not replay urg ptr.
4864 *
4865 * NOTE: interesting situation not covered by specs.
4866 * Misbehaving sender may send urg ptr, pointing to segment,
4867 * which we already have in ofo queue. We are not able to fetch
4868 * such data and will stay in TCP_URG_NOTYET until will be eaten
4869 * by recvmsg(). Seems, we are not obliged to handle such wicked
4870 * situations. But it is worth to think about possibility of some
4871 * DoSes using some hypothetical application level deadlock.
4872 */
4873 if (before(ptr, tp->rcv_nxt))
4874 return;
4875
4876 /* Do we already have a newer (or duplicate) urgent pointer? */
4877 if (tp->urg_data && !after(ptr, tp->urg_seq))
4878 return;
4879
4880 /* Tell the world about our new urgent pointer. */
4881 sk_send_sigurg(sk);
4882
4883 /* We may be adding urgent data when the last byte read was
4884 * urgent. To do this requires some care. We cannot just ignore
4885 * tp->copied_seq since we would read the last urgent byte again
4886 * as data, nor can we alter copied_seq until this data arrives
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08004887 * or we break the semantics of SIOCATMARK (and thus sockatmark())
Linus Torvalds1da177e2005-04-16 15:20:36 -07004888 *
4889 * NOTE. Double Dutch. Rendering to plain English: author of comment
4890 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4891 * and expect that both A and B disappear from stream. This is _wrong_.
4892 * Though this happens in BSD with high probability, this is occasional.
4893 * Any application relying on this is buggy. Note also, that fix "works"
4894 * only in this artificial test. Insert some normal data between A and B and we will
4895 * decline of BSD again. Verdict: it is better to remove to trap
4896 * buggy users.
4897 */
4898 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004899 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004900 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4901 tp->copied_seq++;
4902 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
David S. Miller8728b832005-08-09 19:25:21 -07004903 __skb_unlink(skb, &sk->sk_receive_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004904 __kfree_skb(skb);
4905 }
4906 }
4907
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004908 tp->urg_data = TCP_URG_NOTYET;
4909 tp->urg_seq = ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004910
4911 /* Disable header prediction. */
4912 tp->pred_flags = 0;
4913}
4914
4915/* This is the 'fast' part of urgent handling. */
4916static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
4917{
4918 struct tcp_sock *tp = tcp_sk(sk);
4919
4920 /* Check if we get a new urgent pointer - normally not. */
4921 if (th->urg)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004922 tcp_check_urg(sk, th);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004923
4924 /* Do we wait for any urgent data? - normally not... */
4925 if (tp->urg_data == TCP_URG_NOTYET) {
4926 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4927 th->syn;
4928
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004929 /* Is the urgent pointer pointing into this packet? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004930 if (ptr < skb->len) {
4931 u8 tmp;
4932 if (skb_copy_bits(skb, ptr, &tmp, 1))
4933 BUG();
4934 tp->urg_data = TCP_URG_VALID | tmp;
4935 if (!sock_flag(sk, SOCK_DEAD))
4936 sk->sk_data_ready(sk, 0);
4937 }
4938 }
4939}
4940
4941static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4942{
4943 struct tcp_sock *tp = tcp_sk(sk);
4944 int chunk = skb->len - hlen;
4945 int err;
4946
4947 local_bh_enable();
Herbert Xu60476372007-04-09 11:59:39 -07004948 if (skb_csum_unnecessary(skb))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004949 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4950 else
4951 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4952 tp->ucopy.iov);
4953
4954 if (!err) {
4955 tp->ucopy.len -= chunk;
4956 tp->copied_seq += chunk;
4957 tcp_rcv_space_adjust(sk);
4958 }
4959
4960 local_bh_disable();
4961 return err;
4962}
4963
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004964static __sum16 __tcp_checksum_complete_user(struct sock *sk,
4965 struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004966{
Al Virob51655b2006-11-14 21:40:42 -08004967 __sum16 result;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004968
4969 if (sock_owned_by_user(sk)) {
4970 local_bh_enable();
4971 result = __tcp_checksum_complete(skb);
4972 local_bh_disable();
4973 } else {
4974 result = __tcp_checksum_complete(skb);
4975 }
4976 return result;
4977}
4978
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004979static inline int tcp_checksum_complete_user(struct sock *sk,
4980 struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004981{
Herbert Xu60476372007-04-09 11:59:39 -07004982 return !skb_csum_unnecessary(skb) &&
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004983 __tcp_checksum_complete_user(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004984}
4985
Chris Leech1a2449a2006-05-23 18:05:53 -07004986#ifdef CONFIG_NET_DMA
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004987static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb,
4988 int hlen)
Chris Leech1a2449a2006-05-23 18:05:53 -07004989{
4990 struct tcp_sock *tp = tcp_sk(sk);
4991 int chunk = skb->len - hlen;
4992 int dma_cookie;
4993 int copied_early = 0;
4994
4995 if (tp->ucopy.wakeup)
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004996 return 0;
Chris Leech1a2449a2006-05-23 18:05:53 -07004997
4998 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
Dan Williamsf67b4592009-01-06 11:38:15 -07004999 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY);
Chris Leech1a2449a2006-05-23 18:05:53 -07005000
Herbert Xu60476372007-04-09 11:59:39 -07005001 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
Chris Leech1a2449a2006-05-23 18:05:53 -07005002
5003 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005004 skb, hlen,
5005 tp->ucopy.iov, chunk,
5006 tp->ucopy.pinned_list);
Chris Leech1a2449a2006-05-23 18:05:53 -07005007
5008 if (dma_cookie < 0)
5009 goto out;
5010
5011 tp->ucopy.dma_cookie = dma_cookie;
5012 copied_early = 1;
5013
5014 tp->ucopy.len -= chunk;
5015 tp->copied_seq += chunk;
5016 tcp_rcv_space_adjust(sk);
5017
5018 if ((tp->ucopy.len == 0) ||
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07005019 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
Chris Leech1a2449a2006-05-23 18:05:53 -07005020 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
5021 tp->ucopy.wakeup = 1;
5022 sk->sk_data_ready(sk, 0);
5023 }
5024 } else if (chunk > 0) {
5025 tp->ucopy.wakeup = 1;
5026 sk->sk_data_ready(sk, 0);
5027 }
5028out:
5029 return copied_early;
5030}
5031#endif /* CONFIG_NET_DMA */
5032
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005033/* Does PAWS and seqno based validation of an incoming segment, flags will
5034 * play significant role here.
5035 */
5036static int tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5037 struct tcphdr *th, int syn_inerr)
5038{
5039 struct tcp_sock *tp = tcp_sk(sk);
5040
5041 /* RFC1323: H1. Apply PAWS check first. */
5042 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
5043 tcp_paws_discard(sk, skb)) {
5044 if (!th->rst) {
5045 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5046 tcp_send_dupack(sk, skb);
5047 goto discard;
5048 }
5049 /* Reset is accepted even if it did not pass PAWS. */
5050 }
5051
5052 /* Step 1: check sequence number */
5053 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5054 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5055 * (RST) segments are validated by checking their SEQ-fields."
5056 * And page 69: "If an incoming segment is not acceptable,
5057 * an acknowledgment should be sent in reply (unless the RST
5058 * bit is set, if so drop the segment and return)".
5059 */
5060 if (!th->rst)
5061 tcp_send_dupack(sk, skb);
5062 goto discard;
5063 }
5064
5065 /* Step 2: check RST bit */
5066 if (th->rst) {
5067 tcp_reset(sk);
5068 goto discard;
5069 }
5070
5071 /* ts_recent update must be made after we are sure that the packet
5072 * is in window.
5073 */
5074 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
5075
5076 /* step 3: check security and precedence [ignored] */
5077
5078 /* step 4: Check for a SYN in window. */
5079 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5080 if (syn_inerr)
5081 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5082 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONSYN);
5083 tcp_reset(sk);
5084 return -1;
5085 }
5086
5087 return 1;
5088
5089discard:
5090 __kfree_skb(skb);
5091 return 0;
5092}
5093
Linus Torvalds1da177e2005-04-16 15:20:36 -07005094/*
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005095 * TCP receive function for the ESTABLISHED state.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005096 *
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005097 * It is split into a fast path and a slow path. The fast path is
Linus Torvalds1da177e2005-04-16 15:20:36 -07005098 * disabled when:
5099 * - A zero window was announced from us - zero window probing
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005100 * is only handled properly in the slow path.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005101 * - Out of order segments arrived.
5102 * - Urgent data is expected.
5103 * - There is no buffer space left
5104 * - Unexpected TCP flags/window values/header lengths are received
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005105 * (detected by checking the TCP header against pred_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005106 * - Data is sent in both directions. Fast path only supports pure senders
5107 * or pure receivers (this means either the sequence number or the ack
5108 * value must stay constant)
5109 * - Unexpected TCP option.
5110 *
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005111 * When these conditions are not satisfied it drops into a standard
Linus Torvalds1da177e2005-04-16 15:20:36 -07005112 * receive procedure patterned after RFC793 to handle all cases.
5113 * The first three cases are guaranteed by proper pred_flags setting,
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005114 * the rest is checked inline. Fast processing is turned on in
Linus Torvalds1da177e2005-04-16 15:20:36 -07005115 * tcp_data_queue when everything is OK.
5116 */
5117int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5118 struct tcphdr *th, unsigned len)
5119{
5120 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005121 int res;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005122
5123 /*
5124 * Header prediction.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005125 * The code loosely follows the one in the famous
Linus Torvalds1da177e2005-04-16 15:20:36 -07005126 * "30 instruction TCP receive" Van Jacobson mail.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005127 *
5128 * Van's trick is to deposit buffers into socket queue
Linus Torvalds1da177e2005-04-16 15:20:36 -07005129 * on a device interrupt, to call tcp_recv function
5130 * on the receive process context and checksum and copy
5131 * the buffer to user space. smart...
5132 *
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005133 * Our current scheme is not silly either but we take the
Linus Torvalds1da177e2005-04-16 15:20:36 -07005134 * extra cost of the net_bh soft interrupt processing...
5135 * We do checksum and copy also but from device to kernel.
5136 */
5137
5138 tp->rx_opt.saw_tstamp = 0;
5139
5140 /* pred_flags is 0xS?10 << 16 + snd_wnd
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08005141 * if header_prediction is to be made
Linus Torvalds1da177e2005-04-16 15:20:36 -07005142 * 'S' will always be tp->tcp_header_len >> 2
5143 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005144 * turn it off (when there are holes in the receive
Linus Torvalds1da177e2005-04-16 15:20:36 -07005145 * space for instance)
5146 * PSH flag is ignored.
5147 */
5148
5149 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
John Dykstra96e0bf42009-03-22 21:49:57 -07005150 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5151 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005152 int tcp_header_len = tp->tcp_header_len;
5153
5154 /* Timestamp header prediction: tcp_header_len
5155 * is automatically equal to th->doff*4 due to pred_flags
5156 * match.
5157 */
5158
5159 /* Check timestamp */
5160 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005161 /* No? Slow path! */
Ilpo Järvinena4356b22008-08-23 05:12:29 -07005162 if (!tcp_parse_aligned_timestamp(tp, th))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005163 goto slow_path;
5164
Linus Torvalds1da177e2005-04-16 15:20:36 -07005165 /* If PAWS failed, check it more carefully in slow path */
5166 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5167 goto slow_path;
5168
5169 /* DO NOT update ts_recent here, if checksum fails
5170 * and timestamp was corrupted part, it will result
5171 * in a hung connection since we will drop all
5172 * future packets due to the PAWS test.
5173 */
5174 }
5175
5176 if (len <= tcp_header_len) {
5177 /* Bulk data transfer: sender */
5178 if (len == tcp_header_len) {
5179 /* Predicted packet is in window by definition.
5180 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5181 * Hence, check seq<=rcv_wup reduces to:
5182 */
5183 if (tcp_header_len ==
5184 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5185 tp->rcv_nxt == tp->rcv_wup)
5186 tcp_store_ts_recent(tp);
5187
Linus Torvalds1da177e2005-04-16 15:20:36 -07005188 /* We know that such packets are checksummed
5189 * on entry.
5190 */
5191 tcp_ack(sk, skb, 0);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005192 __kfree_skb(skb);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005193 tcp_data_snd_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005194 return 0;
5195 } else { /* Header too small */
Pavel Emelyanov63231bd2008-07-16 20:22:25 -07005196 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005197 goto discard;
5198 }
5199 } else {
5200 int eaten = 0;
Chris Leech1a2449a2006-05-23 18:05:53 -07005201 int copied_early = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005202
Chris Leech1a2449a2006-05-23 18:05:53 -07005203 if (tp->copied_seq == tp->rcv_nxt &&
5204 len - tcp_header_len <= tp->ucopy.len) {
5205#ifdef CONFIG_NET_DMA
5206 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
5207 copied_early = 1;
5208 eaten = 1;
5209 }
5210#endif
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005211 if (tp->ucopy.task == current &&
5212 sock_owned_by_user(sk) && !copied_early) {
Chris Leech1a2449a2006-05-23 18:05:53 -07005213 __set_current_state(TASK_RUNNING);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005214
Chris Leech1a2449a2006-05-23 18:05:53 -07005215 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
5216 eaten = 1;
5217 }
5218 if (eaten) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005219 /* Predicted packet is in window by definition.
5220 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5221 * Hence, check seq<=rcv_wup reduces to:
5222 */
5223 if (tcp_header_len ==
5224 (sizeof(struct tcphdr) +
5225 TCPOLEN_TSTAMP_ALIGNED) &&
5226 tp->rcv_nxt == tp->rcv_wup)
5227 tcp_store_ts_recent(tp);
5228
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005229 tcp_rcv_rtt_measure_ts(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005230
5231 __skb_pull(skb, tcp_header_len);
5232 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07005233 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005234 }
Chris Leech1a2449a2006-05-23 18:05:53 -07005235 if (copied_early)
5236 tcp_cleanup_rbuf(sk, skb->len);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005237 }
5238 if (!eaten) {
5239 if (tcp_checksum_complete_user(sk, skb))
5240 goto csum_error;
5241
5242 /* Predicted packet is in window by definition.
5243 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5244 * Hence, check seq<=rcv_wup reduces to:
5245 */
5246 if (tcp_header_len ==
5247 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5248 tp->rcv_nxt == tp->rcv_wup)
5249 tcp_store_ts_recent(tp);
5250
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005251 tcp_rcv_rtt_measure_ts(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005252
5253 if ((int)skb->truesize > sk->sk_forward_alloc)
5254 goto step5;
5255
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07005256 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005257
5258 /* Bulk data transfer: receiver */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005259 __skb_pull(skb, tcp_header_len);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005260 __skb_queue_tail(&sk->sk_receive_queue, skb);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08005261 skb_set_owner_r(skb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005262 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5263 }
5264
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005265 tcp_event_data_recv(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005266
5267 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5268 /* Well, only one small jumplet in fast path... */
5269 tcp_ack(sk, skb, FLAG_DATA);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005270 tcp_data_snd_check(sk);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005271 if (!inet_csk_ack_scheduled(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005272 goto no_ack;
5273 }
5274
Ali Saidi53240c22008-10-07 15:31:19 -07005275 if (!copied_early || tp->rcv_nxt != tp->rcv_wup)
5276 __tcp_ack_snd_check(sk, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005277no_ack:
Chris Leech1a2449a2006-05-23 18:05:53 -07005278#ifdef CONFIG_NET_DMA
5279 if (copied_early)
5280 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
5281 else
5282#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07005283 if (eaten)
5284 __kfree_skb(skb);
5285 else
5286 sk->sk_data_ready(sk, 0);
5287 return 0;
5288 }
5289 }
5290
5291slow_path:
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005292 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005293 goto csum_error;
5294
5295 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07005296 * Standard slow path.
5297 */
5298
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005299 res = tcp_validate_incoming(sk, skb, th, 1);
5300 if (res <= 0)
5301 return -res;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005302
5303step5:
John Dykstra96e0bf42009-03-22 21:49:57 -07005304 if (th->ack && tcp_ack(sk, skb, FLAG_SLOWPATH) < 0)
5305 goto discard;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005306
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005307 tcp_rcv_rtt_measure_ts(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005308
5309 /* Process urgent data. */
5310 tcp_urg(sk, skb, th);
5311
5312 /* step 7: process the segment text */
5313 tcp_data_queue(sk, skb);
5314
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005315 tcp_data_snd_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005316 tcp_ack_snd_check(sk);
5317 return 0;
5318
5319csum_error:
Pavel Emelyanov63231bd2008-07-16 20:22:25 -07005320 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005321
5322discard:
5323 __kfree_skb(skb);
5324 return 0;
5325}
5326
5327static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5328 struct tcphdr *th, unsigned len)
5329{
5330 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08005331 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005332 int saved_clamp = tp->rx_opt.mss_clamp;
5333
5334 tcp_parse_options(skb, &tp->rx_opt, 0);
5335
5336 if (th->ack) {
5337 /* rfc793:
5338 * "If the state is SYN-SENT then
5339 * first check the ACK bit
5340 * If the ACK bit is set
5341 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5342 * a reset (unless the RST bit is set, if so drop
5343 * the segment and return)"
5344 *
5345 * We do not send data with SYN, so that RFC-correct
5346 * test reduces to:
5347 */
5348 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
5349 goto reset_and_undo;
5350
5351 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5352 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5353 tcp_time_stamp)) {
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07005354 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005355 goto reset_and_undo;
5356 }
5357
5358 /* Now ACK is acceptable.
5359 *
5360 * "If the RST bit is set
5361 * If the ACK was acceptable then signal the user "error:
5362 * connection reset", drop the segment, enter CLOSED state,
5363 * delete TCB, and return."
5364 */
5365
5366 if (th->rst) {
5367 tcp_reset(sk);
5368 goto discard;
5369 }
5370
5371 /* rfc793:
5372 * "fifth, if neither of the SYN or RST bits is set then
5373 * drop the segment and return."
5374 *
5375 * See note below!
5376 * --ANK(990513)
5377 */
5378 if (!th->syn)
5379 goto discard_and_undo;
5380
5381 /* rfc793:
5382 * "If the SYN bit is on ...
5383 * are acceptable then ...
5384 * (our SYN has been ACKed), change the connection
5385 * state to ESTABLISHED..."
5386 */
5387
5388 TCP_ECN_rcv_synack(tp, th);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005389
5390 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5391 tcp_ack(sk, skb, FLAG_SLOWPATH);
5392
5393 /* Ok.. it's good. Set up sequence numbers and
5394 * move to established.
5395 */
5396 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5397 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5398
5399 /* RFC1323: The window in SYN & SYN/ACK segments is
5400 * never scaled.
5401 */
5402 tp->snd_wnd = ntohs(th->window);
Hantzis Fotisee7537b2009-03-02 22:42:02 -08005403 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005404
5405 if (!tp->rx_opt.wscale_ok) {
5406 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5407 tp->window_clamp = min(tp->window_clamp, 65535U);
5408 }
5409
5410 if (tp->rx_opt.saw_tstamp) {
5411 tp->rx_opt.tstamp_ok = 1;
5412 tp->tcp_header_len =
5413 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5414 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5415 tcp_store_ts_recent(tp);
5416 } else {
5417 tp->tcp_header_len = sizeof(struct tcphdr);
5418 }
5419
Ilpo Järvinene60402d2007-08-09 15:14:46 +03005420 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5421 tcp_enable_fack(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005422
John Heffner5d424d52006-03-20 17:53:41 -08005423 tcp_mtup_init(sk);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08005424 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005425 tcp_initialize_rcv_mss(sk);
5426
5427 /* Remember, tcp_poll() does not lock socket!
5428 * Change state from SYN-SENT only after copied_seq
5429 * is initialized. */
5430 tp->copied_seq = tp->rcv_nxt;
Ralf Baechlee16aa202006-12-07 00:11:33 -08005431 smp_mb();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005432 tcp_set_state(sk, TCP_ESTABLISHED);
5433
Venkat Yekkirala6b877692006-11-08 17:04:09 -06005434 security_inet_conn_established(sk, skb);
5435
Linus Torvalds1da177e2005-04-16 15:20:36 -07005436 /* Make sure socket is routed, for correct metrics. */
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -08005437 icsk->icsk_af_ops->rebuild_header(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005438
5439 tcp_init_metrics(sk);
5440
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03005441 tcp_init_congestion_control(sk);
Stephen Hemminger317a76f2005-06-23 12:19:55 -07005442
Linus Torvalds1da177e2005-04-16 15:20:36 -07005443 /* Prevent spurious tcp_cwnd_restart() on first data
5444 * packet.
5445 */
5446 tp->lsndtime = tcp_time_stamp;
5447
5448 tcp_init_buffer_space(sk);
5449
5450 if (sock_flag(sk, SOCK_KEEPOPEN))
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005451 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005452
5453 if (!tp->rx_opt.snd_wscale)
5454 __tcp_fast_path_on(tp, tp->snd_wnd);
5455 else
5456 tp->pred_flags = 0;
5457
5458 if (!sock_flag(sk, SOCK_DEAD)) {
5459 sk->sk_state_change(sk);
Pavel Emelyanov8d8ad9d2007-11-26 20:10:50 +08005460 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005461 }
5462
Arnaldo Carvalho de Melo295f7322005-08-09 20:11:56 -07005463 if (sk->sk_write_pending ||
5464 icsk->icsk_accept_queue.rskq_defer_accept ||
5465 icsk->icsk_ack.pingpong) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005466 /* Save one ACK. Data will be ready after
5467 * several ticks, if write_pending is set.
5468 *
5469 * It may be deleted, but with this feature tcpdumps
5470 * look so _wonderfully_ clever, that I was not able
5471 * to stand against the temptation 8) --ANK
5472 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005473 inet_csk_schedule_ack(sk);
Arnaldo Carvalho de Melo295f7322005-08-09 20:11:56 -07005474 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5475 icsk->icsk_ack.ato = TCP_ATO_MIN;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005476 tcp_incr_quickack(sk);
5477 tcp_enter_quickack_mode(sk);
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07005478 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5479 TCP_DELACK_MAX, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005480
5481discard:
5482 __kfree_skb(skb);
5483 return 0;
5484 } else {
5485 tcp_send_ack(sk);
5486 }
5487 return -1;
5488 }
5489
5490 /* No ACK in the segment */
5491
5492 if (th->rst) {
5493 /* rfc793:
5494 * "If the RST bit is set
5495 *
5496 * Otherwise (no ACK) drop the segment and return."
5497 */
5498
5499 goto discard_and_undo;
5500 }
5501
5502 /* PAWS check. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005503 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
Ilpo Järvinenc887e6d2009-03-14 14:23:03 +00005504 tcp_paws_reject(&tp->rx_opt, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005505 goto discard_and_undo;
5506
5507 if (th->syn) {
5508 /* We see SYN without ACK. It is attempt of
5509 * simultaneous connect with crossed SYNs.
5510 * Particularly, it can be connect to self.
5511 */
5512 tcp_set_state(sk, TCP_SYN_RECV);
5513
5514 if (tp->rx_opt.saw_tstamp) {
5515 tp->rx_opt.tstamp_ok = 1;
5516 tcp_store_ts_recent(tp);
5517 tp->tcp_header_len =
5518 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5519 } else {
5520 tp->tcp_header_len = sizeof(struct tcphdr);
5521 }
5522
5523 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5524 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5525
5526 /* RFC1323: The window in SYN & SYN/ACK segments is
5527 * never scaled.
5528 */
5529 tp->snd_wnd = ntohs(th->window);
5530 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5531 tp->max_window = tp->snd_wnd;
5532
5533 TCP_ECN_rcv_syn(tp, th);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005534
John Heffner5d424d52006-03-20 17:53:41 -08005535 tcp_mtup_init(sk);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08005536 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005537 tcp_initialize_rcv_mss(sk);
5538
Linus Torvalds1da177e2005-04-16 15:20:36 -07005539 tcp_send_synack(sk);
5540#if 0
5541 /* Note, we could accept data and URG from this segment.
5542 * There are no obstacles to make this.
5543 *
5544 * However, if we ignore data in ACKless segments sometimes,
5545 * we have no reasons to accept it sometimes.
5546 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5547 * is not flawless. So, discard packet for sanity.
5548 * Uncomment this return to process the data.
5549 */
5550 return -1;
5551#else
5552 goto discard;
5553#endif
5554 }
5555 /* "fifth, if neither of the SYN or RST bits is set then
5556 * drop the segment and return."
5557 */
5558
5559discard_and_undo:
5560 tcp_clear_options(&tp->rx_opt);
5561 tp->rx_opt.mss_clamp = saved_clamp;
5562 goto discard;
5563
5564reset_and_undo:
5565 tcp_clear_options(&tp->rx_opt);
5566 tp->rx_opt.mss_clamp = saved_clamp;
5567 return 1;
5568}
5569
Linus Torvalds1da177e2005-04-16 15:20:36 -07005570/*
5571 * This function implements the receiving procedure of RFC 793 for
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005572 * all states except ESTABLISHED and TIME_WAIT.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005573 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5574 * address independent.
5575 */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005576
Linus Torvalds1da177e2005-04-16 15:20:36 -07005577int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5578 struct tcphdr *th, unsigned len)
5579{
5580 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -08005581 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005582 int queued = 0;
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005583 int res;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005584
5585 tp->rx_opt.saw_tstamp = 0;
5586
5587 switch (sk->sk_state) {
5588 case TCP_CLOSE:
5589 goto discard;
5590
5591 case TCP_LISTEN:
Stephen Hemminger2de979b2007-03-08 20:45:19 -08005592 if (th->ack)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005593 return 1;
5594
Stephen Hemminger2de979b2007-03-08 20:45:19 -08005595 if (th->rst)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005596 goto discard;
5597
Stephen Hemminger2de979b2007-03-08 20:45:19 -08005598 if (th->syn) {
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -08005599 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005600 return 1;
5601
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005602 /* Now we have several options: In theory there is
5603 * nothing else in the frame. KA9Q has an option to
Linus Torvalds1da177e2005-04-16 15:20:36 -07005604 * send data with the syn, BSD accepts data with the
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005605 * syn up to the [to be] advertised window and
5606 * Solaris 2.1 gives you a protocol error. For now
5607 * we just ignore it, that fits the spec precisely
Linus Torvalds1da177e2005-04-16 15:20:36 -07005608 * and avoids incompatibilities. It would be nice in
5609 * future to drop through and process the data.
5610 *
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005611 * Now that TTCP is starting to be used we ought to
Linus Torvalds1da177e2005-04-16 15:20:36 -07005612 * queue this data.
5613 * But, this leaves one open to an easy denial of
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005614 * service attack, and SYN cookies can't defend
Linus Torvalds1da177e2005-04-16 15:20:36 -07005615 * against this problem. So, we drop the data
Masayuki Nakagawafb7e2392007-01-23 20:15:06 -08005616 * in the interest of security over speed unless
5617 * it's still in use.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005618 */
Masayuki Nakagawafb7e2392007-01-23 20:15:06 -08005619 kfree_skb(skb);
5620 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005621 }
5622 goto discard;
5623
5624 case TCP_SYN_SENT:
Linus Torvalds1da177e2005-04-16 15:20:36 -07005625 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5626 if (queued >= 0)
5627 return queued;
5628
5629 /* Do step6 onward by hand. */
5630 tcp_urg(sk, skb, th);
5631 __kfree_skb(skb);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005632 tcp_data_snd_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005633 return 0;
5634 }
5635
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005636 res = tcp_validate_incoming(sk, skb, th, 0);
5637 if (res <= 0)
5638 return -res;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005639
5640 /* step 5: check the ACK field */
5641 if (th->ack) {
John Dykstra96e0bf42009-03-22 21:49:57 -07005642 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH) > 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005643
Stephen Hemminger2de979b2007-03-08 20:45:19 -08005644 switch (sk->sk_state) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005645 case TCP_SYN_RECV:
5646 if (acceptable) {
5647 tp->copied_seq = tp->rcv_nxt;
Ralf Baechlee16aa202006-12-07 00:11:33 -08005648 smp_mb();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005649 tcp_set_state(sk, TCP_ESTABLISHED);
5650 sk->sk_state_change(sk);
5651
5652 /* Note, that this wakeup is only for marginal
5653 * crossed SYN case. Passively open sockets
5654 * are not waked up, because sk->sk_sleep ==
5655 * NULL and sk->sk_socket == NULL.
5656 */
Pavel Emelyanov8d8ad9d2007-11-26 20:10:50 +08005657 if (sk->sk_socket)
5658 sk_wake_async(sk,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005659 SOCK_WAKE_IO, POLL_OUT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005660
5661 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5662 tp->snd_wnd = ntohs(th->window) <<
5663 tp->rx_opt.snd_wscale;
Hantzis Fotisee7537b2009-03-02 22:42:02 -08005664 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005665
5666 /* tcp_ack considers this ACK as duplicate
5667 * and does not calculate rtt.
5668 * Fix it at least with timestamps.
5669 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005670 if (tp->rx_opt.saw_tstamp &&
5671 tp->rx_opt.rcv_tsecr && !tp->srtt)
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08005672 tcp_ack_saw_tstamp(sk, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005673
5674 if (tp->rx_opt.tstamp_ok)
5675 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5676
5677 /* Make sure socket is routed, for
5678 * correct metrics.
5679 */
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -08005680 icsk->icsk_af_ops->rebuild_header(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005681
5682 tcp_init_metrics(sk);
5683
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03005684 tcp_init_congestion_control(sk);
Stephen Hemminger317a76f2005-06-23 12:19:55 -07005685
Linus Torvalds1da177e2005-04-16 15:20:36 -07005686 /* Prevent spurious tcp_cwnd_restart() on
5687 * first data packet.
5688 */
5689 tp->lsndtime = tcp_time_stamp;
5690
John Heffner5d424d52006-03-20 17:53:41 -08005691 tcp_mtup_init(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005692 tcp_initialize_rcv_mss(sk);
5693 tcp_init_buffer_space(sk);
5694 tcp_fast_path_on(tp);
5695 } else {
5696 return 1;
5697 }
5698 break;
5699
5700 case TCP_FIN_WAIT1:
5701 if (tp->snd_una == tp->write_seq) {
5702 tcp_set_state(sk, TCP_FIN_WAIT2);
5703 sk->sk_shutdown |= SEND_SHUTDOWN;
5704 dst_confirm(sk->sk_dst_cache);
5705
5706 if (!sock_flag(sk, SOCK_DEAD))
5707 /* Wake up lingering close() */
5708 sk->sk_state_change(sk);
5709 else {
5710 int tmo;
5711
5712 if (tp->linger2 < 0 ||
5713 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5714 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5715 tcp_done(sk);
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07005716 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005717 return 1;
5718 }
5719
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005720 tmo = tcp_fin_time(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005721 if (tmo > TCP_TIMEWAIT_LEN) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005722 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005723 } else if (th->fin || sock_owned_by_user(sk)) {
5724 /* Bad case. We could lose such FIN otherwise.
5725 * It is not a big problem, but it looks confusing
5726 * and not so rare event. We still can lose it now,
5727 * if it spins in bh_lock_sock(), but it is really
5728 * marginal case.
5729 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005730 inet_csk_reset_keepalive_timer(sk, tmo);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005731 } else {
5732 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5733 goto discard;
5734 }
5735 }
5736 }
5737 break;
5738
5739 case TCP_CLOSING:
5740 if (tp->snd_una == tp->write_seq) {
5741 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5742 goto discard;
5743 }
5744 break;
5745
5746 case TCP_LAST_ACK:
5747 if (tp->snd_una == tp->write_seq) {
5748 tcp_update_metrics(sk);
5749 tcp_done(sk);
5750 goto discard;
5751 }
5752 break;
5753 }
5754 } else
5755 goto discard;
5756
5757 /* step 6: check the URG bit */
5758 tcp_urg(sk, skb, th);
5759
5760 /* step 7: process the segment text */
5761 switch (sk->sk_state) {
5762 case TCP_CLOSE_WAIT:
5763 case TCP_CLOSING:
5764 case TCP_LAST_ACK:
5765 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5766 break;
5767 case TCP_FIN_WAIT1:
5768 case TCP_FIN_WAIT2:
5769 /* RFC 793 says to queue data in these states,
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005770 * RFC 1122 says we MUST send a reset.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005771 * BSD 4.4 also does reset.
5772 */
5773 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5774 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5775 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07005776 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005777 tcp_reset(sk);
5778 return 1;
5779 }
5780 }
5781 /* Fall through */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005782 case TCP_ESTABLISHED:
Linus Torvalds1da177e2005-04-16 15:20:36 -07005783 tcp_data_queue(sk, skb);
5784 queued = 1;
5785 break;
5786 }
5787
5788 /* tcp_data could move socket to TIME-WAIT */
5789 if (sk->sk_state != TCP_CLOSE) {
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005790 tcp_data_snd_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005791 tcp_ack_snd_check(sk);
5792 }
5793
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005794 if (!queued) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005795discard:
5796 __kfree_skb(skb);
5797 }
5798 return 0;
5799}
5800
5801EXPORT_SYMBOL(sysctl_tcp_ecn);
5802EXPORT_SYMBOL(sysctl_tcp_reordering);
Glenn Griffinc6aefaf2008-02-07 21:49:26 -08005803EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005804EXPORT_SYMBOL(tcp_parse_options);
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09005805#ifdef CONFIG_TCP_MD5SIG
5806EXPORT_SYMBOL(tcp_parse_md5sig_option);
5807#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07005808EXPORT_SYMBOL(tcp_rcv_established);
5809EXPORT_SYMBOL(tcp_rcv_state_process);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -08005810EXPORT_SYMBOL(tcp_initialize_rcv_mss);