blob: 57ae96a042206274a467e6f8ac5fb5d4899b06bd [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;
Ilpo Järvinen255cac92009-05-04 11:07:36 -070080int sysctl_tcp_ecn __read_mostly = 2;
Brian Haleyab32ea52006-09-22 14:15:41 -070081int 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 Takano564262c12007-10-25 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);
William Allen Simpsonbee7ca92009-11-10 09:51:18 +0000143 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -0700144 /* 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);
William Allen Simpsonbee7ca92009-11-10 09:51:18 +0000414 hint = min(hint, TCP_MSS_DEFAULT);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800415 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
Linus Torvalds1da177e2005-04-16 15:20:36 -0700600/* Called to compute a smoothed rtt estimate. The data fed to this
601 * routine either comes from timestamps, or from segments that were
602 * known _not_ to have been retransmitted [see Karn/Partridge
603 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
604 * piece by Van Jacobson.
605 * NOTE: the next three routines used to be one big routine.
606 * To save cycles in the RFC 1323 implementation it was better to break
607 * it up into three procedures. -- erics
608 */
Stephen Hemminger2d2abba2005-11-10 16:56:12 -0800609static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700610{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300611 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700612 long m = mrtt; /* RTT */
613
Linus Torvalds1da177e2005-04-16 15:20:36 -0700614 /* The following amusing code comes from Jacobson's
615 * article in SIGCOMM '88. Note that rtt and mdev
616 * are scaled versions of rtt and mean deviation.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900617 * This is designed to be as fast as possible
Linus Torvalds1da177e2005-04-16 15:20:36 -0700618 * m stands for "measurement".
619 *
620 * On a 1990 paper the rto value is changed to:
621 * RTO = rtt + 4 * mdev
622 *
623 * Funny. This algorithm seems to be very broken.
624 * These formulae increase RTO, when it should be decreased, increase
Stephen Hemminger31f34262005-11-15 15:17:10 -0800625 * too slowly, when it should be increased quickly, decrease too quickly
Linus Torvalds1da177e2005-04-16 15:20:36 -0700626 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
627 * does not matter how to _calculate_ it. Seems, it was trap
628 * that VJ failed to avoid. 8)
629 */
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800630 if (m == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700631 m = 1;
632 if (tp->srtt != 0) {
633 m -= (tp->srtt >> 3); /* m is now error in rtt est */
634 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
635 if (m < 0) {
636 m = -m; /* m is now abs(error) */
637 m -= (tp->mdev >> 2); /* similar update on mdev */
638 /* This is similar to one of Eifel findings.
639 * Eifel blocks mdev updates when rtt decreases.
640 * This solution is a bit different: we use finer gain
641 * for mdev in this case (alpha*beta).
642 * Like Eifel it also prevents growth of rto,
643 * but also it limits too fast rto decreases,
644 * happening in pure Eifel.
645 */
646 if (m > 0)
647 m >>= 3;
648 } else {
649 m -= (tp->mdev >> 2); /* similar update on mdev */
650 }
651 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
652 if (tp->mdev > tp->mdev_max) {
653 tp->mdev_max = tp->mdev;
654 if (tp->mdev_max > tp->rttvar)
655 tp->rttvar = tp->mdev_max;
656 }
657 if (after(tp->snd_una, tp->rtt_seq)) {
658 if (tp->mdev_max < tp->rttvar)
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800659 tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700660 tp->rtt_seq = tp->snd_nxt;
David S. Miller05bb1fa2007-08-30 22:10:28 -0700661 tp->mdev_max = tcp_rto_min(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700662 }
663 } else {
664 /* no previous measure. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800665 tp->srtt = m << 3; /* take the measured time to be rtt */
666 tp->mdev = m << 1; /* make sure rto = 3*rtt */
David S. Miller05bb1fa2007-08-30 22:10:28 -0700667 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700668 tp->rtt_seq = tp->snd_nxt;
669 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700670}
671
672/* Calculate rto without backoff. This is the second half of Van Jacobson's
673 * routine referred to above.
674 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700675static inline void tcp_set_rto(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700676{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700677 const struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700678 /* Old crap is replaced with new one. 8)
679 *
680 * More seriously:
681 * 1. If rtt variance happened to be less 50msec, it is hallucination.
682 * It cannot be less due to utterly erratic ACK generation made
683 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
684 * to do with delayed acks, because at cwnd>2 true delack timeout
685 * is invisible. Actually, Linux-2.4 also generates erratic
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800686 * ACKs in some circumstances.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700687 */
Damian Lukowskif1ecd5d2009-08-26 00:16:31 +0000688 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700689
690 /* 2. Fixups made earlier cannot be right.
691 * If we do not estimate RTO correctly without them,
692 * all the algo is pure shit and should be replaced
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800693 * with correct one. It is exactly, which we pretend to do.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700694 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700695
Ilpo Järvinenee6aac52008-12-05 22:43:08 -0800696 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
697 * guarantees that rto is higher.
698 */
Damian Lukowskif1ecd5d2009-08-26 00:16:31 +0000699 tcp_bound_rto(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700700}
701
702/* Save metrics learned by this TCP session.
703 This function is called only, when TCP finishes successfully
704 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
705 */
706void tcp_update_metrics(struct sock *sk)
707{
708 struct tcp_sock *tp = tcp_sk(sk);
709 struct dst_entry *dst = __sk_dst_get(sk);
710
711 if (sysctl_tcp_nometrics_save)
712 return;
713
714 dst_confirm(dst);
715
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800716 if (dst && (dst->flags & DST_HOST)) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300717 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700718 int m;
Stephen Hemmingerc1e20f72008-07-18 23:02:15 -0700719 unsigned long rtt;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700720
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300721 if (icsk->icsk_backoff || !tp->srtt) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700722 /* This session failed to estimate rtt. Why?
723 * Probably, no packets returned in time.
724 * Reset our results.
725 */
726 if (!(dst_metric_locked(dst, RTAX_RTT)))
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800727 dst->metrics[RTAX_RTT - 1] = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700728 return;
729 }
730
Stephen Hemmingerc1e20f72008-07-18 23:02:15 -0700731 rtt = dst_metric_rtt(dst, RTAX_RTT);
732 m = rtt - tp->srtt;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700733
734 /* If newly calculated rtt larger than stored one,
735 * store new one. Otherwise, use EWMA. Remember,
736 * rtt overestimation is always better than underestimation.
737 */
738 if (!(dst_metric_locked(dst, RTAX_RTT))) {
739 if (m <= 0)
Stephen Hemmingerc1e20f72008-07-18 23:02:15 -0700740 set_dst_metric_rtt(dst, RTAX_RTT, tp->srtt);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700741 else
Stephen Hemmingerc1e20f72008-07-18 23:02:15 -0700742 set_dst_metric_rtt(dst, RTAX_RTT, rtt - (m >> 3));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700743 }
744
745 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
Stephen Hemmingerc1e20f72008-07-18 23:02:15 -0700746 unsigned long var;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700747 if (m < 0)
748 m = -m;
749
750 /* Scale deviation to rttvar fixed point */
751 m >>= 1;
752 if (m < tp->mdev)
753 m = tp->mdev;
754
Stephen Hemmingerc1e20f72008-07-18 23:02:15 -0700755 var = dst_metric_rtt(dst, RTAX_RTTVAR);
756 if (m >= var)
757 var = m;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700758 else
Stephen Hemmingerc1e20f72008-07-18 23:02:15 -0700759 var -= (var - m) >> 2;
760
761 set_dst_metric_rtt(dst, RTAX_RTTVAR, var);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700762 }
763
Ilpo Järvinen0b6a05c2009-09-15 01:30:10 -0700764 if (tcp_in_initial_slowstart(tp)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700765 /* Slow start still did not finish. */
766 if (dst_metric(dst, RTAX_SSTHRESH) &&
767 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
768 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
769 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
770 if (!dst_metric_locked(dst, RTAX_CWND) &&
771 tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800772 dst->metrics[RTAX_CWND - 1] = tp->snd_cwnd;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700773 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300774 icsk->icsk_ca_state == TCP_CA_Open) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700775 /* Cong. avoidance phase, cwnd is reliable. */
776 if (!dst_metric_locked(dst, RTAX_SSTHRESH))
777 dst->metrics[RTAX_SSTHRESH-1] =
778 max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
779 if (!dst_metric_locked(dst, RTAX_CWND))
Satoru SATOH5ffc02a2008-05-04 22:14:42 -0700780 dst->metrics[RTAX_CWND-1] = (dst_metric(dst, RTAX_CWND) + tp->snd_cwnd) >> 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700781 } else {
782 /* Else slow start did not finish, cwnd is non-sense,
783 ssthresh may be also invalid.
784 */
785 if (!dst_metric_locked(dst, RTAX_CWND))
Satoru SATOH5ffc02a2008-05-04 22:14:42 -0700786 dst->metrics[RTAX_CWND-1] = (dst_metric(dst, RTAX_CWND) + tp->snd_ssthresh) >> 1;
787 if (dst_metric(dst, RTAX_SSTHRESH) &&
Linus Torvalds1da177e2005-04-16 15:20:36 -0700788 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
Satoru SATOH5ffc02a2008-05-04 22:14:42 -0700789 tp->snd_ssthresh > dst_metric(dst, RTAX_SSTHRESH))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700790 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
791 }
792
793 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
Satoru SATOH5ffc02a2008-05-04 22:14:42 -0700794 if (dst_metric(dst, RTAX_REORDERING) < tp->reordering &&
Linus Torvalds1da177e2005-04-16 15:20:36 -0700795 tp->reordering != sysctl_tcp_reordering)
796 dst->metrics[RTAX_REORDERING-1] = tp->reordering;
797 }
798 }
799}
800
Gerrit Renker410e27a2008-09-09 13:27:22 +0200801/* Numbers are taken from RFC3390.
802 *
803 * John Heffner states:
804 *
805 * The RFC specifies a window of no more than 4380 bytes
806 * unless 2*MSS > 4380. Reading the pseudocode in the RFC
807 * is a bit misleading because they use a clamp at 4380 bytes
808 * rather than use a multiplier in the relevant range.
809 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700810__u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
811{
812 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
813
Gerrit Renker410e27a2008-09-09 13:27:22 +0200814 if (!cwnd) {
815 if (tp->mss_cache > 1460)
816 cwnd = 2;
817 else
818 cwnd = (tp->mss_cache > 1095) ? 3 : 4;
819 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700820 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
821}
822
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800823/* Set slow start threshold and cwnd not falling to slow start */
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -0800824void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800825{
826 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -0800827 const struct inet_connection_sock *icsk = inet_csk(sk);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800828
829 tp->prior_ssthresh = 0;
830 tp->bytes_acked = 0;
Ilpo Järvinene01f9d72007-03-02 13:27:25 -0800831 if (icsk->icsk_ca_state < TCP_CA_CWR) {
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800832 tp->undo_marker = 0;
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -0800833 if (set_ssthresh)
834 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800835 tp->snd_cwnd = min(tp->snd_cwnd,
836 tcp_packets_in_flight(tp) + 1U);
837 tp->snd_cwnd_cnt = 0;
838 tp->high_seq = tp->snd_nxt;
839 tp->snd_cwnd_stamp = tcp_time_stamp;
840 TCP_ECN_queue_cwr(tp);
841
842 tcp_set_ca_state(sk, TCP_CA_CWR);
843 }
844}
845
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300846/*
847 * Packet counting of FACK is based on in-order assumptions, therefore TCP
848 * disables it when reordering is detected
849 */
850static void tcp_disable_fack(struct tcp_sock *tp)
851{
Ilpo Järvinen85cc3912007-11-15 19:39:31 -0800852 /* RFC3517 uses different metric in lost marker => reset on change */
853 if (tcp_is_fack(tp))
854 tp->lost_skb_hint = NULL;
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300855 tp->rx_opt.sack_ok &= ~2;
856}
857
Ryousei Takano564262c12007-10-25 23:03:52 -0700858/* Take a notice that peer is sending D-SACKs */
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300859static void tcp_dsack_seen(struct tcp_sock *tp)
860{
861 tp->rx_opt.sack_ok |= 4;
862}
863
Linus Torvalds1da177e2005-04-16 15:20:36 -0700864/* Initialize metrics on socket. */
865
866static void tcp_init_metrics(struct sock *sk)
867{
868 struct tcp_sock *tp = tcp_sk(sk);
869 struct dst_entry *dst = __sk_dst_get(sk);
870
871 if (dst == NULL)
872 goto reset;
873
874 dst_confirm(dst);
875
876 if (dst_metric_locked(dst, RTAX_CWND))
877 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
878 if (dst_metric(dst, RTAX_SSTHRESH)) {
879 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
880 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
881 tp->snd_ssthresh = tp->snd_cwnd_clamp;
882 }
883 if (dst_metric(dst, RTAX_REORDERING) &&
884 tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300885 tcp_disable_fack(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700886 tp->reordering = dst_metric(dst, RTAX_REORDERING);
887 }
888
889 if (dst_metric(dst, RTAX_RTT) == 0)
890 goto reset;
891
Stephen Hemmingerc1e20f72008-07-18 23:02:15 -0700892 if (!tp->srtt && dst_metric_rtt(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700893 goto reset;
894
895 /* Initial rtt is determined from SYN,SYN-ACK.
896 * The segment is small and rtt may appear much
897 * less than real one. Use per-dst memory
898 * to make it more realistic.
899 *
900 * A bit of theory. RTT is time passed after "normal" sized packet
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800901 * is sent until it is ACKed. In normal circumstances sending small
Linus Torvalds1da177e2005-04-16 15:20:36 -0700902 * packets force peer to delay ACKs and calculation is correct too.
903 * The algorithm is adaptive and, provided we follow specs, it
904 * NEVER underestimate RTT. BUT! If peer tries to make some clever
905 * tricks sort of "quick acks" for time long enough to decrease RTT
906 * to low value, and then abruptly stops to do it and starts to delay
907 * ACKs, wait for troubles.
908 */
Stephen Hemmingerc1e20f72008-07-18 23:02:15 -0700909 if (dst_metric_rtt(dst, RTAX_RTT) > tp->srtt) {
910 tp->srtt = dst_metric_rtt(dst, RTAX_RTT);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700911 tp->rtt_seq = tp->snd_nxt;
912 }
Stephen Hemmingerc1e20f72008-07-18 23:02:15 -0700913 if (dst_metric_rtt(dst, RTAX_RTTVAR) > tp->mdev) {
914 tp->mdev = dst_metric_rtt(dst, RTAX_RTTVAR);
Satoru SATOH488faa22007-12-16 14:00:19 -0800915 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700916 }
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700917 tcp_set_rto(sk);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700918 if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700919 goto reset;
Ilpo Järvinen86bceba2009-04-14 02:08:53 -0700920
921cwnd:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700922 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
923 tp->snd_cwnd_stamp = tcp_time_stamp;
924 return;
925
926reset:
927 /* Play conservative. If timestamps are not
928 * supported, TCP will fail to recalculate correct
929 * rtt, if initial rto is too small. FORGET ALL AND RESET!
930 */
931 if (!tp->rx_opt.saw_tstamp && tp->srtt) {
932 tp->srtt = 0;
933 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700934 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700935 }
Ilpo Järvinen86bceba2009-04-14 02:08:53 -0700936 goto cwnd;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700937}
938
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300939static void tcp_update_reordering(struct sock *sk, const int metric,
940 const int ts)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700941{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300942 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700943 if (metric > tp->reordering) {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -0700944 int mib_idx;
945
Linus Torvalds1da177e2005-04-16 15:20:36 -0700946 tp->reordering = min(TCP_MAX_REORDERING, metric);
947
948 /* This exciting event is worth to be remembered. 8) */
949 if (ts)
Pavel Emelyanov40b215e2008-07-03 01:05:41 -0700950 mib_idx = LINUX_MIB_TCPTSREORDER;
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300951 else if (tcp_is_reno(tp))
Pavel Emelyanov40b215e2008-07-03 01:05:41 -0700952 mib_idx = LINUX_MIB_TCPRENOREORDER;
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300953 else if (tcp_is_fack(tp))
Pavel Emelyanov40b215e2008-07-03 01:05:41 -0700954 mib_idx = LINUX_MIB_TCPFACKREORDER;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700955 else
Pavel Emelyanov40b215e2008-07-03 01:05:41 -0700956 mib_idx = LINUX_MIB_TCPSACKREORDER;
957
Pavel Emelyanovde0744a2008-07-16 20:31:16 -0700958 NET_INC_STATS_BH(sock_net(sk), mib_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700959#if FASTRETRANS_DEBUG > 1
960 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300961 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700962 tp->reordering,
963 tp->fackets_out,
964 tp->sacked_out,
965 tp->undo_marker ? tp->undo_retrans : 0);
966#endif
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300967 tcp_disable_fack(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700968 }
969}
970
Ilpo Järvinen006f5822008-09-20 21:20:20 -0700971/* This must be called before lost_out is incremented */
Ilpo Järvinenc8c213f2008-09-20 21:18:55 -0700972static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
973{
Ilpo Järvinen006f5822008-09-20 21:20:20 -0700974 if ((tp->retransmit_skb_hint == NULL) ||
Ilpo Järvinenc8c213f2008-09-20 21:18:55 -0700975 before(TCP_SKB_CB(skb)->seq,
976 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
Ilpo Järvinen006f5822008-09-20 21:20:20 -0700977 tp->retransmit_skb_hint = skb;
978
979 if (!tp->lost_out ||
980 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
981 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
Ilpo Järvinenc8c213f2008-09-20 21:18:55 -0700982}
983
Ilpo Järvinen41ea36e2008-09-20 21:19:22 -0700984static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
985{
986 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
987 tcp_verify_retransmit_hint(tp, skb);
988
989 tp->lost_out += tcp_skb_pcount(skb);
990 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
991 }
992}
993
Ilpo Järvinene1aa6802008-11-24 21:11:55 -0800994static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
995 struct sk_buff *skb)
Ilpo Järvinen006f5822008-09-20 21:20:20 -0700996{
997 tcp_verify_retransmit_hint(tp, skb);
998
999 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1000 tp->lost_out += tcp_skb_pcount(skb);
1001 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1002 }
1003}
1004
Linus Torvalds1da177e2005-04-16 15:20:36 -07001005/* This procedure tags the retransmission queue when SACKs arrive.
1006 *
1007 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1008 * Packets in queue with these bits set are counted in variables
1009 * sacked_out, retrans_out and lost_out, correspondingly.
1010 *
1011 * Valid combinations are:
1012 * Tag InFlight Description
1013 * 0 1 - orig segment is in flight.
1014 * S 0 - nothing flies, orig reached receiver.
1015 * L 0 - nothing flies, orig lost by net.
1016 * R 2 - both orig and retransmit are in flight.
1017 * L|R 1 - orig is lost, retransmit is in flight.
1018 * S|R 1 - orig reached receiver, retrans is still in flight.
1019 * (L|S|R is logically valid, it could occur when L|R is sacked,
1020 * but it is equivalent to plain S and code short-curcuits it to S.
1021 * L|S is logically invalid, it would mean -1 packet in flight 8))
1022 *
1023 * These 6 states form finite state machine, controlled by the following events:
1024 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1025 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1026 * 3. Loss detection event of one of three flavors:
1027 * A. Scoreboard estimator decided the packet is lost.
1028 * A'. Reno "three dupacks" marks head of queue lost.
1029 * A''. Its FACK modfication, head until snd.fack is lost.
1030 * B. SACK arrives sacking data transmitted after never retransmitted
1031 * hole was sent out.
1032 * C. SACK arrives sacking SND.NXT at the moment, when the
1033 * segment was retransmitted.
1034 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1035 *
1036 * It is pleasant to note, that state diagram turns out to be commutative,
1037 * so that we are allowed not to be bothered by order of our actions,
1038 * when multiple events arrive simultaneously. (see the function below).
1039 *
1040 * Reordering detection.
1041 * --------------------
1042 * Reordering metric is maximal distance, which a packet can be displaced
1043 * in packet stream. With SACKs we can estimate it:
1044 *
1045 * 1. SACK fills old hole and the corresponding segment was not
1046 * ever retransmitted -> reordering. Alas, we cannot use it
1047 * when segment was retransmitted.
1048 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1049 * for retransmitted and already SACKed segment -> reordering..
1050 * Both of these heuristics are not used in Loss state, when we cannot
1051 * account for retransmits accurately.
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001052 *
1053 * SACK block validation.
1054 * ----------------------
1055 *
1056 * SACK block range validation checks that the received SACK block fits to
1057 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1058 * Note that SND.UNA is not included to the range though being valid because
Ilpo Järvinen0e835332007-10-01 15:28:17 -07001059 * it means that the receiver is rather inconsistent with itself reporting
1060 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1061 * perfectly valid, however, in light of RFC2018 which explicitly states
1062 * that "SACK block MUST reflect the newest segment. Even if the newest
1063 * segment is going to be discarded ...", not that it looks very clever
1064 * in case of head skb. Due to potentional receiver driven attacks, we
1065 * choose to avoid immediate execution of a walk in write queue due to
1066 * reneging and defer head skb's loss recovery to standard loss recovery
1067 * procedure that will eventually trigger (nothing forbids us doing this).
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001068 *
1069 * Implements also blockage to start_seq wrap-around. Problem lies in the
1070 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1071 * there's no guarantee that it will be before snd_nxt (n). The problem
1072 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1073 * wrap (s_w):
1074 *
1075 * <- outs wnd -> <- wrapzone ->
1076 * u e n u_w e_w s n_w
1077 * | | | | | | |
1078 * |<------------+------+----- TCP seqno space --------------+---------->|
1079 * ...-- <2^31 ->| |<--------...
1080 * ...---- >2^31 ------>| |<--------...
1081 *
1082 * Current code wouldn't be vulnerable but it's better still to discard such
1083 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1084 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1085 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1086 * equal to the ideal case (infinite seqno space without wrap caused issues).
1087 *
1088 * With D-SACK the lower bound is extended to cover sequence space below
1089 * SND.UNA down to undo_marker, which is the last point of interest. Yet
Ryousei Takano564262c12007-10-25 23:03:52 -07001090 * again, D-SACK block must not to go across snd_una (for the same reason as
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001091 * for the normal SACK blocks, explained above). But there all simplicity
1092 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1093 * fully below undo_marker they do not affect behavior in anyway and can
1094 * therefore be safely ignored. In rare cases (which are more or less
1095 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1096 * fragmentation and packet reordering past skb's retransmission. To consider
1097 * them correctly, the acceptable range must be extended even more though
1098 * the exact amount is rather hard to quantify. However, tp->max_window can
1099 * be used as an exaggerated estimate.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001100 */
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001101static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack,
1102 u32 start_seq, u32 end_seq)
1103{
1104 /* Too far in future, or reversed (interpretation is ambiguous) */
1105 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1106 return 0;
1107
1108 /* Nasty start_seq wrap-around check (see comments above) */
1109 if (!before(start_seq, tp->snd_nxt))
1110 return 0;
1111
Ryousei Takano564262c12007-10-25 23:03:52 -07001112 /* In outstanding window? ...This is valid exit for D-SACKs too.
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001113 * start_seq == snd_una is non-sensical (see comments above)
1114 */
1115 if (after(start_seq, tp->snd_una))
1116 return 1;
1117
1118 if (!is_dsack || !tp->undo_marker)
1119 return 0;
1120
1121 /* ...Then it's D-SACK, and must reside below snd_una completely */
1122 if (!after(end_seq, tp->snd_una))
1123 return 0;
1124
1125 if (!before(start_seq, tp->undo_marker))
1126 return 1;
1127
1128 /* Too old */
1129 if (!after(end_seq, tp->undo_marker))
1130 return 0;
1131
1132 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1133 * start_seq < undo_marker and end_seq >= undo_marker.
1134 */
1135 return !before(start_seq, end_seq - tp->max_window);
1136}
1137
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001138/* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1139 * Event "C". Later note: FACK people cheated me again 8), we have to account
1140 * for reordering! Ugly, but should help.
Ilpo Järvinenf785a8e2007-10-11 17:35:41 -07001141 *
1142 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1143 * less than what is now known to be received by the other end (derived from
Ilpo Järvinen9f58f3b2007-11-15 19:42:54 -08001144 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1145 * retransmitted skbs to avoid some costly processing per ACKs.
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001146 */
Ilpo Järvinen407ef1d2007-12-02 00:47:57 +02001147static void tcp_mark_lost_retrans(struct sock *sk)
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001148{
Ilpo Järvinen9f58f3b2007-11-15 19:42:54 -08001149 const struct inet_connection_sock *icsk = inet_csk(sk);
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001150 struct tcp_sock *tp = tcp_sk(sk);
1151 struct sk_buff *skb;
Ilpo Järvinenf785a8e2007-10-11 17:35:41 -07001152 int cnt = 0;
Ilpo Järvinendf2e0142007-10-18 05:07:57 -07001153 u32 new_low_seq = tp->snd_nxt;
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001154 u32 received_upto = tcp_highest_sack_seq(tp);
Ilpo Järvinen9f58f3b2007-11-15 19:42:54 -08001155
1156 if (!tcp_is_fack(tp) || !tp->retrans_out ||
1157 !after(received_upto, tp->lost_retrans_low) ||
1158 icsk->icsk_ca_state != TCP_CA_Recovery)
Ilpo Järvinen407ef1d2007-12-02 00:47:57 +02001159 return;
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001160
1161 tcp_for_write_queue(skb, sk) {
1162 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1163
1164 if (skb == tcp_send_head(sk))
1165 break;
Ilpo Järvinenf785a8e2007-10-11 17:35:41 -07001166 if (cnt == tp->retrans_out)
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001167 break;
1168 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1169 continue;
1170
Ilpo Järvinenf785a8e2007-10-11 17:35:41 -07001171 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1172 continue;
1173
Ilpo Järvinend0af4162009-02-28 04:44:32 +00001174 /* TODO: We would like to get rid of tcp_is_fack(tp) only
1175 * constraint here (see above) but figuring out that at
1176 * least tp->reordering SACK blocks reside between ack_seq
1177 * and received_upto is not easy task to do cheaply with
1178 * the available datastructures.
1179 *
1180 * Whether FACK should check here for tp->reordering segs
1181 * in-between one could argue for either way (it would be
1182 * rather simple to implement as we could count fack_count
1183 * during the walk and do tp->fackets_out - fack_count).
1184 */
1185 if (after(received_upto, ack_seq)) {
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001186 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1187 tp->retrans_out -= tcp_skb_pcount(skb);
1188
Ilpo Järvinen006f5822008-09-20 21:20:20 -07001189 tcp_skb_mark_lost_uncond_verify(tp, skb);
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07001190 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
Ilpo Järvinenf785a8e2007-10-11 17:35:41 -07001191 } else {
Ilpo Järvinendf2e0142007-10-18 05:07:57 -07001192 if (before(ack_seq, new_low_seq))
Ilpo Järvinenb08d6cb2007-10-11 17:36:13 -07001193 new_low_seq = ack_seq;
Ilpo Järvinenf785a8e2007-10-11 17:35:41 -07001194 cnt += tcp_skb_pcount(skb);
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001195 }
1196 }
Ilpo Järvinenb08d6cb2007-10-11 17:36:13 -07001197
1198 if (tp->retrans_out)
1199 tp->lost_retrans_low = new_low_seq;
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001200}
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001201
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07001202static int tcp_check_dsack(struct sock *sk, struct sk_buff *ack_skb,
David S. Millerd06e0212007-06-18 22:43:06 -07001203 struct tcp_sack_block_wire *sp, int num_sacks,
1204 u32 prior_snd_una)
1205{
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07001206 struct tcp_sock *tp = tcp_sk(sk);
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07001207 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1208 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
David S. Millerd06e0212007-06-18 22:43:06 -07001209 int dup_sack = 0;
1210
1211 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1212 dup_sack = 1;
Ilpo Järvinene60402d2007-08-09 15:14:46 +03001213 tcp_dsack_seen(tp);
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07001214 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
David S. Millerd06e0212007-06-18 22:43:06 -07001215 } else if (num_sacks > 1) {
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07001216 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1217 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
David S. Millerd06e0212007-06-18 22:43:06 -07001218
1219 if (!after(end_seq_0, end_seq_1) &&
1220 !before(start_seq_0, start_seq_1)) {
1221 dup_sack = 1;
Ilpo Järvinene60402d2007-08-09 15:14:46 +03001222 tcp_dsack_seen(tp);
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07001223 NET_INC_STATS_BH(sock_net(sk),
1224 LINUX_MIB_TCPDSACKOFORECV);
David S. Millerd06e0212007-06-18 22:43:06 -07001225 }
1226 }
1227
1228 /* D-SACK for already forgotten data... Do dumb counting. */
1229 if (dup_sack &&
1230 !after(end_seq_0, prior_snd_una) &&
1231 after(end_seq_0, tp->undo_marker))
1232 tp->undo_retrans--;
1233
1234 return dup_sack;
1235}
1236
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001237struct tcp_sacktag_state {
1238 int reord;
1239 int fack_count;
1240 int flag;
1241};
1242
Ilpo Järvinend1935942007-10-11 17:34:25 -07001243/* Check if skb is fully within the SACK block. In presence of GSO skbs,
1244 * the incoming SACK may not exactly match but we can find smaller MSS
1245 * aligned portion of it that matches. Therefore we might need to fragment
1246 * which may fail and creates some hassle (caller must handle error case
1247 * returns).
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001248 *
1249 * FIXME: this could be merged to shift decision code
Ilpo Järvinend1935942007-10-11 17:34:25 -07001250 */
Adrian Bunk0f79efd2007-10-26 03:57:36 -07001251static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1252 u32 start_seq, u32 end_seq)
Ilpo Järvinend1935942007-10-11 17:34:25 -07001253{
1254 int in_sack, err;
1255 unsigned int pkt_len;
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001256 unsigned int mss;
Ilpo Järvinend1935942007-10-11 17:34:25 -07001257
1258 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1259 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1260
1261 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1262 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001263 mss = tcp_skb_mss(skb);
Ilpo Järvinend1935942007-10-11 17:34:25 -07001264 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1265
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001266 if (!in_sack) {
Ilpo Järvinend1935942007-10-11 17:34:25 -07001267 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001268 if (pkt_len < mss)
1269 pkt_len = mss;
1270 } else {
Ilpo Järvinend1935942007-10-11 17:34:25 -07001271 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001272 if (pkt_len < mss)
1273 return -EINVAL;
1274 }
1275
1276 /* Round if necessary so that SACKs cover only full MSSes
1277 * and/or the remaining small portion (if present)
1278 */
1279 if (pkt_len > mss) {
1280 unsigned int new_len = (pkt_len / mss) * mss;
1281 if (!in_sack && new_len < pkt_len) {
1282 new_len += mss;
1283 if (new_len > skb->len)
1284 return 0;
1285 }
1286 pkt_len = new_len;
1287 }
1288 err = tcp_fragment(sk, skb, pkt_len, mss);
Ilpo Järvinend1935942007-10-11 17:34:25 -07001289 if (err < 0)
1290 return err;
1291 }
1292
1293 return in_sack;
1294}
1295
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001296static u8 tcp_sacktag_one(struct sk_buff *skb, struct sock *sk,
1297 struct tcp_sacktag_state *state,
1298 int dup_sack, int pcount)
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001299{
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001300 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001301 u8 sacked = TCP_SKB_CB(skb)->sacked;
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001302 int fack_count = state->fack_count;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001303
1304 /* Account D-SACK for retransmitted packet. */
1305 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1306 if (after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1307 tp->undo_retrans--;
Ilpo Järvinenede9f3b2007-12-02 00:47:58 +02001308 if (sacked & TCPCB_SACKED_ACKED)
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001309 state->reord = min(fack_count, state->reord);
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001310 }
1311
1312 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1313 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001314 return sacked;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001315
1316 if (!(sacked & TCPCB_SACKED_ACKED)) {
1317 if (sacked & TCPCB_SACKED_RETRANS) {
1318 /* If the segment is not tagged as lost,
1319 * we do not clear RETRANS, believing
1320 * that retransmission is still in flight.
1321 */
1322 if (sacked & TCPCB_LOST) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001323 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001324 tp->lost_out -= pcount;
1325 tp->retrans_out -= pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001326 }
1327 } else {
1328 if (!(sacked & TCPCB_RETRANS)) {
1329 /* New sack for not retransmitted frame,
1330 * which was in hole. It is reordering.
1331 */
1332 if (before(TCP_SKB_CB(skb)->seq,
1333 tcp_highest_sack_seq(tp)))
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001334 state->reord = min(fack_count,
1335 state->reord);
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001336
1337 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1338 if (!after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark))
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001339 state->flag |= FLAG_ONLY_ORIG_SACKED;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001340 }
1341
1342 if (sacked & TCPCB_LOST) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001343 sacked &= ~TCPCB_LOST;
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001344 tp->lost_out -= pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001345 }
1346 }
1347
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001348 sacked |= TCPCB_SACKED_ACKED;
1349 state->flag |= FLAG_DATA_SACKED;
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001350 tp->sacked_out += pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001351
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001352 fack_count += pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001353
1354 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1355 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1356 before(TCP_SKB_CB(skb)->seq,
1357 TCP_SKB_CB(tp->lost_skb_hint)->seq))
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001358 tp->lost_cnt_hint += pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001359
1360 if (fack_count > tp->fackets_out)
1361 tp->fackets_out = fack_count;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001362 }
1363
1364 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1365 * frames and clear it. undo_retrans is decreased above, L|R frames
1366 * are accounted above as well.
1367 */
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001368 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1369 sacked &= ~TCPCB_SACKED_RETRANS;
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001370 tp->retrans_out -= pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001371 }
1372
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001373 return sacked;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001374}
1375
Ilpo Järvinen50133162008-12-05 22:42:41 -08001376static int tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001377 struct tcp_sacktag_state *state,
Ilpo Järvinen9ec06ff2009-03-01 00:21:36 -08001378 unsigned int pcount, int shifted, int mss,
1379 int dup_sack)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001380{
1381 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen50133162008-12-05 22:42:41 -08001382 struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001383
1384 BUG_ON(!pcount);
1385
Ilpo Järvinen92ee76b2008-11-24 21:26:56 -08001386 /* Tweak before seqno plays */
1387 if (!tcp_is_fack(tp) && tcp_is_sack(tp) && tp->lost_skb_hint &&
1388 !before(TCP_SKB_CB(tp->lost_skb_hint)->seq, TCP_SKB_CB(skb)->seq))
1389 tp->lost_cnt_hint += pcount;
1390
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001391 TCP_SKB_CB(prev)->end_seq += shifted;
1392 TCP_SKB_CB(skb)->seq += shifted;
1393
1394 skb_shinfo(prev)->gso_segs += pcount;
1395 BUG_ON(skb_shinfo(skb)->gso_segs < pcount);
1396 skb_shinfo(skb)->gso_segs -= pcount;
1397
1398 /* When we're adding to gso_segs == 1, gso_size will be zero,
1399 * in theory this shouldn't be necessary but as long as DSACK
1400 * code can come after this skb later on it's better to keep
1401 * setting gso_size to something.
1402 */
1403 if (!skb_shinfo(prev)->gso_size) {
1404 skb_shinfo(prev)->gso_size = mss;
1405 skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1406 }
1407
1408 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1409 if (skb_shinfo(skb)->gso_segs <= 1) {
1410 skb_shinfo(skb)->gso_size = 0;
1411 skb_shinfo(skb)->gso_type = 0;
1412 }
1413
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001414 /* We discard results */
Ilpo Järvinen9ec06ff2009-03-01 00:21:36 -08001415 tcp_sacktag_one(skb, sk, state, dup_sack, pcount);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001416
1417 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1418 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1419
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001420 if (skb->len > 0) {
1421 BUG_ON(!tcp_skb_pcount(skb));
Ilpo Järvinen111cc8b2008-11-24 21:27:22 -08001422 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001423 return 0;
1424 }
1425
1426 /* Whole SKB was eaten :-) */
1427
Ilpo Järvinen92ee76b2008-11-24 21:26:56 -08001428 if (skb == tp->retransmit_skb_hint)
1429 tp->retransmit_skb_hint = prev;
1430 if (skb == tp->scoreboard_skb_hint)
1431 tp->scoreboard_skb_hint = prev;
1432 if (skb == tp->lost_skb_hint) {
1433 tp->lost_skb_hint = prev;
1434 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1435 }
1436
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001437 TCP_SKB_CB(skb)->flags |= TCP_SKB_CB(prev)->flags;
1438 if (skb == tcp_highest_sack(sk))
1439 tcp_advance_highest_sack(sk, skb);
1440
1441 tcp_unlink_write_queue(skb, sk);
1442 sk_wmem_free_skb(sk, skb);
1443
Ilpo Järvinen111cc8b2008-11-24 21:27:22 -08001444 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1445
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001446 return 1;
1447}
1448
1449/* I wish gso_size would have a bit more sane initialization than
1450 * something-or-zero which complicates things
1451 */
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001452static int tcp_skb_seglen(struct sk_buff *skb)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001453{
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001454 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001455}
1456
1457/* Shifting pages past head area doesn't work */
1458static int skb_can_shift(struct sk_buff *skb)
1459{
1460 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1461}
1462
1463/* Try collapsing SACK blocks spanning across multiple skbs to a single
1464 * skb.
1465 */
1466static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001467 struct tcp_sacktag_state *state,
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001468 u32 start_seq, u32 end_seq,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001469 int dup_sack)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001470{
1471 struct tcp_sock *tp = tcp_sk(sk);
1472 struct sk_buff *prev;
1473 int mss;
1474 int pcount = 0;
1475 int len;
1476 int in_sack;
1477
1478 if (!sk_can_gso(sk))
1479 goto fallback;
1480
1481 /* Normally R but no L won't result in plain S */
1482 if (!dup_sack &&
Ilpo Järvinen9969ca52008-12-05 22:41:06 -08001483 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001484 goto fallback;
1485 if (!skb_can_shift(skb))
1486 goto fallback;
1487 /* This frame is about to be dropped (was ACKed). */
1488 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1489 goto fallback;
1490
1491 /* Can only happen with delayed DSACK + discard craziness */
1492 if (unlikely(skb == tcp_write_queue_head(sk)))
1493 goto fallback;
1494 prev = tcp_write_queue_prev(sk, skb);
1495
1496 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1497 goto fallback;
1498
1499 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1500 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1501
1502 if (in_sack) {
1503 len = skb->len;
1504 pcount = tcp_skb_pcount(skb);
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001505 mss = tcp_skb_seglen(skb);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001506
1507 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1508 * drop this restriction as unnecessary
1509 */
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001510 if (mss != tcp_skb_seglen(prev))
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001511 goto fallback;
1512 } else {
1513 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1514 goto noop;
1515 /* CHECKME: This is non-MSS split case only?, this will
1516 * cause skipped skbs due to advancing loop btw, original
1517 * has that feature too
1518 */
1519 if (tcp_skb_pcount(skb) <= 1)
1520 goto noop;
1521
1522 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1523 if (!in_sack) {
1524 /* TODO: head merge to next could be attempted here
1525 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1526 * though it might not be worth of the additional hassle
1527 *
1528 * ...we can probably just fallback to what was done
1529 * previously. We could try merging non-SACKed ones
1530 * as well but it probably isn't going to buy off
1531 * because later SACKs might again split them, and
1532 * it would make skb timestamp tracking considerably
1533 * harder problem.
1534 */
1535 goto fallback;
1536 }
1537
1538 len = end_seq - TCP_SKB_CB(skb)->seq;
1539 BUG_ON(len < 0);
1540 BUG_ON(len > skb->len);
1541
1542 /* MSS boundaries should be honoured or else pcount will
1543 * severely break even though it makes things bit trickier.
1544 * Optimize common case to avoid most of the divides
1545 */
1546 mss = tcp_skb_mss(skb);
1547
1548 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1549 * drop this restriction as unnecessary
1550 */
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001551 if (mss != tcp_skb_seglen(prev))
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001552 goto fallback;
1553
1554 if (len == mss) {
1555 pcount = 1;
1556 } else if (len < mss) {
1557 goto noop;
1558 } else {
1559 pcount = len / mss;
1560 len = pcount * mss;
1561 }
1562 }
1563
1564 if (!skb_shift(prev, skb, len))
1565 goto fallback;
Ilpo Järvinen9ec06ff2009-03-01 00:21:36 -08001566 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001567 goto out;
1568
1569 /* Hole filled allows collapsing with the next as well, this is very
1570 * useful when hole on every nth skb pattern happens
1571 */
1572 if (prev == tcp_write_queue_tail(sk))
1573 goto out;
1574 skb = tcp_write_queue_next(sk, prev);
1575
Ilpo Järvinenf0bc52f2008-12-05 22:40:47 -08001576 if (!skb_can_shift(skb) ||
1577 (skb == tcp_send_head(sk)) ||
1578 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001579 (mss != tcp_skb_seglen(skb)))
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001580 goto out;
1581
1582 len = skb->len;
1583 if (skb_shift(prev, skb, len)) {
1584 pcount += tcp_skb_pcount(skb);
Ilpo Järvinen9ec06ff2009-03-01 00:21:36 -08001585 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001586 }
1587
1588out:
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001589 state->fack_count += pcount;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001590 return prev;
1591
1592noop:
1593 return skb;
1594
1595fallback:
Ilpo Järvinen111cc8b2008-11-24 21:27:22 -08001596 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001597 return NULL;
1598}
1599
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001600static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1601 struct tcp_sack_block *next_dup,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001602 struct tcp_sacktag_state *state,
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001603 u32 start_seq, u32 end_seq,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001604 int dup_sack_in)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001605{
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001606 struct tcp_sock *tp = tcp_sk(sk);
1607 struct sk_buff *tmp;
1608
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001609 tcp_for_write_queue_from(skb, sk) {
1610 int in_sack = 0;
1611 int dup_sack = dup_sack_in;
1612
1613 if (skb == tcp_send_head(sk))
1614 break;
1615
1616 /* queue is in-order => we can short-circuit the walk early */
1617 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1618 break;
1619
1620 if ((next_dup != NULL) &&
1621 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1622 in_sack = tcp_match_skb_to_sack(sk, skb,
1623 next_dup->start_seq,
1624 next_dup->end_seq);
1625 if (in_sack > 0)
1626 dup_sack = 1;
1627 }
1628
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001629 /* skb reference here is a bit tricky to get right, since
1630 * shifting can eat and free both this skb and the next,
1631 * so not even _safe variant of the loop is enough.
1632 */
1633 if (in_sack <= 0) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001634 tmp = tcp_shift_skb_data(sk, skb, state,
1635 start_seq, end_seq, dup_sack);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001636 if (tmp != NULL) {
1637 if (tmp != skb) {
1638 skb = tmp;
1639 continue;
1640 }
1641
1642 in_sack = 0;
1643 } else {
1644 in_sack = tcp_match_skb_to_sack(sk, skb,
1645 start_seq,
1646 end_seq);
1647 }
1648 }
1649
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001650 if (unlikely(in_sack < 0))
1651 break;
1652
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001653 if (in_sack) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001654 TCP_SKB_CB(skb)->sacked = tcp_sacktag_one(skb, sk,
1655 state,
1656 dup_sack,
1657 tcp_skb_pcount(skb));
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001658
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001659 if (!before(TCP_SKB_CB(skb)->seq,
1660 tcp_highest_sack_seq(tp)))
1661 tcp_advance_highest_sack(sk, skb);
1662 }
1663
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001664 state->fack_count += tcp_skb_pcount(skb);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001665 }
1666 return skb;
1667}
1668
1669/* Avoid all extra work that is being done by sacktag while walking in
1670 * a normal way
1671 */
1672static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001673 struct tcp_sacktag_state *state,
1674 u32 skip_to_seq)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001675{
1676 tcp_for_write_queue_from(skb, sk) {
1677 if (skb == tcp_send_head(sk))
1678 break;
1679
Ilpo Järvinene8bae272008-11-24 21:12:28 -08001680 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001681 break;
Ilpo Järvinend152a7d2008-03-03 12:10:16 -08001682
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001683 state->fack_count += tcp_skb_pcount(skb);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001684 }
1685 return skb;
1686}
1687
1688static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1689 struct sock *sk,
1690 struct tcp_sack_block *next_dup,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001691 struct tcp_sacktag_state *state,
1692 u32 skip_to_seq)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001693{
1694 if (next_dup == NULL)
1695 return skb;
1696
1697 if (before(next_dup->start_seq, skip_to_seq)) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001698 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1699 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1700 next_dup->start_seq, next_dup->end_seq,
1701 1);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001702 }
1703
1704 return skb;
1705}
1706
1707static int tcp_sack_cache_ok(struct tcp_sock *tp, struct tcp_sack_block *cache)
1708{
1709 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1710}
1711
Linus Torvalds1da177e2005-04-16 15:20:36 -07001712static int
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001713tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb,
1714 u32 prior_snd_una)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001715{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001716 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001717 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo9c702202007-04-25 18:04:18 -07001718 unsigned char *ptr = (skb_transport_header(ack_skb) +
1719 TCP_SKB_CB(ack_skb)->sacked);
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001720 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
Adam Langley4389dde2008-07-19 00:07:02 -07001721 struct tcp_sack_block sp[TCP_NUM_SACKS];
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001722 struct tcp_sack_block *cache;
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001723 struct tcp_sacktag_state state;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001724 struct sk_buff *skb;
Adam Langley4389dde2008-07-19 00:07:02 -07001725 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001726 int used_sacks;
Ilpo Järvinen7769f402007-06-15 15:14:04 -07001727 int found_dup_sack = 0;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001728 int i, j;
Baruch Evenfda03fb2007-02-04 23:35:57 -08001729 int first_sack_index;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001730
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001731 state.flag = 0;
1732 state.reord = tp->packets_out;
1733
Ilpo Järvinend738cd82007-03-24 21:03:23 -07001734 if (!tp->sacked_out) {
Ilpo Järvinende83c052007-10-07 23:37:55 -07001735 if (WARN_ON(tp->fackets_out))
1736 tp->fackets_out = 0;
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001737 tcp_highest_sack_reset(sk);
Ilpo Järvinend738cd82007-03-24 21:03:23 -07001738 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001739
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07001740 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
David S. Millerd06e0212007-06-18 22:43:06 -07001741 num_sacks, prior_snd_una);
1742 if (found_dup_sack)
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001743 state.flag |= FLAG_DSACKING_ACK;
Baruch Even6f746512007-02-04 23:36:42 -08001744
1745 /* Eliminate too old ACKs, but take into
1746 * account more or less fresh ones, they can
1747 * contain valid SACK info.
1748 */
1749 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1750 return 0;
1751
Ilpo Järvinen96a2d412007-11-14 15:47:18 -08001752 if (!tp->packets_out)
1753 goto out;
1754
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001755 used_sacks = 0;
1756 first_sack_index = 0;
1757 for (i = 0; i < num_sacks; i++) {
1758 int dup_sack = !i && found_dup_sack;
1759
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07001760 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1761 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001762
1763 if (!tcp_is_sackblock_valid(tp, dup_sack,
1764 sp[used_sacks].start_seq,
1765 sp[used_sacks].end_seq)) {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001766 int mib_idx;
1767
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001768 if (dup_sack) {
1769 if (!tp->undo_marker)
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001770 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001771 else
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001772 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001773 } else {
1774 /* Don't count olds caused by ACK reordering */
1775 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1776 !after(sp[used_sacks].end_seq, tp->snd_una))
1777 continue;
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001778 mib_idx = LINUX_MIB_TCPSACKDISCARD;
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001779 }
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001780
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07001781 NET_INC_STATS_BH(sock_net(sk), mib_idx);
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001782 if (i == 0)
1783 first_sack_index = -1;
1784 continue;
1785 }
1786
1787 /* Ignore very old stuff early */
1788 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1789 continue;
1790
1791 used_sacks++;
1792 }
1793
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001794 /* order SACK blocks to allow in order walk of the retrans queue */
1795 for (i = used_sacks - 1; i > 0; i--) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001796 for (j = 0; j < i; j++) {
1797 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
Ilpo Järvinena0bffff2009-03-21 13:36:17 -07001798 swap(sp[j], sp[j + 1]);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001799
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001800 /* Track where the first SACK block goes to */
1801 if (j == first_sack_index)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001802 first_sack_index = j + 1;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001803 }
1804 }
1805 }
1806
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001807 skb = tcp_write_queue_head(sk);
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001808 state.fack_count = 0;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001809 i = 0;
1810
1811 if (!tp->sacked_out) {
1812 /* It's already past, so skip checking against it */
1813 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1814 } else {
1815 cache = tp->recv_sack_cache;
1816 /* Skip empty blocks in at head of the cache */
1817 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1818 !cache->end_seq)
1819 cache++;
Baruch Evenfda03fb2007-02-04 23:35:57 -08001820 }
1821
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001822 while (i < used_sacks) {
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001823 u32 start_seq = sp[i].start_seq;
1824 u32 end_seq = sp[i].end_seq;
Ilpo Järvinen7769f402007-06-15 15:14:04 -07001825 int dup_sack = (found_dup_sack && (i == first_sack_index));
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001826 struct tcp_sack_block *next_dup = NULL;
Ilpo Järvinene56d6cd2007-11-01 00:09:37 -07001827
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001828 if (found_dup_sack && ((i + 1) == first_sack_index))
1829 next_dup = &sp[i + 1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001830
1831 /* Event "B" in the comment above. */
1832 if (after(end_seq, tp->high_seq))
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001833 state.flag |= FLAG_DATA_LOST;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001834
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001835 /* Skip too early cached blocks */
1836 while (tcp_sack_cache_ok(tp, cache) &&
1837 !before(start_seq, cache->end_seq))
1838 cache++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001839
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001840 /* Can skip some work by looking recv_sack_cache? */
1841 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1842 after(end_seq, cache->start_seq)) {
David S. Millerfe067e82007-03-07 12:12:44 -08001843
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001844 /* Head todo? */
1845 if (before(start_seq, cache->start_seq)) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001846 skb = tcp_sacktag_skip(skb, sk, &state,
1847 start_seq);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001848 skb = tcp_sacktag_walk(skb, sk, next_dup,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001849 &state,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001850 start_seq,
1851 cache->start_seq,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001852 dup_sack);
Baruch Evenfda03fb2007-02-04 23:35:57 -08001853 }
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001854
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001855 /* Rest of the block already fully processed? */
Ilpo Järvinen20de20b2007-11-16 16:17:05 -08001856 if (!after(end_seq, cache->end_seq))
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001857 goto advance_sp;
Ilpo Järvinen20de20b2007-11-16 16:17:05 -08001858
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001859 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001860 &state,
1861 cache->end_seq);
Ilpo Järvinene56d6cd2007-11-01 00:09:37 -07001862
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001863 /* ...tail remains todo... */
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001864 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
Ilpo Järvinen20de20b2007-11-16 16:17:05 -08001865 /* ...but better entrypoint exists! */
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001866 skb = tcp_highest_sack(sk);
1867 if (skb == NULL)
1868 break;
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001869 state.fack_count = tp->fackets_out;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001870 cache++;
1871 goto walk;
1872 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001873
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001874 skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001875 /* Check overlap against next cached too (past this one already) */
1876 cache++;
1877 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001878 }
Ilpo Järvinenfbd52eb2007-11-10 21:24:19 -08001879
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001880 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1881 skb = tcp_highest_sack(sk);
1882 if (skb == NULL)
1883 break;
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001884 state.fack_count = tp->fackets_out;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001885 }
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001886 skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001887
1888walk:
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001889 skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1890 start_seq, end_seq, dup_sack);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001891
1892advance_sp:
Ilpo Järvinenfbd52eb2007-11-10 21:24:19 -08001893 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1894 * due to in-order walk
1895 */
1896 if (after(end_seq, tp->frto_highmark))
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001897 state.flag &= ~FLAG_ONLY_ORIG_SACKED;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001898
1899 i++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001900 }
1901
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001902 /* Clear the head of the cache sack blocks so we can skip it next time */
1903 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1904 tp->recv_sack_cache[i].start_seq = 0;
1905 tp->recv_sack_cache[i].end_seq = 0;
1906 }
1907 for (j = 0; j < used_sacks; j++)
1908 tp->recv_sack_cache[i++] = sp[j];
1909
Ilpo Järvinen407ef1d2007-12-02 00:47:57 +02001910 tcp_mark_lost_retrans(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001911
Ilpo Järvinen86426c22007-08-09 14:45:17 +03001912 tcp_verify_left_out(tp);
1913
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001914 if ((state.reord < tp->fackets_out) &&
Ilpo Järvinenf5771112007-11-15 19:35:11 -08001915 ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) &&
Ilpo Järvinenc5e7af02007-02-23 16:22:06 -08001916 (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001917 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001918
Ilpo Järvinen96a2d412007-11-14 15:47:18 -08001919out:
1920
Linus Torvalds1da177e2005-04-16 15:20:36 -07001921#if FASTRETRANS_DEBUG > 0
Ilpo Järvinen547b7922008-07-25 21:43:18 -07001922 WARN_ON((int)tp->sacked_out < 0);
1923 WARN_ON((int)tp->lost_out < 0);
1924 WARN_ON((int)tp->retrans_out < 0);
1925 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001926#endif
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001927 return state.flag;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001928}
1929
Ilpo Järvinen882beba2008-04-07 22:33:07 -07001930/* Limits sacked_out so that sum with lost_out isn't ever larger than
1931 * packets_out. Returns zero if sacked_out adjustement wasn't necessary.
Ilpo Järvinen30935cf2007-02-21 23:01:36 -08001932 */
Ilpo Järvinen8eecaba2008-11-25 13:45:29 -08001933static int tcp_limit_reno_sacked(struct tcp_sock *tp)
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001934{
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001935 u32 holes;
1936
1937 holes = max(tp->lost_out, 1U);
1938 holes = min(holes, tp->packets_out);
1939
1940 if ((tp->sacked_out + holes) > tp->packets_out) {
1941 tp->sacked_out = tp->packets_out - holes;
Ilpo Järvinen882beba2008-04-07 22:33:07 -07001942 return 1;
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001943 }
Ilpo Järvinen882beba2008-04-07 22:33:07 -07001944 return 0;
1945}
1946
1947/* If we receive more dupacks than we expected counting segments
1948 * in assumption of absent reordering, interpret this as reordering.
1949 * The only another reason could be bug in receiver TCP.
1950 */
1951static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1952{
1953 struct tcp_sock *tp = tcp_sk(sk);
1954 if (tcp_limit_reno_sacked(tp))
1955 tcp_update_reordering(sk, tp->packets_out + addend, 0);
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001956}
1957
1958/* Emulate SACKs for SACKless connection: account for a new dupack. */
1959
1960static void tcp_add_reno_sack(struct sock *sk)
1961{
1962 struct tcp_sock *tp = tcp_sk(sk);
1963 tp->sacked_out++;
1964 tcp_check_reno_reordering(sk, 0);
Ilpo Järvinen005903b2007-08-09 14:44:16 +03001965 tcp_verify_left_out(tp);
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001966}
1967
1968/* Account for ACK, ACKing some data in Reno Recovery phase. */
1969
1970static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1971{
1972 struct tcp_sock *tp = tcp_sk(sk);
1973
1974 if (acked > 0) {
1975 /* One ACK acked hole. The rest eat duplicate ACKs. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001976 if (acked - 1 >= tp->sacked_out)
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001977 tp->sacked_out = 0;
1978 else
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001979 tp->sacked_out -= acked - 1;
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001980 }
1981 tcp_check_reno_reordering(sk, acked);
Ilpo Järvinen005903b2007-08-09 14:44:16 +03001982 tcp_verify_left_out(tp);
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001983}
1984
1985static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1986{
1987 tp->sacked_out = 0;
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001988}
1989
Ilpo Järvinen62ab2222008-05-08 01:09:11 -07001990static int tcp_is_sackfrto(const struct tcp_sock *tp)
1991{
1992 return (sysctl_tcp_frto == 0x2) && !tcp_is_reno(tp);
1993}
1994
Ilpo Järvinen95eacd22007-10-01 15:27:42 -07001995/* F-RTO can only be used if TCP has never retransmitted anything other than
1996 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1997 */
Ilpo Järvinen46d0de42007-02-21 23:10:39 -08001998int tcp_use_frto(struct sock *sk)
Ilpo Järvinenbdaae172007-02-21 22:59:58 -08001999{
2000 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen6adb4f72008-04-07 22:33:57 -07002001 const struct inet_connection_sock *icsk = inet_csk(sk);
Ilpo Järvinen46d0de42007-02-21 23:10:39 -08002002 struct sk_buff *skb;
Ilpo Järvinenbdaae172007-02-21 22:59:58 -08002003
Ilpo Järvinen575ee712007-04-30 00:39:55 -07002004 if (!sysctl_tcp_frto)
Ilpo Järvinen46d0de42007-02-21 23:10:39 -08002005 return 0;
2006
Ilpo Järvinen6adb4f72008-04-07 22:33:57 -07002007 /* MTU probe and F-RTO won't really play nicely along currently */
2008 if (icsk->icsk_mtup.probe_size)
2009 return 0;
2010
Ilpo Järvinen62ab2222008-05-08 01:09:11 -07002011 if (tcp_is_sackfrto(tp))
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08002012 return 1;
2013
Ilpo Järvinen46d0de42007-02-21 23:10:39 -08002014 /* Avoid expensive walking of rexmit queue if possible */
2015 if (tp->retrans_out > 1)
2016 return 0;
2017
David S. Millerfe067e82007-03-07 12:12:44 -08002018 skb = tcp_write_queue_head(sk);
David S. Miller28e34872008-09-23 02:51:41 -07002019 if (tcp_skb_is_last(sk, skb))
2020 return 1;
David S. Millerfe067e82007-03-07 12:12:44 -08002021 skb = tcp_write_queue_next(sk, skb); /* Skips head */
2022 tcp_for_write_queue_from(skb, sk) {
2023 if (skb == tcp_send_head(sk))
2024 break;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002025 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
Ilpo Järvinen46d0de42007-02-21 23:10:39 -08002026 return 0;
2027 /* Short-circuit when first non-SACKed skb has been checked */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002028 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
Ilpo Järvinen46d0de42007-02-21 23:10:39 -08002029 break;
2030 }
2031 return 1;
Ilpo Järvinenbdaae172007-02-21 22:59:58 -08002032}
2033
Ilpo Järvinen30935cf2007-02-21 23:01:36 -08002034/* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
2035 * recovery a bit and use heuristics in tcp_process_frto() to detect if
Ilpo Järvinend1a54c62007-02-21 23:11:57 -08002036 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
2037 * keep retrans_out counting accurate (with SACK F-RTO, other than head
2038 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
2039 * bits are handled if the Loss state is really to be entered (in
2040 * tcp_enter_frto_loss).
Ilpo Järvinen7487c482007-02-21 23:02:30 -08002041 *
2042 * Do like tcp_enter_loss() would; when RTO expires the second time it
2043 * does:
2044 * "Reduce ssthresh if it has not yet been made inside this window."
Linus Torvalds1da177e2005-04-16 15:20:36 -07002045 */
2046void tcp_enter_frto(struct sock *sk)
2047{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002048 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002049 struct tcp_sock *tp = tcp_sk(sk);
2050 struct sk_buff *skb;
2051
Ilpo Järvinen7487c482007-02-21 23:02:30 -08002052 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002053 tp->snd_una == tp->high_seq ||
Ilpo Järvinen7487c482007-02-21 23:02:30 -08002054 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
2055 !icsk->icsk_retransmits)) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002056 tp->prior_ssthresh = tcp_current_ssthresh(sk);
Ilpo Järvinen66e93e42007-02-21 23:13:47 -08002057 /* Our state is too optimistic in ssthresh() call because cwnd
Ryousei Takano564262c12007-10-25 23:03:52 -07002058 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
Ilpo Järvinen66e93e42007-02-21 23:13:47 -08002059 * recovery has not yet completed. Pattern would be this: RTO,
2060 * Cumulative ACK, RTO (2xRTO for the same segment does not end
2061 * up here twice).
2062 * RFC4138 should be more specific on what to do, even though
2063 * RTO is quite unlikely to occur after the first Cumulative ACK
2064 * due to back-off and complexity of triggering events ...
2065 */
2066 if (tp->frto_counter) {
2067 u32 stored_cwnd;
2068 stored_cwnd = tp->snd_cwnd;
2069 tp->snd_cwnd = 2;
2070 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2071 tp->snd_cwnd = stored_cwnd;
2072 } else {
2073 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2074 }
2075 /* ... in theory, cong.control module could do "any tricks" in
2076 * ssthresh(), which means that ca_state, lost bits and lost_out
2077 * counter would have to be faked before the call occurs. We
2078 * consider that too expensive, unlikely and hacky, so modules
2079 * using these in ssthresh() must deal these incompatibility
2080 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
2081 */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002082 tcp_ca_event(sk, CA_EVENT_FRTO);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002083 }
2084
Linus Torvalds1da177e2005-04-16 15:20:36 -07002085 tp->undo_marker = tp->snd_una;
2086 tp->undo_retrans = 0;
2087
David S. Millerfe067e82007-03-07 12:12:44 -08002088 skb = tcp_write_queue_head(sk);
Ilpo Järvinen009a2e32007-09-20 11:34:38 -07002089 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2090 tp->undo_marker = 0;
Ilpo Järvinend1a54c62007-02-21 23:11:57 -08002091 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
Ilpo Järvinen522e7542007-02-21 22:54:52 -08002092 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
Ilpo Järvinend1a54c62007-02-21 23:11:57 -08002093 tp->retrans_out -= tcp_skb_pcount(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002094 }
Ilpo Järvinen005903b2007-08-09 14:44:16 +03002095 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002096
Ilpo Järvinen746aa322007-11-13 21:01:23 -08002097 /* Too bad if TCP was application limited */
2098 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
2099
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08002100 /* Earlier loss recovery underway (see RFC4138; Appendix B).
2101 * The last condition is necessary at least in tp->frto_counter case.
2102 */
Ilpo Järvinen62ab2222008-05-08 01:09:11 -07002103 if (tcp_is_sackfrto(tp) && (tp->frto_counter ||
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08002104 ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
2105 after(tp->high_seq, tp->snd_una)) {
2106 tp->frto_highmark = tp->high_seq;
2107 } else {
2108 tp->frto_highmark = tp->snd_nxt;
2109 }
Ilpo Järvinen7b0eb222007-02-21 23:03:35 -08002110 tcp_set_ca_state(sk, TCP_CA_Disorder);
2111 tp->high_seq = tp->snd_nxt;
Ilpo Järvinen7487c482007-02-21 23:02:30 -08002112 tp->frto_counter = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002113}
2114
2115/* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
2116 * which indicates that we should follow the traditional RTO recovery,
2117 * i.e. mark everything lost and do go-back-N retransmission.
2118 */
Ilpo Järvinend1a54c62007-02-21 23:11:57 -08002119static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002120{
2121 struct tcp_sock *tp = tcp_sk(sk);
2122 struct sk_buff *skb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002123
Linus Torvalds1da177e2005-04-16 15:20:36 -07002124 tp->lost_out = 0;
Ilpo Järvinend1a54c62007-02-21 23:11:57 -08002125 tp->retrans_out = 0;
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002126 if (tcp_is_reno(tp))
Ilpo Järvinen9bff40f2007-05-27 01:53:49 -07002127 tcp_reset_reno_sack(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002128
David S. Millerfe067e82007-03-07 12:12:44 -08002129 tcp_for_write_queue(skb, sk) {
2130 if (skb == tcp_send_head(sk))
2131 break;
Ilpo Järvinen23aeeec2007-11-13 21:03:13 -08002132
2133 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
Ilpo Järvinend1a54c62007-02-21 23:11:57 -08002134 /*
2135 * Count the retransmission made on RTO correctly (only when
2136 * waiting for the first ACK and did not get it)...
2137 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002138 if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) {
Ilpo Järvinen0a9f2a42007-07-15 00:19:29 -07002139 /* For some reason this R-bit might get cleared? */
2140 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
2141 tp->retrans_out += tcp_skb_pcount(skb);
Ilpo Järvinend1a54c62007-02-21 23:11:57 -08002142 /* ...enter this if branch just for the first segment */
2143 flag |= FLAG_DATA_ACKED;
2144 } else {
Ilpo Järvinen009a2e32007-09-20 11:34:38 -07002145 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
2146 tp->undo_marker = 0;
Ilpo Järvinen23aeeec2007-11-13 21:03:13 -08002147 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
Ilpo Järvinend1a54c62007-02-21 23:11:57 -08002148 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002149
Ilpo Järvinen79d445162008-05-13 02:54:19 -07002150 /* Marking forward transmissions that were made after RTO lost
2151 * can cause unnecessary retransmissions in some scenarios,
2152 * SACK blocks will mitigate that in some but not in all cases.
2153 * We used to not mark them but it was causing break-ups with
2154 * receivers that do only in-order receival.
2155 *
2156 * TODO: we could detect presence of such receiver and select
2157 * different behavior per flow.
2158 */
2159 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
Ilpo Järvinen9bff40f2007-05-27 01:53:49 -07002160 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2161 tp->lost_out += tcp_skb_pcount(skb);
Ilpo Järvinen006f5822008-09-20 21:20:20 -07002162 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002163 }
2164 }
Ilpo Järvinen005903b2007-08-09 14:44:16 +03002165 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002166
Ilpo Järvinen95c49222007-02-21 23:05:18 -08002167 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002168 tp->snd_cwnd_cnt = 0;
2169 tp->snd_cwnd_stamp = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002170 tp->frto_counter = 0;
Ilpo Järvinen16e90682007-10-11 17:32:31 -07002171 tp->bytes_acked = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002172
2173 tp->reordering = min_t(unsigned int, tp->reordering,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002174 sysctl_tcp_reordering);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002175 tcp_set_ca_state(sk, TCP_CA_Loss);
Ilpo Järvinen79d445162008-05-13 02:54:19 -07002176 tp->high_seq = tp->snd_nxt;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002177 TCP_ECN_queue_cwr(tp);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002178
Ilpo Järvinen64edc272008-09-20 21:18:32 -07002179 tcp_clear_all_retrans_hints(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002180}
2181
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07002182static void tcp_clear_retrans_partial(struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002183{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002184 tp->retrans_out = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002185 tp->lost_out = 0;
2186
2187 tp->undo_marker = 0;
2188 tp->undo_retrans = 0;
2189}
2190
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07002191void tcp_clear_retrans(struct tcp_sock *tp)
2192{
2193 tcp_clear_retrans_partial(tp);
2194
2195 tp->fackets_out = 0;
2196 tp->sacked_out = 0;
2197}
2198
Linus Torvalds1da177e2005-04-16 15:20:36 -07002199/* Enter Loss state. If "how" is not zero, forget all SACK information
2200 * and reset tags completely, otherwise preserve SACKs. If receiver
2201 * dropped its ofo queue, we will know this due to reneging detection.
2202 */
2203void tcp_enter_loss(struct sock *sk, int how)
2204{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002205 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002206 struct tcp_sock *tp = tcp_sk(sk);
2207 struct sk_buff *skb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002208
2209 /* Reduce ssthresh if it has not yet been made inside this window. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002210 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
2211 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2212 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2213 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2214 tcp_ca_event(sk, CA_EVENT_LOSS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002215 }
2216 tp->snd_cwnd = 1;
2217 tp->snd_cwnd_cnt = 0;
2218 tp->snd_cwnd_stamp = tcp_time_stamp;
2219
Stephen Hemminger9772efb2005-11-10 17:09:53 -08002220 tp->bytes_acked = 0;
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07002221 tcp_clear_retrans_partial(tp);
2222
2223 if (tcp_is_reno(tp))
2224 tcp_reset_reno_sack(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002225
Ilpo Järvinenb7689202007-09-20 11:37:19 -07002226 if (!how) {
2227 /* Push undo marker, if it was plain RTO and nothing
2228 * was retransmitted. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002229 tp->undo_marker = tp->snd_una;
Ilpo Järvinenb7689202007-09-20 11:37:19 -07002230 } else {
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07002231 tp->sacked_out = 0;
2232 tp->fackets_out = 0;
Ilpo Järvinenb7689202007-09-20 11:37:19 -07002233 }
Ilpo Järvinen64edc272008-09-20 21:18:32 -07002234 tcp_clear_all_retrans_hints(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002235
David S. Millerfe067e82007-03-07 12:12:44 -08002236 tcp_for_write_queue(skb, sk) {
2237 if (skb == tcp_send_head(sk))
2238 break;
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07002239
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002240 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002241 tp->undo_marker = 0;
2242 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
2243 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
2244 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2245 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2246 tp->lost_out += tcp_skb_pcount(skb);
Ilpo Järvinen006f5822008-09-20 21:20:20 -07002247 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002248 }
2249 }
Ilpo Järvinen005903b2007-08-09 14:44:16 +03002250 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002251
2252 tp->reordering = min_t(unsigned int, tp->reordering,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002253 sysctl_tcp_reordering);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002254 tcp_set_ca_state(sk, TCP_CA_Loss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002255 tp->high_seq = tp->snd_nxt;
2256 TCP_ECN_queue_cwr(tp);
Ryousei Takano564262c12007-10-25 23:03:52 -07002257 /* Abort F-RTO algorithm if one is in progress */
Ilpo Järvinen580e5722007-05-19 13:56:57 -07002258 tp->frto_counter = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002259}
2260
Ilpo Järvinencadbd032007-12-31 04:49:21 -08002261/* If ACK arrived pointing to a remembered SACK, it means that our
2262 * remembered SACKs do not reflect real state of receiver i.e.
2263 * receiver _host_ is heavily congested (or buggy).
2264 *
2265 * Do processing similar to RTO timeout.
2266 */
2267static int tcp_check_sack_reneging(struct sock *sk, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002268{
Ilpo Järvinencadbd032007-12-31 04:49:21 -08002269 if (flag & FLAG_SACK_RENEGING) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002270 struct inet_connection_sock *icsk = inet_csk(sk);
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07002271 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002272
2273 tcp_enter_loss(sk, 1);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002274 icsk->icsk_retransmits++;
David S. Millerfe067e82007-03-07 12:12:44 -08002275 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002276 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002277 icsk->icsk_rto, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002278 return 1;
2279 }
2280 return 0;
2281}
2282
2283static inline int tcp_fackets_out(struct tcp_sock *tp)
2284{
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002285 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002286}
2287
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002288/* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2289 * counter when SACK is enabled (without SACK, sacked_out is used for
2290 * that purpose).
2291 *
2292 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2293 * segments up to the highest received SACK block so far and holes in
2294 * between them.
2295 *
2296 * With reordering, holes may still be in flight, so RFC3517 recovery
2297 * uses pure sacked_out (total number of SACKed segments) even though
2298 * it violates the RFC that uses duplicate ACKs, often these are equal
2299 * but when e.g. out-of-window ACKs or packet duplication occurs,
2300 * they differ. Since neither occurs due to loss, TCP should really
2301 * ignore them.
2302 */
Andreas Petlundea84e552009-10-27 03:27:21 +00002303static inline int tcp_dupack_heuristics(struct tcp_sock *tp)
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002304{
2305 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2306}
2307
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002308static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002309{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002310 return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002311}
2312
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002313static inline int tcp_head_timedout(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002314{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002315 struct tcp_sock *tp = tcp_sk(sk);
2316
Linus Torvalds1da177e2005-04-16 15:20:36 -07002317 return tp->packets_out &&
David S. Millerfe067e82007-03-07 12:12:44 -08002318 tcp_skb_timedout(sk, tcp_write_queue_head(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002319}
2320
2321/* Linux NewReno/SACK/FACK/ECN state machine.
2322 * --------------------------------------
2323 *
2324 * "Open" Normal state, no dubious events, fast path.
2325 * "Disorder" In all the respects it is "Open",
2326 * but requires a bit more attention. It is entered when
2327 * we see some SACKs or dupacks. It is split of "Open"
2328 * mainly to move some processing from fast path to slow one.
2329 * "CWR" CWND was reduced due to some Congestion Notification event.
2330 * It can be ECN, ICMP source quench, local device congestion.
2331 * "Recovery" CWND was reduced, we are fast-retransmitting.
2332 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2333 *
2334 * tcp_fastretrans_alert() is entered:
2335 * - each incoming ACK, if state is not "Open"
2336 * - when arrived ACK is unusual, namely:
2337 * * SACK
2338 * * Duplicate ACK.
2339 * * ECN ECE.
2340 *
2341 * Counting packets in flight is pretty simple.
2342 *
2343 * in_flight = packets_out - left_out + retrans_out
2344 *
2345 * packets_out is SND.NXT-SND.UNA counted in packets.
2346 *
2347 * retrans_out is number of retransmitted segments.
2348 *
2349 * left_out is number of segments left network, but not ACKed yet.
2350 *
2351 * left_out = sacked_out + lost_out
2352 *
2353 * sacked_out: Packets, which arrived to receiver out of order
2354 * and hence not ACKed. With SACKs this number is simply
2355 * amount of SACKed data. Even without SACKs
2356 * it is easy to give pretty reliable estimate of this number,
2357 * counting duplicate ACKs.
2358 *
2359 * lost_out: Packets lost by network. TCP has no explicit
2360 * "loss notification" feedback from network (for now).
2361 * It means that this number can be only _guessed_.
2362 * Actually, it is the heuristics to predict lossage that
2363 * distinguishes different algorithms.
2364 *
2365 * F.e. after RTO, when all the queue is considered as lost,
2366 * lost_out = packets_out and in_flight = retrans_out.
2367 *
2368 * Essentially, we have now two algorithms counting
2369 * lost packets.
2370 *
2371 * FACK: It is the simplest heuristics. As soon as we decided
2372 * that something is lost, we decide that _all_ not SACKed
2373 * packets until the most forward SACK are lost. I.e.
2374 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2375 * It is absolutely correct estimate, if network does not reorder
2376 * packets. And it loses any connection to reality when reordering
2377 * takes place. We use FACK by default until reordering
2378 * is suspected on the path to this destination.
2379 *
2380 * NewReno: when Recovery is entered, we assume that one segment
2381 * is lost (classic Reno). While we are in Recovery and
2382 * a partial ACK arrives, we assume that one more packet
2383 * is lost (NewReno). This heuristics are the same in NewReno
2384 * and SACK.
2385 *
2386 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2387 * deflation etc. CWND is real congestion window, never inflated, changes
2388 * only according to classic VJ rules.
2389 *
2390 * Really tricky (and requiring careful tuning) part of algorithm
2391 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2392 * The first determines the moment _when_ we should reduce CWND and,
2393 * hence, slow down forward transmission. In fact, it determines the moment
2394 * when we decide that hole is caused by loss, rather than by a reorder.
2395 *
2396 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2397 * holes, caused by lost packets.
2398 *
2399 * And the most logically complicated part of algorithm is undo
2400 * heuristics. We detect false retransmits due to both too early
2401 * fast retransmit (reordering) and underestimated RTO, analyzing
2402 * timestamps and D-SACKs. When we detect that some segments were
2403 * retransmitted by mistake and CWND reduction was wrong, we undo
2404 * window reduction and abort recovery phase. This logic is hidden
2405 * inside several functions named tcp_try_undo_<something>.
2406 */
2407
2408/* This function decides, when we should leave Disordered state
2409 * and enter Recovery phase, reducing congestion window.
2410 *
2411 * Main question: may we further continue forward transmission
2412 * with the same cwnd?
2413 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002414static int tcp_time_to_recover(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002415{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002416 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002417 __u32 packets_out;
2418
Ryousei Takano564262c12007-10-25 23:03:52 -07002419 /* Do not perform any recovery during F-RTO algorithm */
Ilpo Järvinen52c63f12007-02-21 23:06:52 -08002420 if (tp->frto_counter)
2421 return 0;
2422
Linus Torvalds1da177e2005-04-16 15:20:36 -07002423 /* Trick#1: The loss is proven. */
2424 if (tp->lost_out)
2425 return 1;
2426
2427 /* Not-A-Trick#2 : Classic rule... */
Andreas Petlundea84e552009-10-27 03:27:21 +00002428 if (tcp_dupack_heuristics(tp) > tp->reordering)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002429 return 1;
2430
2431 /* Trick#3 : when we use RFC2988 timer restart, fast
2432 * retransmit can be triggered by timeout of queue head.
2433 */
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002434 if (tcp_is_fack(tp) && tcp_head_timedout(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002435 return 1;
2436
2437 /* Trick#4: It is still not OK... But will it be useful to delay
2438 * recovery more?
2439 */
2440 packets_out = tp->packets_out;
2441 if (packets_out <= tp->reordering &&
2442 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002443 !tcp_may_send_now(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002444 /* We have nothing to send. This connection is limited
2445 * either by receiver window or by application.
2446 */
2447 return 1;
2448 }
2449
2450 return 0;
2451}
2452
Ilpo Järvinen7363a5b2009-02-28 04:44:33 +00002453/* New heuristics: it is possible only after we switched to restart timer
2454 * each time when something is ACKed. Hence, we can detect timed out packets
2455 * during fast retransmit without falling to slow start.
2456 *
2457 * Usefulness of this as is very questionable, since we should know which of
2458 * the segments is the next to timeout which is relatively expensive to find
2459 * in general case unless we add some data structure just for that. The
2460 * current approach certainly won't find the right one too often and when it
2461 * finally does find _something_ it usually marks large part of the window
2462 * right away (because a retransmission with a larger timestamp blocks the
2463 * loop from advancing). -ij
2464 */
2465static void tcp_timeout_skbs(struct sock *sk)
2466{
2467 struct tcp_sock *tp = tcp_sk(sk);
2468 struct sk_buff *skb;
2469
2470 if (!tcp_is_fack(tp) || !tcp_head_timedout(sk))
2471 return;
2472
2473 skb = tp->scoreboard_skb_hint;
2474 if (tp->scoreboard_skb_hint == NULL)
2475 skb = tcp_write_queue_head(sk);
2476
2477 tcp_for_write_queue_from(skb, sk) {
2478 if (skb == tcp_send_head(sk))
2479 break;
2480 if (!tcp_skb_timedout(sk, skb))
2481 break;
2482
2483 tcp_skb_mark_lost(tp, skb);
2484 }
2485
2486 tp->scoreboard_skb_hint = skb;
2487
2488 tcp_verify_left_out(tp);
2489}
2490
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002491/* Mark head of queue up as lost. With RFC3517 SACK, the packets is
2492 * is against sacked "cnt", otherwise it's against facked "cnt"
2493 */
Ilpo Järvinen1b69d742008-04-07 22:31:38 -07002494static void tcp_mark_head_lost(struct sock *sk, int packets)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002495{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002496 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002497 struct sk_buff *skb;
Ilpo Järvinenc137f3d2008-04-07 22:32:38 -07002498 int cnt, oldcnt;
2499 int err;
2500 unsigned int mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002501
Ilpo Järvinen547b7922008-07-25 21:43:18 -07002502 WARN_ON(packets > tp->packets_out);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002503 if (tp->lost_skb_hint) {
2504 skb = tp->lost_skb_hint;
2505 cnt = tp->lost_cnt_hint;
2506 } else {
David S. Millerfe067e82007-03-07 12:12:44 -08002507 skb = tcp_write_queue_head(sk);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002508 cnt = 0;
2509 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002510
David S. Millerfe067e82007-03-07 12:12:44 -08002511 tcp_for_write_queue_from(skb, sk) {
2512 if (skb == tcp_send_head(sk))
2513 break;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002514 /* TODO: do this better */
2515 /* this is not the most efficient way to do this... */
2516 tp->lost_skb_hint = skb;
2517 tp->lost_cnt_hint = cnt;
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002518
Ilpo Järvinenc137f3d2008-04-07 22:32:38 -07002519 if (after(TCP_SKB_CB(skb)->end_seq, tp->high_seq))
2520 break;
2521
2522 oldcnt = cnt;
Ilpo Järvinenad1984e2008-01-30 20:06:02 -08002523 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002524 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2525 cnt += tcp_skb_pcount(skb);
2526
Ilpo Järvinenc137f3d2008-04-07 22:32:38 -07002527 if (cnt > packets) {
2528 if (tcp_is_sack(tp) || (oldcnt >= packets))
2529 break;
2530
2531 mss = skb_shinfo(skb)->gso_size;
2532 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2533 if (err < 0)
2534 break;
2535 cnt = packets;
2536 }
2537
Ilpo Järvinen41ea36e2008-09-20 21:19:22 -07002538 tcp_skb_mark_lost(tp, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002539 }
Ilpo Järvinen005903b2007-08-09 14:44:16 +03002540 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002541}
2542
2543/* Account newly detected lost packet(s) */
2544
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002545static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002546{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002547 struct tcp_sock *tp = tcp_sk(sk);
2548
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002549 if (tcp_is_reno(tp)) {
Ilpo Järvinen1b69d742008-04-07 22:31:38 -07002550 tcp_mark_head_lost(sk, 1);
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002551 } else if (tcp_is_fack(tp)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002552 int lost = tp->fackets_out - tp->reordering;
2553 if (lost <= 0)
2554 lost = 1;
Ilpo Järvinen1b69d742008-04-07 22:31:38 -07002555 tcp_mark_head_lost(sk, lost);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002556 } else {
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002557 int sacked_upto = tp->sacked_out - tp->reordering;
Ilpo Järvinen1b69d742008-04-07 22:31:38 -07002558 if (sacked_upto < fast_rexmit)
2559 sacked_upto = fast_rexmit;
2560 tcp_mark_head_lost(sk, sacked_upto);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002561 }
2562
Ilpo Järvinen7363a5b2009-02-28 04:44:33 +00002563 tcp_timeout_skbs(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002564}
2565
2566/* CWND moderation, preventing bursts due to too big ACKs
2567 * in dubious situations.
2568 */
2569static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2570{
2571 tp->snd_cwnd = min(tp->snd_cwnd,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002572 tcp_packets_in_flight(tp) + tcp_max_burst(tp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002573 tp->snd_cwnd_stamp = tcp_time_stamp;
2574}
2575
Stephen Hemminger72dc5b92006-06-05 17:30:08 -07002576/* Lower bound on congestion window is slow start threshold
2577 * unless congestion avoidance choice decides to overide it.
2578 */
2579static inline u32 tcp_cwnd_min(const struct sock *sk)
2580{
2581 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2582
2583 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
2584}
2585
Linus Torvalds1da177e2005-04-16 15:20:36 -07002586/* Decrease cwnd each second ack. */
Ilpo Järvinen1e757f92007-07-30 19:48:37 -07002587static void tcp_cwnd_down(struct sock *sk, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002588{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002589 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002590 int decr = tp->snd_cwnd_cnt + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002591
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002592 if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK)) ||
2593 (tcp_is_reno(tp) && !(flag & FLAG_NOT_DUP))) {
2594 tp->snd_cwnd_cnt = decr & 1;
Ilpo Järvinen1e757f92007-07-30 19:48:37 -07002595 decr >>= 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002596
Ilpo Järvinen1e757f92007-07-30 19:48:37 -07002597 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
2598 tp->snd_cwnd -= decr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002599
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002600 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
Ilpo Järvinen1e757f92007-07-30 19:48:37 -07002601 tp->snd_cwnd_stamp = tcp_time_stamp;
2602 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002603}
2604
2605/* Nothing was retransmitted or returned timestamp is less
2606 * than timestamp of the first retransmission.
2607 */
2608static inline int tcp_packet_delayed(struct tcp_sock *tp)
2609{
2610 return !tp->retrans_stamp ||
2611 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
Arnd Hannemannd7ee1472008-04-21 14:46:22 -07002612 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002613}
2614
2615/* Undo procedures. */
2616
2617#if FASTRETRANS_DEBUG > 1
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002618static void DBGUNDO(struct sock *sk, const char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002619{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002620 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002621 struct inet_sock *inet = inet_sk(sk);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002622
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002623 if (sk->sk_family == AF_INET) {
Harvey Harrison673d57e2008-10-31 00:53:57 -07002624 printk(KERN_DEBUG "Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002625 msg,
Harvey Harrison673d57e2008-10-31 00:53:57 -07002626 &inet->daddr, ntohs(inet->dport),
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002627 tp->snd_cwnd, tcp_left_out(tp),
2628 tp->snd_ssthresh, tp->prior_ssthresh,
2629 tp->packets_out);
2630 }
2631#if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2632 else if (sk->sk_family == AF_INET6) {
2633 struct ipv6_pinfo *np = inet6_sk(sk);
Harvey Harrison5b095d9892008-10-29 12:52:50 -07002634 printk(KERN_DEBUG "Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002635 msg,
Harvey Harrison0c6ce782008-10-28 16:09:23 -07002636 &np->daddr, ntohs(inet->dport),
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002637 tp->snd_cwnd, tcp_left_out(tp),
2638 tp->snd_ssthresh, tp->prior_ssthresh,
2639 tp->packets_out);
2640 }
2641#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002642}
2643#else
2644#define DBGUNDO(x...) do { } while (0)
2645#endif
2646
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002647static void tcp_undo_cwr(struct sock *sk, const int undo)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002648{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002649 struct tcp_sock *tp = tcp_sk(sk);
2650
Linus Torvalds1da177e2005-04-16 15:20:36 -07002651 if (tp->prior_ssthresh) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002652 const struct inet_connection_sock *icsk = inet_csk(sk);
2653
2654 if (icsk->icsk_ca_ops->undo_cwnd)
2655 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002656 else
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002657 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002658
2659 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
2660 tp->snd_ssthresh = tp->prior_ssthresh;
2661 TCP_ECN_withdraw_cwr(tp);
2662 }
2663 } else {
2664 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2665 }
2666 tcp_moderate_cwnd(tp);
2667 tp->snd_cwnd_stamp = tcp_time_stamp;
2668}
2669
2670static inline int tcp_may_undo(struct tcp_sock *tp)
2671{
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002672 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002673}
2674
2675/* People celebrate: "We love our President!" */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002676static int tcp_try_undo_recovery(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002677{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002678 struct tcp_sock *tp = tcp_sk(sk);
2679
Linus Torvalds1da177e2005-04-16 15:20:36 -07002680 if (tcp_may_undo(tp)) {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07002681 int mib_idx;
2682
Linus Torvalds1da177e2005-04-16 15:20:36 -07002683 /* Happy end! We did not retransmit anything
2684 * or our original transmission succeeded.
2685 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002686 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002687 tcp_undo_cwr(sk, 1);
2688 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07002689 mib_idx = LINUX_MIB_TCPLOSSUNDO;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002690 else
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07002691 mib_idx = LINUX_MIB_TCPFULLUNDO;
2692
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07002693 NET_INC_STATS_BH(sock_net(sk), mib_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002694 tp->undo_marker = 0;
2695 }
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002696 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002697 /* Hold old state until something *above* high_seq
2698 * is ACKed. For Reno it is MUST to prevent false
2699 * fast retransmits (RFC2582). SACK TCP is safe. */
2700 tcp_moderate_cwnd(tp);
2701 return 1;
2702 }
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002703 tcp_set_ca_state(sk, TCP_CA_Open);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002704 return 0;
2705}
2706
2707/* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002708static void tcp_try_undo_dsack(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002709{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002710 struct tcp_sock *tp = tcp_sk(sk);
2711
Linus Torvalds1da177e2005-04-16 15:20:36 -07002712 if (tp->undo_marker && !tp->undo_retrans) {
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002713 DBGUNDO(sk, "D-SACK");
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002714 tcp_undo_cwr(sk, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002715 tp->undo_marker = 0;
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07002716 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002717 }
2718}
2719
2720/* Undo during fast recovery after partial ACK. */
2721
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002722static int tcp_try_undo_partial(struct sock *sk, int acked)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002723{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002724 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002725 /* Partial ACK arrived. Force Hoe's retransmit. */
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002726 int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002727
2728 if (tcp_may_undo(tp)) {
2729 /* Plain luck! Hole if filled with delayed
2730 * packet, rather than with a retransmit.
2731 */
2732 if (tp->retrans_out == 0)
2733 tp->retrans_stamp = 0;
2734
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002735 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002736
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002737 DBGUNDO(sk, "Hoe");
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002738 tcp_undo_cwr(sk, 0);
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07002739 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002740
2741 /* So... Do not make Hoe's retransmit yet.
2742 * If the first packet was delayed, the rest
2743 * ones are most probably delayed as well.
2744 */
2745 failed = 0;
2746 }
2747 return failed;
2748}
2749
2750/* Undo during loss recovery after partial ACK. */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002751static int tcp_try_undo_loss(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002752{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002753 struct tcp_sock *tp = tcp_sk(sk);
2754
Linus Torvalds1da177e2005-04-16 15:20:36 -07002755 if (tcp_may_undo(tp)) {
2756 struct sk_buff *skb;
David S. Millerfe067e82007-03-07 12:12:44 -08002757 tcp_for_write_queue(skb, sk) {
2758 if (skb == tcp_send_head(sk))
2759 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002760 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2761 }
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002762
Ilpo Järvinen5af4ec22007-09-20 11:30:48 -07002763 tcp_clear_all_retrans_hints(tp);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002764
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002765 DBGUNDO(sk, "partial loss");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002766 tp->lost_out = 0;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002767 tcp_undo_cwr(sk, 1);
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07002768 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002769 inet_csk(sk)->icsk_retransmits = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002770 tp->undo_marker = 0;
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002771 if (tcp_is_sack(tp))
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002772 tcp_set_ca_state(sk, TCP_CA_Open);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002773 return 1;
2774 }
2775 return 0;
2776}
2777
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002778static inline void tcp_complete_cwr(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002779{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002780 struct tcp_sock *tp = tcp_sk(sk);
Stephen Hemminger317a76f2005-06-23 12:19:55 -07002781 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002782 tp->snd_cwnd_stamp = tcp_time_stamp;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002783 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002784}
2785
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07002786static void tcp_try_keep_open(struct sock *sk)
2787{
2788 struct tcp_sock *tp = tcp_sk(sk);
2789 int state = TCP_CA_Open;
2790
2791 if (tcp_left_out(tp) || tp->retrans_out || tp->undo_marker)
2792 state = TCP_CA_Disorder;
2793
2794 if (inet_csk(sk)->icsk_ca_state != state) {
2795 tcp_set_ca_state(sk, state);
2796 tp->high_seq = tp->snd_nxt;
2797 }
2798}
2799
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002800static void tcp_try_to_open(struct sock *sk, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002801{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002802 struct tcp_sock *tp = tcp_sk(sk);
2803
Ilpo Järvinen86426c22007-08-09 14:45:17 +03002804 tcp_verify_left_out(tp);
2805
Ilpo Järvinena1c1f282008-05-13 02:53:26 -07002806 if (!tp->frto_counter && tp->retrans_out == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002807 tp->retrans_stamp = 0;
2808
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002809 if (flag & FLAG_ECE)
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -08002810 tcp_enter_cwr(sk, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002811
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002812 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07002813 tcp_try_keep_open(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002814 tcp_moderate_cwnd(tp);
2815 } else {
Ilpo Järvinen1e757f92007-07-30 19:48:37 -07002816 tcp_cwnd_down(sk, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002817 }
2818}
2819
John Heffner5d424d52006-03-20 17:53:41 -08002820static void tcp_mtup_probe_failed(struct sock *sk)
2821{
2822 struct inet_connection_sock *icsk = inet_csk(sk);
2823
2824 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2825 icsk->icsk_mtup.probe_size = 0;
2826}
2827
Ilpo Järvinen72211e92009-03-14 14:23:04 +00002828static void tcp_mtup_probe_success(struct sock *sk)
John Heffner5d424d52006-03-20 17:53:41 -08002829{
2830 struct tcp_sock *tp = tcp_sk(sk);
2831 struct inet_connection_sock *icsk = inet_csk(sk);
2832
2833 /* FIXME: breaks with very large cwnd */
2834 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2835 tp->snd_cwnd = tp->snd_cwnd *
2836 tcp_mss_to_mtu(sk, tp->mss_cache) /
2837 icsk->icsk_mtup.probe_size;
2838 tp->snd_cwnd_cnt = 0;
2839 tp->snd_cwnd_stamp = tcp_time_stamp;
2840 tp->rcv_ssthresh = tcp_current_ssthresh(sk);
2841
2842 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2843 icsk->icsk_mtup.probe_size = 0;
2844 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2845}
2846
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08002847/* Do a simple retransmit without using the backoff mechanisms in
2848 * tcp_timer. This is used for path mtu discovery.
2849 * The socket is already locked here.
2850 */
2851void tcp_simple_retransmit(struct sock *sk)
2852{
2853 const struct inet_connection_sock *icsk = inet_csk(sk);
2854 struct tcp_sock *tp = tcp_sk(sk);
2855 struct sk_buff *skb;
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00002856 unsigned int mss = tcp_current_mss(sk);
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08002857 u32 prior_lost = tp->lost_out;
2858
2859 tcp_for_write_queue(skb, sk) {
2860 if (skb == tcp_send_head(sk))
2861 break;
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08002862 if (tcp_skb_seglen(skb) > mss &&
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08002863 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2864 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2865 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2866 tp->retrans_out -= tcp_skb_pcount(skb);
2867 }
2868 tcp_skb_mark_lost_uncond_verify(tp, skb);
2869 }
2870 }
2871
2872 tcp_clear_retrans_hints_partial(tp);
2873
2874 if (prior_lost == tp->lost_out)
2875 return;
2876
2877 if (tcp_is_reno(tp))
2878 tcp_limit_reno_sacked(tp);
2879
2880 tcp_verify_left_out(tp);
2881
2882 /* Don't muck with the congestion window here.
2883 * Reason is that we do not increase amount of _data_
2884 * in network, but units changed and effective
2885 * cwnd/ssthresh really reduced now.
2886 */
2887 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2888 tp->high_seq = tp->snd_nxt;
2889 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2890 tp->prior_ssthresh = 0;
2891 tp->undo_marker = 0;
2892 tcp_set_ca_state(sk, TCP_CA_Loss);
2893 }
2894 tcp_xmit_retransmit_queue(sk);
2895}
2896
Linus Torvalds1da177e2005-04-16 15:20:36 -07002897/* Process an event, which can update packets-in-flight not trivially.
2898 * Main goal of this function is to calculate new estimate for left_out,
2899 * taking into account both packets sitting in receiver's buffer and
2900 * packets lost by network.
2901 *
2902 * Besides that it does CWND reduction, when packet loss is detected
2903 * and changes state of machine.
2904 *
2905 * It does _not_ decide what to send, it is made in function
2906 * tcp_xmit_retransmit_queue().
2907 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002908static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002909{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002910 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002911 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002912 int is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
2913 int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002914 (tcp_fackets_out(tp) > tp->reordering));
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07002915 int fast_rexmit = 0, mib_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002916
Ilpo Järvinen3ccd3132007-12-31 04:43:32 -08002917 if (WARN_ON(!tp->packets_out && tp->sacked_out))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002918 tp->sacked_out = 0;
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07002919 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002920 tp->fackets_out = 0;
2921
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002922 /* Now state machine starts.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002923 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002924 if (flag & FLAG_ECE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002925 tp->prior_ssthresh = 0;
2926
2927 /* B. In all the states check for reneging SACKs. */
Ilpo Järvinencadbd032007-12-31 04:49:21 -08002928 if (tcp_check_sack_reneging(sk, flag))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002929 return;
2930
2931 /* C. Process data loss notification, provided it is valid. */
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002932 if (tcp_is_fack(tp) && (flag & FLAG_DATA_LOST) &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07002933 before(tp->snd_una, tp->high_seq) &&
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002934 icsk->icsk_ca_state != TCP_CA_Open &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07002935 tp->fackets_out > tp->reordering) {
Ilpo Järvinen1b69d742008-04-07 22:31:38 -07002936 tcp_mark_head_lost(sk, tp->fackets_out - tp->reordering);
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07002937 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002938 }
2939
Ilpo Järvinen005903b2007-08-09 14:44:16 +03002940 /* D. Check consistency of the current state. */
2941 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002942
2943 /* E. Check state exit conditions. State can be terminated
2944 * when high_seq is ACKed. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002945 if (icsk->icsk_ca_state == TCP_CA_Open) {
Ilpo Järvinen547b7922008-07-25 21:43:18 -07002946 WARN_ON(tp->retrans_out != 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002947 tp->retrans_stamp = 0;
2948 } else if (!before(tp->snd_una, tp->high_seq)) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002949 switch (icsk->icsk_ca_state) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002950 case TCP_CA_Loss:
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002951 icsk->icsk_retransmits = 0;
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002952 if (tcp_try_undo_recovery(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002953 return;
2954 break;
2955
2956 case TCP_CA_CWR:
2957 /* CWR is to be held something *above* high_seq
2958 * is ACKed for CWR bit to reach receiver. */
2959 if (tp->snd_una != tp->high_seq) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002960 tcp_complete_cwr(sk);
2961 tcp_set_ca_state(sk, TCP_CA_Open);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002962 }
2963 break;
2964
2965 case TCP_CA_Disorder:
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002966 tcp_try_undo_dsack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002967 if (!tp->undo_marker ||
2968 /* For SACK case do not Open to allow to undo
2969 * catching for all duplicate ACKs. */
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002970 tcp_is_reno(tp) || tp->snd_una != tp->high_seq) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002971 tp->undo_marker = 0;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002972 tcp_set_ca_state(sk, TCP_CA_Open);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002973 }
2974 break;
2975
2976 case TCP_CA_Recovery:
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002977 if (tcp_is_reno(tp))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002978 tcp_reset_reno_sack(tp);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002979 if (tcp_try_undo_recovery(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002980 return;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002981 tcp_complete_cwr(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002982 break;
2983 }
2984 }
2985
2986 /* F. Process state. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002987 switch (icsk->icsk_ca_state) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002988 case TCP_CA_Recovery:
Ilpo Järvinen2e605292007-08-02 19:46:58 -07002989 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002990 if (tcp_is_reno(tp) && is_dupack)
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002991 tcp_add_reno_sack(sk);
Ilpo Järvinen1b6d427b2007-08-09 14:53:36 +03002992 } else
2993 do_lost = tcp_try_undo_partial(sk, pkts_acked);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002994 break;
2995 case TCP_CA_Loss:
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002996 if (flag & FLAG_DATA_ACKED)
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002997 icsk->icsk_retransmits = 0;
Ilpo Järvinen882beba2008-04-07 22:33:07 -07002998 if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED)
2999 tcp_reset_reno_sack(tp);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003000 if (!tcp_try_undo_loss(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003001 tcp_moderate_cwnd(tp);
3002 tcp_xmit_retransmit_queue(sk);
3003 return;
3004 }
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003005 if (icsk->icsk_ca_state != TCP_CA_Open)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003006 return;
3007 /* Loss is undone; fall through to processing in Open state. */
3008 default:
Ilpo Järvinene60402d2007-08-09 15:14:46 +03003009 if (tcp_is_reno(tp)) {
Ilpo Järvinen2e605292007-08-02 19:46:58 -07003010 if (flag & FLAG_SND_UNA_ADVANCED)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003011 tcp_reset_reno_sack(tp);
3012 if (is_dupack)
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003013 tcp_add_reno_sack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003014 }
3015
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003016 if (icsk->icsk_ca_state == TCP_CA_Disorder)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003017 tcp_try_undo_dsack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003018
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003019 if (!tcp_time_to_recover(sk)) {
3020 tcp_try_to_open(sk, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003021 return;
3022 }
3023
John Heffner5d424d52006-03-20 17:53:41 -08003024 /* MTU probe failure: don't reduce cwnd */
3025 if (icsk->icsk_ca_state < TCP_CA_CWR &&
3026 icsk->icsk_mtup.probe_size &&
John Heffner0e7b1362006-03-20 21:32:58 -08003027 tp->snd_una == tp->mtu_probe.probe_seq_start) {
John Heffner5d424d52006-03-20 17:53:41 -08003028 tcp_mtup_probe_failed(sk);
3029 /* Restores the reduction we did in tcp_mtup_probe() */
3030 tp->snd_cwnd++;
3031 tcp_simple_retransmit(sk);
3032 return;
3033 }
3034
Linus Torvalds1da177e2005-04-16 15:20:36 -07003035 /* Otherwise enter Recovery state */
3036
Ilpo Järvinene60402d2007-08-09 15:14:46 +03003037 if (tcp_is_reno(tp))
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07003038 mib_idx = LINUX_MIB_TCPRENORECOVERY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003039 else
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07003040 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
3041
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07003042 NET_INC_STATS_BH(sock_net(sk), mib_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003043
3044 tp->high_seq = tp->snd_nxt;
3045 tp->prior_ssthresh = 0;
3046 tp->undo_marker = tp->snd_una;
3047 tp->undo_retrans = tp->retrans_out;
3048
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003049 if (icsk->icsk_ca_state < TCP_CA_CWR) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003050 if (!(flag & FLAG_ECE))
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003051 tp->prior_ssthresh = tcp_current_ssthresh(sk);
3052 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003053 TCP_ECN_queue_cwr(tp);
3054 }
3055
Stephen Hemminger9772efb2005-11-10 17:09:53 -08003056 tp->bytes_acked = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003057 tp->snd_cwnd_cnt = 0;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003058 tcp_set_ca_state(sk, TCP_CA_Recovery);
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08003059 fast_rexmit = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003060 }
3061
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08003062 if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk)))
3063 tcp_update_scoreboard(sk, fast_rexmit);
Ilpo Järvinen1e757f92007-07-30 19:48:37 -07003064 tcp_cwnd_down(sk, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003065 tcp_xmit_retransmit_queue(sk);
3066}
3067
Ilpo Järvinen41834b72008-12-05 22:43:26 -08003068static void tcp_valid_rtt_meas(struct sock *sk, u32 seq_rtt)
3069{
3070 tcp_rtt_estimator(sk, seq_rtt);
3071 tcp_set_rto(sk);
3072 inet_csk(sk)->icsk_backoff = 0;
3073}
3074
Linus Torvalds1da177e2005-04-16 15:20:36 -07003075/* Read draft-ietf-tcplw-high-performance before mucking
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08003076 * with this code. (Supersedes RFC1323)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003077 */
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08003078static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003079{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003080 /* RTTM Rule: A TSecr value received in a segment is used to
3081 * update the averaged RTT measurement only if the segment
3082 * acknowledges some new data, i.e., only if it advances the
3083 * left edge of the send window.
3084 *
3085 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3086 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
3087 *
3088 * Changed: reset backoff as soon as we see the first valid sample.
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08003089 * If we do not, we get strongly overestimated rto. With timestamps
Linus Torvalds1da177e2005-04-16 15:20:36 -07003090 * samples are accepted even from very old segments: f.e., when rtt=1
3091 * increases to 8, we retransmit 5 times and after 8 seconds delayed
3092 * answer arrives rto becomes 120 seconds! If at least one of segments
3093 * in window is lost... Voila. --ANK (010210)
3094 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003095 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen41834b72008-12-05 22:43:26 -08003096
3097 tcp_valid_rtt_meas(sk, tcp_time_stamp - tp->rx_opt.rcv_tsecr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003098}
3099
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08003100static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003101{
3102 /* We don't have a timestamp. Can only use
3103 * packets that are not retransmitted to determine
3104 * rtt estimates. Also, we must not reset the
3105 * backoff for rto until we get a non-retransmitted
3106 * packet. This allows us to deal with a situation
3107 * where the network delay has increased suddenly.
3108 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
3109 */
3110
3111 if (flag & FLAG_RETRANS_DATA_ACKED)
3112 return;
3113
Ilpo Järvinen41834b72008-12-05 22:43:26 -08003114 tcp_valid_rtt_meas(sk, seq_rtt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003115}
3116
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003117static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08003118 const s32 seq_rtt)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003119{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003120 const struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003121 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
3122 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08003123 tcp_ack_saw_tstamp(sk, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003124 else if (seq_rtt >= 0)
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08003125 tcp_ack_no_tstamp(sk, seq_rtt, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003126}
3127
Ilpo Järvinenc3a05c62007-12-02 00:47:59 +02003128static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003129{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003130 const struct inet_connection_sock *icsk = inet_csk(sk);
Ilpo Järvinenc3a05c62007-12-02 00:47:59 +02003131 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003132 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003133}
3134
Linus Torvalds1da177e2005-04-16 15:20:36 -07003135/* Restart timer after forward progress on connection.
3136 * RFC2988 recommends to restart timer to now+rto.
3137 */
Ilpo Järvinen6728e7d2007-08-24 22:53:26 -07003138static void tcp_rearm_rto(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003139{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003140 struct tcp_sock *tp = tcp_sk(sk);
3141
Linus Torvalds1da177e2005-04-16 15:20:36 -07003142 if (!tp->packets_out) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003143 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003144 } else {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003145 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3146 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003147 }
3148}
3149
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003150/* If we get here, the whole TSO packet has not been acked. */
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003151static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003152{
3153 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003154 u32 packets_acked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003155
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003156 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003157
3158 packets_acked = tcp_skb_pcount(skb);
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003159 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003160 return 0;
3161 packets_acked -= tcp_skb_pcount(skb);
3162
3163 if (packets_acked) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003164 BUG_ON(tcp_skb_pcount(skb) == 0);
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003165 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003166 }
3167
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003168 return packets_acked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003169}
3170
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003171/* Remove acknowledged frames from the retransmission queue. If our packet
3172 * is before the ack sequence we can discard it as it's confirmed to have
3173 * arrived at the other end.
3174 */
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07003175static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3176 u32 prior_snd_una)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003177{
3178 struct tcp_sock *tp = tcp_sk(sk);
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08003179 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003180 struct sk_buff *skb;
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003181 u32 now = tcp_time_stamp;
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003182 int fully_acked = 1;
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003183 int flag = 0;
Ilpo Järvinen72018832007-12-30 04:37:55 -08003184 u32 pkts_acked = 0;
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003185 u32 reord = tp->packets_out;
Ilpo Järvinen90638a02008-09-20 21:25:52 -07003186 u32 prior_sacked = tp->sacked_out;
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003187 s32 seq_rtt = -1;
Gavin McCullagh2072c222007-12-29 19:11:21 -08003188 s32 ca_seq_rtt = -1;
Ilpo Järvinenb9ce2042007-06-15 15:08:43 -07003189 ktime_t last_ackt = net_invalid_timestamp();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003190
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003191 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09003192 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
Ilpo Järvinen72018832007-12-30 04:37:55 -08003193 u32 acked_pcount;
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003194 u8 sacked = scb->sacked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003195
Gavin McCullagh2072c222007-12-29 19:11:21 -08003196 /* Determine how many packets and what bytes were acked, tso and else */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003197 if (after(scb->end_seq, tp->snd_una)) {
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003198 if (tcp_skb_pcount(skb) == 1 ||
3199 !after(tp->snd_una, scb->seq))
3200 break;
3201
Ilpo Järvinen72018832007-12-30 04:37:55 -08003202 acked_pcount = tcp_tso_acked(sk, skb);
3203 if (!acked_pcount)
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003204 break;
3205
3206 fully_acked = 0;
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003207 } else {
Ilpo Järvinen72018832007-12-30 04:37:55 -08003208 acked_pcount = tcp_skb_pcount(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003209 }
3210
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003211 if (sacked & TCPCB_RETRANS) {
3212 if (sacked & TCPCB_SACKED_RETRANS)
Ilpo Järvinen72018832007-12-30 04:37:55 -08003213 tp->retrans_out -= acked_pcount;
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003214 flag |= FLAG_RETRANS_DATA_ACKED;
3215 ca_seq_rtt = -1;
3216 seq_rtt = -1;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003217 if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1))
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003218 flag |= FLAG_NONHEAD_RETRANS_ACKED;
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003219 } else {
Gavin McCullagh2072c222007-12-29 19:11:21 -08003220 ca_seq_rtt = now - scb->when;
3221 last_ackt = skb->tstamp;
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003222 if (seq_rtt < 0) {
Gavin McCullagh2072c222007-12-29 19:11:21 -08003223 seq_rtt = ca_seq_rtt;
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003224 }
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003225 if (!(sacked & TCPCB_SACKED_ACKED))
Ilpo Järvinen72018832007-12-30 04:37:55 -08003226 reord = min(pkts_acked, reord);
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08003227 }
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003228
3229 if (sacked & TCPCB_SACKED_ACKED)
Ilpo Järvinen72018832007-12-30 04:37:55 -08003230 tp->sacked_out -= acked_pcount;
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003231 if (sacked & TCPCB_LOST)
Ilpo Järvinen72018832007-12-30 04:37:55 -08003232 tp->lost_out -= acked_pcount;
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003233
Ilpo Järvinen72018832007-12-30 04:37:55 -08003234 tp->packets_out -= acked_pcount;
3235 pkts_acked += acked_pcount;
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003236
Ilpo Järvinen009a2e32007-09-20 11:34:38 -07003237 /* Initial outgoing SYN's get put onto the write_queue
3238 * just like anything else we transmit. It is not
3239 * true data, and if we misinform our callers that
3240 * this ACK acks real data, we will erroneously exit
3241 * connection startup slow start one packet too
3242 * quickly. This is severely frowned upon behavior.
3243 */
3244 if (!(scb->flags & TCPCB_FLAG_SYN)) {
3245 flag |= FLAG_DATA_ACKED;
3246 } else {
3247 flag |= FLAG_SYN_ACKED;
3248 tp->retrans_stamp = 0;
3249 }
3250
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003251 if (!fully_acked)
3252 break;
3253
David S. Millerfe067e82007-03-07 12:12:44 -08003254 tcp_unlink_write_queue(skb, sk);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08003255 sk_wmem_free_skb(sk, skb);
Ilpo Järvinen90638a02008-09-20 21:25:52 -07003256 tp->scoreboard_skb_hint = NULL;
Ilpo Järvinenef9da472008-09-20 21:25:15 -07003257 if (skb == tp->retransmit_skb_hint)
3258 tp->retransmit_skb_hint = NULL;
Ilpo Järvinen90638a02008-09-20 21:25:52 -07003259 if (skb == tp->lost_skb_hint)
3260 tp->lost_skb_hint = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003261 }
3262
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07003263 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3264 tp->snd_up = tp->snd_una;
3265
Ilpo Järvinencadbd032007-12-31 04:49:21 -08003266 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3267 flag |= FLAG_SACK_RENEGING;
3268
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003269 if (flag & FLAG_ACKED) {
Stephen Hemminger164891a2007-04-23 22:26:16 -07003270 const struct tcp_congestion_ops *ca_ops
3271 = inet_csk(sk)->icsk_ca_ops;
3272
Ilpo Järvinen72211e92009-03-14 14:23:04 +00003273 if (unlikely(icsk->icsk_mtup.probe_size &&
3274 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3275 tcp_mtup_probe_success(sk);
3276 }
3277
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003278 tcp_ack_update_rtt(sk, flag, seq_rtt);
Ilpo Järvinen6728e7d2007-08-24 22:53:26 -07003279 tcp_rearm_rto(sk);
Stephen Hemminger317a76f2005-06-23 12:19:55 -07003280
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003281 if (tcp_is_reno(tp)) {
3282 tcp_remove_reno_sacks(sk, pkts_acked);
3283 } else {
Ilpo Järvinen59a08cb2009-02-28 04:44:28 +00003284 int delta;
3285
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003286 /* Non-retransmitted hole got filled? That's reordering */
3287 if (reord < prior_fackets)
3288 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
Ilpo Järvinen90638a02008-09-20 21:25:52 -07003289
Ilpo Järvinen59a08cb2009-02-28 04:44:28 +00003290 delta = tcp_is_fack(tp) ? pkts_acked :
3291 prior_sacked - tp->sacked_out;
3292 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003293 }
3294
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07003295 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08003296
Stephen Hemminger30cfd0b2007-07-25 23:49:34 -07003297 if (ca_ops->pkts_acked) {
3298 s32 rtt_us = -1;
Ilpo Järvinenb9ce2042007-06-15 15:08:43 -07003299
Stephen Hemminger30cfd0b2007-07-25 23:49:34 -07003300 /* Is the ACK triggering packet unambiguous? */
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003301 if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
Stephen Hemminger30cfd0b2007-07-25 23:49:34 -07003302 /* High resolution needed and available? */
3303 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
3304 !ktime_equal(last_ackt,
3305 net_invalid_timestamp()))
3306 rtt_us = ktime_us_delta(ktime_get_real(),
3307 last_ackt);
Gavin McCullagh2072c222007-12-29 19:11:21 -08003308 else if (ca_seq_rtt > 0)
3309 rtt_us = jiffies_to_usecs(ca_seq_rtt);
Stephen Hemminger30cfd0b2007-07-25 23:49:34 -07003310 }
3311
3312 ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
3313 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003314 }
3315
3316#if FASTRETRANS_DEBUG > 0
Ilpo Järvinen547b7922008-07-25 21:43:18 -07003317 WARN_ON((int)tp->sacked_out < 0);
3318 WARN_ON((int)tp->lost_out < 0);
3319 WARN_ON((int)tp->retrans_out < 0);
Ilpo Järvinene60402d2007-08-09 15:14:46 +03003320 if (!tp->packets_out && tcp_is_sack(tp)) {
Stephen Hemmingercfcabdc2007-10-09 01:59:42 -07003321 icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003322 if (tp->lost_out) {
3323 printk(KERN_DEBUG "Leak l=%u %d\n",
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003324 tp->lost_out, icsk->icsk_ca_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003325 tp->lost_out = 0;
3326 }
3327 if (tp->sacked_out) {
3328 printk(KERN_DEBUG "Leak s=%u %d\n",
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003329 tp->sacked_out, icsk->icsk_ca_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003330 tp->sacked_out = 0;
3331 }
3332 if (tp->retrans_out) {
3333 printk(KERN_DEBUG "Leak r=%u %d\n",
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003334 tp->retrans_out, icsk->icsk_ca_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003335 tp->retrans_out = 0;
3336 }
3337 }
3338#endif
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003339 return flag;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003340}
3341
3342static void tcp_ack_probe(struct sock *sk)
3343{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003344 const struct tcp_sock *tp = tcp_sk(sk);
3345 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003346
3347 /* Was it a usable window open? */
3348
Ilpo Järvinen90840de2007-12-31 04:48:41 -08003349 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003350 icsk->icsk_backoff = 0;
3351 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003352 /* Socket must be waked up by subsequent tcp_data_snd_check().
3353 * This function is not for random using!
3354 */
3355 } else {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003356 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07003357 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3358 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003359 }
3360}
3361
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003362static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003363{
3364 return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003365 inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003366}
3367
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003368static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003369{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003370 const struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003371 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003372 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003373}
3374
3375/* Check that window update is acceptable.
3376 * The function assumes that snd_una<=ack<=snd_next.
3377 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003378static inline int tcp_may_update_window(const struct tcp_sock *tp,
3379 const u32 ack, const u32 ack_seq,
3380 const u32 nwin)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003381{
3382 return (after(ack, tp->snd_una) ||
3383 after(ack_seq, tp->snd_wl1) ||
3384 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
3385}
3386
3387/* Update our send window.
3388 *
3389 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3390 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3391 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003392static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack,
3393 u32 ack_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003394{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003395 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003396 int flag = 0;
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07003397 u32 nwin = ntohs(tcp_hdr(skb)->window);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003398
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07003399 if (likely(!tcp_hdr(skb)->syn))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003400 nwin <<= tp->rx_opt.snd_wscale;
3401
3402 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3403 flag |= FLAG_WIN_UPDATE;
Hantzis Fotisee7537b2009-03-02 22:42:02 -08003404 tcp_update_wl(tp, ack_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003405
3406 if (tp->snd_wnd != nwin) {
3407 tp->snd_wnd = nwin;
3408
3409 /* Note, it is the only place, where
3410 * fast path is recovered for sending TCP.
3411 */
Herbert Xu2ad41062005-10-27 18:47:46 +10003412 tp->pred_flags = 0;
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003413 tcp_fast_path_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003414
3415 if (nwin > tp->max_window) {
3416 tp->max_window = nwin;
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08003417 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003418 }
3419 }
3420 }
3421
3422 tp->snd_una = ack;
3423
3424 return flag;
3425}
3426
Ilpo Järvinen9ead9a12007-02-21 22:56:19 -08003427/* A very conservative spurious RTO response algorithm: reduce cwnd and
3428 * continue in congestion avoidance.
3429 */
3430static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
3431{
3432 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
Ilpo Järvinenaa8b6a72007-02-21 23:06:03 -08003433 tp->snd_cwnd_cnt = 0;
Ilpo Järvinen16e90682007-10-11 17:32:31 -07003434 tp->bytes_acked = 0;
Ilpo Järvinen46323652007-05-19 13:56:23 -07003435 TCP_ECN_queue_cwr(tp);
Ilpo Järvinen9ead9a12007-02-21 22:56:19 -08003436 tcp_moderate_cwnd(tp);
3437}
3438
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -08003439/* A conservative spurious RTO response algorithm: reduce cwnd using
3440 * rate halving and continue in congestion avoidance.
3441 */
3442static void tcp_ratehalving_spur_to_response(struct sock *sk)
3443{
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -08003444 tcp_enter_cwr(sk, 0);
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -08003445}
3446
Ilpo Järvinene317f6f2007-03-02 13:34:19 -08003447static void tcp_undo_spur_to_response(struct sock *sk, int flag)
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -08003448{
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003449 if (flag & FLAG_ECE)
Ilpo Järvinene317f6f2007-03-02 13:34:19 -08003450 tcp_ratehalving_spur_to_response(sk);
3451 else
3452 tcp_undo_cwr(sk, 1);
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -08003453}
3454
Ilpo Järvinen30935cf2007-02-21 23:01:36 -08003455/* F-RTO spurious RTO detection algorithm (RFC4138)
3456 *
Ilpo Järvinen6408d202007-02-21 23:04:11 -08003457 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3458 * comments). State (ACK number) is kept in frto_counter. When ACK advances
3459 * window (but not to or beyond highest sequence sent before RTO):
Ilpo Järvinen30935cf2007-02-21 23:01:36 -08003460 * On First ACK, send two new segments out.
3461 * On Second ACK, RTO was likely spurious. Do spurious response (response
3462 * algorithm is not part of the F-RTO detection algorithm
3463 * given in RFC4138 but can be selected separately).
3464 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
Ilpo Järvinend551e452007-04-30 00:42:20 -07003465 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3466 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3467 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
Ilpo Järvinen30935cf2007-02-21 23:01:36 -08003468 *
3469 * Rationale: if the RTO was spurious, new ACKs should arrive from the
3470 * original window even after we transmit two new data segments.
3471 *
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08003472 * SACK version:
3473 * on first step, wait until first cumulative ACK arrives, then move to
3474 * the second step. In second step, the next ACK decides.
3475 *
Ilpo Järvinen30935cf2007-02-21 23:01:36 -08003476 * F-RTO is implemented (mainly) in four functions:
3477 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
3478 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3479 * called when tcp_use_frto() showed green light
3480 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3481 * - tcp_enter_frto_loss() is called if there is not enough evidence
3482 * to prove that the RTO is indeed spurious. It transfers the control
3483 * from F-RTO to the conventional RTO recovery
3484 */
Ilpo Järvinen2e605292007-08-02 19:46:58 -07003485static int tcp_process_frto(struct sock *sk, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003486{
3487 struct tcp_sock *tp = tcp_sk(sk);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09003488
Ilpo Järvinen005903b2007-08-09 14:44:16 +03003489 tcp_verify_left_out(tp);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09003490
Ilpo Järvinen7487c482007-02-21 23:02:30 -08003491 /* Duplicate the behavior from Loss state (fastretrans_alert) */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003492 if (flag & FLAG_DATA_ACKED)
Ilpo Järvinen7487c482007-02-21 23:02:30 -08003493 inet_csk(sk)->icsk_retransmits = 0;
3494
Ilpo Järvinen009a2e32007-09-20 11:34:38 -07003495 if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||
3496 ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
3497 tp->undo_marker = 0;
3498
Ilpo Järvinen95c49222007-02-21 23:05:18 -08003499 if (!before(tp->snd_una, tp->frto_highmark)) {
Ilpo Järvinend551e452007-04-30 00:42:20 -07003500 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
Ilpo Järvinen7c9a4a52007-02-21 23:08:34 -08003501 return 1;
Ilpo Järvinen95c49222007-02-21 23:05:18 -08003502 }
3503
Ilpo Järvinen62ab2222008-05-08 01:09:11 -07003504 if (!tcp_is_sackfrto(tp)) {
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08003505 /* RFC4138 shortcoming in step 2; should also have case c):
3506 * ACK isn't duplicate nor advances window, e.g., opposite dir
3507 * data, winupdate
3508 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003509 if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP))
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08003510 return 1;
Ilpo Järvinen6408d202007-02-21 23:04:11 -08003511
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003512 if (!(flag & FLAG_DATA_ACKED)) {
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08003513 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
3514 flag);
3515 return 1;
3516 }
3517 } else {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003518 if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08003519 /* Prevent sending of new data. */
3520 tp->snd_cwnd = min(tp->snd_cwnd,
3521 tcp_packets_in_flight(tp));
3522 return 1;
3523 }
3524
Ilpo Järvinend551e452007-04-30 00:42:20 -07003525 if ((tp->frto_counter >= 2) &&
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003526 (!(flag & FLAG_FORWARD_PROGRESS) ||
3527 ((flag & FLAG_DATA_SACKED) &&
3528 !(flag & FLAG_ONLY_ORIG_SACKED)))) {
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08003529 /* RFC4138 shortcoming (see comment above) */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003530 if (!(flag & FLAG_FORWARD_PROGRESS) &&
3531 (flag & FLAG_NOT_DUP))
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08003532 return 1;
3533
3534 tcp_enter_frto_loss(sk, 3, flag);
3535 return 1;
3536 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003537 }
3538
3539 if (tp->frto_counter == 1) {
Ilpo Järvinen3e6f0492007-12-05 02:20:21 -08003540 /* tcp_may_send_now needs to see updated state */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003541 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
Ilpo Järvinen94d0ea72007-02-21 23:07:27 -08003542 tp->frto_counter = 2;
Ilpo Järvinen3e6f0492007-12-05 02:20:21 -08003543
3544 if (!tcp_may_send_now(sk))
3545 tcp_enter_frto_loss(sk, 2, flag);
3546
Ilpo Järvinen7c9a4a52007-02-21 23:08:34 -08003547 return 1;
Ilpo Järvinend551e452007-04-30 00:42:20 -07003548 } else {
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -08003549 switch (sysctl_tcp_frto_response) {
3550 case 2:
Ilpo Järvinene317f6f2007-03-02 13:34:19 -08003551 tcp_undo_spur_to_response(sk, flag);
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -08003552 break;
3553 case 1:
3554 tcp_conservative_spur_to_response(tp);
3555 break;
3556 default:
3557 tcp_ratehalving_spur_to_response(sk);
3558 break;
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07003559 }
Ilpo Järvinen94d0ea72007-02-21 23:07:27 -08003560 tp->frto_counter = 0;
Ilpo Järvinen009a2e32007-09-20 11:34:38 -07003561 tp->undo_marker = 0;
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07003562 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSPURIOUSRTOS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003563 }
Ilpo Järvinen7c9a4a52007-02-21 23:08:34 -08003564 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003565}
3566
Linus Torvalds1da177e2005-04-16 15:20:36 -07003567/* This routine deals with incoming acks, but not outgoing ones. */
3568static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
3569{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003570 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003571 struct tcp_sock *tp = tcp_sk(sk);
3572 u32 prior_snd_una = tp->snd_una;
3573 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3574 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3575 u32 prior_in_flight;
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003576 u32 prior_fackets;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003577 int prior_packets;
Ilpo Järvinen7c9a4a52007-02-21 23:08:34 -08003578 int frto_cwnd = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003579
John Dykstra96e0bf42009-03-22 21:49:57 -07003580 /* If the ack is older than previous acks
Linus Torvalds1da177e2005-04-16 15:20:36 -07003581 * then we can probably ignore it.
3582 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003583 if (before(ack, prior_snd_una))
3584 goto old_ack;
3585
John Dykstra96e0bf42009-03-22 21:49:57 -07003586 /* If the ack includes data we haven't sent yet, discard
3587 * this segment (RFC793 Section 3.9).
3588 */
3589 if (after(ack, tp->snd_nxt))
3590 goto invalid_ack;
3591
Ilpo Järvinen2e605292007-08-02 19:46:58 -07003592 if (after(ack, prior_snd_una))
3593 flag |= FLAG_SND_UNA_ADVANCED;
3594
Daikichi Osuga3fdf3f02006-08-29 02:01:44 -07003595 if (sysctl_tcp_abc) {
3596 if (icsk->icsk_ca_state < TCP_CA_CWR)
3597 tp->bytes_acked += ack - prior_snd_una;
3598 else if (icsk->icsk_ca_state == TCP_CA_Loss)
3599 /* we assume just one segment left network */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003600 tp->bytes_acked += min(ack - prior_snd_una,
3601 tp->mss_cache);
Daikichi Osuga3fdf3f02006-08-29 02:01:44 -07003602 }
Stephen Hemminger9772efb2005-11-10 17:09:53 -08003603
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003604 prior_fackets = tp->fackets_out;
Ilpo Järvinen52d34082007-12-05 02:21:35 -08003605 prior_in_flight = tcp_packets_in_flight(tp);
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003606
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003607 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003608 /* Window is constant, pure forward advance.
3609 * No more checks are required.
3610 * Note, we use the fact that SND.UNA>=SND.WL2.
3611 */
Hantzis Fotisee7537b2009-03-02 22:42:02 -08003612 tcp_update_wl(tp, ack_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003613 tp->snd_una = ack;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003614 flag |= FLAG_WIN_UPDATE;
3615
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003616 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
Stephen Hemminger317a76f2005-06-23 12:19:55 -07003617
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07003618 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003619 } else {
3620 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3621 flag |= FLAG_DATA;
3622 else
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07003623 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003624
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003625 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003626
3627 if (TCP_SKB_CB(skb)->sacked)
3628 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3629
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07003630 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003631 flag |= FLAG_ECE;
3632
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003633 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003634 }
3635
3636 /* We passed data and got it acked, remove any soft error
3637 * log. Something worked...
3638 */
3639 sk->sk_err_soft = 0;
David S. Miller4b53fb62008-07-23 16:38:45 -07003640 icsk->icsk_probes_out = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003641 tp->rcv_tstamp = tcp_time_stamp;
3642 prior_packets = tp->packets_out;
3643 if (!prior_packets)
3644 goto no_queue;
3645
Linus Torvalds1da177e2005-04-16 15:20:36 -07003646 /* See if we can take anything off of the retransmit queue. */
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07003647 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003648
Ilpo Järvinene1cd8f72007-11-14 15:55:09 -08003649 if (tp->frto_counter)
3650 frto_cwnd = tcp_process_frto(sk, flag);
Ilpo Järvinen3de96472007-10-01 15:28:48 -07003651 /* Guarantee sacktag reordering detection against wrap-arounds */
3652 if (before(tp->frto_highmark, tp->snd_una))
3653 tp->frto_highmark = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003654
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003655 if (tcp_ack_is_dubious(sk, flag)) {
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08003656 /* Advance CWND, if state allows this. */
Ilpo Järvinen7c9a4a52007-02-21 23:08:34 -08003657 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
3658 tcp_may_raise_cwnd(sk, flag))
Ilpo Järvinenc3a05c62007-12-02 00:47:59 +02003659 tcp_cong_avoid(sk, ack, prior_in_flight);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003660 tcp_fastretrans_alert(sk, prior_packets - tp->packets_out,
3661 flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003662 } else {
Ilpo Järvinen7c9a4a52007-02-21 23:08:34 -08003663 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
Ilpo Järvinenc3a05c62007-12-02 00:47:59 +02003664 tcp_cong_avoid(sk, ack, prior_in_flight);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003665 }
3666
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003667 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003668 dst_confirm(sk->sk_dst_cache);
3669
3670 return 1;
3671
3672no_queue:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003673 /* If this ack opens up a zero window, clear backoff. It was
3674 * being used to time the probes, and is probably far higher than
3675 * it needs to be for normal retransmission.
3676 */
David S. Millerfe067e82007-03-07 12:12:44 -08003677 if (tcp_send_head(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003678 tcp_ack_probe(sk);
3679 return 1;
3680
John Dykstra96e0bf42009-03-22 21:49:57 -07003681invalid_ack:
3682 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3683 return -1;
3684
Linus Torvalds1da177e2005-04-16 15:20:36 -07003685old_ack:
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07003686 if (TCP_SKB_CB(skb)->sacked) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003687 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07003688 if (icsk->icsk_ca_state == TCP_CA_Open)
3689 tcp_try_keep_open(sk);
3690 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003691
John Dykstra96e0bf42009-03-22 21:49:57 -07003692 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003693 return 0;
3694}
3695
Linus Torvalds1da177e2005-04-16 15:20:36 -07003696/* Look for tcp options. Normally only called on SYN and SYNACK packets.
3697 * But, this can also be called on packets in the established flow when
3698 * the fast version below fails.
3699 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003700void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx,
William Allen Simpson4957faade2009-12-02 18:25:27 +00003701 u8 **hvpp, int estab, struct dst_entry *dst)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003702{
3703 unsigned char *ptr;
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07003704 struct tcphdr *th = tcp_hdr(skb);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003705 int length = (th->doff * 4) - sizeof(struct tcphdr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003706
3707 ptr = (unsigned char *)(th + 1);
3708 opt_rx->saw_tstamp = 0;
3709
Stephen Hemminger2de979b2007-03-08 20:45:19 -08003710 while (length > 0) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003711 int opcode = *ptr++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003712 int opsize;
3713
3714 switch (opcode) {
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003715 case TCPOPT_EOL:
3716 return;
3717 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3718 length--;
3719 continue;
3720 default:
3721 opsize = *ptr++;
3722 if (opsize < 2) /* "silly options" */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003723 return;
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003724 if (opsize > length)
3725 return; /* don't parse partial options */
3726 switch (opcode) {
3727 case TCPOPT_MSS:
3728 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07003729 u16 in_mss = get_unaligned_be16(ptr);
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003730 if (in_mss) {
3731 if (opt_rx->user_mss &&
3732 opt_rx->user_mss < in_mss)
3733 in_mss = opt_rx->user_mss;
3734 opt_rx->mss_clamp = in_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003735 }
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07003736 }
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003737 break;
3738 case TCPOPT_WINDOW:
3739 if (opsize == TCPOLEN_WINDOW && th->syn &&
Gilad Ben-Yossef345cda22009-10-28 04:15:26 +00003740 !estab && sysctl_tcp_window_scaling &&
3741 !dst_feature(dst, RTAX_FEATURE_NO_WSCALE)) {
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003742 __u8 snd_wscale = *(__u8 *)ptr;
3743 opt_rx->wscale_ok = 1;
3744 if (snd_wscale > 14) {
3745 if (net_ratelimit())
3746 printk(KERN_INFO "tcp_parse_options: Illegal window "
3747 "scaling value %d >14 received.\n",
3748 snd_wscale);
3749 snd_wscale = 14;
3750 }
3751 opt_rx->snd_wscale = snd_wscale;
3752 }
3753 break;
3754 case TCPOPT_TIMESTAMP:
3755 if ((opsize == TCPOLEN_TIMESTAMP) &&
3756 ((estab && opt_rx->tstamp_ok) ||
Gilad Ben-Yossefcda42eb2009-10-28 04:15:25 +00003757 (!estab && sysctl_tcp_timestamps &&
3758 !dst_feature(dst, RTAX_FEATURE_NO_TSTAMP)))) {
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003759 opt_rx->saw_tstamp = 1;
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07003760 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3761 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003762 }
3763 break;
3764 case TCPOPT_SACK_PERM:
3765 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
Gilad Ben-Yossef1aba7212009-10-28 04:15:24 +00003766 !estab && sysctl_tcp_sack &&
3767 !dst_feature(dst, RTAX_FEATURE_NO_SACK)) {
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003768 opt_rx->sack_ok = 1;
3769 tcp_sack_reset(opt_rx);
3770 }
3771 break;
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07003772
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003773 case TCPOPT_SACK:
3774 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3775 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3776 opt_rx->sack_ok) {
3777 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3778 }
3779 break;
3780#ifdef CONFIG_TCP_MD5SIG
3781 case TCPOPT_MD5SIG:
3782 /*
3783 * The MD5 Hash has already been
3784 * checked (see tcp_v{4,6}_do_rcv()).
3785 */
3786 break;
3787#endif
William Allen Simpson4957faade2009-12-02 18:25:27 +00003788 case TCPOPT_COOKIE:
3789 /* This option is variable length.
3790 */
3791 switch (opsize) {
3792 case TCPOLEN_COOKIE_BASE:
3793 /* not yet implemented */
3794 break;
3795 case TCPOLEN_COOKIE_PAIR:
3796 /* not yet implemented */
3797 break;
3798 case TCPOLEN_COOKIE_MIN+0:
3799 case TCPOLEN_COOKIE_MIN+2:
3800 case TCPOLEN_COOKIE_MIN+4:
3801 case TCPOLEN_COOKIE_MIN+6:
3802 case TCPOLEN_COOKIE_MAX:
3803 /* 16-bit multiple */
3804 opt_rx->cookie_plus = opsize;
3805 *hvpp = ptr;
3806 default:
3807 /* ignore option */
3808 break;
3809 };
3810 break;
3811 };
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003812
3813 ptr += opsize-2;
3814 length -= opsize;
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07003815 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003816 }
3817}
3818
Ilpo Järvinena4356b22008-08-23 05:12:29 -07003819static int tcp_parse_aligned_timestamp(struct tcp_sock *tp, struct tcphdr *th)
3820{
3821 __be32 *ptr = (__be32 *)(th + 1);
3822
3823 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3824 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3825 tp->rx_opt.saw_tstamp = 1;
3826 ++ptr;
3827 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3828 ++ptr;
3829 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3830 return 1;
3831 }
3832 return 0;
3833}
3834
Linus Torvalds1da177e2005-04-16 15:20:36 -07003835/* Fast parse options. This hopes to only see timestamps.
3836 * If it is wrong it falls back on tcp_parse_options().
3837 */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -08003838static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
William Allen Simpson4957faade2009-12-02 18:25:27 +00003839 struct tcp_sock *tp, u8 **hvpp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003840{
William Allen Simpson4957faade2009-12-02 18:25:27 +00003841 /* In the spirit of fast parsing, compare doff directly to constant
3842 * values. Because equality is used, short doff can be ignored here.
3843 */
3844 if (th->doff == (sizeof(*th) / 4)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003845 tp->rx_opt.saw_tstamp = 0;
3846 return 0;
3847 } else if (tp->rx_opt.tstamp_ok &&
William Allen Simpson4957faade2009-12-02 18:25:27 +00003848 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
Ilpo Järvinena4356b22008-08-23 05:12:29 -07003849 if (tcp_parse_aligned_timestamp(tp, th))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003850 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003851 }
William Allen Simpson4957faade2009-12-02 18:25:27 +00003852 tcp_parse_options(skb, &tp->rx_opt, hvpp, 1, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003853 return 1;
3854}
3855
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09003856#ifdef CONFIG_TCP_MD5SIG
3857/*
3858 * Parse MD5 Signature option
3859 */
3860u8 *tcp_parse_md5sig_option(struct tcphdr *th)
3861{
3862 int length = (th->doff << 2) - sizeof (*th);
3863 u8 *ptr = (u8*)(th + 1);
3864
3865 /* If the TCP option is too short, we can short cut */
3866 if (length < TCPOLEN_MD5SIG)
3867 return NULL;
3868
3869 while (length > 0) {
3870 int opcode = *ptr++;
3871 int opsize;
3872
3873 switch(opcode) {
3874 case TCPOPT_EOL:
3875 return NULL;
3876 case TCPOPT_NOP:
3877 length--;
3878 continue;
3879 default:
3880 opsize = *ptr++;
3881 if (opsize < 2 || opsize > length)
3882 return NULL;
3883 if (opcode == TCPOPT_MD5SIG)
3884 return ptr;
3885 }
3886 ptr += opsize - 2;
3887 length -= opsize;
3888 }
3889 return NULL;
3890}
3891#endif
3892
Linus Torvalds1da177e2005-04-16 15:20:36 -07003893static inline void tcp_store_ts_recent(struct tcp_sock *tp)
3894{
3895 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
James Morris9d729f72007-03-04 16:12:44 -08003896 tp->rx_opt.ts_recent_stamp = get_seconds();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003897}
3898
3899static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3900{
3901 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3902 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3903 * extra check below makes sure this can only happen
3904 * for pure ACK frames. -DaveM
3905 *
3906 * Not only, also it occurs for expired timestamps.
3907 */
3908
Ilpo Järvinenc887e6d2009-03-14 14:23:03 +00003909 if (tcp_paws_check(&tp->rx_opt, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003910 tcp_store_ts_recent(tp);
3911 }
3912}
3913
3914/* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3915 *
3916 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3917 * it can pass through stack. So, the following predicate verifies that
3918 * this segment is not used for anything but congestion avoidance or
3919 * fast retransmit. Moreover, we even are able to eliminate most of such
3920 * second order effects, if we apply some small "replay" window (~RTO)
3921 * to timestamp space.
3922 *
3923 * All these measures still do not guarantee that we reject wrapped ACKs
3924 * on networks with high bandwidth, when sequence space is recycled fastly,
3925 * but it guarantees that such events will be very rare and do not affect
3926 * connection seriously. This doesn't look nice, but alas, PAWS is really
3927 * buggy extension.
3928 *
3929 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3930 * states that events when retransmit arrives after original data are rare.
3931 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3932 * the biggest problem on large power networks even with minor reordering.
3933 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3934 * up to bandwidth of 18Gigabit/sec. 8) ]
3935 */
3936
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003937static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003938{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003939 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07003940 struct tcphdr *th = tcp_hdr(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003941 u32 seq = TCP_SKB_CB(skb)->seq;
3942 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3943
3944 return (/* 1. Pure ACK with correct sequence number. */
3945 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3946
3947 /* 2. ... and duplicate ACK. */
3948 ack == tp->snd_una &&
3949
3950 /* 3. ... and does not update window. */
3951 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3952
3953 /* 4. ... and sits in replay window. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003954 (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 -07003955}
3956
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003957static inline int tcp_paws_discard(const struct sock *sk,
3958 const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003959{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003960 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinenc887e6d2009-03-14 14:23:03 +00003961
3962 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
3963 !tcp_disordered_ack(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003964}
3965
3966/* Check segment sequence number for validity.
3967 *
3968 * Segment controls are considered valid, if the segment
3969 * fits to the window after truncation to the window. Acceptability
3970 * of data (and SYN, FIN, of course) is checked separately.
3971 * See tcp_data_queue(), for example.
3972 *
3973 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3974 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3975 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3976 * (borrowed from freebsd)
3977 */
3978
3979static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
3980{
3981 return !before(end_seq, tp->rcv_wup) &&
3982 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3983}
3984
3985/* When we get a reset we do this. */
3986static void tcp_reset(struct sock *sk)
3987{
3988 /* We want the right error as BSD sees it (and indeed as we do). */
3989 switch (sk->sk_state) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003990 case TCP_SYN_SENT:
3991 sk->sk_err = ECONNREFUSED;
3992 break;
3993 case TCP_CLOSE_WAIT:
3994 sk->sk_err = EPIPE;
3995 break;
3996 case TCP_CLOSE:
3997 return;
3998 default:
3999 sk->sk_err = ECONNRESET;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004000 }
4001
4002 if (!sock_flag(sk, SOCK_DEAD))
4003 sk->sk_error_report(sk);
4004
4005 tcp_done(sk);
4006}
4007
4008/*
4009 * Process the FIN bit. This now behaves as it is supposed to work
4010 * and the FIN takes effect when it is validly part of sequence
4011 * space. Not before when we get holes.
4012 *
4013 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4014 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4015 * TIME-WAIT)
4016 *
4017 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4018 * close and we go into CLOSING (and later onto TIME-WAIT)
4019 *
4020 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4021 */
4022static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
4023{
4024 struct tcp_sock *tp = tcp_sk(sk);
4025
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004026 inet_csk_schedule_ack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004027
4028 sk->sk_shutdown |= RCV_SHUTDOWN;
4029 sock_set_flag(sk, SOCK_DONE);
4030
4031 switch (sk->sk_state) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004032 case TCP_SYN_RECV:
4033 case TCP_ESTABLISHED:
4034 /* Move to CLOSE_WAIT */
4035 tcp_set_state(sk, TCP_CLOSE_WAIT);
4036 inet_csk(sk)->icsk_ack.pingpong = 1;
4037 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004038
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004039 case TCP_CLOSE_WAIT:
4040 case TCP_CLOSING:
4041 /* Received a retransmission of the FIN, do
4042 * nothing.
4043 */
4044 break;
4045 case TCP_LAST_ACK:
4046 /* RFC793: Remain in the LAST-ACK state. */
4047 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004048
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004049 case TCP_FIN_WAIT1:
4050 /* This case occurs when a simultaneous close
4051 * happens, we must ack the received FIN and
4052 * enter the CLOSING state.
4053 */
4054 tcp_send_ack(sk);
4055 tcp_set_state(sk, TCP_CLOSING);
4056 break;
4057 case TCP_FIN_WAIT2:
4058 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4059 tcp_send_ack(sk);
4060 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4061 break;
4062 default:
4063 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4064 * cases we should never reach this piece of code.
4065 */
4066 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
Harvey Harrison0dc47872008-03-05 20:47:47 -08004067 __func__, sk->sk_state);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004068 break;
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07004069 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004070
4071 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4072 * Probably, we should reset in this case. For now drop them.
4073 */
4074 __skb_queue_purge(&tp->out_of_order_queue);
Ilpo Järvinene60402d2007-08-09 15:14:46 +03004075 if (tcp_is_sack(tp))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004076 tcp_sack_reset(&tp->rx_opt);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08004077 sk_mem_reclaim(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004078
4079 if (!sock_flag(sk, SOCK_DEAD)) {
4080 sk->sk_state_change(sk);
4081
4082 /* Do not send POLL_HUP for half duplex close. */
4083 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4084 sk->sk_state == TCP_CLOSE)
Pavel Emelyanov8d8ad9d2007-11-26 20:10:50 +08004085 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004086 else
Pavel Emelyanov8d8ad9d2007-11-26 20:10:50 +08004087 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004088 }
4089}
4090
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004091static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4092 u32 end_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004093{
4094 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4095 if (before(seq, sp->start_seq))
4096 sp->start_seq = seq;
4097 if (after(end_seq, sp->end_seq))
4098 sp->end_seq = end_seq;
4099 return 1;
4100 }
4101 return 0;
4102}
4103
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004104static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004105{
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004106 struct tcp_sock *tp = tcp_sk(sk);
Gilad Ben-Yossefdc343472009-10-28 04:15:27 +00004107 struct dst_entry *dst = __sk_dst_get(sk);
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004108
Gilad Ben-Yossefdc343472009-10-28 04:15:27 +00004109 if (tcp_is_sack(tp) && sysctl_tcp_dsack &&
4110 !dst_feature(dst, RTAX_FEATURE_NO_DSACK)) {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07004111 int mib_idx;
4112
Linus Torvalds1da177e2005-04-16 15:20:36 -07004113 if (before(seq, tp->rcv_nxt))
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07004114 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004115 else
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07004116 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4117
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07004118 NET_INC_STATS_BH(sock_net(sk), mib_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004119
4120 tp->rx_opt.dsack = 1;
4121 tp->duplicate_sack[0].start_seq = seq;
4122 tp->duplicate_sack[0].end_seq = end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004123 }
4124}
4125
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004126static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004127{
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004128 struct tcp_sock *tp = tcp_sk(sk);
4129
Linus Torvalds1da177e2005-04-16 15:20:36 -07004130 if (!tp->rx_opt.dsack)
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004131 tcp_dsack_set(sk, seq, end_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004132 else
4133 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4134}
4135
4136static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
4137{
4138 struct tcp_sock *tp = tcp_sk(sk);
Gilad Ben-Yossefdc343472009-10-28 04:15:27 +00004139 struct dst_entry *dst = __sk_dst_get(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004140
4141 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4142 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07004143 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004144 tcp_enter_quickack_mode(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004145
Gilad Ben-Yossefdc343472009-10-28 04:15:27 +00004146 if (tcp_is_sack(tp) && sysctl_tcp_dsack &&
4147 !dst_feature(dst, RTAX_FEATURE_NO_DSACK)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004148 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4149
4150 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4151 end_seq = tp->rcv_nxt;
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004152 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004153 }
4154 }
4155
4156 tcp_send_ack(sk);
4157}
4158
4159/* These routines update the SACK block as out-of-order packets arrive or
4160 * in-order packets close up the sequence space.
4161 */
4162static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4163{
4164 int this_sack;
4165 struct tcp_sack_block *sp = &tp->selective_acks[0];
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004166 struct tcp_sack_block *swalk = sp + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004167
4168 /* See if the recent change to the first SACK eats into
4169 * or hits the sequence space of other SACK blocks, if so coalesce.
4170 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004171 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004172 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4173 int i;
4174
4175 /* Zap SWALK, by moving every further SACK up by one slot.
4176 * Decrease num_sacks.
4177 */
4178 tp->rx_opt.num_sacks--;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004179 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4180 sp[i] = sp[i + 1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07004181 continue;
4182 }
4183 this_sack++, swalk++;
4184 }
4185}
4186
Linus Torvalds1da177e2005-04-16 15:20:36 -07004187static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4188{
4189 struct tcp_sock *tp = tcp_sk(sk);
4190 struct tcp_sack_block *sp = &tp->selective_acks[0];
4191 int cur_sacks = tp->rx_opt.num_sacks;
4192 int this_sack;
4193
4194 if (!cur_sacks)
4195 goto new_sack;
4196
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004197 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004198 if (tcp_sack_extend(sp, seq, end_seq)) {
4199 /* Rotate this_sack to the first one. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004200 for (; this_sack > 0; this_sack--, sp--)
Ilpo Järvinena0bffff2009-03-21 13:36:17 -07004201 swap(*sp, *(sp - 1));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004202 if (cur_sacks > 1)
4203 tcp_sack_maybe_coalesce(tp);
4204 return;
4205 }
4206 }
4207
4208 /* Could not find an adjacent existing SACK, build a new one,
4209 * put it at the front, and shift everyone else down. We
4210 * always know there is at least one SACK present already here.
4211 *
4212 * If the sack array is full, forget about the last one.
4213 */
Adam Langley4389dde2008-07-19 00:07:02 -07004214 if (this_sack >= TCP_NUM_SACKS) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004215 this_sack--;
4216 tp->rx_opt.num_sacks--;
4217 sp--;
4218 }
Stephen Hemminger2de979b2007-03-08 20:45:19 -08004219 for (; this_sack > 0; this_sack--, sp--)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004220 *sp = *(sp - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004221
4222new_sack:
4223 /* Build the new head SACK, and we're done. */
4224 sp->start_seq = seq;
4225 sp->end_seq = end_seq;
4226 tp->rx_opt.num_sacks++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004227}
4228
4229/* RCV.NXT advances, some SACKs should be eaten. */
4230
4231static void tcp_sack_remove(struct tcp_sock *tp)
4232{
4233 struct tcp_sack_block *sp = &tp->selective_acks[0];
4234 int num_sacks = tp->rx_opt.num_sacks;
4235 int this_sack;
4236
4237 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
David S. Millerb03efcf2005-07-08 14:57:23 -07004238 if (skb_queue_empty(&tp->out_of_order_queue)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004239 tp->rx_opt.num_sacks = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004240 return;
4241 }
4242
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004243 for (this_sack = 0; this_sack < num_sacks;) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004244 /* Check if the start of the sack is covered by RCV.NXT. */
4245 if (!before(tp->rcv_nxt, sp->start_seq)) {
4246 int i;
4247
4248 /* RCV.NXT must cover all the block! */
Ilpo Järvinen547b7922008-07-25 21:43:18 -07004249 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004250
4251 /* Zap this SACK, by moving forward any other SACKS. */
4252 for (i=this_sack+1; i < num_sacks; i++)
4253 tp->selective_acks[i-1] = tp->selective_acks[i];
4254 num_sacks--;
4255 continue;
4256 }
4257 this_sack++;
4258 sp++;
4259 }
Ilpo Järvinen5861f8e2009-03-14 14:23:01 +00004260 tp->rx_opt.num_sacks = num_sacks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004261}
4262
4263/* This one checks to see if we can put data from the
4264 * out_of_order queue into the receive_queue.
4265 */
4266static void tcp_ofo_queue(struct sock *sk)
4267{
4268 struct tcp_sock *tp = tcp_sk(sk);
4269 __u32 dsack_high = tp->rcv_nxt;
4270 struct sk_buff *skb;
4271
4272 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4273 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4274 break;
4275
4276 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4277 __u32 dsack = dsack_high;
4278 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4279 dsack_high = TCP_SKB_CB(skb)->end_seq;
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004280 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004281 }
4282
4283 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4284 SOCK_DEBUG(sk, "ofo packet was already received \n");
David S. Miller8728b832005-08-09 19:25:21 -07004285 __skb_unlink(skb, &tp->out_of_order_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004286 __kfree_skb(skb);
4287 continue;
4288 }
4289 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4290 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4291 TCP_SKB_CB(skb)->end_seq);
4292
David S. Miller8728b832005-08-09 19:25:21 -07004293 __skb_unlink(skb, &tp->out_of_order_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004294 __skb_queue_tail(&sk->sk_receive_queue, skb);
4295 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07004296 if (tcp_hdr(skb)->fin)
4297 tcp_fin(skb, sk, tcp_hdr(skb));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004298 }
4299}
4300
Vitaliy Gusev56f367b2008-04-15 20:26:34 -07004301static int tcp_prune_ofo_queue(struct sock *sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004302static int tcp_prune_queue(struct sock *sk);
4303
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004304static inline int tcp_try_rmem_schedule(struct sock *sk, unsigned int size)
4305{
4306 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4307 !sk_rmem_schedule(sk, size)) {
4308
4309 if (tcp_prune_queue(sk) < 0)
4310 return -1;
4311
4312 if (!sk_rmem_schedule(sk, size)) {
Vitaliy Gusev56f367b2008-04-15 20:26:34 -07004313 if (!tcp_prune_ofo_queue(sk))
4314 return -1;
4315
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004316 if (!sk_rmem_schedule(sk, size))
4317 return -1;
4318 }
4319 }
4320 return 0;
4321}
4322
Linus Torvalds1da177e2005-04-16 15:20:36 -07004323static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4324{
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07004325 struct tcphdr *th = tcp_hdr(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004326 struct tcp_sock *tp = tcp_sk(sk);
4327 int eaten = -1;
4328
4329 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4330 goto drop;
4331
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004332 __skb_pull(skb, th->doff * 4);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004333
4334 TCP_ECN_accept_cwr(tp, skb);
4335
Ilpo Järvinen5861f8e2009-03-14 14:23:01 +00004336 tp->rx_opt.dsack = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004337
4338 /* Queue data for delivery to the user.
4339 * Packets in sequence go to the receive queue.
4340 * Out of sequence packets to the out_of_order_queue.
4341 */
4342 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4343 if (tcp_receive_window(tp) == 0)
4344 goto out_of_window;
4345
4346 /* Ok. In sequence. In window. */
4347 if (tp->ucopy.task == current &&
4348 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4349 sock_owned_by_user(sk) && !tp->urg_data) {
4350 int chunk = min_t(unsigned int, skb->len,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004351 tp->ucopy.len);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004352
4353 __set_current_state(TASK_RUNNING);
4354
4355 local_bh_enable();
4356 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
4357 tp->ucopy.len -= chunk;
4358 tp->copied_seq += chunk;
4359 eaten = (chunk == skb->len && !th->fin);
4360 tcp_rcv_space_adjust(sk);
4361 }
4362 local_bh_disable();
4363 }
4364
4365 if (eaten <= 0) {
4366queue_and_out:
4367 if (eaten < 0 &&
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004368 tcp_try_rmem_schedule(sk, skb->truesize))
4369 goto drop;
4370
Hideo Aoki3ab224b2007-12-31 00:11:19 -08004371 skb_set_owner_r(skb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004372 __skb_queue_tail(&sk->sk_receive_queue, skb);
4373 }
4374 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
Stephen Hemminger2de979b2007-03-08 20:45:19 -08004375 if (skb->len)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004376 tcp_event_data_recv(sk, skb);
Stephen Hemminger2de979b2007-03-08 20:45:19 -08004377 if (th->fin)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004378 tcp_fin(skb, sk, th);
4379
David S. Millerb03efcf2005-07-08 14:57:23 -07004380 if (!skb_queue_empty(&tp->out_of_order_queue)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004381 tcp_ofo_queue(sk);
4382
4383 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4384 * gap in queue is filled.
4385 */
David S. Millerb03efcf2005-07-08 14:57:23 -07004386 if (skb_queue_empty(&tp->out_of_order_queue))
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004387 inet_csk(sk)->icsk_ack.pingpong = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004388 }
4389
4390 if (tp->rx_opt.num_sacks)
4391 tcp_sack_remove(tp);
4392
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004393 tcp_fast_path_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004394
4395 if (eaten > 0)
4396 __kfree_skb(skb);
4397 else if (!sock_flag(sk, SOCK_DEAD))
4398 sk->sk_data_ready(sk, 0);
4399 return;
4400 }
4401
4402 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4403 /* A retransmit, 2nd most common case. Force an immediate ack. */
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07004404 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004405 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004406
4407out_of_window:
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004408 tcp_enter_quickack_mode(sk);
4409 inet_csk_schedule_ack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004410drop:
4411 __kfree_skb(skb);
4412 return;
4413 }
4414
4415 /* Out of window. F.e. zero window probe. */
4416 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4417 goto out_of_window;
4418
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004419 tcp_enter_quickack_mode(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004420
4421 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4422 /* Partial packet, seq < rcv_next < end_seq */
4423 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4424 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4425 TCP_SKB_CB(skb)->end_seq);
4426
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004427 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004428
Linus Torvalds1da177e2005-04-16 15:20:36 -07004429 /* If window is closed, drop tail of packet. But after
4430 * remembering D-SACK for its head made in previous line.
4431 */
4432 if (!tcp_receive_window(tp))
4433 goto out_of_window;
4434 goto queue_and_out;
4435 }
4436
4437 TCP_ECN_check_ce(tp, skb);
4438
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004439 if (tcp_try_rmem_schedule(sk, skb->truesize))
4440 goto drop;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004441
4442 /* Disable header prediction. */
4443 tp->pred_flags = 0;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004444 inet_csk_schedule_ack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004445
4446 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4447 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4448
Hideo Aoki3ab224b2007-12-31 00:11:19 -08004449 skb_set_owner_r(skb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004450
4451 if (!skb_peek(&tp->out_of_order_queue)) {
4452 /* Initial out of order segment, build 1 SACK. */
Ilpo Järvinene60402d2007-08-09 15:14:46 +03004453 if (tcp_is_sack(tp)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004454 tp->rx_opt.num_sacks = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004455 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4456 tp->selective_acks[0].end_seq =
4457 TCP_SKB_CB(skb)->end_seq;
4458 }
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004459 __skb_queue_head(&tp->out_of_order_queue, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004460 } else {
David S. Miller91521942009-05-28 21:35:47 -07004461 struct sk_buff *skb1 = skb_peek_tail(&tp->out_of_order_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004462 u32 seq = TCP_SKB_CB(skb)->seq;
4463 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4464
4465 if (seq == TCP_SKB_CB(skb1)->end_seq) {
Gerrit Renker7de6c032008-04-14 00:05:09 -07004466 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004467
4468 if (!tp->rx_opt.num_sacks ||
4469 tp->selective_acks[0].end_seq != seq)
4470 goto add_sack;
4471
4472 /* Common case: data arrive in order after hole. */
4473 tp->selective_acks[0].end_seq = end_seq;
4474 return;
4475 }
4476
4477 /* Find place to insert this segment. */
David S. Miller91521942009-05-28 21:35:47 -07004478 while (1) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004479 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4480 break;
David S. Miller91521942009-05-28 21:35:47 -07004481 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4482 skb1 = NULL;
4483 break;
4484 }
4485 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4486 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004487
4488 /* Do skb overlap to previous one? */
David S. Miller91521942009-05-28 21:35:47 -07004489 if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004490 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4491 /* All the bits are present. Drop. */
4492 __kfree_skb(skb);
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004493 tcp_dsack_set(sk, seq, end_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004494 goto add_sack;
4495 }
4496 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4497 /* Partial overlap. */
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004498 tcp_dsack_set(sk, seq,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004499 TCP_SKB_CB(skb1)->end_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004500 } else {
David S. Miller91521942009-05-28 21:35:47 -07004501 if (skb_queue_is_first(&tp->out_of_order_queue,
4502 skb1))
4503 skb1 = NULL;
4504 else
4505 skb1 = skb_queue_prev(
4506 &tp->out_of_order_queue,
4507 skb1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004508 }
4509 }
David S. Miller91521942009-05-28 21:35:47 -07004510 if (!skb1)
4511 __skb_queue_head(&tp->out_of_order_queue, skb);
4512 else
4513 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004514
Linus Torvalds1da177e2005-04-16 15:20:36 -07004515 /* And clean segments covered by new one as whole. */
Ilpo Järvinen2df90012009-05-29 15:02:29 -07004516 while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4517 skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
David S. Miller91521942009-05-28 21:35:47 -07004518
Ilpo Järvinen2df90012009-05-29 15:02:29 -07004519 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4520 break;
4521 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004522 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
Ilpo Järvinen2df90012009-05-29 15:02:29 -07004523 end_seq);
4524 break;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004525 }
Ilpo Järvinen2df90012009-05-29 15:02:29 -07004526 __skb_unlink(skb1, &tp->out_of_order_queue);
4527 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4528 TCP_SKB_CB(skb1)->end_seq);
4529 __kfree_skb(skb1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004530 }
4531
4532add_sack:
Ilpo Järvinene60402d2007-08-09 15:14:46 +03004533 if (tcp_is_sack(tp))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004534 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4535 }
4536}
4537
Ilpo Järvinen2cf46632008-08-23 05:11:41 -07004538static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4539 struct sk_buff_head *list)
4540{
David S. Miller91521942009-05-28 21:35:47 -07004541 struct sk_buff *next = NULL;
4542
4543 if (!skb_queue_is_last(list, skb))
4544 next = skb_queue_next(list, skb);
Ilpo Järvinen2cf46632008-08-23 05:11:41 -07004545
4546 __skb_unlink(skb, list);
4547 __kfree_skb(skb);
4548 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4549
4550 return next;
4551}
4552
Linus Torvalds1da177e2005-04-16 15:20:36 -07004553/* Collapse contiguous sequence of skbs head..tail with
4554 * sequence numbers start..end.
David S. Miller91521942009-05-28 21:35:47 -07004555 *
4556 * If tail is NULL, this means until the end of the list.
4557 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07004558 * Segments with FIN/SYN are not collapsed (only because this
4559 * simplifies code)
4560 */
4561static void
David S. Miller8728b832005-08-09 19:25:21 -07004562tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4563 struct sk_buff *head, struct sk_buff *tail,
4564 u32 start, u32 end)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004565{
David S. Miller91521942009-05-28 21:35:47 -07004566 struct sk_buff *skb, *n;
4567 bool end_of_skbs;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004568
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08004569 /* First, check that queue is collapsible and find
Linus Torvalds1da177e2005-04-16 15:20:36 -07004570 * the point where collapsing can be useful. */
David S. Miller91521942009-05-28 21:35:47 -07004571 skb = head;
4572restart:
4573 end_of_skbs = true;
4574 skb_queue_walk_from_safe(list, skb, n) {
4575 if (skb == tail)
4576 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004577 /* No new bits? It is possible on ofo queue. */
4578 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
Ilpo Järvinen2cf46632008-08-23 05:11:41 -07004579 skb = tcp_collapse_one(sk, skb, list);
David S. Miller91521942009-05-28 21:35:47 -07004580 if (!skb)
4581 break;
4582 goto restart;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004583 }
4584
4585 /* The first skb to collapse is:
4586 * - not SYN/FIN and
4587 * - bloated or contains data before "start" or
4588 * overlaps to the next one.
4589 */
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07004590 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07004591 (tcp_win_from_space(skb->truesize) > skb->len ||
David S. Miller91521942009-05-28 21:35:47 -07004592 before(TCP_SKB_CB(skb)->seq, start))) {
4593 end_of_skbs = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004594 break;
David S. Miller91521942009-05-28 21:35:47 -07004595 }
4596
4597 if (!skb_queue_is_last(list, skb)) {
4598 struct sk_buff *next = skb_queue_next(list, skb);
4599 if (next != tail &&
4600 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4601 end_of_skbs = false;
4602 break;
4603 }
4604 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004605
4606 /* Decided to skip this, advance start seq. */
4607 start = TCP_SKB_CB(skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004608 }
David S. Miller91521942009-05-28 21:35:47 -07004609 if (end_of_skbs || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004610 return;
4611
4612 while (before(start, end)) {
4613 struct sk_buff *nskb;
Chuck Leverc2636b42007-10-23 21:07:32 -07004614 unsigned int header = skb_headroom(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004615 int copy = SKB_MAX_ORDER(header, 0);
4616
4617 /* Too big header? This can happen with IPv6. */
4618 if (copy < 0)
4619 return;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004620 if (end - start < copy)
4621 copy = end - start;
4622 nskb = alloc_skb(copy + header, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004623 if (!nskb)
4624 return;
Arnaldo Carvalho de Meloc51957d2007-03-10 12:47:22 -03004625
Arnaldo Carvalho de Melo98e399f2007-03-19 15:33:04 -07004626 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
Arnaldo Carvalho de Melo9c702202007-04-25 18:04:18 -07004627 skb_set_network_header(nskb, (skb_network_header(skb) -
4628 skb->head));
4629 skb_set_transport_header(nskb, (skb_transport_header(skb) -
4630 skb->head));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004631 skb_reserve(nskb, header);
4632 memcpy(nskb->head, skb->head, header);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004633 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4634 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
David S. Miller43f59c82008-09-21 21:28:51 -07004635 __skb_queue_before(list, skb, nskb);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08004636 skb_set_owner_r(nskb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004637
4638 /* Copy data, releasing collapsed skbs. */
4639 while (copy > 0) {
4640 int offset = start - TCP_SKB_CB(skb)->seq;
4641 int size = TCP_SKB_CB(skb)->end_seq - start;
4642
Kris Katterjohn09a62662006-01-08 22:24:28 -08004643 BUG_ON(offset < 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004644 if (size > 0) {
4645 size = min(copy, size);
4646 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4647 BUG();
4648 TCP_SKB_CB(nskb)->end_seq += size;
4649 copy -= size;
4650 start += size;
4651 }
4652 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
Ilpo Järvinen2cf46632008-08-23 05:11:41 -07004653 skb = tcp_collapse_one(sk, skb, list);
David S. Miller91521942009-05-28 21:35:47 -07004654 if (!skb ||
4655 skb == tail ||
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07004656 tcp_hdr(skb)->syn ||
4657 tcp_hdr(skb)->fin)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004658 return;
4659 }
4660 }
4661 }
4662}
4663
4664/* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4665 * and tcp_collapse() them until all the queue is collapsed.
4666 */
4667static void tcp_collapse_ofo_queue(struct sock *sk)
4668{
4669 struct tcp_sock *tp = tcp_sk(sk);
4670 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4671 struct sk_buff *head;
4672 u32 start, end;
4673
4674 if (skb == NULL)
4675 return;
4676
4677 start = TCP_SKB_CB(skb)->seq;
4678 end = TCP_SKB_CB(skb)->end_seq;
4679 head = skb;
4680
4681 for (;;) {
David S. Miller91521942009-05-28 21:35:47 -07004682 struct sk_buff *next = NULL;
4683
4684 if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4685 next = skb_queue_next(&tp->out_of_order_queue, skb);
4686 skb = next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004687
4688 /* Segment is terminated when we see gap or when
4689 * we are at the end of all the queue. */
David S. Miller91521942009-05-28 21:35:47 -07004690 if (!skb ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07004691 after(TCP_SKB_CB(skb)->seq, end) ||
4692 before(TCP_SKB_CB(skb)->end_seq, start)) {
David S. Miller8728b832005-08-09 19:25:21 -07004693 tcp_collapse(sk, &tp->out_of_order_queue,
4694 head, skb, start, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004695 head = skb;
David S. Miller91521942009-05-28 21:35:47 -07004696 if (!skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004697 break;
4698 /* Start new segment */
4699 start = TCP_SKB_CB(skb)->seq;
4700 end = TCP_SKB_CB(skb)->end_seq;
4701 } else {
4702 if (before(TCP_SKB_CB(skb)->seq, start))
4703 start = TCP_SKB_CB(skb)->seq;
4704 if (after(TCP_SKB_CB(skb)->end_seq, end))
4705 end = TCP_SKB_CB(skb)->end_seq;
4706 }
4707 }
4708}
4709
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004710/*
4711 * Purge the out-of-order queue.
Vitaliy Gusev56f367b2008-04-15 20:26:34 -07004712 * Return true if queue was pruned.
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004713 */
Vitaliy Gusev56f367b2008-04-15 20:26:34 -07004714static int tcp_prune_ofo_queue(struct sock *sk)
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004715{
4716 struct tcp_sock *tp = tcp_sk(sk);
Vitaliy Gusev56f367b2008-04-15 20:26:34 -07004717 int res = 0;
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004718
4719 if (!skb_queue_empty(&tp->out_of_order_queue)) {
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07004720 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004721 __skb_queue_purge(&tp->out_of_order_queue);
4722
4723 /* Reset SACK state. A conforming SACK implementation will
4724 * do the same at a timeout based retransmit. When a connection
4725 * is in a sad state like this, we care only about integrity
4726 * of the connection not performance.
4727 */
4728 if (tp->rx_opt.sack_ok)
4729 tcp_sack_reset(&tp->rx_opt);
4730 sk_mem_reclaim(sk);
Vitaliy Gusev56f367b2008-04-15 20:26:34 -07004731 res = 1;
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004732 }
Vitaliy Gusev56f367b2008-04-15 20:26:34 -07004733 return res;
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004734}
4735
Linus Torvalds1da177e2005-04-16 15:20:36 -07004736/* Reduce allocated memory if we can, trying to get
4737 * the socket within its memory limits again.
4738 *
4739 * Return less than zero if we should start dropping frames
4740 * until the socket owning process reads some of the data
4741 * to stabilize the situation.
4742 */
4743static int tcp_prune_queue(struct sock *sk)
4744{
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004745 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004746
4747 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4748
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07004749 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004750
4751 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004752 tcp_clamp_window(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004753 else if (tcp_memory_pressure)
4754 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4755
4756 tcp_collapse_ofo_queue(sk);
David S. Miller91521942009-05-28 21:35:47 -07004757 if (!skb_queue_empty(&sk->sk_receive_queue))
4758 tcp_collapse(sk, &sk->sk_receive_queue,
4759 skb_peek(&sk->sk_receive_queue),
4760 NULL,
4761 tp->copied_seq, tp->rcv_nxt);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08004762 sk_mem_reclaim(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004763
4764 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4765 return 0;
4766
4767 /* Collapsing did not help, destructive actions follow.
4768 * This must not ever occur. */
4769
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07004770 tcp_prune_ofo_queue(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004771
4772 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4773 return 0;
4774
4775 /* If we are really being abused, tell the caller to silently
4776 * drop receive data on the floor. It will get retransmitted
4777 * and hopefully then we'll have sufficient space.
4778 */
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07004779 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004780
4781 /* Massive buffer overcommit. */
4782 tp->pred_flags = 0;
4783 return -1;
4784}
4785
Linus Torvalds1da177e2005-04-16 15:20:36 -07004786/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4787 * As additional protections, we do not touch cwnd in retransmission phases,
4788 * and if application hit its sndbuf limit recently.
4789 */
4790void tcp_cwnd_application_limited(struct sock *sk)
4791{
4792 struct tcp_sock *tp = tcp_sk(sk);
4793
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03004794 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07004795 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4796 /* Limited by application or receiver window. */
Ilpo Järvinend254bcd2006-08-04 16:57:42 -07004797 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4798 u32 win_used = max(tp->snd_cwnd_used, init_win);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004799 if (win_used < tp->snd_cwnd) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03004800 tp->snd_ssthresh = tcp_current_ssthresh(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004801 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4802 }
4803 tp->snd_cwnd_used = 0;
4804 }
4805 tp->snd_cwnd_stamp = tcp_time_stamp;
4806}
4807
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004808static int tcp_should_expand_sndbuf(struct sock *sk)
David S. Miller0d9901d2005-07-05 15:21:10 -07004809{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004810 struct tcp_sock *tp = tcp_sk(sk);
4811
David S. Miller0d9901d2005-07-05 15:21:10 -07004812 /* If the user specified a specific send buffer setting, do
4813 * not modify it.
4814 */
4815 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4816 return 0;
4817
4818 /* If we are under global TCP memory pressure, do not expand. */
4819 if (tcp_memory_pressure)
4820 return 0;
4821
4822 /* If we are under soft global TCP memory pressure, do not expand. */
4823 if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
4824 return 0;
4825
4826 /* If we filled the congestion window, do not expand. */
4827 if (tp->packets_out >= tp->snd_cwnd)
4828 return 0;
4829
4830 return 1;
4831}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004832
4833/* When incoming ACK allowed to free some skb from write_queue,
4834 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4835 * on the exit from tcp input handler.
4836 *
4837 * PROBLEM: sndbuf expansion does not work well with largesend.
4838 */
4839static void tcp_new_space(struct sock *sk)
4840{
4841 struct tcp_sock *tp = tcp_sk(sk);
4842
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004843 if (tcp_should_expand_sndbuf(sk)) {
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004844 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
Ilpo Järvinen4a7e5602008-10-07 14:43:31 -07004845 MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
4846 int demanded = max_t(unsigned int, tp->snd_cwnd,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004847 tp->reordering + 1);
4848 sndmem *= 2 * demanded;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004849 if (sndmem > sk->sk_sndbuf)
4850 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
4851 tp->snd_cwnd_stamp = tcp_time_stamp;
4852 }
4853
4854 sk->sk_write_space(sk);
4855}
4856
Stephen Hemminger40efc6f2006-01-03 16:03:49 -08004857static void tcp_check_space(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004858{
4859 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4860 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4861 if (sk->sk_socket &&
4862 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4863 tcp_new_space(sk);
4864 }
4865}
4866
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004867static inline void tcp_data_snd_check(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004868{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004869 tcp_push_pending_frames(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004870 tcp_check_space(sk);
4871}
4872
4873/*
4874 * Check if sending an ack is needed.
4875 */
4876static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4877{
4878 struct tcp_sock *tp = tcp_sk(sk);
4879
4880 /* More than one full frame received... */
Joe Perches9d4fb272009-11-23 10:41:23 -08004881 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07004882 /* ... and right edge of window advances far enough.
4883 * (tcp_recvmsg() will send ACK otherwise). Or...
4884 */
Joe Perches9d4fb272009-11-23 10:41:23 -08004885 __tcp_select_window(sk) >= tp->rcv_wnd) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07004886 /* We ACK each frame or... */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004887 tcp_in_quickack_mode(sk) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07004888 /* We have out of order data. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004889 (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004890 /* Then ack it now */
4891 tcp_send_ack(sk);
4892 } else {
4893 /* Else, send delayed ack. */
4894 tcp_send_delayed_ack(sk);
4895 }
4896}
4897
Stephen Hemminger40efc6f2006-01-03 16:03:49 -08004898static inline void tcp_ack_snd_check(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004899{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004900 if (!inet_csk_ack_scheduled(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004901 /* We sent a data segment already. */
4902 return;
4903 }
4904 __tcp_ack_snd_check(sk, 1);
4905}
4906
4907/*
4908 * This routine is only called when we have urgent data
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08004909 * signaled. Its the 'slow' part of tcp_urg. It could be
Linus Torvalds1da177e2005-04-16 15:20:36 -07004910 * moved inline now as tcp_urg is only called from one
4911 * place. We handle URGent data wrong. We have to - as
4912 * BSD still doesn't use the correction from RFC961.
4913 * For 1003.1g we should support a new option TCP_STDURG to permit
4914 * either form (or just set the sysctl tcp_stdurg).
4915 */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004916
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004917static void tcp_check_urg(struct sock *sk, struct tcphdr *th)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004918{
4919 struct tcp_sock *tp = tcp_sk(sk);
4920 u32 ptr = ntohs(th->urg_ptr);
4921
4922 if (ptr && !sysctl_tcp_stdurg)
4923 ptr--;
4924 ptr += ntohl(th->seq);
4925
4926 /* Ignore urgent data that we've already seen and read. */
4927 if (after(tp->copied_seq, ptr))
4928 return;
4929
4930 /* Do not replay urg ptr.
4931 *
4932 * NOTE: interesting situation not covered by specs.
4933 * Misbehaving sender may send urg ptr, pointing to segment,
4934 * which we already have in ofo queue. We are not able to fetch
4935 * such data and will stay in TCP_URG_NOTYET until will be eaten
4936 * by recvmsg(). Seems, we are not obliged to handle such wicked
4937 * situations. But it is worth to think about possibility of some
4938 * DoSes using some hypothetical application level deadlock.
4939 */
4940 if (before(ptr, tp->rcv_nxt))
4941 return;
4942
4943 /* Do we already have a newer (or duplicate) urgent pointer? */
4944 if (tp->urg_data && !after(ptr, tp->urg_seq))
4945 return;
4946
4947 /* Tell the world about our new urgent pointer. */
4948 sk_send_sigurg(sk);
4949
4950 /* We may be adding urgent data when the last byte read was
4951 * urgent. To do this requires some care. We cannot just ignore
4952 * tp->copied_seq since we would read the last urgent byte again
4953 * as data, nor can we alter copied_seq until this data arrives
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08004954 * or we break the semantics of SIOCATMARK (and thus sockatmark())
Linus Torvalds1da177e2005-04-16 15:20:36 -07004955 *
4956 * NOTE. Double Dutch. Rendering to plain English: author of comment
4957 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4958 * and expect that both A and B disappear from stream. This is _wrong_.
4959 * Though this happens in BSD with high probability, this is occasional.
4960 * Any application relying on this is buggy. Note also, that fix "works"
4961 * only in this artificial test. Insert some normal data between A and B and we will
4962 * decline of BSD again. Verdict: it is better to remove to trap
4963 * buggy users.
4964 */
4965 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004966 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004967 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4968 tp->copied_seq++;
4969 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
David S. Miller8728b832005-08-09 19:25:21 -07004970 __skb_unlink(skb, &sk->sk_receive_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004971 __kfree_skb(skb);
4972 }
4973 }
4974
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004975 tp->urg_data = TCP_URG_NOTYET;
4976 tp->urg_seq = ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004977
4978 /* Disable header prediction. */
4979 tp->pred_flags = 0;
4980}
4981
4982/* This is the 'fast' part of urgent handling. */
4983static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
4984{
4985 struct tcp_sock *tp = tcp_sk(sk);
4986
4987 /* Check if we get a new urgent pointer - normally not. */
4988 if (th->urg)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004989 tcp_check_urg(sk, th);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004990
4991 /* Do we wait for any urgent data? - normally not... */
4992 if (tp->urg_data == TCP_URG_NOTYET) {
4993 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4994 th->syn;
4995
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004996 /* Is the urgent pointer pointing into this packet? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004997 if (ptr < skb->len) {
4998 u8 tmp;
4999 if (skb_copy_bits(skb, ptr, &tmp, 1))
5000 BUG();
5001 tp->urg_data = TCP_URG_VALID | tmp;
5002 if (!sock_flag(sk, SOCK_DEAD))
5003 sk->sk_data_ready(sk, 0);
5004 }
5005 }
5006}
5007
5008static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
5009{
5010 struct tcp_sock *tp = tcp_sk(sk);
5011 int chunk = skb->len - hlen;
5012 int err;
5013
5014 local_bh_enable();
Herbert Xu60476372007-04-09 11:59:39 -07005015 if (skb_csum_unnecessary(skb))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005016 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
5017 else
5018 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
5019 tp->ucopy.iov);
5020
5021 if (!err) {
5022 tp->ucopy.len -= chunk;
5023 tp->copied_seq += chunk;
5024 tcp_rcv_space_adjust(sk);
5025 }
5026
5027 local_bh_disable();
5028 return err;
5029}
5030
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005031static __sum16 __tcp_checksum_complete_user(struct sock *sk,
5032 struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005033{
Al Virob51655b2006-11-14 21:40:42 -08005034 __sum16 result;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005035
5036 if (sock_owned_by_user(sk)) {
5037 local_bh_enable();
5038 result = __tcp_checksum_complete(skb);
5039 local_bh_disable();
5040 } else {
5041 result = __tcp_checksum_complete(skb);
5042 }
5043 return result;
5044}
5045
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005046static inline int tcp_checksum_complete_user(struct sock *sk,
5047 struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005048{
Herbert Xu60476372007-04-09 11:59:39 -07005049 return !skb_csum_unnecessary(skb) &&
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005050 __tcp_checksum_complete_user(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005051}
5052
Chris Leech1a2449a2006-05-23 18:05:53 -07005053#ifdef CONFIG_NET_DMA
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005054static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb,
5055 int hlen)
Chris Leech1a2449a2006-05-23 18:05:53 -07005056{
5057 struct tcp_sock *tp = tcp_sk(sk);
5058 int chunk = skb->len - hlen;
5059 int dma_cookie;
5060 int copied_early = 0;
5061
5062 if (tp->ucopy.wakeup)
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005063 return 0;
Chris Leech1a2449a2006-05-23 18:05:53 -07005064
5065 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
Dan Williamsf67b4592009-01-06 11:38:15 -07005066 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY);
Chris Leech1a2449a2006-05-23 18:05:53 -07005067
Herbert Xu60476372007-04-09 11:59:39 -07005068 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
Chris Leech1a2449a2006-05-23 18:05:53 -07005069
5070 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005071 skb, hlen,
5072 tp->ucopy.iov, chunk,
5073 tp->ucopy.pinned_list);
Chris Leech1a2449a2006-05-23 18:05:53 -07005074
5075 if (dma_cookie < 0)
5076 goto out;
5077
5078 tp->ucopy.dma_cookie = dma_cookie;
5079 copied_early = 1;
5080
5081 tp->ucopy.len -= chunk;
5082 tp->copied_seq += chunk;
5083 tcp_rcv_space_adjust(sk);
5084
5085 if ((tp->ucopy.len == 0) ||
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07005086 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
Chris Leech1a2449a2006-05-23 18:05:53 -07005087 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
5088 tp->ucopy.wakeup = 1;
5089 sk->sk_data_ready(sk, 0);
5090 }
5091 } else if (chunk > 0) {
5092 tp->ucopy.wakeup = 1;
5093 sk->sk_data_ready(sk, 0);
5094 }
5095out:
5096 return copied_early;
5097}
5098#endif /* CONFIG_NET_DMA */
5099
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005100/* Does PAWS and seqno based validation of an incoming segment, flags will
5101 * play significant role here.
5102 */
5103static int tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5104 struct tcphdr *th, int syn_inerr)
5105{
William Allen Simpson4957faade2009-12-02 18:25:27 +00005106 u8 *hash_location;
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005107 struct tcp_sock *tp = tcp_sk(sk);
5108
5109 /* RFC1323: H1. Apply PAWS check first. */
William Allen Simpson4957faade2009-12-02 18:25:27 +00005110 if (tcp_fast_parse_options(skb, th, tp, &hash_location) &&
5111 tp->rx_opt.saw_tstamp &&
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005112 tcp_paws_discard(sk, skb)) {
5113 if (!th->rst) {
5114 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5115 tcp_send_dupack(sk, skb);
5116 goto discard;
5117 }
5118 /* Reset is accepted even if it did not pass PAWS. */
5119 }
5120
5121 /* Step 1: check sequence number */
5122 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5123 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5124 * (RST) segments are validated by checking their SEQ-fields."
5125 * And page 69: "If an incoming segment is not acceptable,
5126 * an acknowledgment should be sent in reply (unless the RST
5127 * bit is set, if so drop the segment and return)".
5128 */
5129 if (!th->rst)
5130 tcp_send_dupack(sk, skb);
5131 goto discard;
5132 }
5133
5134 /* Step 2: check RST bit */
5135 if (th->rst) {
5136 tcp_reset(sk);
5137 goto discard;
5138 }
5139
5140 /* ts_recent update must be made after we are sure that the packet
5141 * is in window.
5142 */
5143 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
5144
5145 /* step 3: check security and precedence [ignored] */
5146
5147 /* step 4: Check for a SYN in window. */
5148 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5149 if (syn_inerr)
5150 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5151 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONSYN);
5152 tcp_reset(sk);
5153 return -1;
5154 }
5155
5156 return 1;
5157
5158discard:
5159 __kfree_skb(skb);
5160 return 0;
5161}
5162
Linus Torvalds1da177e2005-04-16 15:20:36 -07005163/*
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005164 * TCP receive function for the ESTABLISHED state.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005165 *
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005166 * It is split into a fast path and a slow path. The fast path is
Linus Torvalds1da177e2005-04-16 15:20:36 -07005167 * disabled when:
5168 * - A zero window was announced from us - zero window probing
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005169 * is only handled properly in the slow path.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005170 * - Out of order segments arrived.
5171 * - Urgent data is expected.
5172 * - There is no buffer space left
5173 * - Unexpected TCP flags/window values/header lengths are received
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005174 * (detected by checking the TCP header against pred_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005175 * - Data is sent in both directions. Fast path only supports pure senders
5176 * or pure receivers (this means either the sequence number or the ack
5177 * value must stay constant)
5178 * - Unexpected TCP option.
5179 *
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005180 * When these conditions are not satisfied it drops into a standard
Linus Torvalds1da177e2005-04-16 15:20:36 -07005181 * receive procedure patterned after RFC793 to handle all cases.
5182 * The first three cases are guaranteed by proper pred_flags setting,
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005183 * the rest is checked inline. Fast processing is turned on in
Linus Torvalds1da177e2005-04-16 15:20:36 -07005184 * tcp_data_queue when everything is OK.
5185 */
5186int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5187 struct tcphdr *th, unsigned len)
5188{
5189 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005190 int res;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005191
5192 /*
5193 * Header prediction.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005194 * The code loosely follows the one in the famous
Linus Torvalds1da177e2005-04-16 15:20:36 -07005195 * "30 instruction TCP receive" Van Jacobson mail.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005196 *
5197 * Van's trick is to deposit buffers into socket queue
Linus Torvalds1da177e2005-04-16 15:20:36 -07005198 * on a device interrupt, to call tcp_recv function
5199 * on the receive process context and checksum and copy
5200 * the buffer to user space. smart...
5201 *
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005202 * Our current scheme is not silly either but we take the
Linus Torvalds1da177e2005-04-16 15:20:36 -07005203 * extra cost of the net_bh soft interrupt processing...
5204 * We do checksum and copy also but from device to kernel.
5205 */
5206
5207 tp->rx_opt.saw_tstamp = 0;
5208
5209 /* pred_flags is 0xS?10 << 16 + snd_wnd
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08005210 * if header_prediction is to be made
Linus Torvalds1da177e2005-04-16 15:20:36 -07005211 * 'S' will always be tp->tcp_header_len >> 2
5212 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005213 * turn it off (when there are holes in the receive
Linus Torvalds1da177e2005-04-16 15:20:36 -07005214 * space for instance)
5215 * PSH flag is ignored.
5216 */
5217
5218 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
John Dykstra96e0bf42009-03-22 21:49:57 -07005219 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5220 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005221 int tcp_header_len = tp->tcp_header_len;
5222
5223 /* Timestamp header prediction: tcp_header_len
5224 * is automatically equal to th->doff*4 due to pred_flags
5225 * match.
5226 */
5227
5228 /* Check timestamp */
5229 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005230 /* No? Slow path! */
Ilpo Järvinena4356b22008-08-23 05:12:29 -07005231 if (!tcp_parse_aligned_timestamp(tp, th))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005232 goto slow_path;
5233
Linus Torvalds1da177e2005-04-16 15:20:36 -07005234 /* If PAWS failed, check it more carefully in slow path */
5235 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5236 goto slow_path;
5237
5238 /* DO NOT update ts_recent here, if checksum fails
5239 * and timestamp was corrupted part, it will result
5240 * in a hung connection since we will drop all
5241 * future packets due to the PAWS test.
5242 */
5243 }
5244
5245 if (len <= tcp_header_len) {
5246 /* Bulk data transfer: sender */
5247 if (len == tcp_header_len) {
5248 /* Predicted packet is in window by definition.
5249 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5250 * Hence, check seq<=rcv_wup reduces to:
5251 */
5252 if (tcp_header_len ==
5253 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5254 tp->rcv_nxt == tp->rcv_wup)
5255 tcp_store_ts_recent(tp);
5256
Linus Torvalds1da177e2005-04-16 15:20:36 -07005257 /* We know that such packets are checksummed
5258 * on entry.
5259 */
5260 tcp_ack(sk, skb, 0);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005261 __kfree_skb(skb);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005262 tcp_data_snd_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005263 return 0;
5264 } else { /* Header too small */
Pavel Emelyanov63231bd2008-07-16 20:22:25 -07005265 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005266 goto discard;
5267 }
5268 } else {
5269 int eaten = 0;
Chris Leech1a2449a2006-05-23 18:05:53 -07005270 int copied_early = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005271
Chris Leech1a2449a2006-05-23 18:05:53 -07005272 if (tp->copied_seq == tp->rcv_nxt &&
5273 len - tcp_header_len <= tp->ucopy.len) {
5274#ifdef CONFIG_NET_DMA
5275 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
5276 copied_early = 1;
5277 eaten = 1;
5278 }
5279#endif
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005280 if (tp->ucopy.task == current &&
5281 sock_owned_by_user(sk) && !copied_early) {
Chris Leech1a2449a2006-05-23 18:05:53 -07005282 __set_current_state(TASK_RUNNING);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005283
Chris Leech1a2449a2006-05-23 18:05:53 -07005284 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
5285 eaten = 1;
5286 }
5287 if (eaten) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005288 /* Predicted packet is in window by definition.
5289 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5290 * Hence, check seq<=rcv_wup reduces to:
5291 */
5292 if (tcp_header_len ==
5293 (sizeof(struct tcphdr) +
5294 TCPOLEN_TSTAMP_ALIGNED) &&
5295 tp->rcv_nxt == tp->rcv_wup)
5296 tcp_store_ts_recent(tp);
5297
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005298 tcp_rcv_rtt_measure_ts(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005299
5300 __skb_pull(skb, tcp_header_len);
5301 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07005302 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005303 }
Chris Leech1a2449a2006-05-23 18:05:53 -07005304 if (copied_early)
5305 tcp_cleanup_rbuf(sk, skb->len);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005306 }
5307 if (!eaten) {
5308 if (tcp_checksum_complete_user(sk, skb))
5309 goto csum_error;
5310
5311 /* Predicted packet is in window by definition.
5312 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5313 * Hence, check seq<=rcv_wup reduces to:
5314 */
5315 if (tcp_header_len ==
5316 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5317 tp->rcv_nxt == tp->rcv_wup)
5318 tcp_store_ts_recent(tp);
5319
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005320 tcp_rcv_rtt_measure_ts(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005321
5322 if ((int)skb->truesize > sk->sk_forward_alloc)
5323 goto step5;
5324
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07005325 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005326
5327 /* Bulk data transfer: receiver */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005328 __skb_pull(skb, tcp_header_len);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005329 __skb_queue_tail(&sk->sk_receive_queue, skb);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08005330 skb_set_owner_r(skb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005331 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5332 }
5333
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005334 tcp_event_data_recv(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005335
5336 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5337 /* Well, only one small jumplet in fast path... */
5338 tcp_ack(sk, skb, FLAG_DATA);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005339 tcp_data_snd_check(sk);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005340 if (!inet_csk_ack_scheduled(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005341 goto no_ack;
5342 }
5343
Ali Saidi53240c22008-10-07 15:31:19 -07005344 if (!copied_early || tp->rcv_nxt != tp->rcv_wup)
5345 __tcp_ack_snd_check(sk, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005346no_ack:
Chris Leech1a2449a2006-05-23 18:05:53 -07005347#ifdef CONFIG_NET_DMA
5348 if (copied_early)
5349 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
5350 else
5351#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07005352 if (eaten)
5353 __kfree_skb(skb);
5354 else
5355 sk->sk_data_ready(sk, 0);
5356 return 0;
5357 }
5358 }
5359
5360slow_path:
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005361 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005362 goto csum_error;
5363
5364 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07005365 * Standard slow path.
5366 */
5367
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005368 res = tcp_validate_incoming(sk, skb, th, 1);
5369 if (res <= 0)
5370 return -res;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005371
5372step5:
John Dykstra96e0bf42009-03-22 21:49:57 -07005373 if (th->ack && tcp_ack(sk, skb, FLAG_SLOWPATH) < 0)
5374 goto discard;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005375
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005376 tcp_rcv_rtt_measure_ts(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005377
5378 /* Process urgent data. */
5379 tcp_urg(sk, skb, th);
5380
5381 /* step 7: process the segment text */
5382 tcp_data_queue(sk, skb);
5383
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005384 tcp_data_snd_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005385 tcp_ack_snd_check(sk);
5386 return 0;
5387
5388csum_error:
Pavel Emelyanov63231bd2008-07-16 20:22:25 -07005389 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005390
5391discard:
5392 __kfree_skb(skb);
5393 return 0;
5394}
5395
5396static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5397 struct tcphdr *th, unsigned len)
5398{
William Allen Simpson4957faade2009-12-02 18:25:27 +00005399 u8 *hash_location;
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08005400 struct inet_connection_sock *icsk = inet_csk(sk);
William Allen Simpson4957faade2009-12-02 18:25:27 +00005401 struct tcp_sock *tp = tcp_sk(sk);
Gilad Ben-Yossef022c3f72009-10-28 04:15:22 +00005402 struct dst_entry *dst = __sk_dst_get(sk);
William Allen Simpson4957faade2009-12-02 18:25:27 +00005403 struct tcp_cookie_values *cvp = tp->cookie_values;
5404 int saved_clamp = tp->rx_opt.mss_clamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005405
William Allen Simpson4957faade2009-12-02 18:25:27 +00005406 tcp_parse_options(skb, &tp->rx_opt, &hash_location, 0, dst);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005407
5408 if (th->ack) {
5409 /* rfc793:
5410 * "If the state is SYN-SENT then
5411 * first check the ACK bit
5412 * If the ACK bit is set
5413 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5414 * a reset (unless the RST bit is set, if so drop
5415 * the segment and return)"
5416 *
5417 * We do not send data with SYN, so that RFC-correct
5418 * test reduces to:
5419 */
5420 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
5421 goto reset_and_undo;
5422
5423 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5424 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5425 tcp_time_stamp)) {
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07005426 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005427 goto reset_and_undo;
5428 }
5429
5430 /* Now ACK is acceptable.
5431 *
5432 * "If the RST bit is set
5433 * If the ACK was acceptable then signal the user "error:
5434 * connection reset", drop the segment, enter CLOSED state,
5435 * delete TCB, and return."
5436 */
5437
5438 if (th->rst) {
5439 tcp_reset(sk);
5440 goto discard;
5441 }
5442
5443 /* rfc793:
5444 * "fifth, if neither of the SYN or RST bits is set then
5445 * drop the segment and return."
5446 *
5447 * See note below!
5448 * --ANK(990513)
5449 */
5450 if (!th->syn)
5451 goto discard_and_undo;
5452
5453 /* rfc793:
5454 * "If the SYN bit is on ...
5455 * are acceptable then ...
5456 * (our SYN has been ACKed), change the connection
5457 * state to ESTABLISHED..."
5458 */
5459
5460 TCP_ECN_rcv_synack(tp, th);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005461
5462 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5463 tcp_ack(sk, skb, FLAG_SLOWPATH);
5464
5465 /* Ok.. it's good. Set up sequence numbers and
5466 * move to established.
5467 */
5468 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5469 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5470
5471 /* RFC1323: The window in SYN & SYN/ACK segments is
5472 * never scaled.
5473 */
5474 tp->snd_wnd = ntohs(th->window);
Hantzis Fotisee7537b2009-03-02 22:42:02 -08005475 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005476
5477 if (!tp->rx_opt.wscale_ok) {
5478 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5479 tp->window_clamp = min(tp->window_clamp, 65535U);
5480 }
5481
5482 if (tp->rx_opt.saw_tstamp) {
5483 tp->rx_opt.tstamp_ok = 1;
5484 tp->tcp_header_len =
5485 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5486 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5487 tcp_store_ts_recent(tp);
5488 } else {
5489 tp->tcp_header_len = sizeof(struct tcphdr);
5490 }
5491
Ilpo Järvinene60402d2007-08-09 15:14:46 +03005492 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5493 tcp_enable_fack(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005494
John Heffner5d424d52006-03-20 17:53:41 -08005495 tcp_mtup_init(sk);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08005496 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005497 tcp_initialize_rcv_mss(sk);
5498
5499 /* Remember, tcp_poll() does not lock socket!
5500 * Change state from SYN-SENT only after copied_seq
5501 * is initialized. */
5502 tp->copied_seq = tp->rcv_nxt;
William Allen Simpson4957faade2009-12-02 18:25:27 +00005503
5504 if (cvp != NULL &&
5505 cvp->cookie_pair_size > 0 &&
5506 tp->rx_opt.cookie_plus > 0) {
5507 int cookie_size = tp->rx_opt.cookie_plus
5508 - TCPOLEN_COOKIE_BASE;
5509 int cookie_pair_size = cookie_size
5510 + cvp->cookie_desired;
5511
5512 /* A cookie extension option was sent and returned.
5513 * Note that each incoming SYNACK replaces the
5514 * Responder cookie. The initial exchange is most
5515 * fragile, as protection against spoofing relies
5516 * entirely upon the sequence and timestamp (above).
5517 * This replacement strategy allows the correct pair to
5518 * pass through, while any others will be filtered via
5519 * Responder verification later.
5520 */
5521 if (sizeof(cvp->cookie_pair) >= cookie_pair_size) {
5522 memcpy(&cvp->cookie_pair[cvp->cookie_desired],
5523 hash_location, cookie_size);
5524 cvp->cookie_pair_size = cookie_pair_size;
5525 }
5526 }
5527
Ralf Baechlee16aa202006-12-07 00:11:33 -08005528 smp_mb();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005529 tcp_set_state(sk, TCP_ESTABLISHED);
5530
Venkat Yekkirala6b877692006-11-08 17:04:09 -06005531 security_inet_conn_established(sk, skb);
5532
Linus Torvalds1da177e2005-04-16 15:20:36 -07005533 /* Make sure socket is routed, for correct metrics. */
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -08005534 icsk->icsk_af_ops->rebuild_header(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005535
5536 tcp_init_metrics(sk);
5537
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03005538 tcp_init_congestion_control(sk);
Stephen Hemminger317a76f2005-06-23 12:19:55 -07005539
Linus Torvalds1da177e2005-04-16 15:20:36 -07005540 /* Prevent spurious tcp_cwnd_restart() on first data
5541 * packet.
5542 */
5543 tp->lsndtime = tcp_time_stamp;
5544
5545 tcp_init_buffer_space(sk);
5546
5547 if (sock_flag(sk, SOCK_KEEPOPEN))
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005548 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005549
5550 if (!tp->rx_opt.snd_wscale)
5551 __tcp_fast_path_on(tp, tp->snd_wnd);
5552 else
5553 tp->pred_flags = 0;
5554
5555 if (!sock_flag(sk, SOCK_DEAD)) {
5556 sk->sk_state_change(sk);
Pavel Emelyanov8d8ad9d2007-11-26 20:10:50 +08005557 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005558 }
5559
Arnaldo Carvalho de Melo295f7322005-08-09 20:11:56 -07005560 if (sk->sk_write_pending ||
5561 icsk->icsk_accept_queue.rskq_defer_accept ||
5562 icsk->icsk_ack.pingpong) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005563 /* Save one ACK. Data will be ready after
5564 * several ticks, if write_pending is set.
5565 *
5566 * It may be deleted, but with this feature tcpdumps
5567 * look so _wonderfully_ clever, that I was not able
5568 * to stand against the temptation 8) --ANK
5569 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005570 inet_csk_schedule_ack(sk);
Arnaldo Carvalho de Melo295f7322005-08-09 20:11:56 -07005571 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5572 icsk->icsk_ack.ato = TCP_ATO_MIN;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005573 tcp_incr_quickack(sk);
5574 tcp_enter_quickack_mode(sk);
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07005575 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5576 TCP_DELACK_MAX, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005577
5578discard:
5579 __kfree_skb(skb);
5580 return 0;
5581 } else {
5582 tcp_send_ack(sk);
5583 }
5584 return -1;
5585 }
5586
5587 /* No ACK in the segment */
5588
5589 if (th->rst) {
5590 /* rfc793:
5591 * "If the RST bit is set
5592 *
5593 * Otherwise (no ACK) drop the segment and return."
5594 */
5595
5596 goto discard_and_undo;
5597 }
5598
5599 /* PAWS check. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005600 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
Ilpo Järvinenc887e6d2009-03-14 14:23:03 +00005601 tcp_paws_reject(&tp->rx_opt, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005602 goto discard_and_undo;
5603
5604 if (th->syn) {
5605 /* We see SYN without ACK. It is attempt of
5606 * simultaneous connect with crossed SYNs.
5607 * Particularly, it can be connect to self.
5608 */
5609 tcp_set_state(sk, TCP_SYN_RECV);
5610
5611 if (tp->rx_opt.saw_tstamp) {
5612 tp->rx_opt.tstamp_ok = 1;
5613 tcp_store_ts_recent(tp);
5614 tp->tcp_header_len =
5615 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5616 } else {
5617 tp->tcp_header_len = sizeof(struct tcphdr);
5618 }
5619
5620 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5621 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5622
5623 /* RFC1323: The window in SYN & SYN/ACK segments is
5624 * never scaled.
5625 */
5626 tp->snd_wnd = ntohs(th->window);
5627 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5628 tp->max_window = tp->snd_wnd;
5629
5630 TCP_ECN_rcv_syn(tp, th);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005631
John Heffner5d424d52006-03-20 17:53:41 -08005632 tcp_mtup_init(sk);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08005633 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005634 tcp_initialize_rcv_mss(sk);
5635
Linus Torvalds1da177e2005-04-16 15:20:36 -07005636 tcp_send_synack(sk);
5637#if 0
5638 /* Note, we could accept data and URG from this segment.
5639 * There are no obstacles to make this.
5640 *
5641 * However, if we ignore data in ACKless segments sometimes,
5642 * we have no reasons to accept it sometimes.
5643 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5644 * is not flawless. So, discard packet for sanity.
5645 * Uncomment this return to process the data.
5646 */
5647 return -1;
5648#else
5649 goto discard;
5650#endif
5651 }
5652 /* "fifth, if neither of the SYN or RST bits is set then
5653 * drop the segment and return."
5654 */
5655
5656discard_and_undo:
5657 tcp_clear_options(&tp->rx_opt);
5658 tp->rx_opt.mss_clamp = saved_clamp;
5659 goto discard;
5660
5661reset_and_undo:
5662 tcp_clear_options(&tp->rx_opt);
5663 tp->rx_opt.mss_clamp = saved_clamp;
5664 return 1;
5665}
5666
Linus Torvalds1da177e2005-04-16 15:20:36 -07005667/*
5668 * This function implements the receiving procedure of RFC 793 for
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005669 * all states except ESTABLISHED and TIME_WAIT.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005670 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5671 * address independent.
5672 */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005673
Linus Torvalds1da177e2005-04-16 15:20:36 -07005674int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5675 struct tcphdr *th, unsigned len)
5676{
5677 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -08005678 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005679 int queued = 0;
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005680 int res;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005681
5682 tp->rx_opt.saw_tstamp = 0;
5683
5684 switch (sk->sk_state) {
5685 case TCP_CLOSE:
5686 goto discard;
5687
5688 case TCP_LISTEN:
Stephen Hemminger2de979b2007-03-08 20:45:19 -08005689 if (th->ack)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005690 return 1;
5691
Stephen Hemminger2de979b2007-03-08 20:45:19 -08005692 if (th->rst)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005693 goto discard;
5694
Stephen Hemminger2de979b2007-03-08 20:45:19 -08005695 if (th->syn) {
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -08005696 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005697 return 1;
5698
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005699 /* Now we have several options: In theory there is
5700 * nothing else in the frame. KA9Q has an option to
Linus Torvalds1da177e2005-04-16 15:20:36 -07005701 * send data with the syn, BSD accepts data with the
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005702 * syn up to the [to be] advertised window and
5703 * Solaris 2.1 gives you a protocol error. For now
5704 * we just ignore it, that fits the spec precisely
Linus Torvalds1da177e2005-04-16 15:20:36 -07005705 * and avoids incompatibilities. It would be nice in
5706 * future to drop through and process the data.
5707 *
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005708 * Now that TTCP is starting to be used we ought to
Linus Torvalds1da177e2005-04-16 15:20:36 -07005709 * queue this data.
5710 * But, this leaves one open to an easy denial of
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005711 * service attack, and SYN cookies can't defend
Linus Torvalds1da177e2005-04-16 15:20:36 -07005712 * against this problem. So, we drop the data
Masayuki Nakagawafb7e2392007-01-23 20:15:06 -08005713 * in the interest of security over speed unless
5714 * it's still in use.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005715 */
Masayuki Nakagawafb7e2392007-01-23 20:15:06 -08005716 kfree_skb(skb);
5717 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005718 }
5719 goto discard;
5720
5721 case TCP_SYN_SENT:
Linus Torvalds1da177e2005-04-16 15:20:36 -07005722 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5723 if (queued >= 0)
5724 return queued;
5725
5726 /* Do step6 onward by hand. */
5727 tcp_urg(sk, skb, th);
5728 __kfree_skb(skb);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005729 tcp_data_snd_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005730 return 0;
5731 }
5732
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005733 res = tcp_validate_incoming(sk, skb, th, 0);
5734 if (res <= 0)
5735 return -res;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005736
5737 /* step 5: check the ACK field */
5738 if (th->ack) {
John Dykstra96e0bf42009-03-22 21:49:57 -07005739 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH) > 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005740
Stephen Hemminger2de979b2007-03-08 20:45:19 -08005741 switch (sk->sk_state) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005742 case TCP_SYN_RECV:
5743 if (acceptable) {
5744 tp->copied_seq = tp->rcv_nxt;
Ralf Baechlee16aa202006-12-07 00:11:33 -08005745 smp_mb();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005746 tcp_set_state(sk, TCP_ESTABLISHED);
5747 sk->sk_state_change(sk);
5748
5749 /* Note, that this wakeup is only for marginal
5750 * crossed SYN case. Passively open sockets
5751 * are not waked up, because sk->sk_sleep ==
5752 * NULL and sk->sk_socket == NULL.
5753 */
Pavel Emelyanov8d8ad9d2007-11-26 20:10:50 +08005754 if (sk->sk_socket)
5755 sk_wake_async(sk,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005756 SOCK_WAKE_IO, POLL_OUT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005757
5758 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5759 tp->snd_wnd = ntohs(th->window) <<
5760 tp->rx_opt.snd_wscale;
Hantzis Fotisee7537b2009-03-02 22:42:02 -08005761 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005762
5763 /* tcp_ack considers this ACK as duplicate
5764 * and does not calculate rtt.
5765 * Fix it at least with timestamps.
5766 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005767 if (tp->rx_opt.saw_tstamp &&
5768 tp->rx_opt.rcv_tsecr && !tp->srtt)
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08005769 tcp_ack_saw_tstamp(sk, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005770
5771 if (tp->rx_opt.tstamp_ok)
5772 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5773
5774 /* Make sure socket is routed, for
5775 * correct metrics.
5776 */
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -08005777 icsk->icsk_af_ops->rebuild_header(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005778
5779 tcp_init_metrics(sk);
5780
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03005781 tcp_init_congestion_control(sk);
Stephen Hemminger317a76f2005-06-23 12:19:55 -07005782
Linus Torvalds1da177e2005-04-16 15:20:36 -07005783 /* Prevent spurious tcp_cwnd_restart() on
5784 * first data packet.
5785 */
5786 tp->lsndtime = tcp_time_stamp;
5787
John Heffner5d424d52006-03-20 17:53:41 -08005788 tcp_mtup_init(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005789 tcp_initialize_rcv_mss(sk);
5790 tcp_init_buffer_space(sk);
5791 tcp_fast_path_on(tp);
5792 } else {
5793 return 1;
5794 }
5795 break;
5796
5797 case TCP_FIN_WAIT1:
5798 if (tp->snd_una == tp->write_seq) {
5799 tcp_set_state(sk, TCP_FIN_WAIT2);
5800 sk->sk_shutdown |= SEND_SHUTDOWN;
5801 dst_confirm(sk->sk_dst_cache);
5802
5803 if (!sock_flag(sk, SOCK_DEAD))
5804 /* Wake up lingering close() */
5805 sk->sk_state_change(sk);
5806 else {
5807 int tmo;
5808
5809 if (tp->linger2 < 0 ||
5810 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5811 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5812 tcp_done(sk);
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07005813 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005814 return 1;
5815 }
5816
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005817 tmo = tcp_fin_time(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005818 if (tmo > TCP_TIMEWAIT_LEN) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005819 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005820 } else if (th->fin || sock_owned_by_user(sk)) {
5821 /* Bad case. We could lose such FIN otherwise.
5822 * It is not a big problem, but it looks confusing
5823 * and not so rare event. We still can lose it now,
5824 * if it spins in bh_lock_sock(), but it is really
5825 * marginal case.
5826 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005827 inet_csk_reset_keepalive_timer(sk, tmo);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005828 } else {
5829 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5830 goto discard;
5831 }
5832 }
5833 }
5834 break;
5835
5836 case TCP_CLOSING:
5837 if (tp->snd_una == tp->write_seq) {
5838 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5839 goto discard;
5840 }
5841 break;
5842
5843 case TCP_LAST_ACK:
5844 if (tp->snd_una == tp->write_seq) {
5845 tcp_update_metrics(sk);
5846 tcp_done(sk);
5847 goto discard;
5848 }
5849 break;
5850 }
5851 } else
5852 goto discard;
5853
5854 /* step 6: check the URG bit */
5855 tcp_urg(sk, skb, th);
5856
5857 /* step 7: process the segment text */
5858 switch (sk->sk_state) {
5859 case TCP_CLOSE_WAIT:
5860 case TCP_CLOSING:
5861 case TCP_LAST_ACK:
5862 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5863 break;
5864 case TCP_FIN_WAIT1:
5865 case TCP_FIN_WAIT2:
5866 /* RFC 793 says to queue data in these states,
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005867 * RFC 1122 says we MUST send a reset.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005868 * BSD 4.4 also does reset.
5869 */
5870 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5871 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5872 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
Pavel Emelyanovde0744a2008-07-16 20:31:16 -07005873 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005874 tcp_reset(sk);
5875 return 1;
5876 }
5877 }
5878 /* Fall through */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005879 case TCP_ESTABLISHED:
Linus Torvalds1da177e2005-04-16 15:20:36 -07005880 tcp_data_queue(sk, skb);
5881 queued = 1;
5882 break;
5883 }
5884
5885 /* tcp_data could move socket to TIME-WAIT */
5886 if (sk->sk_state != TCP_CLOSE) {
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005887 tcp_data_snd_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005888 tcp_ack_snd_check(sk);
5889 }
5890
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005891 if (!queued) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005892discard:
5893 __kfree_skb(skb);
5894 }
5895 return 0;
5896}
5897
5898EXPORT_SYMBOL(sysctl_tcp_ecn);
5899EXPORT_SYMBOL(sysctl_tcp_reordering);
Glenn Griffinc6aefaf2008-02-07 21:49:26 -08005900EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005901EXPORT_SYMBOL(tcp_parse_options);
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09005902#ifdef CONFIG_TCP_MD5SIG
5903EXPORT_SYMBOL(tcp_parse_md5sig_option);
5904#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07005905EXPORT_SYMBOL(tcp_rcv_established);
5906EXPORT_SYMBOL(tcp_rcv_state_process);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -08005907EXPORT_SYMBOL(tcp_initialize_rcv_mss);