blob: 7facdb0f69608be661e409c2ead641f5d0fb1bc8 [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 *
8 * Version: $Id: tcp_input.c,v 1.243 2002/02/01 22:01:04 davem Exp $
9 *
Jesper Juhl02c30a82005-05-05 16:16:16 -070010 * Authors: Ross Biro
Linus Torvalds1da177e2005-04-16 15:20:36 -070011 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12 * Mark Evans, <evansmp@uhura.aston.ac.uk>
13 * Corey Minyard <wf-rch!minyard@relay.EU.net>
14 * Florian La Roche, <flla@stud.uni-sb.de>
15 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
16 * Linus Torvalds, <torvalds@cs.helsinki.fi>
17 * Alan Cox, <gw4pts@gw4pts.ampr.org>
18 * Matthew Dillon, <dillon@apollo.west.oic.com>
19 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
20 * Jorge Cwik, <jorge@laser.satlink.net>
21 */
22
23/*
24 * Changes:
25 * Pedro Roque : Fast Retransmit/Recovery.
26 * Two receive queues.
27 * Retransmit queue handled by TCP.
28 * Better retransmit timer handling.
29 * New congestion avoidance.
30 * Header prediction.
31 * Variable renaming.
32 *
33 * Eric : Fast Retransmit.
34 * Randy Scott : MSS option defines.
35 * Eric Schenk : Fixes to slow start algorithm.
36 * Eric Schenk : Yet another double ACK bug.
37 * Eric Schenk : Delayed ACK bug fixes.
38 * Eric Schenk : Floyd style fast retrans war avoidance.
39 * David S. Miller : Don't allow zero congestion window.
40 * Eric Schenk : Fix retransmitter so that it sends
41 * next packet on ack of previous packet.
42 * Andi Kleen : Moved open_request checking here
43 * and process RSTs for open_requests.
44 * Andi Kleen : Better prune_queue, and other fixes.
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -080045 * Andrey Savochkin: Fix RTT measurements in the presence of
Linus Torvalds1da177e2005-04-16 15:20:36 -070046 * timestamps.
47 * Andrey Savochkin: Check sequence numbers correctly when
48 * removing SACKs due to in sequence incoming
49 * data segments.
50 * Andi Kleen: Make sure we never ack data there is not
51 * enough room for. Also make this condition
52 * a fatal error if it might still happen.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +090053 * Andi Kleen: Add tcp_measure_rcv_mss to make
Linus Torvalds1da177e2005-04-16 15:20:36 -070054 * connections with MSS<min(MTU,ann. MSS)
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +090055 * work without delayed acks.
Linus Torvalds1da177e2005-04-16 15:20:36 -070056 * Andi Kleen: Process packets with PSH set in the
57 * fast path.
58 * J Hadi Salim: ECN support
59 * Andrei Gurtov,
60 * Pasi Sarolahti,
61 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
62 * engine. Lots of bugs are found.
63 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
Linus Torvalds1da177e2005-04-16 15:20:36 -070064 */
65
Linus Torvalds1da177e2005-04-16 15:20:36 -070066#include <linux/mm.h>
67#include <linux/module.h>
68#include <linux/sysctl.h>
69#include <net/tcp.h>
70#include <net/inet_common.h>
71#include <linux/ipsec.h>
72#include <asm/unaligned.h>
Chris Leech1a2449a2006-05-23 18:05:53 -070073#include <net/netdma.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070074
Brian Haleyab32ea52006-09-22 14:15:41 -070075int sysctl_tcp_timestamps __read_mostly = 1;
76int sysctl_tcp_window_scaling __read_mostly = 1;
77int sysctl_tcp_sack __read_mostly = 1;
78int sysctl_tcp_fack __read_mostly = 1;
79int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
80int sysctl_tcp_ecn __read_mostly;
81int sysctl_tcp_dsack __read_mostly = 1;
82int sysctl_tcp_app_win __read_mostly = 31;
83int sysctl_tcp_adv_win_scale __read_mostly = 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -070084
Brian Haleyab32ea52006-09-22 14:15:41 -070085int sysctl_tcp_stdurg __read_mostly;
86int sysctl_tcp_rfc1337 __read_mostly;
87int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
Ilpo Järvinenc96fd3d2007-09-20 11:36:37 -070088int sysctl_tcp_frto __read_mostly = 2;
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -080089int sysctl_tcp_frto_response __read_mostly;
Brian Haleyab32ea52006-09-22 14:15:41 -070090int sysctl_tcp_nometrics_save __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -070091
Brian Haleyab32ea52006-09-22 14:15:41 -070092int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
93int sysctl_tcp_abc __read_mostly;
Linus Torvalds1da177e2005-04-16 15:20:36 -070094
Linus Torvalds1da177e2005-04-16 15:20:36 -070095#define FLAG_DATA 0x01 /* Incoming frame contained data. */
96#define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
97#define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
98#define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
99#define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
100#define FLAG_DATA_SACKED 0x20 /* New SACK. */
101#define FLAG_ECE 0x40 /* ECE in this ACK */
102#define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */
103#define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
Ilpo Järvinen4dc26652007-02-21 23:16:11 -0800104#define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */
Ilpo Järvinen2e605292007-08-02 19:46:58 -0700105#define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
Ryousei 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
Ilpo Järvinen4dc26652007-02-21 23:16:11 -0800116#define IsSackFrto() (sysctl_tcp_frto == 0x2)
117
Linus Torvalds1da177e2005-04-16 15:20:36 -0700118#define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700119#define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700120
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900121/* Adapt the MSS value used to make delayed ack decision to the
Linus Torvalds1da177e2005-04-16 15:20:36 -0700122 * real world.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900123 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800124static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700125{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700126 struct inet_connection_sock *icsk = inet_csk(sk);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900127 const unsigned int lss = icsk->icsk_ack.last_seg_size;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700128 unsigned int len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700129
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900130 icsk->icsk_ack.last_seg_size = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700131
132 /* skb->len may jitter because of SACKs, even if peer
133 * sends good full-sized frames.
134 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800135 len = skb_shinfo(skb)->gso_size ? : skb->len;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700136 if (len >= icsk->icsk_ack.rcv_mss) {
137 icsk->icsk_ack.rcv_mss = len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700138 } else {
139 /* Otherwise, we make more careful check taking into account,
140 * that SACKs block is variable.
141 *
142 * "len" is invariant segment length, including TCP header.
143 */
Arnaldo Carvalho de Melo9c702202007-04-25 18:04:18 -0700144 len += skb->data - skb_transport_header(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700145 if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) ||
146 /* If PSH is not set, packet should be
147 * full sized, provided peer TCP is not badly broken.
148 * This observation (if it is correct 8)) allows
149 * to handle super-low mtu links fairly.
150 */
151 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -0700152 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700153 /* Subtract also invariant (if peer is RFC compliant),
154 * tcp header plus fixed timestamp option length.
155 * Resulting "len" is MSS free of SACK jitter.
156 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700157 len -= tcp_sk(sk)->tcp_header_len;
158 icsk->icsk_ack.last_seg_size = len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700159 if (len == lss) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700160 icsk->icsk_ack.rcv_mss = len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700161 return;
162 }
163 }
Alexey Kuznetsov1ef96962006-09-19 12:52:50 -0700164 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
165 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700166 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700167 }
168}
169
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700170static void tcp_incr_quickack(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700171{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700172 struct inet_connection_sock *icsk = inet_csk(sk);
173 unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700174
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800175 if (quickacks == 0)
176 quickacks = 2;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700177 if (quickacks > icsk->icsk_ack.quick)
178 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700179}
180
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700181void tcp_enter_quickack_mode(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700182{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700183 struct inet_connection_sock *icsk = inet_csk(sk);
184 tcp_incr_quickack(sk);
185 icsk->icsk_ack.pingpong = 0;
186 icsk->icsk_ack.ato = TCP_ATO_MIN;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700187}
188
189/* Send ACKs quickly, if "quick" count is not exhausted
190 * and the session is not interactive.
191 */
192
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700193static inline int tcp_in_quickack_mode(const struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700194{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700195 const struct inet_connection_sock *icsk = inet_csk(sk);
196 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700197}
198
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700199static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
200{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800201 if (tp->ecn_flags & TCP_ECN_OK)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700202 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
203}
204
205static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, struct sk_buff *skb)
206{
207 if (tcp_hdr(skb)->cwr)
208 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
209}
210
211static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
212{
213 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
214}
215
216static inline void TCP_ECN_check_ce(struct tcp_sock *tp, struct sk_buff *skb)
217{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800218 if (tp->ecn_flags & TCP_ECN_OK) {
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700219 if (INET_ECN_is_ce(TCP_SKB_CB(skb)->flags))
220 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
221 /* Funny extension: if ECT is not set on a segment,
222 * it is surely retransmit. It is not in ECN RFC,
223 * but Linux follows this rule. */
224 else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb)->flags)))
225 tcp_enter_quickack_mode((struct sock *)tp);
226 }
227}
228
229static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, struct tcphdr *th)
230{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800231 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700232 tp->ecn_flags &= ~TCP_ECN_OK;
233}
234
235static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, struct tcphdr *th)
236{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800237 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700238 tp->ecn_flags &= ~TCP_ECN_OK;
239}
240
241static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock *tp, struct tcphdr *th)
242{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800243 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700244 return 1;
245 return 0;
246}
247
Linus Torvalds1da177e2005-04-16 15:20:36 -0700248/* Buffer size and advertised window tuning.
249 *
250 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
251 */
252
253static void tcp_fixup_sndbuf(struct sock *sk)
254{
255 int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 +
256 sizeof(struct sk_buff);
257
258 if (sk->sk_sndbuf < 3 * sndmem)
259 sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]);
260}
261
262/* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
263 *
264 * All tcp_full_space() is split to two parts: "network" buffer, allocated
265 * forward and advertised in receiver window (tp->rcv_wnd) and
266 * "application buffer", required to isolate scheduling/application
267 * latencies from network.
268 * window_clamp is maximal advertised window. It can be less than
269 * tcp_full_space(), in this case tcp_full_space() - window_clamp
270 * is reserved for "application" buffer. The less window_clamp is
271 * the smoother our behaviour from viewpoint of network, but the lower
272 * throughput and the higher sensitivity of the connection to losses. 8)
273 *
274 * rcv_ssthresh is more strict window_clamp used at "slow start"
275 * phase to predict further behaviour of this connection.
276 * It is used for two goals:
277 * - to enforce header prediction at sender, even when application
278 * requires some significant "application buffer". It is check #1.
279 * - to prevent pruning of receive queue because of misprediction
280 * of receiver window. Check #2.
281 *
282 * The scheme does not work when sender sends good segments opening
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800283 * window and then starts to feed us spaghetti. But it should work
Linus Torvalds1da177e2005-04-16 15:20:36 -0700284 * in common situations. Otherwise, we have to rely on queue collapsing.
285 */
286
287/* Slow part of check#2. */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700288static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700289{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700290 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700291 /* Optimize this! */
Eric Dumazetdfd4f0a2007-12-21 06:07:53 -0800292 int truesize = tcp_win_from_space(skb->truesize) >> 1;
293 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700294
295 while (tp->rcv_ssthresh <= window) {
296 if (truesize <= skb->len)
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700297 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700298
299 truesize >>= 1;
300 window >>= 1;
301 }
302 return 0;
303}
304
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800305static void tcp_grow_window(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700306{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700307 struct tcp_sock *tp = tcp_sk(sk);
308
Linus Torvalds1da177e2005-04-16 15:20:36 -0700309 /* Check #1 */
310 if (tp->rcv_ssthresh < tp->window_clamp &&
311 (int)tp->rcv_ssthresh < tcp_space(sk) &&
312 !tcp_memory_pressure) {
313 int incr;
314
315 /* Check #2. Increase window, if skb with such overhead
316 * will fit to rcvbuf in future.
317 */
318 if (tcp_win_from_space(skb->truesize) <= skb->len)
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800319 incr = 2 * tp->advmss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700320 else
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700321 incr = __tcp_grow_window(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700322
323 if (incr) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800324 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
325 tp->window_clamp);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700326 inet_csk(sk)->icsk_ack.quick |= 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700327 }
328 }
329}
330
331/* 3. Tuning rcvbuf, when connection enters established state. */
332
333static void tcp_fixup_rcvbuf(struct sock *sk)
334{
335 struct tcp_sock *tp = tcp_sk(sk);
336 int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
337
338 /* Try to select rcvbuf so that 4 mss-sized segments
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800339 * will fit to window and corresponding skbs will fit to our rcvbuf.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700340 * (was 3; 4 is minimum to allow fast retransmit to work.)
341 */
342 while (tcp_win_from_space(rcvmem) < tp->advmss)
343 rcvmem += 128;
344 if (sk->sk_rcvbuf < 4 * rcvmem)
345 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]);
346}
347
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800348/* 4. Try to fixup all. It is made immediately after connection enters
Linus Torvalds1da177e2005-04-16 15:20:36 -0700349 * established state.
350 */
351static void tcp_init_buffer_space(struct sock *sk)
352{
353 struct tcp_sock *tp = tcp_sk(sk);
354 int maxwin;
355
356 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
357 tcp_fixup_rcvbuf(sk);
358 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
359 tcp_fixup_sndbuf(sk);
360
361 tp->rcvq_space.space = tp->rcv_wnd;
362
363 maxwin = tcp_full_space(sk);
364
365 if (tp->window_clamp >= maxwin) {
366 tp->window_clamp = maxwin;
367
368 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
369 tp->window_clamp = max(maxwin -
370 (maxwin >> sysctl_tcp_app_win),
371 4 * tp->advmss);
372 }
373
374 /* Force reservation of one segment. */
375 if (sysctl_tcp_app_win &&
376 tp->window_clamp > 2 * tp->advmss &&
377 tp->window_clamp + tp->advmss > maxwin)
378 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
379
380 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
381 tp->snd_cwnd_stamp = tcp_time_stamp;
382}
383
Linus Torvalds1da177e2005-04-16 15:20:36 -0700384/* 5. Recalculate window clamp after socket hit its memory bounds. */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700385static void tcp_clamp_window(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700386{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700387 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300388 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700389
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300390 icsk->icsk_ack.quick = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700391
John Heffner326f36e2005-11-10 17:11:48 -0800392 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
393 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
394 !tcp_memory_pressure &&
395 atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
396 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
397 sysctl_tcp_rmem[2]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700398 }
John Heffner326f36e2005-11-10 17:11:48 -0800399 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800400 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700401}
402
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800403/* Initialize RCV_MSS value.
404 * RCV_MSS is an our guess about MSS used by the peer.
405 * We haven't any direct information about the MSS.
406 * It's better to underestimate the RCV_MSS rather than overestimate.
407 * Overestimations make us ACKing less frequently than needed.
408 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
409 */
410void tcp_initialize_rcv_mss(struct sock *sk)
411{
412 struct tcp_sock *tp = tcp_sk(sk);
413 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
414
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800415 hint = min(hint, tp->rcv_wnd / 2);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800416 hint = min(hint, TCP_MIN_RCVMSS);
417 hint = max(hint, TCP_MIN_MSS);
418
419 inet_csk(sk)->icsk_ack.rcv_mss = hint;
420}
421
Linus Torvalds1da177e2005-04-16 15:20:36 -0700422/* Receiver "autotuning" code.
423 *
424 * The algorithm for RTT estimation w/o timestamps is based on
425 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
426 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
427 *
428 * More detail on this code can be found at
429 * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
430 * though this reference is out of date. A new paper
431 * is pending.
432 */
433static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
434{
435 u32 new_sample = tp->rcv_rtt_est.rtt;
436 long m = sample;
437
438 if (m == 0)
439 m = 1;
440
441 if (new_sample != 0) {
442 /* If we sample in larger samples in the non-timestamp
443 * case, we could grossly overestimate the RTT especially
444 * with chatty applications or bulk transfer apps which
445 * are stalled on filesystem I/O.
446 *
447 * Also, since we are only going for a minimum in the
Stephen Hemminger31f34262005-11-15 15:17:10 -0800448 * non-timestamp case, we do not smooth things out
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800449 * else with timestamps disabled convergence takes too
Linus Torvalds1da177e2005-04-16 15:20:36 -0700450 * long.
451 */
452 if (!win_dep) {
453 m -= (new_sample >> 3);
454 new_sample += m;
455 } else if (m < new_sample)
456 new_sample = m << 3;
457 } else {
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800458 /* No previous measure. */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700459 new_sample = m << 3;
460 }
461
462 if (tp->rcv_rtt_est.rtt != new_sample)
463 tp->rcv_rtt_est.rtt = new_sample;
464}
465
466static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
467{
468 if (tp->rcv_rtt_est.time == 0)
469 goto new_measure;
470 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
471 return;
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800472 tcp_rcv_rtt_update(tp, jiffies - tp->rcv_rtt_est.time, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700473
474new_measure:
475 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
476 tp->rcv_rtt_est.time = tcp_time_stamp;
477}
478
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800479static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
480 const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700481{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700482 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700483 if (tp->rx_opt.rcv_tsecr &&
484 (TCP_SKB_CB(skb)->end_seq -
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700485 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700486 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
487}
488
489/*
490 * This function should be called every time data is copied to user space.
491 * It calculates the appropriate TCP receive buffer space.
492 */
493void tcp_rcv_space_adjust(struct sock *sk)
494{
495 struct tcp_sock *tp = tcp_sk(sk);
496 int time;
497 int space;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900498
Linus Torvalds1da177e2005-04-16 15:20:36 -0700499 if (tp->rcvq_space.time == 0)
500 goto new_measure;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900501
Linus Torvalds1da177e2005-04-16 15:20:36 -0700502 time = tcp_time_stamp - tp->rcvq_space.time;
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800503 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700504 return;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900505
Linus Torvalds1da177e2005-04-16 15:20:36 -0700506 space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
507
508 space = max(tp->rcvq_space.space, space);
509
510 if (tp->rcvq_space.space != space) {
511 int rcvmem;
512
513 tp->rcvq_space.space = space;
514
John Heffner6fcf9412006-02-09 17:06:57 -0800515 if (sysctl_tcp_moderate_rcvbuf &&
516 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700517 int new_clamp = space;
518
519 /* Receive space grows, normalize in order to
520 * take into account packet headers and sk_buff
521 * structure overhead.
522 */
523 space /= tp->advmss;
524 if (!space)
525 space = 1;
526 rcvmem = (tp->advmss + MAX_TCP_HEADER +
527 16 + sizeof(struct sk_buff));
528 while (tcp_win_from_space(rcvmem) < tp->advmss)
529 rcvmem += 128;
530 space *= rcvmem;
531 space = min(space, sysctl_tcp_rmem[2]);
532 if (space > sk->sk_rcvbuf) {
533 sk->sk_rcvbuf = space;
534
535 /* Make the window clamp follow along. */
536 tp->window_clamp = new_clamp;
537 }
538 }
539 }
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900540
Linus Torvalds1da177e2005-04-16 15:20:36 -0700541new_measure:
542 tp->rcvq_space.seq = tp->copied_seq;
543 tp->rcvq_space.time = tcp_time_stamp;
544}
545
546/* There is something which you must keep in mind when you analyze the
547 * behavior of the tp->ato delayed ack timeout interval. When a
548 * connection starts up, we want to ack as quickly as possible. The
549 * problem is that "good" TCP's do slow start at the beginning of data
550 * transmission. The means that until we send the first few ACK's the
551 * sender will sit on his end and only queue most of his data, because
552 * he can only send snd_cwnd unacked packets at any given time. For
553 * each ACK we send, he increments snd_cwnd and transmits more of his
554 * queue. -DaveM
555 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700556static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700557{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700558 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700559 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700560 u32 now;
561
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700562 inet_csk_schedule_ack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700563
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700564 tcp_measure_rcv_mss(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700565
566 tcp_rcv_rtt_measure(tp);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900567
Linus Torvalds1da177e2005-04-16 15:20:36 -0700568 now = tcp_time_stamp;
569
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700570 if (!icsk->icsk_ack.ato) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700571 /* The _first_ data packet received, initialize
572 * delayed ACK engine.
573 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700574 tcp_incr_quickack(sk);
575 icsk->icsk_ack.ato = TCP_ATO_MIN;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700576 } else {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700577 int m = now - icsk->icsk_ack.lrcvtime;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700578
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800579 if (m <= TCP_ATO_MIN / 2) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700580 /* The fastest case is the first. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700581 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
582 } else if (m < icsk->icsk_ack.ato) {
583 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
584 if (icsk->icsk_ack.ato > icsk->icsk_rto)
585 icsk->icsk_ack.ato = icsk->icsk_rto;
586 } else if (m > icsk->icsk_rto) {
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800587 /* Too long gap. Apparently sender failed to
Linus Torvalds1da177e2005-04-16 15:20:36 -0700588 * restart window, so that we send ACKs quickly.
589 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700590 tcp_incr_quickack(sk);
Hideo Aoki3ab224b2007-12-31 00:11:19 -0800591 sk_mem_reclaim(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700592 }
593 }
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700594 icsk->icsk_ack.lrcvtime = now;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700595
596 TCP_ECN_check_ce(tp, skb);
597
598 if (skb->len >= 128)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700599 tcp_grow_window(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700600}
601
David S. Miller05bb1fa2007-08-30 22:10:28 -0700602static u32 tcp_rto_min(struct sock *sk)
603{
604 struct dst_entry *dst = __sk_dst_get(sk);
605 u32 rto_min = TCP_RTO_MIN;
606
David S. Miller5c127c52007-08-31 14:39:44 -0700607 if (dst && dst_metric_locked(dst, RTAX_RTO_MIN))
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800608 rto_min = dst->metrics[RTAX_RTO_MIN - 1];
David S. Miller05bb1fa2007-08-30 22:10:28 -0700609 return rto_min;
610}
611
Linus Torvalds1da177e2005-04-16 15:20:36 -0700612/* Called to compute a smoothed rtt estimate. The data fed to this
613 * routine either comes from timestamps, or from segments that were
614 * known _not_ to have been retransmitted [see Karn/Partridge
615 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
616 * piece by Van Jacobson.
617 * NOTE: the next three routines used to be one big routine.
618 * To save cycles in the RFC 1323 implementation it was better to break
619 * it up into three procedures. -- erics
620 */
Stephen Hemminger2d2abba2005-11-10 16:56:12 -0800621static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700622{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300623 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700624 long m = mrtt; /* RTT */
625
Linus Torvalds1da177e2005-04-16 15:20:36 -0700626 /* The following amusing code comes from Jacobson's
627 * article in SIGCOMM '88. Note that rtt and mdev
628 * are scaled versions of rtt and mean deviation.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900629 * This is designed to be as fast as possible
Linus Torvalds1da177e2005-04-16 15:20:36 -0700630 * m stands for "measurement".
631 *
632 * On a 1990 paper the rto value is changed to:
633 * RTO = rtt + 4 * mdev
634 *
635 * Funny. This algorithm seems to be very broken.
636 * These formulae increase RTO, when it should be decreased, increase
Stephen Hemminger31f34262005-11-15 15:17:10 -0800637 * too slowly, when it should be increased quickly, decrease too quickly
Linus Torvalds1da177e2005-04-16 15:20:36 -0700638 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
639 * does not matter how to _calculate_ it. Seems, it was trap
640 * that VJ failed to avoid. 8)
641 */
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800642 if (m == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700643 m = 1;
644 if (tp->srtt != 0) {
645 m -= (tp->srtt >> 3); /* m is now error in rtt est */
646 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
647 if (m < 0) {
648 m = -m; /* m is now abs(error) */
649 m -= (tp->mdev >> 2); /* similar update on mdev */
650 /* This is similar to one of Eifel findings.
651 * Eifel blocks mdev updates when rtt decreases.
652 * This solution is a bit different: we use finer gain
653 * for mdev in this case (alpha*beta).
654 * Like Eifel it also prevents growth of rto,
655 * but also it limits too fast rto decreases,
656 * happening in pure Eifel.
657 */
658 if (m > 0)
659 m >>= 3;
660 } else {
661 m -= (tp->mdev >> 2); /* similar update on mdev */
662 }
663 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
664 if (tp->mdev > tp->mdev_max) {
665 tp->mdev_max = tp->mdev;
666 if (tp->mdev_max > tp->rttvar)
667 tp->rttvar = tp->mdev_max;
668 }
669 if (after(tp->snd_una, tp->rtt_seq)) {
670 if (tp->mdev_max < tp->rttvar)
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800671 tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700672 tp->rtt_seq = tp->snd_nxt;
David S. Miller05bb1fa2007-08-30 22:10:28 -0700673 tp->mdev_max = tcp_rto_min(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700674 }
675 } else {
676 /* no previous measure. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800677 tp->srtt = m << 3; /* take the measured time to be rtt */
678 tp->mdev = m << 1; /* make sure rto = 3*rtt */
David S. Miller05bb1fa2007-08-30 22:10:28 -0700679 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700680 tp->rtt_seq = tp->snd_nxt;
681 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700682}
683
684/* Calculate rto without backoff. This is the second half of Van Jacobson's
685 * routine referred to above.
686 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700687static inline void tcp_set_rto(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700688{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700689 const struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700690 /* Old crap is replaced with new one. 8)
691 *
692 * More seriously:
693 * 1. If rtt variance happened to be less 50msec, it is hallucination.
694 * It cannot be less due to utterly erratic ACK generation made
695 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
696 * to do with delayed acks, because at cwnd>2 true delack timeout
697 * is invisible. Actually, Linux-2.4 also generates erratic
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800698 * ACKs in some circumstances.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700699 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700700 inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700701
702 /* 2. Fixups made earlier cannot be right.
703 * If we do not estimate RTO correctly without them,
704 * all the algo is pure shit and should be replaced
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800705 * with correct one. It is exactly, which we pretend to do.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700706 */
707}
708
709/* NOTE: clamping at TCP_RTO_MIN is not required, current algo
710 * guarantees that rto is higher.
711 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700712static inline void tcp_bound_rto(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700713{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700714 if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX)
715 inet_csk(sk)->icsk_rto = TCP_RTO_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700716}
717
718/* Save metrics learned by this TCP session.
719 This function is called only, when TCP finishes successfully
720 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
721 */
722void tcp_update_metrics(struct sock *sk)
723{
724 struct tcp_sock *tp = tcp_sk(sk);
725 struct dst_entry *dst = __sk_dst_get(sk);
726
727 if (sysctl_tcp_nometrics_save)
728 return;
729
730 dst_confirm(dst);
731
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800732 if (dst && (dst->flags & DST_HOST)) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300733 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700734 int m;
735
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300736 if (icsk->icsk_backoff || !tp->srtt) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700737 /* This session failed to estimate rtt. Why?
738 * Probably, no packets returned in time.
739 * Reset our results.
740 */
741 if (!(dst_metric_locked(dst, RTAX_RTT)))
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800742 dst->metrics[RTAX_RTT - 1] = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700743 return;
744 }
745
746 m = dst_metric(dst, RTAX_RTT) - tp->srtt;
747
748 /* If newly calculated rtt larger than stored one,
749 * store new one. Otherwise, use EWMA. Remember,
750 * rtt overestimation is always better than underestimation.
751 */
752 if (!(dst_metric_locked(dst, RTAX_RTT))) {
753 if (m <= 0)
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800754 dst->metrics[RTAX_RTT - 1] = tp->srtt;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700755 else
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800756 dst->metrics[RTAX_RTT - 1] -= (m >> 3);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700757 }
758
759 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
760 if (m < 0)
761 m = -m;
762
763 /* Scale deviation to rttvar fixed point */
764 m >>= 1;
765 if (m < tp->mdev)
766 m = tp->mdev;
767
768 if (m >= dst_metric(dst, RTAX_RTTVAR))
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800769 dst->metrics[RTAX_RTTVAR - 1] = m;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700770 else
771 dst->metrics[RTAX_RTTVAR-1] -=
772 (dst->metrics[RTAX_RTTVAR-1] - m)>>2;
773 }
774
775 if (tp->snd_ssthresh >= 0xFFFF) {
776 /* Slow start still did not finish. */
777 if (dst_metric(dst, RTAX_SSTHRESH) &&
778 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
779 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
780 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
781 if (!dst_metric_locked(dst, RTAX_CWND) &&
782 tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800783 dst->metrics[RTAX_CWND - 1] = tp->snd_cwnd;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700784 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300785 icsk->icsk_ca_state == TCP_CA_Open) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700786 /* Cong. avoidance phase, cwnd is reliable. */
787 if (!dst_metric_locked(dst, RTAX_SSTHRESH))
788 dst->metrics[RTAX_SSTHRESH-1] =
789 max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
790 if (!dst_metric_locked(dst, RTAX_CWND))
791 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1;
792 } else {
793 /* Else slow start did not finish, cwnd is non-sense,
794 ssthresh may be also invalid.
795 */
796 if (!dst_metric_locked(dst, RTAX_CWND))
797 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1;
798 if (dst->metrics[RTAX_SSTHRESH-1] &&
799 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
800 tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1])
801 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
802 }
803
804 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
805 if (dst->metrics[RTAX_REORDERING-1] < tp->reordering &&
806 tp->reordering != sysctl_tcp_reordering)
807 dst->metrics[RTAX_REORDERING-1] = tp->reordering;
808 }
809 }
810}
811
David S. Miller26722872007-08-24 22:21:50 -0700812/* Numbers are taken from RFC3390.
813 *
814 * John Heffner states:
815 *
816 * The RFC specifies a window of no more than 4380 bytes
817 * unless 2*MSS > 4380. Reading the pseudocode in the RFC
818 * is a bit misleading because they use a clamp at 4380 bytes
819 * rather than use a multiplier in the relevant range.
820 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700821__u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
822{
823 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
824
825 if (!cwnd) {
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700826 if (tp->mss_cache > 1460)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700827 cwnd = 2;
828 else
David S. Millerc1b4a7e2005-07-05 15:24:38 -0700829 cwnd = (tp->mss_cache > 1095) ? 3 : 4;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700830 }
831 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
832}
833
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800834/* Set slow start threshold and cwnd not falling to slow start */
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -0800835void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800836{
837 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -0800838 const struct inet_connection_sock *icsk = inet_csk(sk);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800839
840 tp->prior_ssthresh = 0;
841 tp->bytes_acked = 0;
Ilpo Järvinene01f9d72007-03-02 13:27:25 -0800842 if (icsk->icsk_ca_state < TCP_CA_CWR) {
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800843 tp->undo_marker = 0;
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -0800844 if (set_ssthresh)
845 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800846 tp->snd_cwnd = min(tp->snd_cwnd,
847 tcp_packets_in_flight(tp) + 1U);
848 tp->snd_cwnd_cnt = 0;
849 tp->high_seq = tp->snd_nxt;
850 tp->snd_cwnd_stamp = tcp_time_stamp;
851 TCP_ECN_queue_cwr(tp);
852
853 tcp_set_ca_state(sk, TCP_CA_CWR);
854 }
855}
856
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300857/*
858 * Packet counting of FACK is based on in-order assumptions, therefore TCP
859 * disables it when reordering is detected
860 */
861static void tcp_disable_fack(struct tcp_sock *tp)
862{
Ilpo Järvinen85cc3912007-11-15 19:39:31 -0800863 /* RFC3517 uses different metric in lost marker => reset on change */
864 if (tcp_is_fack(tp))
865 tp->lost_skb_hint = NULL;
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300866 tp->rx_opt.sack_ok &= ~2;
867}
868
Ryousei Takano564262c12007-10-25 23:03:52 -0700869/* Take a notice that peer is sending D-SACKs */
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300870static void tcp_dsack_seen(struct tcp_sock *tp)
871{
872 tp->rx_opt.sack_ok |= 4;
873}
874
Linus Torvalds1da177e2005-04-16 15:20:36 -0700875/* Initialize metrics on socket. */
876
877static void tcp_init_metrics(struct sock *sk)
878{
879 struct tcp_sock *tp = tcp_sk(sk);
880 struct dst_entry *dst = __sk_dst_get(sk);
881
882 if (dst == NULL)
883 goto reset;
884
885 dst_confirm(dst);
886
887 if (dst_metric_locked(dst, RTAX_CWND))
888 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
889 if (dst_metric(dst, RTAX_SSTHRESH)) {
890 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
891 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
892 tp->snd_ssthresh = tp->snd_cwnd_clamp;
893 }
894 if (dst_metric(dst, RTAX_REORDERING) &&
895 tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300896 tcp_disable_fack(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700897 tp->reordering = dst_metric(dst, RTAX_REORDERING);
898 }
899
900 if (dst_metric(dst, RTAX_RTT) == 0)
901 goto reset;
902
903 if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
904 goto reset;
905
906 /* Initial rtt is determined from SYN,SYN-ACK.
907 * The segment is small and rtt may appear much
908 * less than real one. Use per-dst memory
909 * to make it more realistic.
910 *
911 * A bit of theory. RTT is time passed after "normal" sized packet
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800912 * is sent until it is ACKed. In normal circumstances sending small
Linus Torvalds1da177e2005-04-16 15:20:36 -0700913 * packets force peer to delay ACKs and calculation is correct too.
914 * The algorithm is adaptive and, provided we follow specs, it
915 * NEVER underestimate RTT. BUT! If peer tries to make some clever
916 * tricks sort of "quick acks" for time long enough to decrease RTT
917 * to low value, and then abruptly stops to do it and starts to delay
918 * ACKs, wait for troubles.
919 */
920 if (dst_metric(dst, RTAX_RTT) > tp->srtt) {
921 tp->srtt = dst_metric(dst, RTAX_RTT);
922 tp->rtt_seq = tp->snd_nxt;
923 }
924 if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) {
925 tp->mdev = dst_metric(dst, RTAX_RTTVAR);
Satoru SATOH488faa22007-12-16 14:00:19 -0800926 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700927 }
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700928 tcp_set_rto(sk);
929 tcp_bound_rto(sk);
930 if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700931 goto reset;
932 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
933 tp->snd_cwnd_stamp = tcp_time_stamp;
934 return;
935
936reset:
937 /* Play conservative. If timestamps are not
938 * supported, TCP will fail to recalculate correct
939 * rtt, if initial rto is too small. FORGET ALL AND RESET!
940 */
941 if (!tp->rx_opt.saw_tstamp && tp->srtt) {
942 tp->srtt = 0;
943 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700944 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700945 }
946}
947
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300948static void tcp_update_reordering(struct sock *sk, const int metric,
949 const int ts)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700950{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300951 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700952 if (metric > tp->reordering) {
953 tp->reordering = min(TCP_MAX_REORDERING, metric);
954
955 /* This exciting event is worth to be remembered. 8) */
956 if (ts)
957 NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER);
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300958 else if (tcp_is_reno(tp))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700959 NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER);
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300960 else if (tcp_is_fack(tp))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700961 NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER);
962 else
963 NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER);
964#if FASTRETRANS_DEBUG > 1
965 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300966 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700967 tp->reordering,
968 tp->fackets_out,
969 tp->sacked_out,
970 tp->undo_marker ? tp->undo_retrans : 0);
971#endif
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300972 tcp_disable_fack(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700973 }
974}
975
976/* This procedure tags the retransmission queue when SACKs arrive.
977 *
978 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
979 * Packets in queue with these bits set are counted in variables
980 * sacked_out, retrans_out and lost_out, correspondingly.
981 *
982 * Valid combinations are:
983 * Tag InFlight Description
984 * 0 1 - orig segment is in flight.
985 * S 0 - nothing flies, orig reached receiver.
986 * L 0 - nothing flies, orig lost by net.
987 * R 2 - both orig and retransmit are in flight.
988 * L|R 1 - orig is lost, retransmit is in flight.
989 * S|R 1 - orig reached receiver, retrans is still in flight.
990 * (L|S|R is logically valid, it could occur when L|R is sacked,
991 * but it is equivalent to plain S and code short-curcuits it to S.
992 * L|S is logically invalid, it would mean -1 packet in flight 8))
993 *
994 * These 6 states form finite state machine, controlled by the following events:
995 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
996 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
997 * 3. Loss detection event of one of three flavors:
998 * A. Scoreboard estimator decided the packet is lost.
999 * A'. Reno "three dupacks" marks head of queue lost.
1000 * A''. Its FACK modfication, head until snd.fack is lost.
1001 * B. SACK arrives sacking data transmitted after never retransmitted
1002 * hole was sent out.
1003 * C. SACK arrives sacking SND.NXT at the moment, when the
1004 * segment was retransmitted.
1005 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1006 *
1007 * It is pleasant to note, that state diagram turns out to be commutative,
1008 * so that we are allowed not to be bothered by order of our actions,
1009 * when multiple events arrive simultaneously. (see the function below).
1010 *
1011 * Reordering detection.
1012 * --------------------
1013 * Reordering metric is maximal distance, which a packet can be displaced
1014 * in packet stream. With SACKs we can estimate it:
1015 *
1016 * 1. SACK fills old hole and the corresponding segment was not
1017 * ever retransmitted -> reordering. Alas, we cannot use it
1018 * when segment was retransmitted.
1019 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1020 * for retransmitted and already SACKed segment -> reordering..
1021 * Both of these heuristics are not used in Loss state, when we cannot
1022 * account for retransmits accurately.
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001023 *
1024 * SACK block validation.
1025 * ----------------------
1026 *
1027 * SACK block range validation checks that the received SACK block fits to
1028 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1029 * Note that SND.UNA is not included to the range though being valid because
Ilpo Järvinen0e835332007-10-01 15:28:17 -07001030 * it means that the receiver is rather inconsistent with itself reporting
1031 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1032 * perfectly valid, however, in light of RFC2018 which explicitly states
1033 * that "SACK block MUST reflect the newest segment. Even if the newest
1034 * segment is going to be discarded ...", not that it looks very clever
1035 * in case of head skb. Due to potentional receiver driven attacks, we
1036 * choose to avoid immediate execution of a walk in write queue due to
1037 * reneging and defer head skb's loss recovery to standard loss recovery
1038 * procedure that will eventually trigger (nothing forbids us doing this).
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001039 *
1040 * Implements also blockage to start_seq wrap-around. Problem lies in the
1041 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1042 * there's no guarantee that it will be before snd_nxt (n). The problem
1043 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1044 * wrap (s_w):
1045 *
1046 * <- outs wnd -> <- wrapzone ->
1047 * u e n u_w e_w s n_w
1048 * | | | | | | |
1049 * |<------------+------+----- TCP seqno space --------------+---------->|
1050 * ...-- <2^31 ->| |<--------...
1051 * ...---- >2^31 ------>| |<--------...
1052 *
1053 * Current code wouldn't be vulnerable but it's better still to discard such
1054 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1055 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1056 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1057 * equal to the ideal case (infinite seqno space without wrap caused issues).
1058 *
1059 * With D-SACK the lower bound is extended to cover sequence space below
1060 * SND.UNA down to undo_marker, which is the last point of interest. Yet
Ryousei Takano564262c12007-10-25 23:03:52 -07001061 * again, D-SACK block must not to go across snd_una (for the same reason as
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001062 * for the normal SACK blocks, explained above). But there all simplicity
1063 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1064 * fully below undo_marker they do not affect behavior in anyway and can
1065 * therefore be safely ignored. In rare cases (which are more or less
1066 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1067 * fragmentation and packet reordering past skb's retransmission. To consider
1068 * them correctly, the acceptable range must be extended even more though
1069 * the exact amount is rather hard to quantify. However, tp->max_window can
1070 * be used as an exaggerated estimate.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001071 */
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001072static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack,
1073 u32 start_seq, u32 end_seq)
1074{
1075 /* Too far in future, or reversed (interpretation is ambiguous) */
1076 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1077 return 0;
1078
1079 /* Nasty start_seq wrap-around check (see comments above) */
1080 if (!before(start_seq, tp->snd_nxt))
1081 return 0;
1082
Ryousei Takano564262c12007-10-25 23:03:52 -07001083 /* In outstanding window? ...This is valid exit for D-SACKs too.
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001084 * start_seq == snd_una is non-sensical (see comments above)
1085 */
1086 if (after(start_seq, tp->snd_una))
1087 return 1;
1088
1089 if (!is_dsack || !tp->undo_marker)
1090 return 0;
1091
1092 /* ...Then it's D-SACK, and must reside below snd_una completely */
1093 if (!after(end_seq, tp->snd_una))
1094 return 0;
1095
1096 if (!before(start_seq, tp->undo_marker))
1097 return 1;
1098
1099 /* Too old */
1100 if (!after(end_seq, tp->undo_marker))
1101 return 0;
1102
1103 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1104 * start_seq < undo_marker and end_seq >= undo_marker.
1105 */
1106 return !before(start_seq, end_seq - tp->max_window);
1107}
1108
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001109/* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1110 * Event "C". Later note: FACK people cheated me again 8), we have to account
1111 * for reordering! Ugly, but should help.
Ilpo Järvinenf785a8e2007-10-11 17:35:41 -07001112 *
1113 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1114 * less than what is now known to be received by the other end (derived from
Ilpo Järvinen9f58f3b2007-11-15 19:42:54 -08001115 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1116 * retransmitted skbs to avoid some costly processing per ACKs.
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001117 */
Ilpo Järvinen407ef1d2007-12-02 00:47:57 +02001118static void tcp_mark_lost_retrans(struct sock *sk)
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001119{
Ilpo Järvinen9f58f3b2007-11-15 19:42:54 -08001120 const struct inet_connection_sock *icsk = inet_csk(sk);
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001121 struct tcp_sock *tp = tcp_sk(sk);
1122 struct sk_buff *skb;
Ilpo Järvinenf785a8e2007-10-11 17:35:41 -07001123 int cnt = 0;
Ilpo Järvinendf2e0142007-10-18 05:07:57 -07001124 u32 new_low_seq = tp->snd_nxt;
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001125 u32 received_upto = tcp_highest_sack_seq(tp);
Ilpo Järvinen9f58f3b2007-11-15 19:42:54 -08001126
1127 if (!tcp_is_fack(tp) || !tp->retrans_out ||
1128 !after(received_upto, tp->lost_retrans_low) ||
1129 icsk->icsk_ca_state != TCP_CA_Recovery)
Ilpo Järvinen407ef1d2007-12-02 00:47:57 +02001130 return;
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001131
1132 tcp_for_write_queue(skb, sk) {
1133 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1134
1135 if (skb == tcp_send_head(sk))
1136 break;
Ilpo Järvinenf785a8e2007-10-11 17:35:41 -07001137 if (cnt == tp->retrans_out)
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001138 break;
1139 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1140 continue;
1141
Ilpo Järvinenf785a8e2007-10-11 17:35:41 -07001142 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1143 continue;
1144
1145 if (after(received_upto, ack_seq) &&
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001146 (tcp_is_fack(tp) ||
Ilpo Järvinenf785a8e2007-10-11 17:35:41 -07001147 !before(received_upto,
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001148 ack_seq + tp->reordering * tp->mss_cache))) {
1149 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1150 tp->retrans_out -= tcp_skb_pcount(skb);
1151
1152 /* clear lost hint */
1153 tp->retransmit_skb_hint = NULL;
1154
1155 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1156 tp->lost_out += tcp_skb_pcount(skb);
1157 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001158 }
Ilpo Järvinenbce392f2007-12-02 00:47:56 +02001159 NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT);
Ilpo Järvinenf785a8e2007-10-11 17:35:41 -07001160 } else {
Ilpo Järvinendf2e0142007-10-18 05:07:57 -07001161 if (before(ack_seq, new_low_seq))
Ilpo Järvinenb08d6cb2007-10-11 17:36:13 -07001162 new_low_seq = ack_seq;
Ilpo Järvinenf785a8e2007-10-11 17:35:41 -07001163 cnt += tcp_skb_pcount(skb);
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001164 }
1165 }
Ilpo Järvinenb08d6cb2007-10-11 17:36:13 -07001166
1167 if (tp->retrans_out)
1168 tp->lost_retrans_low = new_low_seq;
Ilpo Järvinen1c1e87e2007-10-10 02:45:32 -07001169}
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001170
David S. Millerd06e0212007-06-18 22:43:06 -07001171static int tcp_check_dsack(struct tcp_sock *tp, struct sk_buff *ack_skb,
1172 struct tcp_sack_block_wire *sp, int num_sacks,
1173 u32 prior_snd_una)
1174{
1175 u32 start_seq_0 = ntohl(get_unaligned(&sp[0].start_seq));
1176 u32 end_seq_0 = ntohl(get_unaligned(&sp[0].end_seq));
1177 int dup_sack = 0;
1178
1179 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1180 dup_sack = 1;
Ilpo Järvinene60402d2007-08-09 15:14:46 +03001181 tcp_dsack_seen(tp);
David S. Millerd06e0212007-06-18 22:43:06 -07001182 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV);
1183 } else if (num_sacks > 1) {
1184 u32 end_seq_1 = ntohl(get_unaligned(&sp[1].end_seq));
1185 u32 start_seq_1 = ntohl(get_unaligned(&sp[1].start_seq));
1186
1187 if (!after(end_seq_0, end_seq_1) &&
1188 !before(start_seq_0, start_seq_1)) {
1189 dup_sack = 1;
Ilpo Järvinene60402d2007-08-09 15:14:46 +03001190 tcp_dsack_seen(tp);
David S. Millerd06e0212007-06-18 22:43:06 -07001191 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV);
1192 }
1193 }
1194
1195 /* D-SACK for already forgotten data... Do dumb counting. */
1196 if (dup_sack &&
1197 !after(end_seq_0, prior_snd_una) &&
1198 after(end_seq_0, tp->undo_marker))
1199 tp->undo_retrans--;
1200
1201 return dup_sack;
1202}
1203
Ilpo Järvinend1935942007-10-11 17:34:25 -07001204/* Check if skb is fully within the SACK block. In presence of GSO skbs,
1205 * the incoming SACK may not exactly match but we can find smaller MSS
1206 * aligned portion of it that matches. Therefore we might need to fragment
1207 * which may fail and creates some hassle (caller must handle error case
1208 * returns).
1209 */
Adrian Bunk0f79efd2007-10-26 03:57:36 -07001210static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1211 u32 start_seq, u32 end_seq)
Ilpo Järvinend1935942007-10-11 17:34:25 -07001212{
1213 int in_sack, err;
1214 unsigned int pkt_len;
1215
1216 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1217 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1218
1219 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1220 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1221
1222 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1223
1224 if (!in_sack)
1225 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1226 else
1227 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1228 err = tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size);
1229 if (err < 0)
1230 return err;
1231 }
1232
1233 return in_sack;
1234}
1235
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001236static int tcp_sacktag_one(struct sk_buff *skb, struct sock *sk,
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001237 int *reord, int dup_sack, int fack_count)
1238{
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001239 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001240 u8 sacked = TCP_SKB_CB(skb)->sacked;
1241 int flag = 0;
1242
1243 /* Account D-SACK for retransmitted packet. */
1244 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1245 if (after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1246 tp->undo_retrans--;
Ilpo Järvinenede9f3b2007-12-02 00:47:58 +02001247 if (sacked & TCPCB_SACKED_ACKED)
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001248 *reord = min(fack_count, *reord);
1249 }
1250
1251 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1252 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1253 return flag;
1254
1255 if (!(sacked & TCPCB_SACKED_ACKED)) {
1256 if (sacked & TCPCB_SACKED_RETRANS) {
1257 /* If the segment is not tagged as lost,
1258 * we do not clear RETRANS, believing
1259 * that retransmission is still in flight.
1260 */
1261 if (sacked & TCPCB_LOST) {
1262 TCP_SKB_CB(skb)->sacked &=
1263 ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1264 tp->lost_out -= tcp_skb_pcount(skb);
1265 tp->retrans_out -= tcp_skb_pcount(skb);
1266
1267 /* clear lost hint */
1268 tp->retransmit_skb_hint = NULL;
1269 }
1270 } else {
1271 if (!(sacked & TCPCB_RETRANS)) {
1272 /* New sack for not retransmitted frame,
1273 * which was in hole. It is reordering.
1274 */
1275 if (before(TCP_SKB_CB(skb)->seq,
1276 tcp_highest_sack_seq(tp)))
1277 *reord = min(fack_count, *reord);
1278
1279 /* SACK enhanced F-RTO (RFC4138; Appendix B) */
1280 if (!after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark))
1281 flag |= FLAG_ONLY_ORIG_SACKED;
1282 }
1283
1284 if (sacked & TCPCB_LOST) {
1285 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1286 tp->lost_out -= tcp_skb_pcount(skb);
1287
1288 /* clear lost hint */
1289 tp->retransmit_skb_hint = NULL;
1290 }
1291 }
1292
1293 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED;
1294 flag |= FLAG_DATA_SACKED;
1295 tp->sacked_out += tcp_skb_pcount(skb);
1296
1297 fack_count += tcp_skb_pcount(skb);
1298
1299 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1300 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1301 before(TCP_SKB_CB(skb)->seq,
1302 TCP_SKB_CB(tp->lost_skb_hint)->seq))
1303 tp->lost_cnt_hint += tcp_skb_pcount(skb);
1304
1305 if (fack_count > tp->fackets_out)
1306 tp->fackets_out = fack_count;
1307
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001308 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
1309 tcp_advance_highest_sack(sk, skb);
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001310 }
1311
1312 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1313 * frames and clear it. undo_retrans is decreased above, L|R frames
1314 * are accounted above as well.
1315 */
1316 if (dup_sack && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)) {
1317 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1318 tp->retrans_out -= tcp_skb_pcount(skb);
1319 tp->retransmit_skb_hint = NULL;
1320 }
1321
1322 return flag;
1323}
1324
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001325static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1326 struct tcp_sack_block *next_dup,
1327 u32 start_seq, u32 end_seq,
1328 int dup_sack_in, int *fack_count,
1329 int *reord, int *flag)
1330{
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001331 tcp_for_write_queue_from(skb, sk) {
1332 int in_sack = 0;
1333 int dup_sack = dup_sack_in;
1334
1335 if (skb == tcp_send_head(sk))
1336 break;
1337
1338 /* queue is in-order => we can short-circuit the walk early */
1339 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1340 break;
1341
1342 if ((next_dup != NULL) &&
1343 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1344 in_sack = tcp_match_skb_to_sack(sk, skb,
1345 next_dup->start_seq,
1346 next_dup->end_seq);
1347 if (in_sack > 0)
1348 dup_sack = 1;
1349 }
1350
1351 if (in_sack <= 0)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001352 in_sack = tcp_match_skb_to_sack(sk, skb, start_seq,
1353 end_seq);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001354 if (unlikely(in_sack < 0))
1355 break;
1356
1357 if (in_sack)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001358 *flag |= tcp_sacktag_one(skb, sk, reord, dup_sack,
1359 *fack_count);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001360
1361 *fack_count += tcp_skb_pcount(skb);
1362 }
1363 return skb;
1364}
1365
1366/* Avoid all extra work that is being done by sacktag while walking in
1367 * a normal way
1368 */
1369static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
Ilpo Järvinend152a7d2008-03-03 12:10:16 -08001370 u32 skip_to_seq, int *fack_count)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001371{
1372 tcp_for_write_queue_from(skb, sk) {
1373 if (skb == tcp_send_head(sk))
1374 break;
1375
Ilpo Järvinenea4f76a2007-11-30 00:59:07 +11001376 if (!before(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001377 break;
Ilpo Järvinend152a7d2008-03-03 12:10:16 -08001378
1379 *fack_count += tcp_skb_pcount(skb);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001380 }
1381 return skb;
1382}
1383
1384static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1385 struct sock *sk,
1386 struct tcp_sack_block *next_dup,
1387 u32 skip_to_seq,
1388 int *fack_count, int *reord,
1389 int *flag)
1390{
1391 if (next_dup == NULL)
1392 return skb;
1393
1394 if (before(next_dup->start_seq, skip_to_seq)) {
Ilpo Järvinend152a7d2008-03-03 12:10:16 -08001395 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq, fack_count);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001396 tcp_sacktag_walk(skb, sk, NULL,
1397 next_dup->start_seq, next_dup->end_seq,
1398 1, fack_count, reord, flag);
1399 }
1400
1401 return skb;
1402}
1403
1404static int tcp_sack_cache_ok(struct tcp_sock *tp, struct tcp_sack_block *cache)
1405{
1406 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1407}
1408
Linus Torvalds1da177e2005-04-16 15:20:36 -07001409static int
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001410tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb,
1411 u32 prior_snd_una)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001412{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001413 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001414 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo9c702202007-04-25 18:04:18 -07001415 unsigned char *ptr = (skb_transport_header(ack_skb) +
1416 TCP_SKB_CB(ack_skb)->sacked);
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001417 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1418 struct tcp_sack_block sp[4];
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001419 struct tcp_sack_block *cache;
1420 struct sk_buff *skb;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001421 int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE) >> 3;
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001422 int used_sacks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001423 int reord = tp->packets_out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001424 int flag = 0;
Ilpo Järvinen7769f402007-06-15 15:14:04 -07001425 int found_dup_sack = 0;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001426 int fack_count;
1427 int i, j;
Baruch Evenfda03fb2007-02-04 23:35:57 -08001428 int first_sack_index;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001429
Ilpo Järvinend738cd82007-03-24 21:03:23 -07001430 if (!tp->sacked_out) {
Ilpo Järvinende83c052007-10-07 23:37:55 -07001431 if (WARN_ON(tp->fackets_out))
1432 tp->fackets_out = 0;
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001433 tcp_highest_sack_reset(sk);
Ilpo Järvinend738cd82007-03-24 21:03:23 -07001434 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001435
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001436 found_dup_sack = tcp_check_dsack(tp, ack_skb, sp_wire,
David S. Millerd06e0212007-06-18 22:43:06 -07001437 num_sacks, prior_snd_una);
1438 if (found_dup_sack)
Ilpo Järvinen49ff4bb2007-08-02 19:47:59 -07001439 flag |= FLAG_DSACKING_ACK;
Baruch Even6f746512007-02-04 23:36:42 -08001440
1441 /* Eliminate too old ACKs, but take into
1442 * account more or less fresh ones, they can
1443 * contain valid SACK info.
1444 */
1445 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1446 return 0;
1447
Ilpo Järvinen96a2d412007-11-14 15:47:18 -08001448 if (!tp->packets_out)
1449 goto out;
1450
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001451 used_sacks = 0;
1452 first_sack_index = 0;
1453 for (i = 0; i < num_sacks; i++) {
1454 int dup_sack = !i && found_dup_sack;
1455
1456 sp[used_sacks].start_seq = ntohl(get_unaligned(&sp_wire[i].start_seq));
1457 sp[used_sacks].end_seq = ntohl(get_unaligned(&sp_wire[i].end_seq));
1458
1459 if (!tcp_is_sackblock_valid(tp, dup_sack,
1460 sp[used_sacks].start_seq,
1461 sp[used_sacks].end_seq)) {
1462 if (dup_sack) {
1463 if (!tp->undo_marker)
1464 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDNOUNDO);
1465 else
1466 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDOLD);
1467 } else {
1468 /* Don't count olds caused by ACK reordering */
1469 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1470 !after(sp[used_sacks].end_seq, tp->snd_una))
1471 continue;
1472 NET_INC_STATS_BH(LINUX_MIB_TCPSACKDISCARD);
1473 }
1474 if (i == 0)
1475 first_sack_index = -1;
1476 continue;
1477 }
1478
1479 /* Ignore very old stuff early */
1480 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1481 continue;
1482
1483 used_sacks++;
1484 }
1485
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001486 /* order SACK blocks to allow in order walk of the retrans queue */
1487 for (i = used_sacks - 1; i > 0; i--) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001488 for (j = 0; j < i; j++) {
1489 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001490 struct tcp_sack_block tmp;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001491
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001492 tmp = sp[j];
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001493 sp[j] = sp[j + 1];
1494 sp[j + 1] = tmp;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001495
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001496 /* Track where the first SACK block goes to */
1497 if (j == first_sack_index)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001498 first_sack_index = j + 1;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001499 }
1500 }
1501 }
1502
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001503 skb = tcp_write_queue_head(sk);
1504 fack_count = 0;
1505 i = 0;
1506
1507 if (!tp->sacked_out) {
1508 /* It's already past, so skip checking against it */
1509 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1510 } else {
1511 cache = tp->recv_sack_cache;
1512 /* Skip empty blocks in at head of the cache */
1513 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1514 !cache->end_seq)
1515 cache++;
Baruch Evenfda03fb2007-02-04 23:35:57 -08001516 }
1517
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001518 while (i < used_sacks) {
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001519 u32 start_seq = sp[i].start_seq;
1520 u32 end_seq = sp[i].end_seq;
Ilpo Järvinen7769f402007-06-15 15:14:04 -07001521 int dup_sack = (found_dup_sack && (i == first_sack_index));
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001522 struct tcp_sack_block *next_dup = NULL;
Ilpo Järvinene56d6cd2007-11-01 00:09:37 -07001523
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001524 if (found_dup_sack && ((i + 1) == first_sack_index))
1525 next_dup = &sp[i + 1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001526
1527 /* Event "B" in the comment above. */
1528 if (after(end_seq, tp->high_seq))
1529 flag |= FLAG_DATA_LOST;
1530
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001531 /* Skip too early cached blocks */
1532 while (tcp_sack_cache_ok(tp, cache) &&
1533 !before(start_seq, cache->end_seq))
1534 cache++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001535
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001536 /* Can skip some work by looking recv_sack_cache? */
1537 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1538 after(end_seq, cache->start_seq)) {
David S. Millerfe067e82007-03-07 12:12:44 -08001539
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001540 /* Head todo? */
1541 if (before(start_seq, cache->start_seq)) {
Ilpo Järvinend152a7d2008-03-03 12:10:16 -08001542 skb = tcp_sacktag_skip(skb, sk, start_seq,
1543 &fack_count);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001544 skb = tcp_sacktag_walk(skb, sk, next_dup,
1545 start_seq,
1546 cache->start_seq,
1547 dup_sack, &fack_count,
1548 &reord, &flag);
Baruch Evenfda03fb2007-02-04 23:35:57 -08001549 }
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001550
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001551 /* Rest of the block already fully processed? */
Ilpo Järvinen20de20b2007-11-16 16:17:05 -08001552 if (!after(end_seq, cache->end_seq))
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001553 goto advance_sp;
Ilpo Järvinen20de20b2007-11-16 16:17:05 -08001554
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001555 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1556 cache->end_seq,
1557 &fack_count, &reord,
1558 &flag);
Ilpo Järvinene56d6cd2007-11-01 00:09:37 -07001559
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001560 /* ...tail remains todo... */
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001561 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
Ilpo Järvinen20de20b2007-11-16 16:17:05 -08001562 /* ...but better entrypoint exists! */
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001563 skb = tcp_highest_sack(sk);
1564 if (skb == NULL)
1565 break;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001566 fack_count = tp->fackets_out;
1567 cache++;
1568 goto walk;
1569 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001570
Ilpo Järvinend152a7d2008-03-03 12:10:16 -08001571 skb = tcp_sacktag_skip(skb, sk, cache->end_seq,
1572 &fack_count);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001573 /* Check overlap against next cached too (past this one already) */
1574 cache++;
1575 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001576 }
Ilpo Järvinenfbd52eb2007-11-10 21:24:19 -08001577
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001578 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1579 skb = tcp_highest_sack(sk);
1580 if (skb == NULL)
1581 break;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001582 fack_count = tp->fackets_out;
1583 }
Ilpo Järvinend152a7d2008-03-03 12:10:16 -08001584 skb = tcp_sacktag_skip(skb, sk, start_seq, &fack_count);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001585
1586walk:
1587 skb = tcp_sacktag_walk(skb, sk, next_dup, start_seq, end_seq,
1588 dup_sack, &fack_count, &reord, &flag);
1589
1590advance_sp:
Ilpo Järvinenfbd52eb2007-11-10 21:24:19 -08001591 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
1592 * due to in-order walk
1593 */
1594 if (after(end_seq, tp->frto_highmark))
1595 flag &= ~FLAG_ONLY_ORIG_SACKED;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001596
1597 i++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001598 }
1599
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001600 /* Clear the head of the cache sack blocks so we can skip it next time */
1601 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1602 tp->recv_sack_cache[i].start_seq = 0;
1603 tp->recv_sack_cache[i].end_seq = 0;
1604 }
1605 for (j = 0; j < used_sacks; j++)
1606 tp->recv_sack_cache[i++] = sp[j];
1607
Ilpo Järvinen407ef1d2007-12-02 00:47:57 +02001608 tcp_mark_lost_retrans(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001609
Ilpo Järvinen86426c22007-08-09 14:45:17 +03001610 tcp_verify_left_out(tp);
1611
Ilpo Järvinenf5771112007-11-15 19:35:11 -08001612 if ((reord < tp->fackets_out) &&
1613 ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) &&
Ilpo Järvinenc5e7af02007-02-23 16:22:06 -08001614 (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
Ilpo Järvinen8dd71c5d2007-11-10 21:20:59 -08001615 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001616
Ilpo Järvinen96a2d412007-11-14 15:47:18 -08001617out:
1618
Linus Torvalds1da177e2005-04-16 15:20:36 -07001619#if FASTRETRANS_DEBUG > 0
1620 BUG_TRAP((int)tp->sacked_out >= 0);
1621 BUG_TRAP((int)tp->lost_out >= 0);
1622 BUG_TRAP((int)tp->retrans_out >= 0);
1623 BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0);
1624#endif
1625 return flag;
1626}
1627
Ilpo Järvinen95eacd22007-10-01 15:27:42 -07001628/* If we receive more dupacks than we expected counting segments
1629 * in assumption of absent reordering, interpret this as reordering.
1630 * The only another reason could be bug in receiver TCP.
Ilpo Järvinen30935cf2007-02-21 23:01:36 -08001631 */
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001632static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1633{
1634 struct tcp_sock *tp = tcp_sk(sk);
1635 u32 holes;
1636
1637 holes = max(tp->lost_out, 1U);
1638 holes = min(holes, tp->packets_out);
1639
1640 if ((tp->sacked_out + holes) > tp->packets_out) {
1641 tp->sacked_out = tp->packets_out - holes;
1642 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1643 }
1644}
1645
1646/* Emulate SACKs for SACKless connection: account for a new dupack. */
1647
1648static void tcp_add_reno_sack(struct sock *sk)
1649{
1650 struct tcp_sock *tp = tcp_sk(sk);
1651 tp->sacked_out++;
1652 tcp_check_reno_reordering(sk, 0);
Ilpo Järvinen005903b2007-08-09 14:44:16 +03001653 tcp_verify_left_out(tp);
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001654}
1655
1656/* Account for ACK, ACKing some data in Reno Recovery phase. */
1657
1658static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1659{
1660 struct tcp_sock *tp = tcp_sk(sk);
1661
1662 if (acked > 0) {
1663 /* One ACK acked hole. The rest eat duplicate ACKs. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001664 if (acked - 1 >= tp->sacked_out)
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001665 tp->sacked_out = 0;
1666 else
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001667 tp->sacked_out -= acked - 1;
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001668 }
1669 tcp_check_reno_reordering(sk, acked);
Ilpo Järvinen005903b2007-08-09 14:44:16 +03001670 tcp_verify_left_out(tp);
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001671}
1672
1673static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1674{
1675 tp->sacked_out = 0;
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001676}
1677
Ilpo Järvinen95eacd22007-10-01 15:27:42 -07001678/* F-RTO can only be used if TCP has never retransmitted anything other than
1679 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1680 */
Ilpo Järvinen46d0de42007-02-21 23:10:39 -08001681int tcp_use_frto(struct sock *sk)
Ilpo Järvinenbdaae172007-02-21 22:59:58 -08001682{
1683 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen46d0de42007-02-21 23:10:39 -08001684 struct sk_buff *skb;
Ilpo Järvinenbdaae172007-02-21 22:59:58 -08001685
Ilpo Järvinen575ee712007-04-30 00:39:55 -07001686 if (!sysctl_tcp_frto)
Ilpo Järvinen46d0de42007-02-21 23:10:39 -08001687 return 0;
1688
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08001689 if (IsSackFrto())
1690 return 1;
1691
Ilpo Järvinen46d0de42007-02-21 23:10:39 -08001692 /* Avoid expensive walking of rexmit queue if possible */
1693 if (tp->retrans_out > 1)
1694 return 0;
1695
David S. Millerfe067e82007-03-07 12:12:44 -08001696 skb = tcp_write_queue_head(sk);
1697 skb = tcp_write_queue_next(sk, skb); /* Skips head */
1698 tcp_for_write_queue_from(skb, sk) {
1699 if (skb == tcp_send_head(sk))
1700 break;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001701 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
Ilpo Järvinen46d0de42007-02-21 23:10:39 -08001702 return 0;
1703 /* Short-circuit when first non-SACKed skb has been checked */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001704 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
Ilpo Järvinen46d0de42007-02-21 23:10:39 -08001705 break;
1706 }
1707 return 1;
Ilpo Järvinenbdaae172007-02-21 22:59:58 -08001708}
1709
Ilpo Järvinen30935cf2007-02-21 23:01:36 -08001710/* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1711 * recovery a bit and use heuristics in tcp_process_frto() to detect if
Ilpo Järvinend1a54c62007-02-21 23:11:57 -08001712 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1713 * keep retrans_out counting accurate (with SACK F-RTO, other than head
1714 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1715 * bits are handled if the Loss state is really to be entered (in
1716 * tcp_enter_frto_loss).
Ilpo Järvinen7487c482007-02-21 23:02:30 -08001717 *
1718 * Do like tcp_enter_loss() would; when RTO expires the second time it
1719 * does:
1720 * "Reduce ssthresh if it has not yet been made inside this window."
Linus Torvalds1da177e2005-04-16 15:20:36 -07001721 */
1722void tcp_enter_frto(struct sock *sk)
1723{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001724 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001725 struct tcp_sock *tp = tcp_sk(sk);
1726 struct sk_buff *skb;
1727
Ilpo Järvinen7487c482007-02-21 23:02:30 -08001728 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001729 tp->snd_una == tp->high_seq ||
Ilpo Järvinen7487c482007-02-21 23:02:30 -08001730 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
1731 !icsk->icsk_retransmits)) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001732 tp->prior_ssthresh = tcp_current_ssthresh(sk);
Ilpo Järvinen66e93e42007-02-21 23:13:47 -08001733 /* Our state is too optimistic in ssthresh() call because cwnd
Ryousei Takano564262c12007-10-25 23:03:52 -07001734 * is not reduced until tcp_enter_frto_loss() when previous F-RTO
Ilpo Järvinen66e93e42007-02-21 23:13:47 -08001735 * recovery has not yet completed. Pattern would be this: RTO,
1736 * Cumulative ACK, RTO (2xRTO for the same segment does not end
1737 * up here twice).
1738 * RFC4138 should be more specific on what to do, even though
1739 * RTO is quite unlikely to occur after the first Cumulative ACK
1740 * due to back-off and complexity of triggering events ...
1741 */
1742 if (tp->frto_counter) {
1743 u32 stored_cwnd;
1744 stored_cwnd = tp->snd_cwnd;
1745 tp->snd_cwnd = 2;
1746 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1747 tp->snd_cwnd = stored_cwnd;
1748 } else {
1749 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1750 }
1751 /* ... in theory, cong.control module could do "any tricks" in
1752 * ssthresh(), which means that ca_state, lost bits and lost_out
1753 * counter would have to be faked before the call occurs. We
1754 * consider that too expensive, unlikely and hacky, so modules
1755 * using these in ssthresh() must deal these incompatibility
1756 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1757 */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001758 tcp_ca_event(sk, CA_EVENT_FRTO);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001759 }
1760
Linus Torvalds1da177e2005-04-16 15:20:36 -07001761 tp->undo_marker = tp->snd_una;
1762 tp->undo_retrans = 0;
1763
David S. Millerfe067e82007-03-07 12:12:44 -08001764 skb = tcp_write_queue_head(sk);
Ilpo Järvinen009a2e32007-09-20 11:34:38 -07001765 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1766 tp->undo_marker = 0;
Ilpo Järvinend1a54c62007-02-21 23:11:57 -08001767 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
Ilpo Järvinen522e7542007-02-21 22:54:52 -08001768 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
Ilpo Järvinend1a54c62007-02-21 23:11:57 -08001769 tp->retrans_out -= tcp_skb_pcount(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001770 }
Ilpo Järvinen005903b2007-08-09 14:44:16 +03001771 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001772
Ilpo Järvinen746aa322007-11-13 21:01:23 -08001773 /* Too bad if TCP was application limited */
1774 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
1775
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08001776 /* Earlier loss recovery underway (see RFC4138; Appendix B).
1777 * The last condition is necessary at least in tp->frto_counter case.
1778 */
1779 if (IsSackFrto() && (tp->frto_counter ||
1780 ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
1781 after(tp->high_seq, tp->snd_una)) {
1782 tp->frto_highmark = tp->high_seq;
1783 } else {
1784 tp->frto_highmark = tp->snd_nxt;
1785 }
Ilpo Järvinen7b0eb222007-02-21 23:03:35 -08001786 tcp_set_ca_state(sk, TCP_CA_Disorder);
1787 tp->high_seq = tp->snd_nxt;
Ilpo Järvinen7487c482007-02-21 23:02:30 -08001788 tp->frto_counter = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001789}
1790
1791/* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1792 * which indicates that we should follow the traditional RTO recovery,
1793 * i.e. mark everything lost and do go-back-N retransmission.
1794 */
Ilpo Järvinend1a54c62007-02-21 23:11:57 -08001795static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001796{
1797 struct tcp_sock *tp = tcp_sk(sk);
1798 struct sk_buff *skb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001799
Linus Torvalds1da177e2005-04-16 15:20:36 -07001800 tp->lost_out = 0;
Ilpo Järvinend1a54c62007-02-21 23:11:57 -08001801 tp->retrans_out = 0;
Ilpo Järvinene60402d2007-08-09 15:14:46 +03001802 if (tcp_is_reno(tp))
Ilpo Järvinen9bff40f2007-05-27 01:53:49 -07001803 tcp_reset_reno_sack(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001804
David S. Millerfe067e82007-03-07 12:12:44 -08001805 tcp_for_write_queue(skb, sk) {
1806 if (skb == tcp_send_head(sk))
1807 break;
Ilpo Järvinen23aeeec2007-11-13 21:03:13 -08001808
1809 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
Ilpo Järvinend1a54c62007-02-21 23:11:57 -08001810 /*
1811 * Count the retransmission made on RTO correctly (only when
1812 * waiting for the first ACK and did not get it)...
1813 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001814 if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) {
Ilpo Järvinen0a9f2a42007-07-15 00:19:29 -07001815 /* For some reason this R-bit might get cleared? */
1816 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1817 tp->retrans_out += tcp_skb_pcount(skb);
Ilpo Järvinend1a54c62007-02-21 23:11:57 -08001818 /* ...enter this if branch just for the first segment */
1819 flag |= FLAG_DATA_ACKED;
1820 } else {
Ilpo Järvinen009a2e32007-09-20 11:34:38 -07001821 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
1822 tp->undo_marker = 0;
Ilpo Järvinen23aeeec2007-11-13 21:03:13 -08001823 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
Ilpo Järvinend1a54c62007-02-21 23:11:57 -08001824 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001825
Ilpo Järvinen9bff40f2007-05-27 01:53:49 -07001826 /* Don't lost mark skbs that were fwd transmitted after RTO */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001827 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) &&
Ilpo Järvinen9bff40f2007-05-27 01:53:49 -07001828 !after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark)) {
1829 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1830 tp->lost_out += tcp_skb_pcount(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001831 }
1832 }
Ilpo Järvinen005903b2007-08-09 14:44:16 +03001833 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001834
Ilpo Järvinen95c49222007-02-21 23:05:18 -08001835 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001836 tp->snd_cwnd_cnt = 0;
1837 tp->snd_cwnd_stamp = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001838 tp->frto_counter = 0;
Ilpo Järvinen16e90682007-10-11 17:32:31 -07001839 tp->bytes_acked = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001840
1841 tp->reordering = min_t(unsigned int, tp->reordering,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001842 sysctl_tcp_reordering);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001843 tcp_set_ca_state(sk, TCP_CA_Loss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001844 tp->high_seq = tp->frto_highmark;
1845 TCP_ECN_queue_cwr(tp);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001846
Ilpo Järvinenb7689202007-09-20 11:37:19 -07001847 tcp_clear_retrans_hints_partial(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001848}
1849
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07001850static void tcp_clear_retrans_partial(struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001851{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001852 tp->retrans_out = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001853 tp->lost_out = 0;
1854
1855 tp->undo_marker = 0;
1856 tp->undo_retrans = 0;
1857}
1858
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07001859void tcp_clear_retrans(struct tcp_sock *tp)
1860{
1861 tcp_clear_retrans_partial(tp);
1862
1863 tp->fackets_out = 0;
1864 tp->sacked_out = 0;
1865}
1866
Linus Torvalds1da177e2005-04-16 15:20:36 -07001867/* Enter Loss state. If "how" is not zero, forget all SACK information
1868 * and reset tags completely, otherwise preserve SACKs. If receiver
1869 * dropped its ofo queue, we will know this due to reneging detection.
1870 */
1871void tcp_enter_loss(struct sock *sk, int how)
1872{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001873 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001874 struct tcp_sock *tp = tcp_sk(sk);
1875 struct sk_buff *skb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001876
1877 /* Reduce ssthresh if it has not yet been made inside this window. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001878 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
1879 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1880 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1881 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1882 tcp_ca_event(sk, CA_EVENT_LOSS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001883 }
1884 tp->snd_cwnd = 1;
1885 tp->snd_cwnd_cnt = 0;
1886 tp->snd_cwnd_stamp = tcp_time_stamp;
1887
Stephen Hemminger9772efb2005-11-10 17:09:53 -08001888 tp->bytes_acked = 0;
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07001889 tcp_clear_retrans_partial(tp);
1890
1891 if (tcp_is_reno(tp))
1892 tcp_reset_reno_sack(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001893
Ilpo Järvinenb7689202007-09-20 11:37:19 -07001894 if (!how) {
1895 /* Push undo marker, if it was plain RTO and nothing
1896 * was retransmitted. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001897 tp->undo_marker = tp->snd_una;
Ilpo Järvinenb7689202007-09-20 11:37:19 -07001898 tcp_clear_retrans_hints_partial(tp);
1899 } else {
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07001900 tp->sacked_out = 0;
1901 tp->fackets_out = 0;
Ilpo Järvinenb7689202007-09-20 11:37:19 -07001902 tcp_clear_all_retrans_hints(tp);
1903 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001904
David S. Millerfe067e82007-03-07 12:12:44 -08001905 tcp_for_write_queue(skb, sk) {
1906 if (skb == tcp_send_head(sk))
1907 break;
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07001908
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001909 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001910 tp->undo_marker = 0;
1911 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1912 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1913 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1914 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1915 tp->lost_out += tcp_skb_pcount(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001916 }
1917 }
Ilpo Järvinen005903b2007-08-09 14:44:16 +03001918 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001919
1920 tp->reordering = min_t(unsigned int, tp->reordering,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001921 sysctl_tcp_reordering);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001922 tcp_set_ca_state(sk, TCP_CA_Loss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001923 tp->high_seq = tp->snd_nxt;
1924 TCP_ECN_queue_cwr(tp);
Ryousei Takano564262c12007-10-25 23:03:52 -07001925 /* Abort F-RTO algorithm if one is in progress */
Ilpo Järvinen580e5722007-05-19 13:56:57 -07001926 tp->frto_counter = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001927}
1928
Ilpo Järvinencadbd032007-12-31 04:49:21 -08001929/* If ACK arrived pointing to a remembered SACK, it means that our
1930 * remembered SACKs do not reflect real state of receiver i.e.
1931 * receiver _host_ is heavily congested (or buggy).
1932 *
1933 * Do processing similar to RTO timeout.
1934 */
1935static int tcp_check_sack_reneging(struct sock *sk, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001936{
Ilpo Järvinencadbd032007-12-31 04:49:21 -08001937 if (flag & FLAG_SACK_RENEGING) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001938 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001939 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING);
1940
1941 tcp_enter_loss(sk, 1);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001942 icsk->icsk_retransmits++;
David S. Millerfe067e82007-03-07 12:12:44 -08001943 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001944 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001945 icsk->icsk_rto, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001946 return 1;
1947 }
1948 return 0;
1949}
1950
1951static inline int tcp_fackets_out(struct tcp_sock *tp)
1952{
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001953 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001954}
1955
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08001956/* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
1957 * counter when SACK is enabled (without SACK, sacked_out is used for
1958 * that purpose).
1959 *
1960 * Instead, with FACK TCP uses fackets_out that includes both SACKed
1961 * segments up to the highest received SACK block so far and holes in
1962 * between them.
1963 *
1964 * With reordering, holes may still be in flight, so RFC3517 recovery
1965 * uses pure sacked_out (total number of SACKed segments) even though
1966 * it violates the RFC that uses duplicate ACKs, often these are equal
1967 * but when e.g. out-of-window ACKs or packet duplication occurs,
1968 * they differ. Since neither occurs due to loss, TCP should really
1969 * ignore them.
1970 */
1971static inline int tcp_dupack_heurestics(struct tcp_sock *tp)
1972{
1973 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
1974}
1975
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001976static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001977{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001978 return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001979}
1980
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001981static inline int tcp_head_timedout(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001982{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001983 struct tcp_sock *tp = tcp_sk(sk);
1984
Linus Torvalds1da177e2005-04-16 15:20:36 -07001985 return tp->packets_out &&
David S. Millerfe067e82007-03-07 12:12:44 -08001986 tcp_skb_timedout(sk, tcp_write_queue_head(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001987}
1988
1989/* Linux NewReno/SACK/FACK/ECN state machine.
1990 * --------------------------------------
1991 *
1992 * "Open" Normal state, no dubious events, fast path.
1993 * "Disorder" In all the respects it is "Open",
1994 * but requires a bit more attention. It is entered when
1995 * we see some SACKs or dupacks. It is split of "Open"
1996 * mainly to move some processing from fast path to slow one.
1997 * "CWR" CWND was reduced due to some Congestion Notification event.
1998 * It can be ECN, ICMP source quench, local device congestion.
1999 * "Recovery" CWND was reduced, we are fast-retransmitting.
2000 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2001 *
2002 * tcp_fastretrans_alert() is entered:
2003 * - each incoming ACK, if state is not "Open"
2004 * - when arrived ACK is unusual, namely:
2005 * * SACK
2006 * * Duplicate ACK.
2007 * * ECN ECE.
2008 *
2009 * Counting packets in flight is pretty simple.
2010 *
2011 * in_flight = packets_out - left_out + retrans_out
2012 *
2013 * packets_out is SND.NXT-SND.UNA counted in packets.
2014 *
2015 * retrans_out is number of retransmitted segments.
2016 *
2017 * left_out is number of segments left network, but not ACKed yet.
2018 *
2019 * left_out = sacked_out + lost_out
2020 *
2021 * sacked_out: Packets, which arrived to receiver out of order
2022 * and hence not ACKed. With SACKs this number is simply
2023 * amount of SACKed data. Even without SACKs
2024 * it is easy to give pretty reliable estimate of this number,
2025 * counting duplicate ACKs.
2026 *
2027 * lost_out: Packets lost by network. TCP has no explicit
2028 * "loss notification" feedback from network (for now).
2029 * It means that this number can be only _guessed_.
2030 * Actually, it is the heuristics to predict lossage that
2031 * distinguishes different algorithms.
2032 *
2033 * F.e. after RTO, when all the queue is considered as lost,
2034 * lost_out = packets_out and in_flight = retrans_out.
2035 *
2036 * Essentially, we have now two algorithms counting
2037 * lost packets.
2038 *
2039 * FACK: It is the simplest heuristics. As soon as we decided
2040 * that something is lost, we decide that _all_ not SACKed
2041 * packets until the most forward SACK are lost. I.e.
2042 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2043 * It is absolutely correct estimate, if network does not reorder
2044 * packets. And it loses any connection to reality when reordering
2045 * takes place. We use FACK by default until reordering
2046 * is suspected on the path to this destination.
2047 *
2048 * NewReno: when Recovery is entered, we assume that one segment
2049 * is lost (classic Reno). While we are in Recovery and
2050 * a partial ACK arrives, we assume that one more packet
2051 * is lost (NewReno). This heuristics are the same in NewReno
2052 * and SACK.
2053 *
2054 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2055 * deflation etc. CWND is real congestion window, never inflated, changes
2056 * only according to classic VJ rules.
2057 *
2058 * Really tricky (and requiring careful tuning) part of algorithm
2059 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2060 * The first determines the moment _when_ we should reduce CWND and,
2061 * hence, slow down forward transmission. In fact, it determines the moment
2062 * when we decide that hole is caused by loss, rather than by a reorder.
2063 *
2064 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2065 * holes, caused by lost packets.
2066 *
2067 * And the most logically complicated part of algorithm is undo
2068 * heuristics. We detect false retransmits due to both too early
2069 * fast retransmit (reordering) and underestimated RTO, analyzing
2070 * timestamps and D-SACKs. When we detect that some segments were
2071 * retransmitted by mistake and CWND reduction was wrong, we undo
2072 * window reduction and abort recovery phase. This logic is hidden
2073 * inside several functions named tcp_try_undo_<something>.
2074 */
2075
2076/* This function decides, when we should leave Disordered state
2077 * and enter Recovery phase, reducing congestion window.
2078 *
2079 * Main question: may we further continue forward transmission
2080 * with the same cwnd?
2081 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002082static int tcp_time_to_recover(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002083{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002084 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002085 __u32 packets_out;
2086
Ryousei Takano564262c12007-10-25 23:03:52 -07002087 /* Do not perform any recovery during F-RTO algorithm */
Ilpo Järvinen52c63f12007-02-21 23:06:52 -08002088 if (tp->frto_counter)
2089 return 0;
2090
Linus Torvalds1da177e2005-04-16 15:20:36 -07002091 /* Trick#1: The loss is proven. */
2092 if (tp->lost_out)
2093 return 1;
2094
2095 /* Not-A-Trick#2 : Classic rule... */
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002096 if (tcp_dupack_heurestics(tp) > tp->reordering)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002097 return 1;
2098
2099 /* Trick#3 : when we use RFC2988 timer restart, fast
2100 * retransmit can be triggered by timeout of queue head.
2101 */
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002102 if (tcp_is_fack(tp) && tcp_head_timedout(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002103 return 1;
2104
2105 /* Trick#4: It is still not OK... But will it be useful to delay
2106 * recovery more?
2107 */
2108 packets_out = tp->packets_out;
2109 if (packets_out <= tp->reordering &&
2110 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002111 !tcp_may_send_now(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002112 /* We have nothing to send. This connection is limited
2113 * either by receiver window or by application.
2114 */
2115 return 1;
2116 }
2117
2118 return 0;
2119}
2120
Ilpo Järvinend8f4f222007-04-20 22:56:38 -07002121/* RFC: This is from the original, I doubt that this is necessary at all:
2122 * clear xmit_retrans hint if seq of this skb is beyond hint. How could we
2123 * retransmitted past LOST markings in the first place? I'm not fully sure
2124 * about undo and end of connection cases, which can cause R without L?
2125 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002126static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
Ilpo Järvinend8f4f222007-04-20 22:56:38 -07002127{
2128 if ((tp->retransmit_skb_hint != NULL) &&
2129 before(TCP_SKB_CB(skb)->seq,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002130 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
Ilpo Järvinen19b2b482007-03-28 12:06:37 -07002131 tp->retransmit_skb_hint = NULL;
Ilpo Järvinend8f4f222007-04-20 22:56:38 -07002132}
2133
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002134/* Mark head of queue up as lost. With RFC3517 SACK, the packets is
2135 * is against sacked "cnt", otherwise it's against facked "cnt"
2136 */
2137static void tcp_mark_head_lost(struct sock *sk, int packets, int fast_rexmit)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002138{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002139 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002140 struct sk_buff *skb;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002141 int cnt;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002142
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002143 BUG_TRAP(packets <= tp->packets_out);
2144 if (tp->lost_skb_hint) {
2145 skb = tp->lost_skb_hint;
2146 cnt = tp->lost_cnt_hint;
2147 } else {
David S. Millerfe067e82007-03-07 12:12:44 -08002148 skb = tcp_write_queue_head(sk);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002149 cnt = 0;
2150 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002151
David S. Millerfe067e82007-03-07 12:12:44 -08002152 tcp_for_write_queue_from(skb, sk) {
2153 if (skb == tcp_send_head(sk))
2154 break;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002155 /* TODO: do this better */
2156 /* this is not the most efficient way to do this... */
2157 tp->lost_skb_hint = skb;
2158 tp->lost_cnt_hint = cnt;
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002159
Ilpo Järvinenad1984e2008-01-30 20:06:02 -08002160 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002161 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2162 cnt += tcp_skb_pcount(skb);
2163
2164 if (((!fast_rexmit || (tp->lost_out > 0)) && (cnt > packets)) ||
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002165 after(TCP_SKB_CB(skb)->end_seq, tp->high_seq))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002166 break;
Ilpo Järvinen3eec0042007-10-11 17:33:11 -07002167 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_SACKED_ACKED|TCPCB_LOST))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002168 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2169 tp->lost_out += tcp_skb_pcount(skb);
Ilpo Järvinend8f4f222007-04-20 22:56:38 -07002170 tcp_verify_retransmit_hint(tp, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002171 }
2172 }
Ilpo Järvinen005903b2007-08-09 14:44:16 +03002173 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002174}
2175
2176/* Account newly detected lost packet(s) */
2177
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002178static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002179{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002180 struct tcp_sock *tp = tcp_sk(sk);
2181
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002182 if (tcp_is_reno(tp)) {
2183 tcp_mark_head_lost(sk, 1, fast_rexmit);
2184 } else if (tcp_is_fack(tp)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002185 int lost = tp->fackets_out - tp->reordering;
2186 if (lost <= 0)
2187 lost = 1;
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002188 tcp_mark_head_lost(sk, lost, fast_rexmit);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002189 } else {
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002190 int sacked_upto = tp->sacked_out - tp->reordering;
2191 if (sacked_upto < 0)
2192 sacked_upto = 0;
2193 tcp_mark_head_lost(sk, sacked_upto, fast_rexmit);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002194 }
2195
2196 /* New heuristics: it is possible only after we switched
2197 * to restart timer each time when something is ACKed.
2198 * Hence, we can detect timed out packets during fast
2199 * retransmit without falling to slow start.
2200 */
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002201 if (tcp_is_fack(tp) && tcp_head_timedout(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002202 struct sk_buff *skb;
2203
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002204 skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint
David S. Millerfe067e82007-03-07 12:12:44 -08002205 : tcp_write_queue_head(sk);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002206
David S. Millerfe067e82007-03-07 12:12:44 -08002207 tcp_for_write_queue_from(skb, sk) {
2208 if (skb == tcp_send_head(sk))
2209 break;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002210 if (!tcp_skb_timedout(sk, skb))
2211 break;
2212
Ilpo Järvinen261ab362007-11-01 00:10:18 -07002213 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_SACKED_ACKED|TCPCB_LOST))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002214 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2215 tp->lost_out += tcp_skb_pcount(skb);
Ilpo Järvinend8f4f222007-04-20 22:56:38 -07002216 tcp_verify_retransmit_hint(tp, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002217 }
2218 }
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002219
2220 tp->scoreboard_skb_hint = skb;
2221
Ilpo Järvinen005903b2007-08-09 14:44:16 +03002222 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002223 }
2224}
2225
2226/* CWND moderation, preventing bursts due to too big ACKs
2227 * in dubious situations.
2228 */
2229static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2230{
2231 tp->snd_cwnd = min(tp->snd_cwnd,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002232 tcp_packets_in_flight(tp) + tcp_max_burst(tp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002233 tp->snd_cwnd_stamp = tcp_time_stamp;
2234}
2235
Stephen Hemminger72dc5b92006-06-05 17:30:08 -07002236/* Lower bound on congestion window is slow start threshold
2237 * unless congestion avoidance choice decides to overide it.
2238 */
2239static inline u32 tcp_cwnd_min(const struct sock *sk)
2240{
2241 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2242
2243 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
2244}
2245
Linus Torvalds1da177e2005-04-16 15:20:36 -07002246/* Decrease cwnd each second ack. */
Ilpo Järvinen1e757f92007-07-30 19:48:37 -07002247static void tcp_cwnd_down(struct sock *sk, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002248{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002249 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002250 int decr = tp->snd_cwnd_cnt + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002251
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002252 if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK)) ||
2253 (tcp_is_reno(tp) && !(flag & FLAG_NOT_DUP))) {
2254 tp->snd_cwnd_cnt = decr & 1;
Ilpo Järvinen1e757f92007-07-30 19:48:37 -07002255 decr >>= 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002256
Ilpo Järvinen1e757f92007-07-30 19:48:37 -07002257 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
2258 tp->snd_cwnd -= decr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002259
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002260 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
Ilpo Järvinen1e757f92007-07-30 19:48:37 -07002261 tp->snd_cwnd_stamp = tcp_time_stamp;
2262 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002263}
2264
2265/* Nothing was retransmitted or returned timestamp is less
2266 * than timestamp of the first retransmission.
2267 */
2268static inline int tcp_packet_delayed(struct tcp_sock *tp)
2269{
2270 return !tp->retrans_stamp ||
2271 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2272 (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0);
2273}
2274
2275/* Undo procedures. */
2276
2277#if FASTRETRANS_DEBUG > 1
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002278static void DBGUNDO(struct sock *sk, const char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002279{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002280 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002281 struct inet_sock *inet = inet_sk(sk);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002282
Linus Torvalds1da177e2005-04-16 15:20:36 -07002283 printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
2284 msg,
2285 NIPQUAD(inet->daddr), ntohs(inet->dport),
Ilpo Järvinen83ae4082007-08-09 14:37:30 +03002286 tp->snd_cwnd, tcp_left_out(tp),
Linus Torvalds1da177e2005-04-16 15:20:36 -07002287 tp->snd_ssthresh, tp->prior_ssthresh,
2288 tp->packets_out);
2289}
2290#else
2291#define DBGUNDO(x...) do { } while (0)
2292#endif
2293
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002294static void tcp_undo_cwr(struct sock *sk, const int undo)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002295{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002296 struct tcp_sock *tp = tcp_sk(sk);
2297
Linus Torvalds1da177e2005-04-16 15:20:36 -07002298 if (tp->prior_ssthresh) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002299 const struct inet_connection_sock *icsk = inet_csk(sk);
2300
2301 if (icsk->icsk_ca_ops->undo_cwnd)
2302 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002303 else
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002304 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002305
2306 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
2307 tp->snd_ssthresh = tp->prior_ssthresh;
2308 TCP_ECN_withdraw_cwr(tp);
2309 }
2310 } else {
2311 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2312 }
2313 tcp_moderate_cwnd(tp);
2314 tp->snd_cwnd_stamp = tcp_time_stamp;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002315
2316 /* There is something screwy going on with the retrans hints after
2317 an undo */
Ilpo Järvinen5af4ec22007-09-20 11:30:48 -07002318 tcp_clear_all_retrans_hints(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002319}
2320
2321static inline int tcp_may_undo(struct tcp_sock *tp)
2322{
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002323 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002324}
2325
2326/* People celebrate: "We love our President!" */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002327static int tcp_try_undo_recovery(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002328{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002329 struct tcp_sock *tp = tcp_sk(sk);
2330
Linus Torvalds1da177e2005-04-16 15:20:36 -07002331 if (tcp_may_undo(tp)) {
2332 /* Happy end! We did not retransmit anything
2333 * or our original transmission succeeded.
2334 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002335 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002336 tcp_undo_cwr(sk, 1);
2337 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002338 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
2339 else
2340 NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO);
2341 tp->undo_marker = 0;
2342 }
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002343 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002344 /* Hold old state until something *above* high_seq
2345 * is ACKed. For Reno it is MUST to prevent false
2346 * fast retransmits (RFC2582). SACK TCP is safe. */
2347 tcp_moderate_cwnd(tp);
2348 return 1;
2349 }
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002350 tcp_set_ca_state(sk, TCP_CA_Open);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002351 return 0;
2352}
2353
2354/* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002355static void tcp_try_undo_dsack(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002356{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002357 struct tcp_sock *tp = tcp_sk(sk);
2358
Linus Torvalds1da177e2005-04-16 15:20:36 -07002359 if (tp->undo_marker && !tp->undo_retrans) {
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002360 DBGUNDO(sk, "D-SACK");
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002361 tcp_undo_cwr(sk, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002362 tp->undo_marker = 0;
2363 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO);
2364 }
2365}
2366
2367/* Undo during fast recovery after partial ACK. */
2368
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002369static int tcp_try_undo_partial(struct sock *sk, int acked)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002370{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002371 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002372 /* Partial ACK arrived. Force Hoe's retransmit. */
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002373 int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002374
2375 if (tcp_may_undo(tp)) {
2376 /* Plain luck! Hole if filled with delayed
2377 * packet, rather than with a retransmit.
2378 */
2379 if (tp->retrans_out == 0)
2380 tp->retrans_stamp = 0;
2381
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002382 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002383
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002384 DBGUNDO(sk, "Hoe");
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002385 tcp_undo_cwr(sk, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002386 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO);
2387
2388 /* So... Do not make Hoe's retransmit yet.
2389 * If the first packet was delayed, the rest
2390 * ones are most probably delayed as well.
2391 */
2392 failed = 0;
2393 }
2394 return failed;
2395}
2396
2397/* Undo during loss recovery after partial ACK. */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002398static int tcp_try_undo_loss(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002399{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002400 struct tcp_sock *tp = tcp_sk(sk);
2401
Linus Torvalds1da177e2005-04-16 15:20:36 -07002402 if (tcp_may_undo(tp)) {
2403 struct sk_buff *skb;
David S. Millerfe067e82007-03-07 12:12:44 -08002404 tcp_for_write_queue(skb, sk) {
2405 if (skb == tcp_send_head(sk))
2406 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002407 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2408 }
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002409
Ilpo Järvinen5af4ec22007-09-20 11:30:48 -07002410 tcp_clear_all_retrans_hints(tp);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002411
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002412 DBGUNDO(sk, "partial loss");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002413 tp->lost_out = 0;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002414 tcp_undo_cwr(sk, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002415 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002416 inet_csk(sk)->icsk_retransmits = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002417 tp->undo_marker = 0;
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002418 if (tcp_is_sack(tp))
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002419 tcp_set_ca_state(sk, TCP_CA_Open);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002420 return 1;
2421 }
2422 return 0;
2423}
2424
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002425static inline void tcp_complete_cwr(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002426{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002427 struct tcp_sock *tp = tcp_sk(sk);
Stephen Hemminger317a76f2005-06-23 12:19:55 -07002428 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002429 tp->snd_cwnd_stamp = tcp_time_stamp;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002430 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002431}
2432
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002433static void tcp_try_to_open(struct sock *sk, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002434{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002435 struct tcp_sock *tp = tcp_sk(sk);
2436
Ilpo Järvinen86426c22007-08-09 14:45:17 +03002437 tcp_verify_left_out(tp);
2438
Linus Torvalds1da177e2005-04-16 15:20:36 -07002439 if (tp->retrans_out == 0)
2440 tp->retrans_stamp = 0;
2441
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002442 if (flag & FLAG_ECE)
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -08002443 tcp_enter_cwr(sk, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002444
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002445 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002446 int state = TCP_CA_Open;
2447
Ilpo Järvinend02596e2007-07-07 13:39:02 +03002448 if (tcp_left_out(tp) || tp->retrans_out || tp->undo_marker)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002449 state = TCP_CA_Disorder;
2450
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002451 if (inet_csk(sk)->icsk_ca_state != state) {
2452 tcp_set_ca_state(sk, state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002453 tp->high_seq = tp->snd_nxt;
2454 }
2455 tcp_moderate_cwnd(tp);
2456 } else {
Ilpo Järvinen1e757f92007-07-30 19:48:37 -07002457 tcp_cwnd_down(sk, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002458 }
2459}
2460
John Heffner5d424d52006-03-20 17:53:41 -08002461static void tcp_mtup_probe_failed(struct sock *sk)
2462{
2463 struct inet_connection_sock *icsk = inet_csk(sk);
2464
2465 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2466 icsk->icsk_mtup.probe_size = 0;
2467}
2468
2469static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb)
2470{
2471 struct tcp_sock *tp = tcp_sk(sk);
2472 struct inet_connection_sock *icsk = inet_csk(sk);
2473
2474 /* FIXME: breaks with very large cwnd */
2475 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2476 tp->snd_cwnd = tp->snd_cwnd *
2477 tcp_mss_to_mtu(sk, tp->mss_cache) /
2478 icsk->icsk_mtup.probe_size;
2479 tp->snd_cwnd_cnt = 0;
2480 tp->snd_cwnd_stamp = tcp_time_stamp;
2481 tp->rcv_ssthresh = tcp_current_ssthresh(sk);
2482
2483 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2484 icsk->icsk_mtup.probe_size = 0;
2485 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2486}
2487
Linus Torvalds1da177e2005-04-16 15:20:36 -07002488/* Process an event, which can update packets-in-flight not trivially.
2489 * Main goal of this function is to calculate new estimate for left_out,
2490 * taking into account both packets sitting in receiver's buffer and
2491 * packets lost by network.
2492 *
2493 * Besides that it does CWND reduction, when packet loss is detected
2494 * and changes state of machine.
2495 *
2496 * It does _not_ decide what to send, it is made in function
2497 * tcp_xmit_retransmit_queue().
2498 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002499static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002500{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002501 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002502 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002503 int is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
2504 int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002505 (tcp_fackets_out(tp) > tp->reordering));
2506 int fast_rexmit = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002507
Ilpo Järvinen3ccd3132007-12-31 04:43:32 -08002508 if (WARN_ON(!tp->packets_out && tp->sacked_out))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002509 tp->sacked_out = 0;
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07002510 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002511 tp->fackets_out = 0;
2512
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002513 /* Now state machine starts.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002514 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002515 if (flag & FLAG_ECE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002516 tp->prior_ssthresh = 0;
2517
2518 /* B. In all the states check for reneging SACKs. */
Ilpo Järvinencadbd032007-12-31 04:49:21 -08002519 if (tcp_check_sack_reneging(sk, flag))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002520 return;
2521
2522 /* C. Process data loss notification, provided it is valid. */
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002523 if (tcp_is_fack(tp) && (flag & FLAG_DATA_LOST) &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07002524 before(tp->snd_una, tp->high_seq) &&
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002525 icsk->icsk_ca_state != TCP_CA_Open &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07002526 tp->fackets_out > tp->reordering) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002527 tcp_mark_head_lost(sk, tp->fackets_out - tp->reordering, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002528 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS);
2529 }
2530
Ilpo Järvinen005903b2007-08-09 14:44:16 +03002531 /* D. Check consistency of the current state. */
2532 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002533
2534 /* E. Check state exit conditions. State can be terminated
2535 * when high_seq is ACKed. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002536 if (icsk->icsk_ca_state == TCP_CA_Open) {
Ilpo Järvinen7b0eb222007-02-21 23:03:35 -08002537 BUG_TRAP(tp->retrans_out == 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002538 tp->retrans_stamp = 0;
2539 } else if (!before(tp->snd_una, tp->high_seq)) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002540 switch (icsk->icsk_ca_state) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002541 case TCP_CA_Loss:
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002542 icsk->icsk_retransmits = 0;
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002543 if (tcp_try_undo_recovery(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002544 return;
2545 break;
2546
2547 case TCP_CA_CWR:
2548 /* CWR is to be held something *above* high_seq
2549 * is ACKed for CWR bit to reach receiver. */
2550 if (tp->snd_una != tp->high_seq) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002551 tcp_complete_cwr(sk);
2552 tcp_set_ca_state(sk, TCP_CA_Open);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002553 }
2554 break;
2555
2556 case TCP_CA_Disorder:
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002557 tcp_try_undo_dsack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002558 if (!tp->undo_marker ||
2559 /* For SACK case do not Open to allow to undo
2560 * catching for all duplicate ACKs. */
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002561 tcp_is_reno(tp) || tp->snd_una != tp->high_seq) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002562 tp->undo_marker = 0;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002563 tcp_set_ca_state(sk, TCP_CA_Open);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002564 }
2565 break;
2566
2567 case TCP_CA_Recovery:
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002568 if (tcp_is_reno(tp))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002569 tcp_reset_reno_sack(tp);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002570 if (tcp_try_undo_recovery(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002571 return;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002572 tcp_complete_cwr(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002573 break;
2574 }
2575 }
2576
2577 /* F. Process state. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002578 switch (icsk->icsk_ca_state) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002579 case TCP_CA_Recovery:
Ilpo Järvinen2e605292007-08-02 19:46:58 -07002580 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002581 if (tcp_is_reno(tp) && is_dupack)
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002582 tcp_add_reno_sack(sk);
Ilpo Järvinen1b6d427b2007-08-09 14:53:36 +03002583 } else
2584 do_lost = tcp_try_undo_partial(sk, pkts_acked);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002585 break;
2586 case TCP_CA_Loss:
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002587 if (flag & FLAG_DATA_ACKED)
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002588 icsk->icsk_retransmits = 0;
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002589 if (!tcp_try_undo_loss(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002590 tcp_moderate_cwnd(tp);
2591 tcp_xmit_retransmit_queue(sk);
2592 return;
2593 }
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002594 if (icsk->icsk_ca_state != TCP_CA_Open)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002595 return;
2596 /* Loss is undone; fall through to processing in Open state. */
2597 default:
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002598 if (tcp_is_reno(tp)) {
Ilpo Järvinen2e605292007-08-02 19:46:58 -07002599 if (flag & FLAG_SND_UNA_ADVANCED)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002600 tcp_reset_reno_sack(tp);
2601 if (is_dupack)
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002602 tcp_add_reno_sack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002603 }
2604
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002605 if (icsk->icsk_ca_state == TCP_CA_Disorder)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002606 tcp_try_undo_dsack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002607
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002608 if (!tcp_time_to_recover(sk)) {
2609 tcp_try_to_open(sk, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002610 return;
2611 }
2612
John Heffner5d424d52006-03-20 17:53:41 -08002613 /* MTU probe failure: don't reduce cwnd */
2614 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2615 icsk->icsk_mtup.probe_size &&
John Heffner0e7b1362006-03-20 21:32:58 -08002616 tp->snd_una == tp->mtu_probe.probe_seq_start) {
John Heffner5d424d52006-03-20 17:53:41 -08002617 tcp_mtup_probe_failed(sk);
2618 /* Restores the reduction we did in tcp_mtup_probe() */
2619 tp->snd_cwnd++;
2620 tcp_simple_retransmit(sk);
2621 return;
2622 }
2623
Linus Torvalds1da177e2005-04-16 15:20:36 -07002624 /* Otherwise enter Recovery state */
2625
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002626 if (tcp_is_reno(tp))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002627 NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY);
2628 else
2629 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY);
2630
2631 tp->high_seq = tp->snd_nxt;
2632 tp->prior_ssthresh = 0;
2633 tp->undo_marker = tp->snd_una;
2634 tp->undo_retrans = tp->retrans_out;
2635
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002636 if (icsk->icsk_ca_state < TCP_CA_CWR) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002637 if (!(flag & FLAG_ECE))
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002638 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2639 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002640 TCP_ECN_queue_cwr(tp);
2641 }
2642
Stephen Hemminger9772efb2005-11-10 17:09:53 -08002643 tp->bytes_acked = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002644 tp->snd_cwnd_cnt = 0;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002645 tcp_set_ca_state(sk, TCP_CA_Recovery);
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002646 fast_rexmit = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002647 }
2648
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002649 if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk)))
2650 tcp_update_scoreboard(sk, fast_rexmit);
Ilpo Järvinen1e757f92007-07-30 19:48:37 -07002651 tcp_cwnd_down(sk, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002652 tcp_xmit_retransmit_queue(sk);
2653}
2654
2655/* Read draft-ietf-tcplw-high-performance before mucking
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08002656 * with this code. (Supersedes RFC1323)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002657 */
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08002658static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002659{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002660 /* RTTM Rule: A TSecr value received in a segment is used to
2661 * update the averaged RTT measurement only if the segment
2662 * acknowledges some new data, i.e., only if it advances the
2663 * left edge of the send window.
2664 *
2665 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2666 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2667 *
2668 * Changed: reset backoff as soon as we see the first valid sample.
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08002669 * If we do not, we get strongly overestimated rto. With timestamps
Linus Torvalds1da177e2005-04-16 15:20:36 -07002670 * samples are accepted even from very old segments: f.e., when rtt=1
2671 * increases to 8, we retransmit 5 times and after 8 seconds delayed
2672 * answer arrives rto becomes 120 seconds! If at least one of segments
2673 * in window is lost... Voila. --ANK (010210)
2674 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002675 struct tcp_sock *tp = tcp_sk(sk);
2676 const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08002677 tcp_rtt_estimator(sk, seq_rtt);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002678 tcp_set_rto(sk);
2679 inet_csk(sk)->icsk_backoff = 0;
2680 tcp_bound_rto(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002681}
2682
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08002683static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002684{
2685 /* We don't have a timestamp. Can only use
2686 * packets that are not retransmitted to determine
2687 * rtt estimates. Also, we must not reset the
2688 * backoff for rto until we get a non-retransmitted
2689 * packet. This allows us to deal with a situation
2690 * where the network delay has increased suddenly.
2691 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2692 */
2693
2694 if (flag & FLAG_RETRANS_DATA_ACKED)
2695 return;
2696
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08002697 tcp_rtt_estimator(sk, seq_rtt);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002698 tcp_set_rto(sk);
2699 inet_csk(sk)->icsk_backoff = 0;
2700 tcp_bound_rto(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002701}
2702
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002703static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08002704 const s32 seq_rtt)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002705{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002706 const struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002707 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2708 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08002709 tcp_ack_saw_tstamp(sk, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002710 else if (seq_rtt >= 0)
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08002711 tcp_ack_no_tstamp(sk, seq_rtt, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002712}
2713
Ilpo Järvinenc3a05c62007-12-02 00:47:59 +02002714static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002715{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002716 const struct inet_connection_sock *icsk = inet_csk(sk);
Ilpo Järvinenc3a05c62007-12-02 00:47:59 +02002717 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002718 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002719}
2720
Linus Torvalds1da177e2005-04-16 15:20:36 -07002721/* Restart timer after forward progress on connection.
2722 * RFC2988 recommends to restart timer to now+rto.
2723 */
Ilpo Järvinen6728e7d2007-08-24 22:53:26 -07002724static void tcp_rearm_rto(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002725{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002726 struct tcp_sock *tp = tcp_sk(sk);
2727
Linus Torvalds1da177e2005-04-16 15:20:36 -07002728 if (!tp->packets_out) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002729 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002730 } else {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002731 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2732 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002733 }
2734}
2735
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07002736/* If we get here, the whole TSO packet has not been acked. */
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07002737static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002738{
2739 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07002740 u32 packets_acked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002741
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07002742 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002743
2744 packets_acked = tcp_skb_pcount(skb);
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07002745 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002746 return 0;
2747 packets_acked -= tcp_skb_pcount(skb);
2748
2749 if (packets_acked) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002750 BUG_ON(tcp_skb_pcount(skb) == 0);
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07002751 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002752 }
2753
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07002754 return packets_acked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002755}
2756
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07002757/* Remove acknowledged frames from the retransmission queue. If our packet
2758 * is before the ack sequence we can discard it as it's confirmed to have
2759 * arrived at the other end.
2760 */
Ilpo Järvinenc776ee02007-12-24 21:55:39 -08002761static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002762{
2763 struct tcp_sock *tp = tcp_sk(sk);
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08002764 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002765 struct sk_buff *skb;
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07002766 u32 now = tcp_time_stamp;
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07002767 int fully_acked = 1;
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07002768 int flag = 0;
Ilpo Järvinen72018832007-12-30 04:37:55 -08002769 u32 pkts_acked = 0;
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08002770 u32 reord = tp->packets_out;
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07002771 s32 seq_rtt = -1;
Gavin McCullagh2072c222007-12-29 19:11:21 -08002772 s32 ca_seq_rtt = -1;
Ilpo Järvinenb9ce2042007-06-15 15:08:43 -07002773 ktime_t last_ackt = net_invalid_timestamp();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002774
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07002775 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002776 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07002777 u32 end_seq;
Ilpo Järvinen72018832007-12-30 04:37:55 -08002778 u32 acked_pcount;
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07002779 u8 sacked = scb->sacked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002780
Gavin McCullagh2072c222007-12-29 19:11:21 -08002781 /* Determine how many packets and what bytes were acked, tso and else */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002782 if (after(scb->end_seq, tp->snd_una)) {
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07002783 if (tcp_skb_pcount(skb) == 1 ||
2784 !after(tp->snd_una, scb->seq))
2785 break;
2786
Ilpo Järvinen72018832007-12-30 04:37:55 -08002787 acked_pcount = tcp_tso_acked(sk, skb);
2788 if (!acked_pcount)
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07002789 break;
2790
2791 fully_acked = 0;
2792 end_seq = tp->snd_una;
2793 } else {
Ilpo Järvinen72018832007-12-30 04:37:55 -08002794 acked_pcount = tcp_skb_pcount(skb);
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07002795 end_seq = scb->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002796 }
2797
John Heffner5d424d52006-03-20 17:53:41 -08002798 /* MTU probing checks */
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07002799 if (fully_acked && icsk->icsk_mtup.probe_size &&
2800 !after(tp->mtu_probe.probe_seq_end, scb->end_seq)) {
2801 tcp_mtup_probe_success(sk, skb);
John Heffner5d424d52006-03-20 17:53:41 -08002802 }
2803
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08002804 if (sacked & TCPCB_RETRANS) {
2805 if (sacked & TCPCB_SACKED_RETRANS)
Ilpo Järvinen72018832007-12-30 04:37:55 -08002806 tp->retrans_out -= acked_pcount;
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08002807 flag |= FLAG_RETRANS_DATA_ACKED;
2808 ca_seq_rtt = -1;
2809 seq_rtt = -1;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002810 if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1))
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08002811 flag |= FLAG_NONHEAD_RETRANS_ACKED;
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08002812 } else {
Gavin McCullagh2072c222007-12-29 19:11:21 -08002813 ca_seq_rtt = now - scb->when;
2814 last_ackt = skb->tstamp;
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08002815 if (seq_rtt < 0) {
Gavin McCullagh2072c222007-12-29 19:11:21 -08002816 seq_rtt = ca_seq_rtt;
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08002817 }
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08002818 if (!(sacked & TCPCB_SACKED_ACKED))
Ilpo Järvinen72018832007-12-30 04:37:55 -08002819 reord = min(pkts_acked, reord);
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08002820 }
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08002821
2822 if (sacked & TCPCB_SACKED_ACKED)
Ilpo Järvinen72018832007-12-30 04:37:55 -08002823 tp->sacked_out -= acked_pcount;
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08002824 if (sacked & TCPCB_LOST)
Ilpo Järvinen72018832007-12-30 04:37:55 -08002825 tp->lost_out -= acked_pcount;
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08002826
Ilpo Järvinen4828e7f2007-12-31 04:50:19 -08002827 if (unlikely(tp->urg_mode && !before(end_seq, tp->snd_up)))
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08002828 tp->urg_mode = 0;
2829
Ilpo Järvinen72018832007-12-30 04:37:55 -08002830 tp->packets_out -= acked_pcount;
2831 pkts_acked += acked_pcount;
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07002832
Ilpo Järvinen009a2e32007-09-20 11:34:38 -07002833 /* Initial outgoing SYN's get put onto the write_queue
2834 * just like anything else we transmit. It is not
2835 * true data, and if we misinform our callers that
2836 * this ACK acks real data, we will erroneously exit
2837 * connection startup slow start one packet too
2838 * quickly. This is severely frowned upon behavior.
2839 */
2840 if (!(scb->flags & TCPCB_FLAG_SYN)) {
2841 flag |= FLAG_DATA_ACKED;
2842 } else {
2843 flag |= FLAG_SYN_ACKED;
2844 tp->retrans_stamp = 0;
2845 }
2846
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07002847 if (!fully_acked)
2848 break;
2849
David S. Millerfe067e82007-03-07 12:12:44 -08002850 tcp_unlink_write_queue(skb, sk);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08002851 sk_wmem_free_skb(sk, skb);
Ilpo Järvinen5af4ec22007-09-20 11:30:48 -07002852 tcp_clear_all_retrans_hints(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002853 }
2854
Ilpo Järvinencadbd032007-12-31 04:49:21 -08002855 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2856 flag |= FLAG_SACK_RENEGING;
2857
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07002858 if (flag & FLAG_ACKED) {
Stephen Hemminger164891a2007-04-23 22:26:16 -07002859 const struct tcp_congestion_ops *ca_ops
2860 = inet_csk(sk)->icsk_ca_ops;
2861
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07002862 tcp_ack_update_rtt(sk, flag, seq_rtt);
Ilpo Järvinen6728e7d2007-08-24 22:53:26 -07002863 tcp_rearm_rto(sk);
Stephen Hemminger317a76f2005-06-23 12:19:55 -07002864
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08002865 if (tcp_is_reno(tp)) {
2866 tcp_remove_reno_sacks(sk, pkts_acked);
2867 } else {
2868 /* Non-retransmitted hole got filled? That's reordering */
2869 if (reord < prior_fackets)
2870 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
2871 }
2872
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07002873 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08002874
Stephen Hemminger30cfd0b2007-07-25 23:49:34 -07002875 if (ca_ops->pkts_acked) {
2876 s32 rtt_us = -1;
Ilpo Järvinenb9ce2042007-06-15 15:08:43 -07002877
Stephen Hemminger30cfd0b2007-07-25 23:49:34 -07002878 /* Is the ACK triggering packet unambiguous? */
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07002879 if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
Stephen Hemminger30cfd0b2007-07-25 23:49:34 -07002880 /* High resolution needed and available? */
2881 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
2882 !ktime_equal(last_ackt,
2883 net_invalid_timestamp()))
2884 rtt_us = ktime_us_delta(ktime_get_real(),
2885 last_ackt);
Gavin McCullagh2072c222007-12-29 19:11:21 -08002886 else if (ca_seq_rtt > 0)
2887 rtt_us = jiffies_to_usecs(ca_seq_rtt);
Stephen Hemminger30cfd0b2007-07-25 23:49:34 -07002888 }
2889
2890 ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
2891 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002892 }
2893
2894#if FASTRETRANS_DEBUG > 0
2895 BUG_TRAP((int)tp->sacked_out >= 0);
2896 BUG_TRAP((int)tp->lost_out >= 0);
2897 BUG_TRAP((int)tp->retrans_out >= 0);
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002898 if (!tp->packets_out && tcp_is_sack(tp)) {
Stephen Hemmingercfcabdc2007-10-09 01:59:42 -07002899 icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002900 if (tp->lost_out) {
2901 printk(KERN_DEBUG "Leak l=%u %d\n",
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002902 tp->lost_out, icsk->icsk_ca_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002903 tp->lost_out = 0;
2904 }
2905 if (tp->sacked_out) {
2906 printk(KERN_DEBUG "Leak s=%u %d\n",
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002907 tp->sacked_out, icsk->icsk_ca_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002908 tp->sacked_out = 0;
2909 }
2910 if (tp->retrans_out) {
2911 printk(KERN_DEBUG "Leak r=%u %d\n",
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002912 tp->retrans_out, icsk->icsk_ca_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002913 tp->retrans_out = 0;
2914 }
2915 }
2916#endif
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07002917 return flag;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002918}
2919
2920static void tcp_ack_probe(struct sock *sk)
2921{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002922 const struct tcp_sock *tp = tcp_sk(sk);
2923 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002924
2925 /* Was it a usable window open? */
2926
Ilpo Järvinen90840de2007-12-31 04:48:41 -08002927 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002928 icsk->icsk_backoff = 0;
2929 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002930 /* Socket must be waked up by subsequent tcp_data_snd_check().
2931 * This function is not for random using!
2932 */
2933 } else {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002934 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07002935 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2936 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002937 }
2938}
2939
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002940static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002941{
2942 return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002943 inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002944}
2945
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002946static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002947{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002948 const struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002949 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002950 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002951}
2952
2953/* Check that window update is acceptable.
2954 * The function assumes that snd_una<=ack<=snd_next.
2955 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002956static inline int tcp_may_update_window(const struct tcp_sock *tp,
2957 const u32 ack, const u32 ack_seq,
2958 const u32 nwin)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002959{
2960 return (after(ack, tp->snd_una) ||
2961 after(ack_seq, tp->snd_wl1) ||
2962 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
2963}
2964
2965/* Update our send window.
2966 *
2967 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2968 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2969 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002970static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack,
2971 u32 ack_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002972{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002973 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002974 int flag = 0;
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07002975 u32 nwin = ntohs(tcp_hdr(skb)->window);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002976
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07002977 if (likely(!tcp_hdr(skb)->syn))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002978 nwin <<= tp->rx_opt.snd_wscale;
2979
2980 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
2981 flag |= FLAG_WIN_UPDATE;
2982 tcp_update_wl(tp, ack, ack_seq);
2983
2984 if (tp->snd_wnd != nwin) {
2985 tp->snd_wnd = nwin;
2986
2987 /* Note, it is the only place, where
2988 * fast path is recovered for sending TCP.
2989 */
Herbert Xu2ad41062005-10-27 18:47:46 +10002990 tp->pred_flags = 0;
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002991 tcp_fast_path_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002992
2993 if (nwin > tp->max_window) {
2994 tp->max_window = nwin;
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08002995 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002996 }
2997 }
2998 }
2999
3000 tp->snd_una = ack;
3001
3002 return flag;
3003}
3004
Ilpo Järvinen9ead9a12007-02-21 22:56:19 -08003005/* A very conservative spurious RTO response algorithm: reduce cwnd and
3006 * continue in congestion avoidance.
3007 */
3008static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
3009{
3010 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
Ilpo Järvinenaa8b6a72007-02-21 23:06:03 -08003011 tp->snd_cwnd_cnt = 0;
Ilpo Järvinen16e90682007-10-11 17:32:31 -07003012 tp->bytes_acked = 0;
Ilpo Järvinen46323652007-05-19 13:56:23 -07003013 TCP_ECN_queue_cwr(tp);
Ilpo Järvinen9ead9a12007-02-21 22:56:19 -08003014 tcp_moderate_cwnd(tp);
3015}
3016
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -08003017/* A conservative spurious RTO response algorithm: reduce cwnd using
3018 * rate halving and continue in congestion avoidance.
3019 */
3020static void tcp_ratehalving_spur_to_response(struct sock *sk)
3021{
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -08003022 tcp_enter_cwr(sk, 0);
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -08003023}
3024
Ilpo Järvinene317f6f2007-03-02 13:34:19 -08003025static void tcp_undo_spur_to_response(struct sock *sk, int flag)
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -08003026{
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003027 if (flag & FLAG_ECE)
Ilpo Järvinene317f6f2007-03-02 13:34:19 -08003028 tcp_ratehalving_spur_to_response(sk);
3029 else
3030 tcp_undo_cwr(sk, 1);
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -08003031}
3032
Ilpo Järvinen30935cf2007-02-21 23:01:36 -08003033/* F-RTO spurious RTO detection algorithm (RFC4138)
3034 *
Ilpo Järvinen6408d202007-02-21 23:04:11 -08003035 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
3036 * comments). State (ACK number) is kept in frto_counter. When ACK advances
3037 * window (but not to or beyond highest sequence sent before RTO):
Ilpo Järvinen30935cf2007-02-21 23:01:36 -08003038 * On First ACK, send two new segments out.
3039 * On Second ACK, RTO was likely spurious. Do spurious response (response
3040 * algorithm is not part of the F-RTO detection algorithm
3041 * given in RFC4138 but can be selected separately).
3042 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
Ilpo Järvinend551e452007-04-30 00:42:20 -07003043 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
3044 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
3045 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
Ilpo Järvinen30935cf2007-02-21 23:01:36 -08003046 *
3047 * Rationale: if the RTO was spurious, new ACKs should arrive from the
3048 * original window even after we transmit two new data segments.
3049 *
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08003050 * SACK version:
3051 * on first step, wait until first cumulative ACK arrives, then move to
3052 * the second step. In second step, the next ACK decides.
3053 *
Ilpo Järvinen30935cf2007-02-21 23:01:36 -08003054 * F-RTO is implemented (mainly) in four functions:
3055 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
3056 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
3057 * called when tcp_use_frto() showed green light
3058 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
3059 * - tcp_enter_frto_loss() is called if there is not enough evidence
3060 * to prove that the RTO is indeed spurious. It transfers the control
3061 * from F-RTO to the conventional RTO recovery
3062 */
Ilpo Järvinen2e605292007-08-02 19:46:58 -07003063static int tcp_process_frto(struct sock *sk, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003064{
3065 struct tcp_sock *tp = tcp_sk(sk);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09003066
Ilpo Järvinen005903b2007-08-09 14:44:16 +03003067 tcp_verify_left_out(tp);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09003068
Ilpo Järvinen7487c482007-02-21 23:02:30 -08003069 /* Duplicate the behavior from Loss state (fastretrans_alert) */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003070 if (flag & FLAG_DATA_ACKED)
Ilpo Järvinen7487c482007-02-21 23:02:30 -08003071 inet_csk(sk)->icsk_retransmits = 0;
3072
Ilpo Järvinen009a2e32007-09-20 11:34:38 -07003073 if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||
3074 ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
3075 tp->undo_marker = 0;
3076
Ilpo Järvinen95c49222007-02-21 23:05:18 -08003077 if (!before(tp->snd_una, tp->frto_highmark)) {
Ilpo Järvinend551e452007-04-30 00:42:20 -07003078 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
Ilpo Järvinen7c9a4a52007-02-21 23:08:34 -08003079 return 1;
Ilpo Järvinen95c49222007-02-21 23:05:18 -08003080 }
3081
Ilpo Järvinene60402d2007-08-09 15:14:46 +03003082 if (!IsSackFrto() || tcp_is_reno(tp)) {
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08003083 /* RFC4138 shortcoming in step 2; should also have case c):
3084 * ACK isn't duplicate nor advances window, e.g., opposite dir
3085 * data, winupdate
3086 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003087 if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP))
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08003088 return 1;
Ilpo Järvinen6408d202007-02-21 23:04:11 -08003089
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003090 if (!(flag & FLAG_DATA_ACKED)) {
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08003091 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
3092 flag);
3093 return 1;
3094 }
3095 } else {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003096 if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08003097 /* Prevent sending of new data. */
3098 tp->snd_cwnd = min(tp->snd_cwnd,
3099 tcp_packets_in_flight(tp));
3100 return 1;
3101 }
3102
Ilpo Järvinend551e452007-04-30 00:42:20 -07003103 if ((tp->frto_counter >= 2) &&
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003104 (!(flag & FLAG_FORWARD_PROGRESS) ||
3105 ((flag & FLAG_DATA_SACKED) &&
3106 !(flag & FLAG_ONLY_ORIG_SACKED)))) {
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08003107 /* RFC4138 shortcoming (see comment above) */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003108 if (!(flag & FLAG_FORWARD_PROGRESS) &&
3109 (flag & FLAG_NOT_DUP))
Ilpo Järvinen4dc26652007-02-21 23:16:11 -08003110 return 1;
3111
3112 tcp_enter_frto_loss(sk, 3, flag);
3113 return 1;
3114 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003115 }
3116
3117 if (tp->frto_counter == 1) {
Ilpo Järvinen3e6f0492007-12-05 02:20:21 -08003118 /* tcp_may_send_now needs to see updated state */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003119 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
Ilpo Järvinen94d0ea72007-02-21 23:07:27 -08003120 tp->frto_counter = 2;
Ilpo Järvinen3e6f0492007-12-05 02:20:21 -08003121
3122 if (!tcp_may_send_now(sk))
3123 tcp_enter_frto_loss(sk, 2, flag);
3124
Ilpo Järvinen7c9a4a52007-02-21 23:08:34 -08003125 return 1;
Ilpo Järvinend551e452007-04-30 00:42:20 -07003126 } else {
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -08003127 switch (sysctl_tcp_frto_response) {
3128 case 2:
Ilpo Järvinene317f6f2007-03-02 13:34:19 -08003129 tcp_undo_spur_to_response(sk, flag);
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -08003130 break;
3131 case 1:
3132 tcp_conservative_spur_to_response(tp);
3133 break;
3134 default:
3135 tcp_ratehalving_spur_to_response(sk);
3136 break;
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07003137 }
Ilpo Järvinen94d0ea72007-02-21 23:07:27 -08003138 tp->frto_counter = 0;
Ilpo Järvinen009a2e32007-09-20 11:34:38 -07003139 tp->undo_marker = 0;
Ilpo Järvinen912d8f02007-09-25 22:47:31 -07003140 NET_INC_STATS_BH(LINUX_MIB_TCPSPURIOUSRTOS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003141 }
Ilpo Järvinen7c9a4a52007-02-21 23:08:34 -08003142 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003143}
3144
Linus Torvalds1da177e2005-04-16 15:20:36 -07003145/* This routine deals with incoming acks, but not outgoing ones. */
3146static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
3147{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003148 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003149 struct tcp_sock *tp = tcp_sk(sk);
3150 u32 prior_snd_una = tp->snd_una;
3151 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3152 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3153 u32 prior_in_flight;
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003154 u32 prior_fackets;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003155 int prior_packets;
Ilpo Järvinen7c9a4a52007-02-21 23:08:34 -08003156 int frto_cwnd = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003157
3158 /* If the ack is newer than sent or older than previous acks
3159 * then we can probably ignore it.
3160 */
3161 if (after(ack, tp->snd_nxt))
3162 goto uninteresting_ack;
3163
3164 if (before(ack, prior_snd_una))
3165 goto old_ack;
3166
Ilpo Järvinen2e605292007-08-02 19:46:58 -07003167 if (after(ack, prior_snd_una))
3168 flag |= FLAG_SND_UNA_ADVANCED;
3169
Daikichi Osuga3fdf3f02006-08-29 02:01:44 -07003170 if (sysctl_tcp_abc) {
3171 if (icsk->icsk_ca_state < TCP_CA_CWR)
3172 tp->bytes_acked += ack - prior_snd_una;
3173 else if (icsk->icsk_ca_state == TCP_CA_Loss)
3174 /* we assume just one segment left network */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003175 tp->bytes_acked += min(ack - prior_snd_una,
3176 tp->mss_cache);
Daikichi Osuga3fdf3f02006-08-29 02:01:44 -07003177 }
Stephen Hemminger9772efb2005-11-10 17:09:53 -08003178
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003179 prior_fackets = tp->fackets_out;
Ilpo Järvinen52d34082007-12-05 02:21:35 -08003180 prior_in_flight = tcp_packets_in_flight(tp);
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003181
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003182 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003183 /* Window is constant, pure forward advance.
3184 * No more checks are required.
3185 * Note, we use the fact that SND.UNA>=SND.WL2.
3186 */
3187 tcp_update_wl(tp, ack, ack_seq);
3188 tp->snd_una = ack;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003189 flag |= FLAG_WIN_UPDATE;
3190
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003191 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
Stephen Hemminger317a76f2005-06-23 12:19:55 -07003192
Linus Torvalds1da177e2005-04-16 15:20:36 -07003193 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS);
3194 } else {
3195 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3196 flag |= FLAG_DATA;
3197 else
3198 NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS);
3199
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003200 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003201
3202 if (TCP_SKB_CB(skb)->sacked)
3203 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3204
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07003205 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003206 flag |= FLAG_ECE;
3207
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003208 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003209 }
3210
3211 /* We passed data and got it acked, remove any soft error
3212 * log. Something worked...
3213 */
3214 sk->sk_err_soft = 0;
3215 tp->rcv_tstamp = tcp_time_stamp;
3216 prior_packets = tp->packets_out;
3217 if (!prior_packets)
3218 goto no_queue;
3219
Linus Torvalds1da177e2005-04-16 15:20:36 -07003220 /* See if we can take anything off of the retransmit queue. */
Ilpo Järvinenc776ee02007-12-24 21:55:39 -08003221 flag |= tcp_clean_rtx_queue(sk, prior_fackets);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003222
Ilpo Järvinene1cd8f72007-11-14 15:55:09 -08003223 if (tp->frto_counter)
3224 frto_cwnd = tcp_process_frto(sk, flag);
Ilpo Järvinen3de96472007-10-01 15:28:48 -07003225 /* Guarantee sacktag reordering detection against wrap-arounds */
3226 if (before(tp->frto_highmark, tp->snd_una))
3227 tp->frto_highmark = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003228
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003229 if (tcp_ack_is_dubious(sk, flag)) {
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08003230 /* Advance CWND, if state allows this. */
Ilpo Järvinen7c9a4a52007-02-21 23:08:34 -08003231 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
3232 tcp_may_raise_cwnd(sk, flag))
Ilpo Järvinenc3a05c62007-12-02 00:47:59 +02003233 tcp_cong_avoid(sk, ack, prior_in_flight);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003234 tcp_fastretrans_alert(sk, prior_packets - tp->packets_out,
3235 flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003236 } else {
Ilpo Järvinen7c9a4a52007-02-21 23:08:34 -08003237 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
Ilpo Järvinenc3a05c62007-12-02 00:47:59 +02003238 tcp_cong_avoid(sk, ack, prior_in_flight);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003239 }
3240
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003241 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003242 dst_confirm(sk->sk_dst_cache);
3243
3244 return 1;
3245
3246no_queue:
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003247 icsk->icsk_probes_out = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003248
3249 /* If this ack opens up a zero window, clear backoff. It was
3250 * being used to time the probes, and is probably far higher than
3251 * it needs to be for normal retransmission.
3252 */
David S. Millerfe067e82007-03-07 12:12:44 -08003253 if (tcp_send_head(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003254 tcp_ack_probe(sk);
3255 return 1;
3256
3257old_ack:
3258 if (TCP_SKB_CB(skb)->sacked)
3259 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
3260
3261uninteresting_ack:
3262 SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3263 return 0;
3264}
3265
Linus Torvalds1da177e2005-04-16 15:20:36 -07003266/* Look for tcp options. Normally only called on SYN and SYNACK packets.
3267 * But, this can also be called on packets in the established flow when
3268 * the fast version below fails.
3269 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003270void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx,
3271 int estab)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003272{
3273 unsigned char *ptr;
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07003274 struct tcphdr *th = tcp_hdr(skb);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003275 int length = (th->doff * 4) - sizeof(struct tcphdr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003276
3277 ptr = (unsigned char *)(th + 1);
3278 opt_rx->saw_tstamp = 0;
3279
Stephen Hemminger2de979b2007-03-08 20:45:19 -08003280 while (length > 0) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003281 int opcode = *ptr++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003282 int opsize;
3283
3284 switch (opcode) {
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003285 case TCPOPT_EOL:
3286 return;
3287 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3288 length--;
3289 continue;
3290 default:
3291 opsize = *ptr++;
3292 if (opsize < 2) /* "silly options" */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003293 return;
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003294 if (opsize > length)
3295 return; /* don't parse partial options */
3296 switch (opcode) {
3297 case TCPOPT_MSS:
3298 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3299 u16 in_mss = ntohs(get_unaligned((__be16 *)ptr));
3300 if (in_mss) {
3301 if (opt_rx->user_mss &&
3302 opt_rx->user_mss < in_mss)
3303 in_mss = opt_rx->user_mss;
3304 opt_rx->mss_clamp = in_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003305 }
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07003306 }
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003307 break;
3308 case TCPOPT_WINDOW:
3309 if (opsize == TCPOLEN_WINDOW && th->syn &&
3310 !estab && sysctl_tcp_window_scaling) {
3311 __u8 snd_wscale = *(__u8 *)ptr;
3312 opt_rx->wscale_ok = 1;
3313 if (snd_wscale > 14) {
3314 if (net_ratelimit())
3315 printk(KERN_INFO "tcp_parse_options: Illegal window "
3316 "scaling value %d >14 received.\n",
3317 snd_wscale);
3318 snd_wscale = 14;
3319 }
3320 opt_rx->snd_wscale = snd_wscale;
3321 }
3322 break;
3323 case TCPOPT_TIMESTAMP:
3324 if ((opsize == TCPOLEN_TIMESTAMP) &&
3325 ((estab && opt_rx->tstamp_ok) ||
3326 (!estab && sysctl_tcp_timestamps))) {
3327 opt_rx->saw_tstamp = 1;
3328 opt_rx->rcv_tsval = ntohl(get_unaligned((__be32 *)ptr));
3329 opt_rx->rcv_tsecr = ntohl(get_unaligned((__be32 *)(ptr+4)));
3330 }
3331 break;
3332 case TCPOPT_SACK_PERM:
3333 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3334 !estab && sysctl_tcp_sack) {
3335 opt_rx->sack_ok = 1;
3336 tcp_sack_reset(opt_rx);
3337 }
3338 break;
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07003339
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08003340 case TCPOPT_SACK:
3341 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3342 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3343 opt_rx->sack_ok) {
3344 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3345 }
3346 break;
3347#ifdef CONFIG_TCP_MD5SIG
3348 case TCPOPT_MD5SIG:
3349 /*
3350 * The MD5 Hash has already been
3351 * checked (see tcp_v{4,6}_do_rcv()).
3352 */
3353 break;
3354#endif
3355 }
3356
3357 ptr += opsize-2;
3358 length -= opsize;
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07003359 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003360 }
3361}
3362
3363/* Fast parse options. This hopes to only see timestamps.
3364 * If it is wrong it falls back on tcp_parse_options().
3365 */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -08003366static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
3367 struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003368{
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003369 if (th->doff == sizeof(struct tcphdr) >> 2) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003370 tp->rx_opt.saw_tstamp = 0;
3371 return 0;
3372 } else if (tp->rx_opt.tstamp_ok &&
3373 th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
Al Viro4f3608b2006-09-27 18:35:09 -07003374 __be32 *ptr = (__be32 *)(th + 1);
3375 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003376 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3377 tp->rx_opt.saw_tstamp = 1;
3378 ++ptr;
3379 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3380 ++ptr;
3381 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3382 return 1;
3383 }
3384 }
3385 tcp_parse_options(skb, &tp->rx_opt, 1);
3386 return 1;
3387}
3388
3389static inline void tcp_store_ts_recent(struct tcp_sock *tp)
3390{
3391 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
James Morris9d729f72007-03-04 16:12:44 -08003392 tp->rx_opt.ts_recent_stamp = get_seconds();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003393}
3394
3395static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3396{
3397 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3398 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3399 * extra check below makes sure this can only happen
3400 * for pure ACK frames. -DaveM
3401 *
3402 * Not only, also it occurs for expired timestamps.
3403 */
3404
Stephen Hemminger2de979b2007-03-08 20:45:19 -08003405 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 ||
James Morris9d729f72007-03-04 16:12:44 -08003406 get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003407 tcp_store_ts_recent(tp);
3408 }
3409}
3410
3411/* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3412 *
3413 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3414 * it can pass through stack. So, the following predicate verifies that
3415 * this segment is not used for anything but congestion avoidance or
3416 * fast retransmit. Moreover, we even are able to eliminate most of such
3417 * second order effects, if we apply some small "replay" window (~RTO)
3418 * to timestamp space.
3419 *
3420 * All these measures still do not guarantee that we reject wrapped ACKs
3421 * on networks with high bandwidth, when sequence space is recycled fastly,
3422 * but it guarantees that such events will be very rare and do not affect
3423 * connection seriously. This doesn't look nice, but alas, PAWS is really
3424 * buggy extension.
3425 *
3426 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3427 * states that events when retransmit arrives after original data are rare.
3428 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3429 * the biggest problem on large power networks even with minor reordering.
3430 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3431 * up to bandwidth of 18Gigabit/sec. 8) ]
3432 */
3433
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003434static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003435{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003436 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07003437 struct tcphdr *th = tcp_hdr(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003438 u32 seq = TCP_SKB_CB(skb)->seq;
3439 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3440
3441 return (/* 1. Pure ACK with correct sequence number. */
3442 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3443
3444 /* 2. ... and duplicate ACK. */
3445 ack == tp->snd_una &&
3446
3447 /* 3. ... and does not update window. */
3448 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3449
3450 /* 4. ... and sits in replay window. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003451 (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 -07003452}
3453
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003454static inline int tcp_paws_discard(const struct sock *sk,
3455 const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003456{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003457 const struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003458 return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW &&
James Morris9d729f72007-03-04 16:12:44 -08003459 get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS &&
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003460 !tcp_disordered_ack(sk, skb));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003461}
3462
3463/* Check segment sequence number for validity.
3464 *
3465 * Segment controls are considered valid, if the segment
3466 * fits to the window after truncation to the window. Acceptability
3467 * of data (and SYN, FIN, of course) is checked separately.
3468 * See tcp_data_queue(), for example.
3469 *
3470 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3471 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3472 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3473 * (borrowed from freebsd)
3474 */
3475
3476static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
3477{
3478 return !before(end_seq, tp->rcv_wup) &&
3479 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3480}
3481
3482/* When we get a reset we do this. */
3483static void tcp_reset(struct sock *sk)
3484{
3485 /* We want the right error as BSD sees it (and indeed as we do). */
3486 switch (sk->sk_state) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003487 case TCP_SYN_SENT:
3488 sk->sk_err = ECONNREFUSED;
3489 break;
3490 case TCP_CLOSE_WAIT:
3491 sk->sk_err = EPIPE;
3492 break;
3493 case TCP_CLOSE:
3494 return;
3495 default:
3496 sk->sk_err = ECONNRESET;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003497 }
3498
3499 if (!sock_flag(sk, SOCK_DEAD))
3500 sk->sk_error_report(sk);
3501
3502 tcp_done(sk);
3503}
3504
3505/*
3506 * Process the FIN bit. This now behaves as it is supposed to work
3507 * and the FIN takes effect when it is validly part of sequence
3508 * space. Not before when we get holes.
3509 *
3510 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3511 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3512 * TIME-WAIT)
3513 *
3514 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3515 * close and we go into CLOSING (and later onto TIME-WAIT)
3516 *
3517 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3518 */
3519static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
3520{
3521 struct tcp_sock *tp = tcp_sk(sk);
3522
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003523 inet_csk_schedule_ack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003524
3525 sk->sk_shutdown |= RCV_SHUTDOWN;
3526 sock_set_flag(sk, SOCK_DONE);
3527
3528 switch (sk->sk_state) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003529 case TCP_SYN_RECV:
3530 case TCP_ESTABLISHED:
3531 /* Move to CLOSE_WAIT */
3532 tcp_set_state(sk, TCP_CLOSE_WAIT);
3533 inet_csk(sk)->icsk_ack.pingpong = 1;
3534 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003535
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003536 case TCP_CLOSE_WAIT:
3537 case TCP_CLOSING:
3538 /* Received a retransmission of the FIN, do
3539 * nothing.
3540 */
3541 break;
3542 case TCP_LAST_ACK:
3543 /* RFC793: Remain in the LAST-ACK state. */
3544 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003545
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003546 case TCP_FIN_WAIT1:
3547 /* This case occurs when a simultaneous close
3548 * happens, we must ack the received FIN and
3549 * enter the CLOSING state.
3550 */
3551 tcp_send_ack(sk);
3552 tcp_set_state(sk, TCP_CLOSING);
3553 break;
3554 case TCP_FIN_WAIT2:
3555 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3556 tcp_send_ack(sk);
3557 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3558 break;
3559 default:
3560 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3561 * cases we should never reach this piece of code.
3562 */
3563 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
3564 __FUNCTION__, sk->sk_state);
3565 break;
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07003566 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003567
3568 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3569 * Probably, we should reset in this case. For now drop them.
3570 */
3571 __skb_queue_purge(&tp->out_of_order_queue);
Ilpo Järvinene60402d2007-08-09 15:14:46 +03003572 if (tcp_is_sack(tp))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003573 tcp_sack_reset(&tp->rx_opt);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08003574 sk_mem_reclaim(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003575
3576 if (!sock_flag(sk, SOCK_DEAD)) {
3577 sk->sk_state_change(sk);
3578
3579 /* Do not send POLL_HUP for half duplex close. */
3580 if (sk->sk_shutdown == SHUTDOWN_MASK ||
3581 sk->sk_state == TCP_CLOSE)
Pavel Emelyanov8d8ad9d2007-11-26 20:10:50 +08003582 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003583 else
Pavel Emelyanov8d8ad9d2007-11-26 20:10:50 +08003584 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003585 }
3586}
3587
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003588static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
3589 u32 end_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003590{
3591 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3592 if (before(seq, sp->start_seq))
3593 sp->start_seq = seq;
3594 if (after(end_seq, sp->end_seq))
3595 sp->end_seq = end_seq;
3596 return 1;
3597 }
3598 return 0;
3599}
3600
Stephen Hemminger40efc6f2006-01-03 16:03:49 -08003601static void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003602{
Ilpo Järvinene60402d2007-08-09 15:14:46 +03003603 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003604 if (before(seq, tp->rcv_nxt))
3605 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT);
3606 else
3607 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT);
3608
3609 tp->rx_opt.dsack = 1;
3610 tp->duplicate_sack[0].start_seq = seq;
3611 tp->duplicate_sack[0].end_seq = end_seq;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003612 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1,
3613 4 - tp->rx_opt.tstamp_ok);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003614 }
3615}
3616
Stephen Hemminger40efc6f2006-01-03 16:03:49 -08003617static void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003618{
3619 if (!tp->rx_opt.dsack)
3620 tcp_dsack_set(tp, seq, end_seq);
3621 else
3622 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3623}
3624
3625static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
3626{
3627 struct tcp_sock *tp = tcp_sk(sk);
3628
3629 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3630 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3631 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003632 tcp_enter_quickack_mode(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003633
Ilpo Järvinene60402d2007-08-09 15:14:46 +03003634 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003635 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3636
3637 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3638 end_seq = tp->rcv_nxt;
3639 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq);
3640 }
3641 }
3642
3643 tcp_send_ack(sk);
3644}
3645
3646/* These routines update the SACK block as out-of-order packets arrive or
3647 * in-order packets close up the sequence space.
3648 */
3649static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
3650{
3651 int this_sack;
3652 struct tcp_sack_block *sp = &tp->selective_acks[0];
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003653 struct tcp_sack_block *swalk = sp + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003654
3655 /* See if the recent change to the first SACK eats into
3656 * or hits the sequence space of other SACK blocks, if so coalesce.
3657 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003658 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003659 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
3660 int i;
3661
3662 /* Zap SWALK, by moving every further SACK up by one slot.
3663 * Decrease num_sacks.
3664 */
3665 tp->rx_opt.num_sacks--;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003666 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks +
3667 tp->rx_opt.dsack,
3668 4 - tp->rx_opt.tstamp_ok);
3669 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
3670 sp[i] = sp[i + 1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003671 continue;
3672 }
3673 this_sack++, swalk++;
3674 }
3675}
3676
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003677static inline void tcp_sack_swap(struct tcp_sack_block *sack1,
3678 struct tcp_sack_block *sack2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003679{
3680 __u32 tmp;
3681
3682 tmp = sack1->start_seq;
3683 sack1->start_seq = sack2->start_seq;
3684 sack2->start_seq = tmp;
3685
3686 tmp = sack1->end_seq;
3687 sack1->end_seq = sack2->end_seq;
3688 sack2->end_seq = tmp;
3689}
3690
3691static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
3692{
3693 struct tcp_sock *tp = tcp_sk(sk);
3694 struct tcp_sack_block *sp = &tp->selective_acks[0];
3695 int cur_sacks = tp->rx_opt.num_sacks;
3696 int this_sack;
3697
3698 if (!cur_sacks)
3699 goto new_sack;
3700
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003701 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003702 if (tcp_sack_extend(sp, seq, end_seq)) {
3703 /* Rotate this_sack to the first one. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003704 for (; this_sack > 0; this_sack--, sp--)
3705 tcp_sack_swap(sp, sp - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003706 if (cur_sacks > 1)
3707 tcp_sack_maybe_coalesce(tp);
3708 return;
3709 }
3710 }
3711
3712 /* Could not find an adjacent existing SACK, build a new one,
3713 * put it at the front, and shift everyone else down. We
3714 * always know there is at least one SACK present already here.
3715 *
3716 * If the sack array is full, forget about the last one.
3717 */
3718 if (this_sack >= 4) {
3719 this_sack--;
3720 tp->rx_opt.num_sacks--;
3721 sp--;
3722 }
Stephen Hemminger2de979b2007-03-08 20:45:19 -08003723 for (; this_sack > 0; this_sack--, sp--)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003724 *sp = *(sp - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003725
3726new_sack:
3727 /* Build the new head SACK, and we're done. */
3728 sp->start_seq = seq;
3729 sp->end_seq = end_seq;
3730 tp->rx_opt.num_sacks++;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003731 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack,
3732 4 - tp->rx_opt.tstamp_ok);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003733}
3734
3735/* RCV.NXT advances, some SACKs should be eaten. */
3736
3737static void tcp_sack_remove(struct tcp_sock *tp)
3738{
3739 struct tcp_sack_block *sp = &tp->selective_acks[0];
3740 int num_sacks = tp->rx_opt.num_sacks;
3741 int this_sack;
3742
3743 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
David S. Millerb03efcf2005-07-08 14:57:23 -07003744 if (skb_queue_empty(&tp->out_of_order_queue)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003745 tp->rx_opt.num_sacks = 0;
3746 tp->rx_opt.eff_sacks = tp->rx_opt.dsack;
3747 return;
3748 }
3749
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003750 for (this_sack = 0; this_sack < num_sacks;) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003751 /* Check if the start of the sack is covered by RCV.NXT. */
3752 if (!before(tp->rcv_nxt, sp->start_seq)) {
3753 int i;
3754
3755 /* RCV.NXT must cover all the block! */
3756 BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq));
3757
3758 /* Zap this SACK, by moving forward any other SACKS. */
3759 for (i=this_sack+1; i < num_sacks; i++)
3760 tp->selective_acks[i-1] = tp->selective_acks[i];
3761 num_sacks--;
3762 continue;
3763 }
3764 this_sack++;
3765 sp++;
3766 }
3767 if (num_sacks != tp->rx_opt.num_sacks) {
3768 tp->rx_opt.num_sacks = num_sacks;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003769 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks +
3770 tp->rx_opt.dsack,
3771 4 - tp->rx_opt.tstamp_ok);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003772 }
3773}
3774
3775/* This one checks to see if we can put data from the
3776 * out_of_order queue into the receive_queue.
3777 */
3778static void tcp_ofo_queue(struct sock *sk)
3779{
3780 struct tcp_sock *tp = tcp_sk(sk);
3781 __u32 dsack_high = tp->rcv_nxt;
3782 struct sk_buff *skb;
3783
3784 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
3785 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
3786 break;
3787
3788 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
3789 __u32 dsack = dsack_high;
3790 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
3791 dsack_high = TCP_SKB_CB(skb)->end_seq;
3792 tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack);
3793 }
3794
3795 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3796 SOCK_DEBUG(sk, "ofo packet was already received \n");
David S. Miller8728b832005-08-09 19:25:21 -07003797 __skb_unlink(skb, &tp->out_of_order_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003798 __kfree_skb(skb);
3799 continue;
3800 }
3801 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
3802 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3803 TCP_SKB_CB(skb)->end_seq);
3804
David S. Miller8728b832005-08-09 19:25:21 -07003805 __skb_unlink(skb, &tp->out_of_order_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003806 __skb_queue_tail(&sk->sk_receive_queue, skb);
3807 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07003808 if (tcp_hdr(skb)->fin)
3809 tcp_fin(skb, sk, tcp_hdr(skb));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003810 }
3811}
3812
3813static int tcp_prune_queue(struct sock *sk);
3814
3815static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
3816{
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07003817 struct tcphdr *th = tcp_hdr(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003818 struct tcp_sock *tp = tcp_sk(sk);
3819 int eaten = -1;
3820
3821 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
3822 goto drop;
3823
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003824 __skb_pull(skb, th->doff * 4);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003825
3826 TCP_ECN_accept_cwr(tp, skb);
3827
3828 if (tp->rx_opt.dsack) {
3829 tp->rx_opt.dsack = 0;
3830 tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003831 4 - tp->rx_opt.tstamp_ok);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003832 }
3833
3834 /* Queue data for delivery to the user.
3835 * Packets in sequence go to the receive queue.
3836 * Out of sequence packets to the out_of_order_queue.
3837 */
3838 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
3839 if (tcp_receive_window(tp) == 0)
3840 goto out_of_window;
3841
3842 /* Ok. In sequence. In window. */
3843 if (tp->ucopy.task == current &&
3844 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
3845 sock_owned_by_user(sk) && !tp->urg_data) {
3846 int chunk = min_t(unsigned int, skb->len,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003847 tp->ucopy.len);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003848
3849 __set_current_state(TASK_RUNNING);
3850
3851 local_bh_enable();
3852 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
3853 tp->ucopy.len -= chunk;
3854 tp->copied_seq += chunk;
3855 eaten = (chunk == skb->len && !th->fin);
3856 tcp_rcv_space_adjust(sk);
3857 }
3858 local_bh_disable();
3859 }
3860
3861 if (eaten <= 0) {
3862queue_and_out:
3863 if (eaten < 0 &&
3864 (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
Hideo Aoki3ab224b2007-12-31 00:11:19 -08003865 !sk_rmem_schedule(sk, skb->truesize))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003866 if (tcp_prune_queue(sk) < 0 ||
Hideo Aoki3ab224b2007-12-31 00:11:19 -08003867 !sk_rmem_schedule(sk, skb->truesize))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003868 goto drop;
3869 }
Hideo Aoki3ab224b2007-12-31 00:11:19 -08003870 skb_set_owner_r(skb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003871 __skb_queue_tail(&sk->sk_receive_queue, skb);
3872 }
3873 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
Stephen Hemminger2de979b2007-03-08 20:45:19 -08003874 if (skb->len)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003875 tcp_event_data_recv(sk, skb);
Stephen Hemminger2de979b2007-03-08 20:45:19 -08003876 if (th->fin)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003877 tcp_fin(skb, sk, th);
3878
David S. Millerb03efcf2005-07-08 14:57:23 -07003879 if (!skb_queue_empty(&tp->out_of_order_queue)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003880 tcp_ofo_queue(sk);
3881
3882 /* RFC2581. 4.2. SHOULD send immediate ACK, when
3883 * gap in queue is filled.
3884 */
David S. Millerb03efcf2005-07-08 14:57:23 -07003885 if (skb_queue_empty(&tp->out_of_order_queue))
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003886 inet_csk(sk)->icsk_ack.pingpong = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003887 }
3888
3889 if (tp->rx_opt.num_sacks)
3890 tcp_sack_remove(tp);
3891
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003892 tcp_fast_path_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003893
3894 if (eaten > 0)
3895 __kfree_skb(skb);
3896 else if (!sock_flag(sk, SOCK_DEAD))
3897 sk->sk_data_ready(sk, 0);
3898 return;
3899 }
3900
3901 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3902 /* A retransmit, 2nd most common case. Force an immediate ack. */
3903 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3904 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3905
3906out_of_window:
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003907 tcp_enter_quickack_mode(sk);
3908 inet_csk_schedule_ack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003909drop:
3910 __kfree_skb(skb);
3911 return;
3912 }
3913
3914 /* Out of window. F.e. zero window probe. */
3915 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
3916 goto out_of_window;
3917
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003918 tcp_enter_quickack_mode(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003919
3920 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3921 /* Partial packet, seq < rcv_next < end_seq */
3922 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
3923 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3924 TCP_SKB_CB(skb)->end_seq);
3925
3926 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09003927
Linus Torvalds1da177e2005-04-16 15:20:36 -07003928 /* If window is closed, drop tail of packet. But after
3929 * remembering D-SACK for its head made in previous line.
3930 */
3931 if (!tcp_receive_window(tp))
3932 goto out_of_window;
3933 goto queue_and_out;
3934 }
3935
3936 TCP_ECN_check_ce(tp, skb);
3937
3938 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
Hideo Aoki3ab224b2007-12-31 00:11:19 -08003939 !sk_rmem_schedule(sk, skb->truesize)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003940 if (tcp_prune_queue(sk) < 0 ||
Hideo Aoki3ab224b2007-12-31 00:11:19 -08003941 !sk_rmem_schedule(sk, skb->truesize))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003942 goto drop;
3943 }
3944
3945 /* Disable header prediction. */
3946 tp->pred_flags = 0;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003947 inet_csk_schedule_ack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003948
3949 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
3950 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3951
Hideo Aoki3ab224b2007-12-31 00:11:19 -08003952 skb_set_owner_r(skb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003953
3954 if (!skb_peek(&tp->out_of_order_queue)) {
3955 /* Initial out of order segment, build 1 SACK. */
Ilpo Järvinene60402d2007-08-09 15:14:46 +03003956 if (tcp_is_sack(tp)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003957 tp->rx_opt.num_sacks = 1;
3958 tp->rx_opt.dsack = 0;
3959 tp->rx_opt.eff_sacks = 1;
3960 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
3961 tp->selective_acks[0].end_seq =
3962 TCP_SKB_CB(skb)->end_seq;
3963 }
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003964 __skb_queue_head(&tp->out_of_order_queue, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003965 } else {
3966 struct sk_buff *skb1 = tp->out_of_order_queue.prev;
3967 u32 seq = TCP_SKB_CB(skb)->seq;
3968 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3969
3970 if (seq == TCP_SKB_CB(skb1)->end_seq) {
David S. Miller8728b832005-08-09 19:25:21 -07003971 __skb_append(skb1, skb, &tp->out_of_order_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003972
3973 if (!tp->rx_opt.num_sacks ||
3974 tp->selective_acks[0].end_seq != seq)
3975 goto add_sack;
3976
3977 /* Common case: data arrive in order after hole. */
3978 tp->selective_acks[0].end_seq = end_seq;
3979 return;
3980 }
3981
3982 /* Find place to insert this segment. */
3983 do {
3984 if (!after(TCP_SKB_CB(skb1)->seq, seq))
3985 break;
3986 } while ((skb1 = skb1->prev) !=
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003987 (struct sk_buff *)&tp->out_of_order_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003988
3989 /* Do skb overlap to previous one? */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003990 if (skb1 != (struct sk_buff *)&tp->out_of_order_queue &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07003991 before(seq, TCP_SKB_CB(skb1)->end_seq)) {
3992 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3993 /* All the bits are present. Drop. */
3994 __kfree_skb(skb);
3995 tcp_dsack_set(tp, seq, end_seq);
3996 goto add_sack;
3997 }
3998 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
3999 /* Partial overlap. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004000 tcp_dsack_set(tp, seq,
4001 TCP_SKB_CB(skb1)->end_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004002 } else {
4003 skb1 = skb1->prev;
4004 }
4005 }
4006 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004007
Linus Torvalds1da177e2005-04-16 15:20:36 -07004008 /* And clean segments covered by new one as whole. */
4009 while ((skb1 = skb->next) !=
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004010 (struct sk_buff *)&tp->out_of_order_queue &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07004011 after(end_seq, TCP_SKB_CB(skb1)->seq)) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004012 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4013 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq,
4014 end_seq);
4015 break;
4016 }
4017 __skb_unlink(skb1, &tp->out_of_order_queue);
4018 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq,
4019 TCP_SKB_CB(skb1)->end_seq);
4020 __kfree_skb(skb1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004021 }
4022
4023add_sack:
Ilpo Järvinene60402d2007-08-09 15:14:46 +03004024 if (tcp_is_sack(tp))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004025 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4026 }
4027}
4028
4029/* Collapse contiguous sequence of skbs head..tail with
4030 * sequence numbers start..end.
4031 * Segments with FIN/SYN are not collapsed (only because this
4032 * simplifies code)
4033 */
4034static void
David S. Miller8728b832005-08-09 19:25:21 -07004035tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4036 struct sk_buff *head, struct sk_buff *tail,
4037 u32 start, u32 end)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004038{
4039 struct sk_buff *skb;
4040
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08004041 /* First, check that queue is collapsible and find
Linus Torvalds1da177e2005-04-16 15:20:36 -07004042 * the point where collapsing can be useful. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004043 for (skb = head; skb != tail;) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004044 /* No new bits? It is possible on ofo queue. */
4045 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4046 struct sk_buff *next = skb->next;
David S. Miller8728b832005-08-09 19:25:21 -07004047 __skb_unlink(skb, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004048 __kfree_skb(skb);
4049 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
4050 skb = next;
4051 continue;
4052 }
4053
4054 /* The first skb to collapse is:
4055 * - not SYN/FIN and
4056 * - bloated or contains data before "start" or
4057 * overlaps to the next one.
4058 */
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07004059 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07004060 (tcp_win_from_space(skb->truesize) > skb->len ||
4061 before(TCP_SKB_CB(skb)->seq, start) ||
4062 (skb->next != tail &&
4063 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
4064 break;
4065
4066 /* Decided to skip this, advance start seq. */
4067 start = TCP_SKB_CB(skb)->end_seq;
4068 skb = skb->next;
4069 }
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07004070 if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004071 return;
4072
4073 while (before(start, end)) {
4074 struct sk_buff *nskb;
Chuck Leverc2636b42007-10-23 21:07:32 -07004075 unsigned int header = skb_headroom(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004076 int copy = SKB_MAX_ORDER(header, 0);
4077
4078 /* Too big header? This can happen with IPv6. */
4079 if (copy < 0)
4080 return;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004081 if (end - start < copy)
4082 copy = end - start;
4083 nskb = alloc_skb(copy + header, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004084 if (!nskb)
4085 return;
Arnaldo Carvalho de Meloc51957d2007-03-10 12:47:22 -03004086
Arnaldo Carvalho de Melo98e399f2007-03-19 15:33:04 -07004087 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
Arnaldo Carvalho de Melo9c702202007-04-25 18:04:18 -07004088 skb_set_network_header(nskb, (skb_network_header(skb) -
4089 skb->head));
4090 skb_set_transport_header(nskb, (skb_transport_header(skb) -
4091 skb->head));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004092 skb_reserve(nskb, header);
4093 memcpy(nskb->head, skb->head, header);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004094 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4095 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
David S. Miller8728b832005-08-09 19:25:21 -07004096 __skb_insert(nskb, skb->prev, skb, list);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08004097 skb_set_owner_r(nskb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004098
4099 /* Copy data, releasing collapsed skbs. */
4100 while (copy > 0) {
4101 int offset = start - TCP_SKB_CB(skb)->seq;
4102 int size = TCP_SKB_CB(skb)->end_seq - start;
4103
Kris Katterjohn09a62662006-01-08 22:24:28 -08004104 BUG_ON(offset < 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004105 if (size > 0) {
4106 size = min(copy, size);
4107 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4108 BUG();
4109 TCP_SKB_CB(nskb)->end_seq += size;
4110 copy -= size;
4111 start += size;
4112 }
4113 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4114 struct sk_buff *next = skb->next;
David S. Miller8728b832005-08-09 19:25:21 -07004115 __skb_unlink(skb, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004116 __kfree_skb(skb);
4117 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
4118 skb = next;
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07004119 if (skb == tail ||
4120 tcp_hdr(skb)->syn ||
4121 tcp_hdr(skb)->fin)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004122 return;
4123 }
4124 }
4125 }
4126}
4127
4128/* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4129 * and tcp_collapse() them until all the queue is collapsed.
4130 */
4131static void tcp_collapse_ofo_queue(struct sock *sk)
4132{
4133 struct tcp_sock *tp = tcp_sk(sk);
4134 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4135 struct sk_buff *head;
4136 u32 start, end;
4137
4138 if (skb == NULL)
4139 return;
4140
4141 start = TCP_SKB_CB(skb)->seq;
4142 end = TCP_SKB_CB(skb)->end_seq;
4143 head = skb;
4144
4145 for (;;) {
4146 skb = skb->next;
4147
4148 /* Segment is terminated when we see gap or when
4149 * we are at the end of all the queue. */
4150 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
4151 after(TCP_SKB_CB(skb)->seq, end) ||
4152 before(TCP_SKB_CB(skb)->end_seq, start)) {
David S. Miller8728b832005-08-09 19:25:21 -07004153 tcp_collapse(sk, &tp->out_of_order_queue,
4154 head, skb, start, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004155 head = skb;
4156 if (skb == (struct sk_buff *)&tp->out_of_order_queue)
4157 break;
4158 /* Start new segment */
4159 start = TCP_SKB_CB(skb)->seq;
4160 end = TCP_SKB_CB(skb)->end_seq;
4161 } else {
4162 if (before(TCP_SKB_CB(skb)->seq, start))
4163 start = TCP_SKB_CB(skb)->seq;
4164 if (after(TCP_SKB_CB(skb)->end_seq, end))
4165 end = TCP_SKB_CB(skb)->end_seq;
4166 }
4167 }
4168}
4169
4170/* Reduce allocated memory if we can, trying to get
4171 * the socket within its memory limits again.
4172 *
4173 * Return less than zero if we should start dropping frames
4174 * until the socket owning process reads some of the data
4175 * to stabilize the situation.
4176 */
4177static int tcp_prune_queue(struct sock *sk)
4178{
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004179 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004180
4181 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4182
4183 NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED);
4184
4185 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004186 tcp_clamp_window(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004187 else if (tcp_memory_pressure)
4188 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4189
4190 tcp_collapse_ofo_queue(sk);
David S. Miller8728b832005-08-09 19:25:21 -07004191 tcp_collapse(sk, &sk->sk_receive_queue,
4192 sk->sk_receive_queue.next,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004193 (struct sk_buff *)&sk->sk_receive_queue,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004194 tp->copied_seq, tp->rcv_nxt);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08004195 sk_mem_reclaim(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004196
4197 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4198 return 0;
4199
4200 /* Collapsing did not help, destructive actions follow.
4201 * This must not ever occur. */
4202
4203 /* First, purge the out_of_order queue. */
David S. Millerb03efcf2005-07-08 14:57:23 -07004204 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4205 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004206 __skb_queue_purge(&tp->out_of_order_queue);
4207
4208 /* Reset SACK state. A conforming SACK implementation will
4209 * do the same at a timeout based retransmit. When a connection
4210 * is in a sad state like this, we care only about integrity
4211 * of the connection not performance.
4212 */
Ilpo Järvinene60402d2007-08-09 15:14:46 +03004213 if (tcp_is_sack(tp))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004214 tcp_sack_reset(&tp->rx_opt);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08004215 sk_mem_reclaim(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004216 }
4217
4218 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4219 return 0;
4220
4221 /* If we are really being abused, tell the caller to silently
4222 * drop receive data on the floor. It will get retransmitted
4223 * and hopefully then we'll have sufficient space.
4224 */
4225 NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED);
4226
4227 /* Massive buffer overcommit. */
4228 tp->pred_flags = 0;
4229 return -1;
4230}
4231
Linus Torvalds1da177e2005-04-16 15:20:36 -07004232/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4233 * As additional protections, we do not touch cwnd in retransmission phases,
4234 * and if application hit its sndbuf limit recently.
4235 */
4236void tcp_cwnd_application_limited(struct sock *sk)
4237{
4238 struct tcp_sock *tp = tcp_sk(sk);
4239
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03004240 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07004241 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4242 /* Limited by application or receiver window. */
Ilpo Järvinend254bcd2006-08-04 16:57:42 -07004243 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4244 u32 win_used = max(tp->snd_cwnd_used, init_win);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004245 if (win_used < tp->snd_cwnd) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03004246 tp->snd_ssthresh = tcp_current_ssthresh(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004247 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4248 }
4249 tp->snd_cwnd_used = 0;
4250 }
4251 tp->snd_cwnd_stamp = tcp_time_stamp;
4252}
4253
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004254static int tcp_should_expand_sndbuf(struct sock *sk)
David S. Miller0d9901d2005-07-05 15:21:10 -07004255{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004256 struct tcp_sock *tp = tcp_sk(sk);
4257
David S. Miller0d9901d2005-07-05 15:21:10 -07004258 /* If the user specified a specific send buffer setting, do
4259 * not modify it.
4260 */
4261 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4262 return 0;
4263
4264 /* If we are under global TCP memory pressure, do not expand. */
4265 if (tcp_memory_pressure)
4266 return 0;
4267
4268 /* If we are under soft global TCP memory pressure, do not expand. */
4269 if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
4270 return 0;
4271
4272 /* If we filled the congestion window, do not expand. */
4273 if (tp->packets_out >= tp->snd_cwnd)
4274 return 0;
4275
4276 return 1;
4277}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004278
4279/* When incoming ACK allowed to free some skb from write_queue,
4280 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4281 * on the exit from tcp input handler.
4282 *
4283 * PROBLEM: sndbuf expansion does not work well with largesend.
4284 */
4285static void tcp_new_space(struct sock *sk)
4286{
4287 struct tcp_sock *tp = tcp_sk(sk);
4288
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004289 if (tcp_should_expand_sndbuf(sk)) {
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004290 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
Linus Torvalds1da177e2005-04-16 15:20:36 -07004291 MAX_TCP_HEADER + 16 + sizeof(struct sk_buff),
4292 demanded = max_t(unsigned int, tp->snd_cwnd,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004293 tp->reordering + 1);
4294 sndmem *= 2 * demanded;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004295 if (sndmem > sk->sk_sndbuf)
4296 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
4297 tp->snd_cwnd_stamp = tcp_time_stamp;
4298 }
4299
4300 sk->sk_write_space(sk);
4301}
4302
Stephen Hemminger40efc6f2006-01-03 16:03:49 -08004303static void tcp_check_space(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004304{
4305 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4306 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4307 if (sk->sk_socket &&
4308 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4309 tcp_new_space(sk);
4310 }
4311}
4312
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004313static inline void tcp_data_snd_check(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004314{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004315 tcp_push_pending_frames(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004316 tcp_check_space(sk);
4317}
4318
4319/*
4320 * Check if sending an ack is needed.
4321 */
4322static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4323{
4324 struct tcp_sock *tp = tcp_sk(sk);
4325
4326 /* More than one full frame received... */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004327 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss
Linus Torvalds1da177e2005-04-16 15:20:36 -07004328 /* ... and right edge of window advances far enough.
4329 * (tcp_recvmsg() will send ACK otherwise). Or...
4330 */
4331 && __tcp_select_window(sk) >= tp->rcv_wnd) ||
4332 /* We ACK each frame or... */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004333 tcp_in_quickack_mode(sk) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07004334 /* We have out of order data. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004335 (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004336 /* Then ack it now */
4337 tcp_send_ack(sk);
4338 } else {
4339 /* Else, send delayed ack. */
4340 tcp_send_delayed_ack(sk);
4341 }
4342}
4343
Stephen Hemminger40efc6f2006-01-03 16:03:49 -08004344static inline void tcp_ack_snd_check(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004345{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004346 if (!inet_csk_ack_scheduled(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004347 /* We sent a data segment already. */
4348 return;
4349 }
4350 __tcp_ack_snd_check(sk, 1);
4351}
4352
4353/*
4354 * This routine is only called when we have urgent data
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08004355 * signaled. Its the 'slow' part of tcp_urg. It could be
Linus Torvalds1da177e2005-04-16 15:20:36 -07004356 * moved inline now as tcp_urg is only called from one
4357 * place. We handle URGent data wrong. We have to - as
4358 * BSD still doesn't use the correction from RFC961.
4359 * For 1003.1g we should support a new option TCP_STDURG to permit
4360 * either form (or just set the sysctl tcp_stdurg).
4361 */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004362
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004363static void tcp_check_urg(struct sock *sk, struct tcphdr *th)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004364{
4365 struct tcp_sock *tp = tcp_sk(sk);
4366 u32 ptr = ntohs(th->urg_ptr);
4367
4368 if (ptr && !sysctl_tcp_stdurg)
4369 ptr--;
4370 ptr += ntohl(th->seq);
4371
4372 /* Ignore urgent data that we've already seen and read. */
4373 if (after(tp->copied_seq, ptr))
4374 return;
4375
4376 /* Do not replay urg ptr.
4377 *
4378 * NOTE: interesting situation not covered by specs.
4379 * Misbehaving sender may send urg ptr, pointing to segment,
4380 * which we already have in ofo queue. We are not able to fetch
4381 * such data and will stay in TCP_URG_NOTYET until will be eaten
4382 * by recvmsg(). Seems, we are not obliged to handle such wicked
4383 * situations. But it is worth to think about possibility of some
4384 * DoSes using some hypothetical application level deadlock.
4385 */
4386 if (before(ptr, tp->rcv_nxt))
4387 return;
4388
4389 /* Do we already have a newer (or duplicate) urgent pointer? */
4390 if (tp->urg_data && !after(ptr, tp->urg_seq))
4391 return;
4392
4393 /* Tell the world about our new urgent pointer. */
4394 sk_send_sigurg(sk);
4395
4396 /* We may be adding urgent data when the last byte read was
4397 * urgent. To do this requires some care. We cannot just ignore
4398 * tp->copied_seq since we would read the last urgent byte again
4399 * as data, nor can we alter copied_seq until this data arrives
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08004400 * or we break the semantics of SIOCATMARK (and thus sockatmark())
Linus Torvalds1da177e2005-04-16 15:20:36 -07004401 *
4402 * NOTE. Double Dutch. Rendering to plain English: author of comment
4403 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4404 * and expect that both A and B disappear from stream. This is _wrong_.
4405 * Though this happens in BSD with high probability, this is occasional.
4406 * Any application relying on this is buggy. Note also, that fix "works"
4407 * only in this artificial test. Insert some normal data between A and B and we will
4408 * decline of BSD again. Verdict: it is better to remove to trap
4409 * buggy users.
4410 */
4411 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004412 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004413 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4414 tp->copied_seq++;
4415 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
David S. Miller8728b832005-08-09 19:25:21 -07004416 __skb_unlink(skb, &sk->sk_receive_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004417 __kfree_skb(skb);
4418 }
4419 }
4420
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004421 tp->urg_data = TCP_URG_NOTYET;
4422 tp->urg_seq = ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004423
4424 /* Disable header prediction. */
4425 tp->pred_flags = 0;
4426}
4427
4428/* This is the 'fast' part of urgent handling. */
4429static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
4430{
4431 struct tcp_sock *tp = tcp_sk(sk);
4432
4433 /* Check if we get a new urgent pointer - normally not. */
4434 if (th->urg)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004435 tcp_check_urg(sk, th);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004436
4437 /* Do we wait for any urgent data? - normally not... */
4438 if (tp->urg_data == TCP_URG_NOTYET) {
4439 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4440 th->syn;
4441
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004442 /* Is the urgent pointer pointing into this packet? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004443 if (ptr < skb->len) {
4444 u8 tmp;
4445 if (skb_copy_bits(skb, ptr, &tmp, 1))
4446 BUG();
4447 tp->urg_data = TCP_URG_VALID | tmp;
4448 if (!sock_flag(sk, SOCK_DEAD))
4449 sk->sk_data_ready(sk, 0);
4450 }
4451 }
4452}
4453
4454static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4455{
4456 struct tcp_sock *tp = tcp_sk(sk);
4457 int chunk = skb->len - hlen;
4458 int err;
4459
4460 local_bh_enable();
Herbert Xu60476372007-04-09 11:59:39 -07004461 if (skb_csum_unnecessary(skb))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004462 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4463 else
4464 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4465 tp->ucopy.iov);
4466
4467 if (!err) {
4468 tp->ucopy.len -= chunk;
4469 tp->copied_seq += chunk;
4470 tcp_rcv_space_adjust(sk);
4471 }
4472
4473 local_bh_disable();
4474 return err;
4475}
4476
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004477static __sum16 __tcp_checksum_complete_user(struct sock *sk,
4478 struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004479{
Al Virob51655b2006-11-14 21:40:42 -08004480 __sum16 result;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004481
4482 if (sock_owned_by_user(sk)) {
4483 local_bh_enable();
4484 result = __tcp_checksum_complete(skb);
4485 local_bh_disable();
4486 } else {
4487 result = __tcp_checksum_complete(skb);
4488 }
4489 return result;
4490}
4491
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004492static inline int tcp_checksum_complete_user(struct sock *sk,
4493 struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004494{
Herbert Xu60476372007-04-09 11:59:39 -07004495 return !skb_csum_unnecessary(skb) &&
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004496 __tcp_checksum_complete_user(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004497}
4498
Chris Leech1a2449a2006-05-23 18:05:53 -07004499#ifdef CONFIG_NET_DMA
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004500static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb,
4501 int hlen)
Chris Leech1a2449a2006-05-23 18:05:53 -07004502{
4503 struct tcp_sock *tp = tcp_sk(sk);
4504 int chunk = skb->len - hlen;
4505 int dma_cookie;
4506 int copied_early = 0;
4507
4508 if (tp->ucopy.wakeup)
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004509 return 0;
Chris Leech1a2449a2006-05-23 18:05:53 -07004510
4511 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
4512 tp->ucopy.dma_chan = get_softnet_dma();
4513
Herbert Xu60476372007-04-09 11:59:39 -07004514 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
Chris Leech1a2449a2006-05-23 18:05:53 -07004515
4516 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004517 skb, hlen,
4518 tp->ucopy.iov, chunk,
4519 tp->ucopy.pinned_list);
Chris Leech1a2449a2006-05-23 18:05:53 -07004520
4521 if (dma_cookie < 0)
4522 goto out;
4523
4524 tp->ucopy.dma_cookie = dma_cookie;
4525 copied_early = 1;
4526
4527 tp->ucopy.len -= chunk;
4528 tp->copied_seq += chunk;
4529 tcp_rcv_space_adjust(sk);
4530
4531 if ((tp->ucopy.len == 0) ||
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07004532 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
Chris Leech1a2449a2006-05-23 18:05:53 -07004533 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
4534 tp->ucopy.wakeup = 1;
4535 sk->sk_data_ready(sk, 0);
4536 }
4537 } else if (chunk > 0) {
4538 tp->ucopy.wakeup = 1;
4539 sk->sk_data_ready(sk, 0);
4540 }
4541out:
4542 return copied_early;
4543}
4544#endif /* CONFIG_NET_DMA */
4545
Linus Torvalds1da177e2005-04-16 15:20:36 -07004546/*
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004547 * TCP receive function for the ESTABLISHED state.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004548 *
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004549 * It is split into a fast path and a slow path. The fast path is
Linus Torvalds1da177e2005-04-16 15:20:36 -07004550 * disabled when:
4551 * - A zero window was announced from us - zero window probing
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004552 * is only handled properly in the slow path.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004553 * - Out of order segments arrived.
4554 * - Urgent data is expected.
4555 * - There is no buffer space left
4556 * - Unexpected TCP flags/window values/header lengths are received
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004557 * (detected by checking the TCP header against pred_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004558 * - Data is sent in both directions. Fast path only supports pure senders
4559 * or pure receivers (this means either the sequence number or the ack
4560 * value must stay constant)
4561 * - Unexpected TCP option.
4562 *
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004563 * When these conditions are not satisfied it drops into a standard
Linus Torvalds1da177e2005-04-16 15:20:36 -07004564 * receive procedure patterned after RFC793 to handle all cases.
4565 * The first three cases are guaranteed by proper pred_flags setting,
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004566 * the rest is checked inline. Fast processing is turned on in
Linus Torvalds1da177e2005-04-16 15:20:36 -07004567 * tcp_data_queue when everything is OK.
4568 */
4569int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
4570 struct tcphdr *th, unsigned len)
4571{
4572 struct tcp_sock *tp = tcp_sk(sk);
4573
4574 /*
4575 * Header prediction.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004576 * The code loosely follows the one in the famous
Linus Torvalds1da177e2005-04-16 15:20:36 -07004577 * "30 instruction TCP receive" Van Jacobson mail.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004578 *
4579 * Van's trick is to deposit buffers into socket queue
Linus Torvalds1da177e2005-04-16 15:20:36 -07004580 * on a device interrupt, to call tcp_recv function
4581 * on the receive process context and checksum and copy
4582 * the buffer to user space. smart...
4583 *
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004584 * Our current scheme is not silly either but we take the
Linus Torvalds1da177e2005-04-16 15:20:36 -07004585 * extra cost of the net_bh soft interrupt processing...
4586 * We do checksum and copy also but from device to kernel.
4587 */
4588
4589 tp->rx_opt.saw_tstamp = 0;
4590
4591 /* pred_flags is 0xS?10 << 16 + snd_wnd
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08004592 * if header_prediction is to be made
Linus Torvalds1da177e2005-04-16 15:20:36 -07004593 * 'S' will always be tp->tcp_header_len >> 2
4594 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004595 * turn it off (when there are holes in the receive
Linus Torvalds1da177e2005-04-16 15:20:36 -07004596 * space for instance)
4597 * PSH flag is ignored.
4598 */
4599
4600 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004601 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004602 int tcp_header_len = tp->tcp_header_len;
4603
4604 /* Timestamp header prediction: tcp_header_len
4605 * is automatically equal to th->doff*4 due to pred_flags
4606 * match.
4607 */
4608
4609 /* Check timestamp */
4610 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
Al Viro4f3608b2006-09-27 18:35:09 -07004611 __be32 *ptr = (__be32 *)(th + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004612
4613 /* No? Slow path! */
Al Viro4f3608b2006-09-27 18:35:09 -07004614 if (*ptr != htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004615 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP))
4616 goto slow_path;
4617
4618 tp->rx_opt.saw_tstamp = 1;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004619 ++ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004620 tp->rx_opt.rcv_tsval = ntohl(*ptr);
4621 ++ptr;
4622 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
4623
4624 /* If PAWS failed, check it more carefully in slow path */
4625 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
4626 goto slow_path;
4627
4628 /* DO NOT update ts_recent here, if checksum fails
4629 * and timestamp was corrupted part, it will result
4630 * in a hung connection since we will drop all
4631 * future packets due to the PAWS test.
4632 */
4633 }
4634
4635 if (len <= tcp_header_len) {
4636 /* Bulk data transfer: sender */
4637 if (len == tcp_header_len) {
4638 /* Predicted packet is in window by definition.
4639 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4640 * Hence, check seq<=rcv_wup reduces to:
4641 */
4642 if (tcp_header_len ==
4643 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4644 tp->rcv_nxt == tp->rcv_wup)
4645 tcp_store_ts_recent(tp);
4646
Linus Torvalds1da177e2005-04-16 15:20:36 -07004647 /* We know that such packets are checksummed
4648 * on entry.
4649 */
4650 tcp_ack(sk, skb, 0);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09004651 __kfree_skb(skb);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004652 tcp_data_snd_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004653 return 0;
4654 } else { /* Header too small */
4655 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4656 goto discard;
4657 }
4658 } else {
4659 int eaten = 0;
Chris Leech1a2449a2006-05-23 18:05:53 -07004660 int copied_early = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004661
Chris Leech1a2449a2006-05-23 18:05:53 -07004662 if (tp->copied_seq == tp->rcv_nxt &&
4663 len - tcp_header_len <= tp->ucopy.len) {
4664#ifdef CONFIG_NET_DMA
4665 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
4666 copied_early = 1;
4667 eaten = 1;
4668 }
4669#endif
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004670 if (tp->ucopy.task == current &&
4671 sock_owned_by_user(sk) && !copied_early) {
Chris Leech1a2449a2006-05-23 18:05:53 -07004672 __set_current_state(TASK_RUNNING);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004673
Chris Leech1a2449a2006-05-23 18:05:53 -07004674 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
4675 eaten = 1;
4676 }
4677 if (eaten) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004678 /* Predicted packet is in window by definition.
4679 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4680 * Hence, check seq<=rcv_wup reduces to:
4681 */
4682 if (tcp_header_len ==
4683 (sizeof(struct tcphdr) +
4684 TCPOLEN_TSTAMP_ALIGNED) &&
4685 tp->rcv_nxt == tp->rcv_wup)
4686 tcp_store_ts_recent(tp);
4687
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004688 tcp_rcv_rtt_measure_ts(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004689
4690 __skb_pull(skb, tcp_header_len);
4691 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4692 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004693 }
Chris Leech1a2449a2006-05-23 18:05:53 -07004694 if (copied_early)
4695 tcp_cleanup_rbuf(sk, skb->len);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004696 }
4697 if (!eaten) {
4698 if (tcp_checksum_complete_user(sk, skb))
4699 goto csum_error;
4700
4701 /* Predicted packet is in window by definition.
4702 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4703 * Hence, check seq<=rcv_wup reduces to:
4704 */
4705 if (tcp_header_len ==
4706 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4707 tp->rcv_nxt == tp->rcv_wup)
4708 tcp_store_ts_recent(tp);
4709
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004710 tcp_rcv_rtt_measure_ts(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004711
4712 if ((int)skb->truesize > sk->sk_forward_alloc)
4713 goto step5;
4714
4715 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS);
4716
4717 /* Bulk data transfer: receiver */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004718 __skb_pull(skb, tcp_header_len);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004719 __skb_queue_tail(&sk->sk_receive_queue, skb);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08004720 skb_set_owner_r(skb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004721 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4722 }
4723
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004724 tcp_event_data_recv(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004725
4726 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
4727 /* Well, only one small jumplet in fast path... */
4728 tcp_ack(sk, skb, FLAG_DATA);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004729 tcp_data_snd_check(sk);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004730 if (!inet_csk_ack_scheduled(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004731 goto no_ack;
4732 }
4733
David S. Miller31432412005-05-23 12:03:06 -07004734 __tcp_ack_snd_check(sk, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004735no_ack:
Chris Leech1a2449a2006-05-23 18:05:53 -07004736#ifdef CONFIG_NET_DMA
4737 if (copied_early)
4738 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
4739 else
4740#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004741 if (eaten)
4742 __kfree_skb(skb);
4743 else
4744 sk->sk_data_ready(sk, 0);
4745 return 0;
4746 }
4747 }
4748
4749slow_path:
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004750 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004751 goto csum_error;
4752
4753 /*
4754 * RFC1323: H1. Apply PAWS check first.
4755 */
4756 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004757 tcp_paws_discard(sk, skb)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004758 if (!th->rst) {
4759 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4760 tcp_send_dupack(sk, skb);
4761 goto discard;
4762 }
4763 /* Resets are accepted even if PAWS failed.
4764
4765 ts_recent update must be made after we are sure
4766 that the packet is in window.
4767 */
4768 }
4769
4770 /*
4771 * Standard slow path.
4772 */
4773
4774 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4775 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4776 * (RST) segments are validated by checking their SEQ-fields."
4777 * And page 69: "If an incoming segment is not acceptable,
4778 * an acknowledgment should be sent in reply (unless the RST bit
4779 * is set, if so drop the segment and return)".
4780 */
4781 if (!th->rst)
4782 tcp_send_dupack(sk, skb);
4783 goto discard;
4784 }
4785
Stephen Hemminger2de979b2007-03-08 20:45:19 -08004786 if (th->rst) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004787 tcp_reset(sk);
4788 goto discard;
4789 }
4790
4791 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4792
4793 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4794 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4795 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4796 tcp_reset(sk);
4797 return 1;
4798 }
4799
4800step5:
Stephen Hemminger2de979b2007-03-08 20:45:19 -08004801 if (th->ack)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004802 tcp_ack(sk, skb, FLAG_SLOWPATH);
4803
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004804 tcp_rcv_rtt_measure_ts(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004805
4806 /* Process urgent data. */
4807 tcp_urg(sk, skb, th);
4808
4809 /* step 7: process the segment text */
4810 tcp_data_queue(sk, skb);
4811
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07004812 tcp_data_snd_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004813 tcp_ack_snd_check(sk);
4814 return 0;
4815
4816csum_error:
4817 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4818
4819discard:
4820 __kfree_skb(skb);
4821 return 0;
4822}
4823
4824static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
4825 struct tcphdr *th, unsigned len)
4826{
4827 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08004828 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004829 int saved_clamp = tp->rx_opt.mss_clamp;
4830
4831 tcp_parse_options(skb, &tp->rx_opt, 0);
4832
4833 if (th->ack) {
4834 /* rfc793:
4835 * "If the state is SYN-SENT then
4836 * first check the ACK bit
4837 * If the ACK bit is set
4838 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4839 * a reset (unless the RST bit is set, if so drop
4840 * the segment and return)"
4841 *
4842 * We do not send data with SYN, so that RFC-correct
4843 * test reduces to:
4844 */
4845 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
4846 goto reset_and_undo;
4847
4848 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4849 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
4850 tcp_time_stamp)) {
4851 NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED);
4852 goto reset_and_undo;
4853 }
4854
4855 /* Now ACK is acceptable.
4856 *
4857 * "If the RST bit is set
4858 * If the ACK was acceptable then signal the user "error:
4859 * connection reset", drop the segment, enter CLOSED state,
4860 * delete TCB, and return."
4861 */
4862
4863 if (th->rst) {
4864 tcp_reset(sk);
4865 goto discard;
4866 }
4867
4868 /* rfc793:
4869 * "fifth, if neither of the SYN or RST bits is set then
4870 * drop the segment and return."
4871 *
4872 * See note below!
4873 * --ANK(990513)
4874 */
4875 if (!th->syn)
4876 goto discard_and_undo;
4877
4878 /* rfc793:
4879 * "If the SYN bit is on ...
4880 * are acceptable then ...
4881 * (our SYN has been ACKed), change the connection
4882 * state to ESTABLISHED..."
4883 */
4884
4885 TCP_ECN_rcv_synack(tp, th);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004886
4887 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
4888 tcp_ack(sk, skb, FLAG_SLOWPATH);
4889
4890 /* Ok.. it's good. Set up sequence numbers and
4891 * move to established.
4892 */
4893 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4894 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4895
4896 /* RFC1323: The window in SYN & SYN/ACK segments is
4897 * never scaled.
4898 */
4899 tp->snd_wnd = ntohs(th->window);
4900 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
4901
4902 if (!tp->rx_opt.wscale_ok) {
4903 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
4904 tp->window_clamp = min(tp->window_clamp, 65535U);
4905 }
4906
4907 if (tp->rx_opt.saw_tstamp) {
4908 tp->rx_opt.tstamp_ok = 1;
4909 tp->tcp_header_len =
4910 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4911 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
4912 tcp_store_ts_recent(tp);
4913 } else {
4914 tp->tcp_header_len = sizeof(struct tcphdr);
4915 }
4916
Ilpo Järvinene60402d2007-08-09 15:14:46 +03004917 if (tcp_is_sack(tp) && sysctl_tcp_fack)
4918 tcp_enable_fack(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004919
John Heffner5d424d52006-03-20 17:53:41 -08004920 tcp_mtup_init(sk);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08004921 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004922 tcp_initialize_rcv_mss(sk);
4923
4924 /* Remember, tcp_poll() does not lock socket!
4925 * Change state from SYN-SENT only after copied_seq
4926 * is initialized. */
4927 tp->copied_seq = tp->rcv_nxt;
Ralf Baechlee16aa202006-12-07 00:11:33 -08004928 smp_mb();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004929 tcp_set_state(sk, TCP_ESTABLISHED);
4930
Venkat Yekkirala6b877692006-11-08 17:04:09 -06004931 security_inet_conn_established(sk, skb);
4932
Linus Torvalds1da177e2005-04-16 15:20:36 -07004933 /* Make sure socket is routed, for correct metrics. */
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -08004934 icsk->icsk_af_ops->rebuild_header(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004935
4936 tcp_init_metrics(sk);
4937
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03004938 tcp_init_congestion_control(sk);
Stephen Hemminger317a76f2005-06-23 12:19:55 -07004939
Linus Torvalds1da177e2005-04-16 15:20:36 -07004940 /* Prevent spurious tcp_cwnd_restart() on first data
4941 * packet.
4942 */
4943 tp->lsndtime = tcp_time_stamp;
4944
4945 tcp_init_buffer_space(sk);
4946
4947 if (sock_flag(sk, SOCK_KEEPOPEN))
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004948 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004949
4950 if (!tp->rx_opt.snd_wscale)
4951 __tcp_fast_path_on(tp, tp->snd_wnd);
4952 else
4953 tp->pred_flags = 0;
4954
4955 if (!sock_flag(sk, SOCK_DEAD)) {
4956 sk->sk_state_change(sk);
Pavel Emelyanov8d8ad9d2007-11-26 20:10:50 +08004957 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004958 }
4959
Arnaldo Carvalho de Melo295f7322005-08-09 20:11:56 -07004960 if (sk->sk_write_pending ||
4961 icsk->icsk_accept_queue.rskq_defer_accept ||
4962 icsk->icsk_ack.pingpong) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004963 /* Save one ACK. Data will be ready after
4964 * several ticks, if write_pending is set.
4965 *
4966 * It may be deleted, but with this feature tcpdumps
4967 * look so _wonderfully_ clever, that I was not able
4968 * to stand against the temptation 8) --ANK
4969 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004970 inet_csk_schedule_ack(sk);
Arnaldo Carvalho de Melo295f7322005-08-09 20:11:56 -07004971 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
4972 icsk->icsk_ack.ato = TCP_ATO_MIN;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004973 tcp_incr_quickack(sk);
4974 tcp_enter_quickack_mode(sk);
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07004975 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
4976 TCP_DELACK_MAX, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004977
4978discard:
4979 __kfree_skb(skb);
4980 return 0;
4981 } else {
4982 tcp_send_ack(sk);
4983 }
4984 return -1;
4985 }
4986
4987 /* No ACK in the segment */
4988
4989 if (th->rst) {
4990 /* rfc793:
4991 * "If the RST bit is set
4992 *
4993 * Otherwise (no ACK) drop the segment and return."
4994 */
4995
4996 goto discard_and_undo;
4997 }
4998
4999 /* PAWS check. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005000 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5001 tcp_paws_check(&tp->rx_opt, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005002 goto discard_and_undo;
5003
5004 if (th->syn) {
5005 /* We see SYN without ACK. It is attempt of
5006 * simultaneous connect with crossed SYNs.
5007 * Particularly, it can be connect to self.
5008 */
5009 tcp_set_state(sk, TCP_SYN_RECV);
5010
5011 if (tp->rx_opt.saw_tstamp) {
5012 tp->rx_opt.tstamp_ok = 1;
5013 tcp_store_ts_recent(tp);
5014 tp->tcp_header_len =
5015 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5016 } else {
5017 tp->tcp_header_len = sizeof(struct tcphdr);
5018 }
5019
5020 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5021 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5022
5023 /* RFC1323: The window in SYN & SYN/ACK segments is
5024 * never scaled.
5025 */
5026 tp->snd_wnd = ntohs(th->window);
5027 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5028 tp->max_window = tp->snd_wnd;
5029
5030 TCP_ECN_rcv_syn(tp, th);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005031
John Heffner5d424d52006-03-20 17:53:41 -08005032 tcp_mtup_init(sk);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08005033 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005034 tcp_initialize_rcv_mss(sk);
5035
Linus Torvalds1da177e2005-04-16 15:20:36 -07005036 tcp_send_synack(sk);
5037#if 0
5038 /* Note, we could accept data and URG from this segment.
5039 * There are no obstacles to make this.
5040 *
5041 * However, if we ignore data in ACKless segments sometimes,
5042 * we have no reasons to accept it sometimes.
5043 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5044 * is not flawless. So, discard packet for sanity.
5045 * Uncomment this return to process the data.
5046 */
5047 return -1;
5048#else
5049 goto discard;
5050#endif
5051 }
5052 /* "fifth, if neither of the SYN or RST bits is set then
5053 * drop the segment and return."
5054 */
5055
5056discard_and_undo:
5057 tcp_clear_options(&tp->rx_opt);
5058 tp->rx_opt.mss_clamp = saved_clamp;
5059 goto discard;
5060
5061reset_and_undo:
5062 tcp_clear_options(&tp->rx_opt);
5063 tp->rx_opt.mss_clamp = saved_clamp;
5064 return 1;
5065}
5066
Linus Torvalds1da177e2005-04-16 15:20:36 -07005067/*
5068 * This function implements the receiving procedure of RFC 793 for
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005069 * all states except ESTABLISHED and TIME_WAIT.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005070 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5071 * address independent.
5072 */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005073
Linus Torvalds1da177e2005-04-16 15:20:36 -07005074int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5075 struct tcphdr *th, unsigned len)
5076{
5077 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -08005078 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005079 int queued = 0;
5080
5081 tp->rx_opt.saw_tstamp = 0;
5082
5083 switch (sk->sk_state) {
5084 case TCP_CLOSE:
5085 goto discard;
5086
5087 case TCP_LISTEN:
Stephen Hemminger2de979b2007-03-08 20:45:19 -08005088 if (th->ack)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005089 return 1;
5090
Stephen Hemminger2de979b2007-03-08 20:45:19 -08005091 if (th->rst)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005092 goto discard;
5093
Stephen Hemminger2de979b2007-03-08 20:45:19 -08005094 if (th->syn) {
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -08005095 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005096 return 1;
5097
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005098 /* Now we have several options: In theory there is
5099 * nothing else in the frame. KA9Q has an option to
Linus Torvalds1da177e2005-04-16 15:20:36 -07005100 * send data with the syn, BSD accepts data with the
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005101 * syn up to the [to be] advertised window and
5102 * Solaris 2.1 gives you a protocol error. For now
5103 * we just ignore it, that fits the spec precisely
Linus Torvalds1da177e2005-04-16 15:20:36 -07005104 * and avoids incompatibilities. It would be nice in
5105 * future to drop through and process the data.
5106 *
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005107 * Now that TTCP is starting to be used we ought to
Linus Torvalds1da177e2005-04-16 15:20:36 -07005108 * queue this data.
5109 * But, this leaves one open to an easy denial of
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005110 * service attack, and SYN cookies can't defend
Linus Torvalds1da177e2005-04-16 15:20:36 -07005111 * against this problem. So, we drop the data
Masayuki Nakagawafb7e2392007-01-23 20:15:06 -08005112 * in the interest of security over speed unless
5113 * it's still in use.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005114 */
Masayuki Nakagawafb7e2392007-01-23 20:15:06 -08005115 kfree_skb(skb);
5116 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005117 }
5118 goto discard;
5119
5120 case TCP_SYN_SENT:
Linus Torvalds1da177e2005-04-16 15:20:36 -07005121 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
5122 if (queued >= 0)
5123 return queued;
5124
5125 /* Do step6 onward by hand. */
5126 tcp_urg(sk, skb, th);
5127 __kfree_skb(skb);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005128 tcp_data_snd_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005129 return 0;
5130 }
5131
5132 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005133 tcp_paws_discard(sk, skb)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005134 if (!th->rst) {
5135 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
5136 tcp_send_dupack(sk, skb);
5137 goto discard;
5138 }
5139 /* Reset is accepted even if it did not pass PAWS. */
5140 }
5141
5142 /* step 1: check sequence number */
5143 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5144 if (!th->rst)
5145 tcp_send_dupack(sk, skb);
5146 goto discard;
5147 }
5148
5149 /* step 2: check RST bit */
Stephen Hemminger2de979b2007-03-08 20:45:19 -08005150 if (th->rst) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005151 tcp_reset(sk);
5152 goto discard;
5153 }
5154
5155 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
5156
5157 /* step 3: check security and precedence [ignored] */
5158
5159 /* step 4:
5160 *
5161 * Check for a SYN in window.
5162 */
5163 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5164 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
5165 tcp_reset(sk);
5166 return 1;
5167 }
5168
5169 /* step 5: check the ACK field */
5170 if (th->ack) {
5171 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
5172
Stephen Hemminger2de979b2007-03-08 20:45:19 -08005173 switch (sk->sk_state) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005174 case TCP_SYN_RECV:
5175 if (acceptable) {
5176 tp->copied_seq = tp->rcv_nxt;
Ralf Baechlee16aa202006-12-07 00:11:33 -08005177 smp_mb();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005178 tcp_set_state(sk, TCP_ESTABLISHED);
5179 sk->sk_state_change(sk);
5180
5181 /* Note, that this wakeup is only for marginal
5182 * crossed SYN case. Passively open sockets
5183 * are not waked up, because sk->sk_sleep ==
5184 * NULL and sk->sk_socket == NULL.
5185 */
Pavel Emelyanov8d8ad9d2007-11-26 20:10:50 +08005186 if (sk->sk_socket)
5187 sk_wake_async(sk,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005188 SOCK_WAKE_IO, POLL_OUT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005189
5190 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5191 tp->snd_wnd = ntohs(th->window) <<
5192 tp->rx_opt.snd_wscale;
5193 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq,
5194 TCP_SKB_CB(skb)->seq);
5195
5196 /* tcp_ack considers this ACK as duplicate
5197 * and does not calculate rtt.
5198 * Fix it at least with timestamps.
5199 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005200 if (tp->rx_opt.saw_tstamp &&
5201 tp->rx_opt.rcv_tsecr && !tp->srtt)
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08005202 tcp_ack_saw_tstamp(sk, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005203
5204 if (tp->rx_opt.tstamp_ok)
5205 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5206
5207 /* Make sure socket is routed, for
5208 * correct metrics.
5209 */
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -08005210 icsk->icsk_af_ops->rebuild_header(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005211
5212 tcp_init_metrics(sk);
5213
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03005214 tcp_init_congestion_control(sk);
Stephen Hemminger317a76f2005-06-23 12:19:55 -07005215
Linus Torvalds1da177e2005-04-16 15:20:36 -07005216 /* Prevent spurious tcp_cwnd_restart() on
5217 * first data packet.
5218 */
5219 tp->lsndtime = tcp_time_stamp;
5220
John Heffner5d424d52006-03-20 17:53:41 -08005221 tcp_mtup_init(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005222 tcp_initialize_rcv_mss(sk);
5223 tcp_init_buffer_space(sk);
5224 tcp_fast_path_on(tp);
5225 } else {
5226 return 1;
5227 }
5228 break;
5229
5230 case TCP_FIN_WAIT1:
5231 if (tp->snd_una == tp->write_seq) {
5232 tcp_set_state(sk, TCP_FIN_WAIT2);
5233 sk->sk_shutdown |= SEND_SHUTDOWN;
5234 dst_confirm(sk->sk_dst_cache);
5235
5236 if (!sock_flag(sk, SOCK_DEAD))
5237 /* Wake up lingering close() */
5238 sk->sk_state_change(sk);
5239 else {
5240 int tmo;
5241
5242 if (tp->linger2 < 0 ||
5243 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5244 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5245 tcp_done(sk);
5246 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
5247 return 1;
5248 }
5249
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005250 tmo = tcp_fin_time(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005251 if (tmo > TCP_TIMEWAIT_LEN) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005252 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005253 } else if (th->fin || sock_owned_by_user(sk)) {
5254 /* Bad case. We could lose such FIN otherwise.
5255 * It is not a big problem, but it looks confusing
5256 * and not so rare event. We still can lose it now,
5257 * if it spins in bh_lock_sock(), but it is really
5258 * marginal case.
5259 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005260 inet_csk_reset_keepalive_timer(sk, tmo);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005261 } else {
5262 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5263 goto discard;
5264 }
5265 }
5266 }
5267 break;
5268
5269 case TCP_CLOSING:
5270 if (tp->snd_una == tp->write_seq) {
5271 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5272 goto discard;
5273 }
5274 break;
5275
5276 case TCP_LAST_ACK:
5277 if (tp->snd_una == tp->write_seq) {
5278 tcp_update_metrics(sk);
5279 tcp_done(sk);
5280 goto discard;
5281 }
5282 break;
5283 }
5284 } else
5285 goto discard;
5286
5287 /* step 6: check the URG bit */
5288 tcp_urg(sk, skb, th);
5289
5290 /* step 7: process the segment text */
5291 switch (sk->sk_state) {
5292 case TCP_CLOSE_WAIT:
5293 case TCP_CLOSING:
5294 case TCP_LAST_ACK:
5295 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
5296 break;
5297 case TCP_FIN_WAIT1:
5298 case TCP_FIN_WAIT2:
5299 /* RFC 793 says to queue data in these states,
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005300 * RFC 1122 says we MUST send a reset.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005301 * BSD 4.4 also does reset.
5302 */
5303 if (sk->sk_shutdown & RCV_SHUTDOWN) {
5304 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5305 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5306 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
5307 tcp_reset(sk);
5308 return 1;
5309 }
5310 }
5311 /* Fall through */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005312 case TCP_ESTABLISHED:
Linus Torvalds1da177e2005-04-16 15:20:36 -07005313 tcp_data_queue(sk, skb);
5314 queued = 1;
5315 break;
5316 }
5317
5318 /* tcp_data could move socket to TIME-WAIT */
5319 if (sk->sk_state != TCP_CLOSE) {
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005320 tcp_data_snd_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005321 tcp_ack_snd_check(sk);
5322 }
5323
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005324 if (!queued) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005325discard:
5326 __kfree_skb(skb);
5327 }
5328 return 0;
5329}
5330
5331EXPORT_SYMBOL(sysctl_tcp_ecn);
5332EXPORT_SYMBOL(sysctl_tcp_reordering);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005333EXPORT_SYMBOL(tcp_parse_options);
5334EXPORT_SYMBOL(tcp_rcv_established);
5335EXPORT_SYMBOL(tcp_rcv_state_process);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -08005336EXPORT_SYMBOL(tcp_initialize_rcv_mss);