Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1 | /* |
| 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 Juhl | 02c30a8 | 2005-05-05 16:16:16 -0700 | [diff] [blame] | 10 | * Authors: Ross Biro |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 11 | * 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 Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 45 | * Andrey Savochkin: Fix RTT measurements in the presence of |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 46 | * 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 Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 53 | * Andi Kleen: Add tcp_measure_rcv_mss to make |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 54 | * connections with MSS<min(MTU,ann. MSS) |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 55 | * work without delayed acks. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 56 | * 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 Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 64 | */ |
| 65 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 66 | #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 Leech | 1a2449a | 2006-05-23 18:05:53 -0700 | [diff] [blame] | 73 | #include <net/netdma.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 74 | |
Brian Haley | ab32ea5 | 2006-09-22 14:15:41 -0700 | [diff] [blame] | 75 | int sysctl_tcp_timestamps __read_mostly = 1; |
| 76 | int sysctl_tcp_window_scaling __read_mostly = 1; |
| 77 | int sysctl_tcp_sack __read_mostly = 1; |
| 78 | int sysctl_tcp_fack __read_mostly = 1; |
| 79 | int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH; |
| 80 | int sysctl_tcp_ecn __read_mostly; |
| 81 | int sysctl_tcp_dsack __read_mostly = 1; |
| 82 | int sysctl_tcp_app_win __read_mostly = 31; |
| 83 | int sysctl_tcp_adv_win_scale __read_mostly = 2; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 84 | |
Brian Haley | ab32ea5 | 2006-09-22 14:15:41 -0700 | [diff] [blame] | 85 | int sysctl_tcp_stdurg __read_mostly; |
| 86 | int sysctl_tcp_rfc1337 __read_mostly; |
| 87 | int sysctl_tcp_max_orphans __read_mostly = NR_FILE; |
Ilpo Järvinen | c96fd3d | 2007-09-20 11:36:37 -0700 | [diff] [blame] | 88 | int sysctl_tcp_frto __read_mostly = 2; |
Ilpo Järvinen | 3cfe3ba | 2007-02-27 10:09:49 -0800 | [diff] [blame] | 89 | int sysctl_tcp_frto_response __read_mostly; |
Brian Haley | ab32ea5 | 2006-09-22 14:15:41 -0700 | [diff] [blame] | 90 | int sysctl_tcp_nometrics_save __read_mostly; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 91 | |
Brian Haley | ab32ea5 | 2006-09-22 14:15:41 -0700 | [diff] [blame] | 92 | int sysctl_tcp_moderate_rcvbuf __read_mostly = 1; |
| 93 | int sysctl_tcp_abc __read_mostly; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 94 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 95 | #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 Järvinen | 4dc2665 | 2007-02-21 23:16:11 -0800 | [diff] [blame] | 104 | #define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */ |
Ilpo Järvinen | 2e60529 | 2007-08-02 19:46:58 -0700 | [diff] [blame] | 105 | #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */ |
Ilpo Järvinen | 49ff4bb | 2007-08-02 19:47:59 -0700 | [diff] [blame] | 106 | #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained DSACK info */ |
Ilpo Järvinen | 009a2e3 | 2007-09-20 11:34:38 -0700 | [diff] [blame] | 107 | #define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 108 | |
| 109 | #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED) |
| 110 | #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED) |
| 111 | #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE) |
| 112 | #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED) |
Ilpo Järvinen | 2e60529 | 2007-08-02 19:46:58 -0700 | [diff] [blame] | 113 | #define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 114 | |
Ilpo Järvinen | 4dc2665 | 2007-02-21 23:16:11 -0800 | [diff] [blame] | 115 | #define IsSackFrto() (sysctl_tcp_frto == 0x2) |
| 116 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 117 | #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH) |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 118 | #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 119 | |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 120 | /* Adapt the MSS value used to make delayed ack decision to the |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 121 | * real world. |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 122 | */ |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 123 | static void tcp_measure_rcv_mss(struct sock *sk, |
| 124 | const struct sk_buff *skb) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 125 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 126 | struct inet_connection_sock *icsk = inet_csk(sk); |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 127 | const unsigned int lss = icsk->icsk_ack.last_seg_size; |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 128 | unsigned int len; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 129 | |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 130 | icsk->icsk_ack.last_seg_size = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 131 | |
| 132 | /* skb->len may jitter because of SACKs, even if peer |
| 133 | * sends good full-sized frames. |
| 134 | */ |
Herbert Xu | ff9b5e0 | 2006-08-31 15:11:02 -0700 | [diff] [blame] | 135 | len = skb_shinfo(skb)->gso_size ?: skb->len; |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 136 | if (len >= icsk->icsk_ack.rcv_mss) { |
| 137 | icsk->icsk_ack.rcv_mss = len; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 138 | } 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 Melo | 9c70220 | 2007-04-25 18:04:18 -0700 | [diff] [blame] | 144 | len += skb->data - skb_transport_header(skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 145 | 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 Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 152 | !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 153 | /* 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 Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 157 | len -= tcp_sk(sk)->tcp_header_len; |
| 158 | icsk->icsk_ack.last_seg_size = len; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 159 | if (len == lss) { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 160 | icsk->icsk_ack.rcv_mss = len; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 161 | return; |
| 162 | } |
| 163 | } |
Alexey Kuznetsov | 1ef9696 | 2006-09-19 12:52:50 -0700 | [diff] [blame] | 164 | if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) |
| 165 | icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2; |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 166 | icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 167 | } |
| 168 | } |
| 169 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 170 | static void tcp_incr_quickack(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 171 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 172 | struct inet_connection_sock *icsk = inet_csk(sk); |
| 173 | unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 174 | |
| 175 | if (quickacks==0) |
| 176 | quickacks=2; |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 177 | if (quickacks > icsk->icsk_ack.quick) |
| 178 | icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 179 | } |
| 180 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 181 | void tcp_enter_quickack_mode(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 182 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 183 | 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 Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 187 | } |
| 188 | |
| 189 | /* Send ACKs quickly, if "quick" count is not exhausted |
| 190 | * and the session is not interactive. |
| 191 | */ |
| 192 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 193 | static inline int tcp_in_quickack_mode(const struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 194 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 195 | const struct inet_connection_sock *icsk = inet_csk(sk); |
| 196 | return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 197 | } |
| 198 | |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 199 | static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp) |
| 200 | { |
| 201 | if (tp->ecn_flags&TCP_ECN_OK) |
| 202 | tp->ecn_flags |= TCP_ECN_QUEUE_CWR; |
| 203 | } |
| 204 | |
| 205 | static 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 | |
| 211 | static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp) |
| 212 | { |
| 213 | tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; |
| 214 | } |
| 215 | |
| 216 | static inline void TCP_ECN_check_ce(struct tcp_sock *tp, struct sk_buff *skb) |
| 217 | { |
| 218 | if (tp->ecn_flags&TCP_ECN_OK) { |
| 219 | 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 | |
| 229 | static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, struct tcphdr *th) |
| 230 | { |
| 231 | if ((tp->ecn_flags&TCP_ECN_OK) && (!th->ece || th->cwr)) |
| 232 | tp->ecn_flags &= ~TCP_ECN_OK; |
| 233 | } |
| 234 | |
| 235 | static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, struct tcphdr *th) |
| 236 | { |
| 237 | if ((tp->ecn_flags&TCP_ECN_OK) && (!th->ece || !th->cwr)) |
| 238 | tp->ecn_flags &= ~TCP_ECN_OK; |
| 239 | } |
| 240 | |
| 241 | static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock *tp, struct tcphdr *th) |
| 242 | { |
| 243 | if (th->ece && !th->syn && (tp->ecn_flags&TCP_ECN_OK)) |
| 244 | return 1; |
| 245 | return 0; |
| 246 | } |
| 247 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 248 | /* Buffer size and advertised window tuning. |
| 249 | * |
| 250 | * 1. Tuning sk->sk_sndbuf, when connection enters established state. |
| 251 | */ |
| 252 | |
| 253 | static 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 Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 283 | * window and then starts to feed us spaghetti. But it should work |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 284 | * in common situations. Otherwise, we have to rely on queue collapsing. |
| 285 | */ |
| 286 | |
| 287 | /* Slow part of check#2. */ |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 288 | static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 289 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 290 | struct tcp_sock *tp = tcp_sk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 291 | /* Optimize this! */ |
| 292 | int truesize = tcp_win_from_space(skb->truesize)/2; |
John Heffner | 326f36e | 2005-11-10 17:11:48 -0800 | [diff] [blame] | 293 | int window = tcp_win_from_space(sysctl_tcp_rmem[2])/2; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 294 | |
| 295 | while (tp->rcv_ssthresh <= window) { |
| 296 | if (truesize <= skb->len) |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 297 | return 2 * inet_csk(sk)->icsk_ack.rcv_mss; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 298 | |
| 299 | truesize >>= 1; |
| 300 | window >>= 1; |
| 301 | } |
| 302 | return 0; |
| 303 | } |
| 304 | |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 305 | static void tcp_grow_window(struct sock *sk, |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 306 | struct sk_buff *skb) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 307 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 308 | struct tcp_sock *tp = tcp_sk(sk); |
| 309 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 310 | /* Check #1 */ |
| 311 | if (tp->rcv_ssthresh < tp->window_clamp && |
| 312 | (int)tp->rcv_ssthresh < tcp_space(sk) && |
| 313 | !tcp_memory_pressure) { |
| 314 | int incr; |
| 315 | |
| 316 | /* Check #2. Increase window, if skb with such overhead |
| 317 | * will fit to rcvbuf in future. |
| 318 | */ |
| 319 | if (tcp_win_from_space(skb->truesize) <= skb->len) |
| 320 | incr = 2*tp->advmss; |
| 321 | else |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 322 | incr = __tcp_grow_window(sk, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 323 | |
| 324 | if (incr) { |
| 325 | tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, tp->window_clamp); |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 326 | inet_csk(sk)->icsk_ack.quick |= 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 327 | } |
| 328 | } |
| 329 | } |
| 330 | |
| 331 | /* 3. Tuning rcvbuf, when connection enters established state. */ |
| 332 | |
| 333 | static 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 Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 339 | * will fit to window and corresponding skbs will fit to our rcvbuf. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 340 | * (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 Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 348 | /* 4. Try to fixup all. It is made immediately after connection enters |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 349 | * established state. |
| 350 | */ |
| 351 | static 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 Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 384 | /* 5. Recalculate window clamp after socket hit its memory bounds. */ |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 385 | static void tcp_clamp_window(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 386 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 387 | struct tcp_sock *tp = tcp_sk(sk); |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 388 | struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 389 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 390 | icsk->icsk_ack.quick = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 391 | |
John Heffner | 326f36e | 2005-11-10 17:11:48 -0800 | [diff] [blame] | 392 | 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 Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 398 | } |
John Heffner | 326f36e | 2005-11-10 17:11:48 -0800 | [diff] [blame] | 399 | if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 400 | tp->rcv_ssthresh = min(tp->window_clamp, 2U*tp->advmss); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 401 | } |
| 402 | |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 403 | |
| 404 | /* Initialize RCV_MSS value. |
| 405 | * RCV_MSS is an our guess about MSS used by the peer. |
| 406 | * We haven't any direct information about the MSS. |
| 407 | * It's better to underestimate the RCV_MSS rather than overestimate. |
| 408 | * Overestimations make us ACKing less frequently than needed. |
| 409 | * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss(). |
| 410 | */ |
| 411 | void tcp_initialize_rcv_mss(struct sock *sk) |
| 412 | { |
| 413 | struct tcp_sock *tp = tcp_sk(sk); |
| 414 | unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache); |
| 415 | |
| 416 | hint = min(hint, tp->rcv_wnd/2); |
| 417 | hint = min(hint, TCP_MIN_RCVMSS); |
| 418 | hint = max(hint, TCP_MIN_MSS); |
| 419 | |
| 420 | inet_csk(sk)->icsk_ack.rcv_mss = hint; |
| 421 | } |
| 422 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 423 | /* Receiver "autotuning" code. |
| 424 | * |
| 425 | * The algorithm for RTT estimation w/o timestamps is based on |
| 426 | * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL. |
| 427 | * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps> |
| 428 | * |
| 429 | * More detail on this code can be found at |
| 430 | * <http://www.psc.edu/~jheffner/senior_thesis.ps>, |
| 431 | * though this reference is out of date. A new paper |
| 432 | * is pending. |
| 433 | */ |
| 434 | static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep) |
| 435 | { |
| 436 | u32 new_sample = tp->rcv_rtt_est.rtt; |
| 437 | long m = sample; |
| 438 | |
| 439 | if (m == 0) |
| 440 | m = 1; |
| 441 | |
| 442 | if (new_sample != 0) { |
| 443 | /* If we sample in larger samples in the non-timestamp |
| 444 | * case, we could grossly overestimate the RTT especially |
| 445 | * with chatty applications or bulk transfer apps which |
| 446 | * are stalled on filesystem I/O. |
| 447 | * |
| 448 | * Also, since we are only going for a minimum in the |
Stephen Hemminger | 31f3426 | 2005-11-15 15:17:10 -0800 | [diff] [blame] | 449 | * non-timestamp case, we do not smooth things out |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 450 | * else with timestamps disabled convergence takes too |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 451 | * long. |
| 452 | */ |
| 453 | if (!win_dep) { |
| 454 | m -= (new_sample >> 3); |
| 455 | new_sample += m; |
| 456 | } else if (m < new_sample) |
| 457 | new_sample = m << 3; |
| 458 | } else { |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 459 | /* No previous measure. */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 460 | new_sample = m << 3; |
| 461 | } |
| 462 | |
| 463 | if (tp->rcv_rtt_est.rtt != new_sample) |
| 464 | tp->rcv_rtt_est.rtt = new_sample; |
| 465 | } |
| 466 | |
| 467 | static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp) |
| 468 | { |
| 469 | if (tp->rcv_rtt_est.time == 0) |
| 470 | goto new_measure; |
| 471 | if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq)) |
| 472 | return; |
| 473 | tcp_rcv_rtt_update(tp, |
| 474 | jiffies - tp->rcv_rtt_est.time, |
| 475 | 1); |
| 476 | |
| 477 | new_measure: |
| 478 | tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; |
| 479 | tp->rcv_rtt_est.time = tcp_time_stamp; |
| 480 | } |
| 481 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 482 | static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, const struct sk_buff *skb) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 483 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 484 | struct tcp_sock *tp = tcp_sk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 485 | if (tp->rx_opt.rcv_tsecr && |
| 486 | (TCP_SKB_CB(skb)->end_seq - |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 487 | TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 488 | tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0); |
| 489 | } |
| 490 | |
| 491 | /* |
| 492 | * This function should be called every time data is copied to user space. |
| 493 | * It calculates the appropriate TCP receive buffer space. |
| 494 | */ |
| 495 | void tcp_rcv_space_adjust(struct sock *sk) |
| 496 | { |
| 497 | struct tcp_sock *tp = tcp_sk(sk); |
| 498 | int time; |
| 499 | int space; |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 500 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 501 | if (tp->rcvq_space.time == 0) |
| 502 | goto new_measure; |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 503 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 504 | time = tcp_time_stamp - tp->rcvq_space.time; |
| 505 | if (time < (tp->rcv_rtt_est.rtt >> 3) || |
| 506 | tp->rcv_rtt_est.rtt == 0) |
| 507 | return; |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 508 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 509 | space = 2 * (tp->copied_seq - tp->rcvq_space.seq); |
| 510 | |
| 511 | space = max(tp->rcvq_space.space, space); |
| 512 | |
| 513 | if (tp->rcvq_space.space != space) { |
| 514 | int rcvmem; |
| 515 | |
| 516 | tp->rcvq_space.space = space; |
| 517 | |
John Heffner | 6fcf941 | 2006-02-09 17:06:57 -0800 | [diff] [blame] | 518 | if (sysctl_tcp_moderate_rcvbuf && |
| 519 | !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 520 | int new_clamp = space; |
| 521 | |
| 522 | /* Receive space grows, normalize in order to |
| 523 | * take into account packet headers and sk_buff |
| 524 | * structure overhead. |
| 525 | */ |
| 526 | space /= tp->advmss; |
| 527 | if (!space) |
| 528 | space = 1; |
| 529 | rcvmem = (tp->advmss + MAX_TCP_HEADER + |
| 530 | 16 + sizeof(struct sk_buff)); |
| 531 | while (tcp_win_from_space(rcvmem) < tp->advmss) |
| 532 | rcvmem += 128; |
| 533 | space *= rcvmem; |
| 534 | space = min(space, sysctl_tcp_rmem[2]); |
| 535 | if (space > sk->sk_rcvbuf) { |
| 536 | sk->sk_rcvbuf = space; |
| 537 | |
| 538 | /* Make the window clamp follow along. */ |
| 539 | tp->window_clamp = new_clamp; |
| 540 | } |
| 541 | } |
| 542 | } |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 543 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 544 | new_measure: |
| 545 | tp->rcvq_space.seq = tp->copied_seq; |
| 546 | tp->rcvq_space.time = tcp_time_stamp; |
| 547 | } |
| 548 | |
| 549 | /* There is something which you must keep in mind when you analyze the |
| 550 | * behavior of the tp->ato delayed ack timeout interval. When a |
| 551 | * connection starts up, we want to ack as quickly as possible. The |
| 552 | * problem is that "good" TCP's do slow start at the beginning of data |
| 553 | * transmission. The means that until we send the first few ACK's the |
| 554 | * sender will sit on his end and only queue most of his data, because |
| 555 | * he can only send snd_cwnd unacked packets at any given time. For |
| 556 | * each ACK we send, he increments snd_cwnd and transmits more of his |
| 557 | * queue. -DaveM |
| 558 | */ |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 559 | static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 560 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 561 | struct tcp_sock *tp = tcp_sk(sk); |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 562 | struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 563 | u32 now; |
| 564 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 565 | inet_csk_schedule_ack(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 566 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 567 | tcp_measure_rcv_mss(sk, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 568 | |
| 569 | tcp_rcv_rtt_measure(tp); |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 570 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 571 | now = tcp_time_stamp; |
| 572 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 573 | if (!icsk->icsk_ack.ato) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 574 | /* The _first_ data packet received, initialize |
| 575 | * delayed ACK engine. |
| 576 | */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 577 | tcp_incr_quickack(sk); |
| 578 | icsk->icsk_ack.ato = TCP_ATO_MIN; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 579 | } else { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 580 | int m = now - icsk->icsk_ack.lrcvtime; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 581 | |
| 582 | if (m <= TCP_ATO_MIN/2) { |
| 583 | /* The fastest case is the first. */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 584 | icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2; |
| 585 | } else if (m < icsk->icsk_ack.ato) { |
| 586 | icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m; |
| 587 | if (icsk->icsk_ack.ato > icsk->icsk_rto) |
| 588 | icsk->icsk_ack.ato = icsk->icsk_rto; |
| 589 | } else if (m > icsk->icsk_rto) { |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 590 | /* Too long gap. Apparently sender failed to |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 591 | * restart window, so that we send ACKs quickly. |
| 592 | */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 593 | tcp_incr_quickack(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 594 | sk_stream_mem_reclaim(sk); |
| 595 | } |
| 596 | } |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 597 | icsk->icsk_ack.lrcvtime = now; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 598 | |
| 599 | TCP_ECN_check_ce(tp, skb); |
| 600 | |
| 601 | if (skb->len >= 128) |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 602 | tcp_grow_window(sk, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 603 | } |
| 604 | |
David S. Miller | 05bb1fa | 2007-08-30 22:10:28 -0700 | [diff] [blame] | 605 | static u32 tcp_rto_min(struct sock *sk) |
| 606 | { |
| 607 | struct dst_entry *dst = __sk_dst_get(sk); |
| 608 | u32 rto_min = TCP_RTO_MIN; |
| 609 | |
David S. Miller | 5c127c5 | 2007-08-31 14:39:44 -0700 | [diff] [blame] | 610 | if (dst && dst_metric_locked(dst, RTAX_RTO_MIN)) |
David S. Miller | 05bb1fa | 2007-08-30 22:10:28 -0700 | [diff] [blame] | 611 | rto_min = dst->metrics[RTAX_RTO_MIN-1]; |
| 612 | return rto_min; |
| 613 | } |
| 614 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 615 | /* Called to compute a smoothed rtt estimate. The data fed to this |
| 616 | * routine either comes from timestamps, or from segments that were |
| 617 | * known _not_ to have been retransmitted [see Karn/Partridge |
| 618 | * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88 |
| 619 | * piece by Van Jacobson. |
| 620 | * NOTE: the next three routines used to be one big routine. |
| 621 | * To save cycles in the RFC 1323 implementation it was better to break |
| 622 | * it up into three procedures. -- erics |
| 623 | */ |
Stephen Hemminger | 2d2abba | 2005-11-10 16:56:12 -0800 | [diff] [blame] | 624 | static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 625 | { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 626 | struct tcp_sock *tp = tcp_sk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 627 | long m = mrtt; /* RTT */ |
| 628 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 629 | /* The following amusing code comes from Jacobson's |
| 630 | * article in SIGCOMM '88. Note that rtt and mdev |
| 631 | * are scaled versions of rtt and mean deviation. |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 632 | * This is designed to be as fast as possible |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 633 | * m stands for "measurement". |
| 634 | * |
| 635 | * On a 1990 paper the rto value is changed to: |
| 636 | * RTO = rtt + 4 * mdev |
| 637 | * |
| 638 | * Funny. This algorithm seems to be very broken. |
| 639 | * These formulae increase RTO, when it should be decreased, increase |
Stephen Hemminger | 31f3426 | 2005-11-15 15:17:10 -0800 | [diff] [blame] | 640 | * too slowly, when it should be increased quickly, decrease too quickly |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 641 | * etc. I guess in BSD RTO takes ONE value, so that it is absolutely |
| 642 | * does not matter how to _calculate_ it. Seems, it was trap |
| 643 | * that VJ failed to avoid. 8) |
| 644 | */ |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 645 | if (m == 0) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 646 | m = 1; |
| 647 | if (tp->srtt != 0) { |
| 648 | m -= (tp->srtt >> 3); /* m is now error in rtt est */ |
| 649 | tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */ |
| 650 | if (m < 0) { |
| 651 | m = -m; /* m is now abs(error) */ |
| 652 | m -= (tp->mdev >> 2); /* similar update on mdev */ |
| 653 | /* This is similar to one of Eifel findings. |
| 654 | * Eifel blocks mdev updates when rtt decreases. |
| 655 | * This solution is a bit different: we use finer gain |
| 656 | * for mdev in this case (alpha*beta). |
| 657 | * Like Eifel it also prevents growth of rto, |
| 658 | * but also it limits too fast rto decreases, |
| 659 | * happening in pure Eifel. |
| 660 | */ |
| 661 | if (m > 0) |
| 662 | m >>= 3; |
| 663 | } else { |
| 664 | m -= (tp->mdev >> 2); /* similar update on mdev */ |
| 665 | } |
| 666 | tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */ |
| 667 | if (tp->mdev > tp->mdev_max) { |
| 668 | tp->mdev_max = tp->mdev; |
| 669 | if (tp->mdev_max > tp->rttvar) |
| 670 | tp->rttvar = tp->mdev_max; |
| 671 | } |
| 672 | if (after(tp->snd_una, tp->rtt_seq)) { |
| 673 | if (tp->mdev_max < tp->rttvar) |
| 674 | tp->rttvar -= (tp->rttvar-tp->mdev_max)>>2; |
| 675 | tp->rtt_seq = tp->snd_nxt; |
David S. Miller | 05bb1fa | 2007-08-30 22:10:28 -0700 | [diff] [blame] | 676 | tp->mdev_max = tcp_rto_min(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 677 | } |
| 678 | } else { |
| 679 | /* no previous measure. */ |
| 680 | tp->srtt = m<<3; /* take the measured time to be rtt */ |
| 681 | tp->mdev = m<<1; /* make sure rto = 3*rtt */ |
David S. Miller | 05bb1fa | 2007-08-30 22:10:28 -0700 | [diff] [blame] | 682 | tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 683 | tp->rtt_seq = tp->snd_nxt; |
| 684 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 685 | } |
| 686 | |
| 687 | /* Calculate rto without backoff. This is the second half of Van Jacobson's |
| 688 | * routine referred to above. |
| 689 | */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 690 | static inline void tcp_set_rto(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 691 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 692 | const struct tcp_sock *tp = tcp_sk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 693 | /* Old crap is replaced with new one. 8) |
| 694 | * |
| 695 | * More seriously: |
| 696 | * 1. If rtt variance happened to be less 50msec, it is hallucination. |
| 697 | * It cannot be less due to utterly erratic ACK generation made |
| 698 | * at least by solaris and freebsd. "Erratic ACKs" has _nothing_ |
| 699 | * to do with delayed acks, because at cwnd>2 true delack timeout |
| 700 | * is invisible. Actually, Linux-2.4 also generates erratic |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 701 | * ACKs in some circumstances. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 702 | */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 703 | inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 704 | |
| 705 | /* 2. Fixups made earlier cannot be right. |
| 706 | * If we do not estimate RTO correctly without them, |
| 707 | * all the algo is pure shit and should be replaced |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 708 | * with correct one. It is exactly, which we pretend to do. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 709 | */ |
| 710 | } |
| 711 | |
| 712 | /* NOTE: clamping at TCP_RTO_MIN is not required, current algo |
| 713 | * guarantees that rto is higher. |
| 714 | */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 715 | static inline void tcp_bound_rto(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 716 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 717 | if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX) |
| 718 | inet_csk(sk)->icsk_rto = TCP_RTO_MAX; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 719 | } |
| 720 | |
| 721 | /* Save metrics learned by this TCP session. |
| 722 | This function is called only, when TCP finishes successfully |
| 723 | i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE. |
| 724 | */ |
| 725 | void tcp_update_metrics(struct sock *sk) |
| 726 | { |
| 727 | struct tcp_sock *tp = tcp_sk(sk); |
| 728 | struct dst_entry *dst = __sk_dst_get(sk); |
| 729 | |
| 730 | if (sysctl_tcp_nometrics_save) |
| 731 | return; |
| 732 | |
| 733 | dst_confirm(dst); |
| 734 | |
| 735 | if (dst && (dst->flags&DST_HOST)) { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 736 | const struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 737 | int m; |
| 738 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 739 | if (icsk->icsk_backoff || !tp->srtt) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 740 | /* This session failed to estimate rtt. Why? |
| 741 | * Probably, no packets returned in time. |
| 742 | * Reset our results. |
| 743 | */ |
| 744 | if (!(dst_metric_locked(dst, RTAX_RTT))) |
| 745 | dst->metrics[RTAX_RTT-1] = 0; |
| 746 | return; |
| 747 | } |
| 748 | |
| 749 | m = dst_metric(dst, RTAX_RTT) - tp->srtt; |
| 750 | |
| 751 | /* If newly calculated rtt larger than stored one, |
| 752 | * store new one. Otherwise, use EWMA. Remember, |
| 753 | * rtt overestimation is always better than underestimation. |
| 754 | */ |
| 755 | if (!(dst_metric_locked(dst, RTAX_RTT))) { |
| 756 | if (m <= 0) |
| 757 | dst->metrics[RTAX_RTT-1] = tp->srtt; |
| 758 | else |
| 759 | dst->metrics[RTAX_RTT-1] -= (m>>3); |
| 760 | } |
| 761 | |
| 762 | if (!(dst_metric_locked(dst, RTAX_RTTVAR))) { |
| 763 | if (m < 0) |
| 764 | m = -m; |
| 765 | |
| 766 | /* Scale deviation to rttvar fixed point */ |
| 767 | m >>= 1; |
| 768 | if (m < tp->mdev) |
| 769 | m = tp->mdev; |
| 770 | |
| 771 | if (m >= dst_metric(dst, RTAX_RTTVAR)) |
| 772 | dst->metrics[RTAX_RTTVAR-1] = m; |
| 773 | else |
| 774 | dst->metrics[RTAX_RTTVAR-1] -= |
| 775 | (dst->metrics[RTAX_RTTVAR-1] - m)>>2; |
| 776 | } |
| 777 | |
| 778 | if (tp->snd_ssthresh >= 0xFFFF) { |
| 779 | /* Slow start still did not finish. */ |
| 780 | if (dst_metric(dst, RTAX_SSTHRESH) && |
| 781 | !dst_metric_locked(dst, RTAX_SSTHRESH) && |
| 782 | (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH)) |
| 783 | dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1; |
| 784 | if (!dst_metric_locked(dst, RTAX_CWND) && |
| 785 | tp->snd_cwnd > dst_metric(dst, RTAX_CWND)) |
| 786 | dst->metrics[RTAX_CWND-1] = tp->snd_cwnd; |
| 787 | } else if (tp->snd_cwnd > tp->snd_ssthresh && |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 788 | icsk->icsk_ca_state == TCP_CA_Open) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 789 | /* Cong. avoidance phase, cwnd is reliable. */ |
| 790 | if (!dst_metric_locked(dst, RTAX_SSTHRESH)) |
| 791 | dst->metrics[RTAX_SSTHRESH-1] = |
| 792 | max(tp->snd_cwnd >> 1, tp->snd_ssthresh); |
| 793 | if (!dst_metric_locked(dst, RTAX_CWND)) |
| 794 | dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1; |
| 795 | } else { |
| 796 | /* Else slow start did not finish, cwnd is non-sense, |
| 797 | ssthresh may be also invalid. |
| 798 | */ |
| 799 | if (!dst_metric_locked(dst, RTAX_CWND)) |
| 800 | dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1; |
| 801 | if (dst->metrics[RTAX_SSTHRESH-1] && |
| 802 | !dst_metric_locked(dst, RTAX_SSTHRESH) && |
| 803 | tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1]) |
| 804 | dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh; |
| 805 | } |
| 806 | |
| 807 | if (!dst_metric_locked(dst, RTAX_REORDERING)) { |
| 808 | if (dst->metrics[RTAX_REORDERING-1] < tp->reordering && |
| 809 | tp->reordering != sysctl_tcp_reordering) |
| 810 | dst->metrics[RTAX_REORDERING-1] = tp->reordering; |
| 811 | } |
| 812 | } |
| 813 | } |
| 814 | |
David S. Miller | 2672287 | 2007-08-24 22:21:50 -0700 | [diff] [blame] | 815 | /* Numbers are taken from RFC3390. |
| 816 | * |
| 817 | * John Heffner states: |
| 818 | * |
| 819 | * The RFC specifies a window of no more than 4380 bytes |
| 820 | * unless 2*MSS > 4380. Reading the pseudocode in the RFC |
| 821 | * is a bit misleading because they use a clamp at 4380 bytes |
| 822 | * rather than use a multiplier in the relevant range. |
| 823 | */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 824 | __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst) |
| 825 | { |
| 826 | __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0); |
| 827 | |
| 828 | if (!cwnd) { |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 829 | if (tp->mss_cache > 1460) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 830 | cwnd = 2; |
| 831 | else |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 832 | cwnd = (tp->mss_cache > 1095) ? 3 : 4; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 833 | } |
| 834 | return min_t(__u32, cwnd, tp->snd_cwnd_clamp); |
| 835 | } |
| 836 | |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 837 | /* Set slow start threshold and cwnd not falling to slow start */ |
Ilpo Järvinen | 3cfe3ba | 2007-02-27 10:09:49 -0800 | [diff] [blame] | 838 | void tcp_enter_cwr(struct sock *sk, const int set_ssthresh) |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 839 | { |
| 840 | struct tcp_sock *tp = tcp_sk(sk); |
Ilpo Järvinen | 3cfe3ba | 2007-02-27 10:09:49 -0800 | [diff] [blame] | 841 | const struct inet_connection_sock *icsk = inet_csk(sk); |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 842 | |
| 843 | tp->prior_ssthresh = 0; |
| 844 | tp->bytes_acked = 0; |
Ilpo Järvinen | e01f9d7 | 2007-03-02 13:27:25 -0800 | [diff] [blame] | 845 | if (icsk->icsk_ca_state < TCP_CA_CWR) { |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 846 | tp->undo_marker = 0; |
Ilpo Järvinen | 3cfe3ba | 2007-02-27 10:09:49 -0800 | [diff] [blame] | 847 | if (set_ssthresh) |
| 848 | tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 849 | tp->snd_cwnd = min(tp->snd_cwnd, |
| 850 | tcp_packets_in_flight(tp) + 1U); |
| 851 | tp->snd_cwnd_cnt = 0; |
| 852 | tp->high_seq = tp->snd_nxt; |
| 853 | tp->snd_cwnd_stamp = tcp_time_stamp; |
| 854 | TCP_ECN_queue_cwr(tp); |
| 855 | |
| 856 | tcp_set_ca_state(sk, TCP_CA_CWR); |
| 857 | } |
| 858 | } |
| 859 | |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 860 | /* |
| 861 | * Packet counting of FACK is based on in-order assumptions, therefore TCP |
| 862 | * disables it when reordering is detected |
| 863 | */ |
| 864 | static void tcp_disable_fack(struct tcp_sock *tp) |
| 865 | { |
| 866 | tp->rx_opt.sack_ok &= ~2; |
| 867 | } |
| 868 | |
| 869 | /* Take a notice that peer is sending DSACKs */ |
| 870 | static void tcp_dsack_seen(struct tcp_sock *tp) |
| 871 | { |
| 872 | tp->rx_opt.sack_ok |= 4; |
| 873 | } |
| 874 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 875 | /* Initialize metrics on socket. */ |
| 876 | |
| 877 | static 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 Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 896 | tcp_disable_fack(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 897 | 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 Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 912 | * is sent until it is ACKed. In normal circumstances sending small |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 913 | * 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); |
| 926 | tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN); |
| 927 | } |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 928 | 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 Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 931 | goto reset; |
| 932 | tp->snd_cwnd = tcp_init_cwnd(tp, dst); |
| 933 | tp->snd_cwnd_stamp = tcp_time_stamp; |
| 934 | return; |
| 935 | |
| 936 | reset: |
| 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 Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 944 | inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 945 | } |
| 946 | } |
| 947 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 948 | static void tcp_update_reordering(struct sock *sk, const int metric, |
| 949 | const int ts) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 950 | { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 951 | struct tcp_sock *tp = tcp_sk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 952 | 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 Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 958 | else if (tcp_is_reno(tp)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 959 | NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER); |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 960 | else if (tcp_is_fack(tp)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 961 | 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 Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 966 | tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 967 | tp->reordering, |
| 968 | tp->fackets_out, |
| 969 | tp->sacked_out, |
| 970 | tp->undo_marker ? tp->undo_retrans : 0); |
| 971 | #endif |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 972 | tcp_disable_fack(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 973 | } |
| 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 Järvinen | 5b3c988 | 2007-08-24 22:54:44 -0700 | [diff] [blame] | 1023 | * |
| 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 |
| 1030 | * it means that the receiver is rather inconsistent with itself (reports |
| 1031 | * SACK reneging when it should advance SND.UNA). |
| 1032 | * |
| 1033 | * Implements also blockage to start_seq wrap-around. Problem lies in the |
| 1034 | * fact that though start_seq (s) is before end_seq (i.e., not reversed), |
| 1035 | * there's no guarantee that it will be before snd_nxt (n). The problem |
| 1036 | * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt |
| 1037 | * wrap (s_w): |
| 1038 | * |
| 1039 | * <- outs wnd -> <- wrapzone -> |
| 1040 | * u e n u_w e_w s n_w |
| 1041 | * | | | | | | | |
| 1042 | * |<------------+------+----- TCP seqno space --------------+---------->| |
| 1043 | * ...-- <2^31 ->| |<--------... |
| 1044 | * ...---- >2^31 ------>| |<--------... |
| 1045 | * |
| 1046 | * Current code wouldn't be vulnerable but it's better still to discard such |
| 1047 | * crazy SACK blocks. Doing this check for start_seq alone closes somewhat |
| 1048 | * similar case (end_seq after snd_nxt wrap) as earlier reversed check in |
| 1049 | * snd_nxt wrap -> snd_una region will then become "well defined", i.e., |
| 1050 | * equal to the ideal case (infinite seqno space without wrap caused issues). |
| 1051 | * |
| 1052 | * With D-SACK the lower bound is extended to cover sequence space below |
| 1053 | * SND.UNA down to undo_marker, which is the last point of interest. Yet |
| 1054 | * again, DSACK block must not to go across snd_una (for the same reason as |
| 1055 | * for the normal SACK blocks, explained above). But there all simplicity |
| 1056 | * ends, TCP might receive valid D-SACKs below that. As long as they reside |
| 1057 | * fully below undo_marker they do not affect behavior in anyway and can |
| 1058 | * therefore be safely ignored. In rare cases (which are more or less |
| 1059 | * theoretical ones), the D-SACK will nicely cross that boundary due to skb |
| 1060 | * fragmentation and packet reordering past skb's retransmission. To consider |
| 1061 | * them correctly, the acceptable range must be extended even more though |
| 1062 | * the exact amount is rather hard to quantify. However, tp->max_window can |
| 1063 | * be used as an exaggerated estimate. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1064 | */ |
Ilpo Järvinen | 5b3c988 | 2007-08-24 22:54:44 -0700 | [diff] [blame] | 1065 | static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack, |
| 1066 | u32 start_seq, u32 end_seq) |
| 1067 | { |
| 1068 | /* Too far in future, or reversed (interpretation is ambiguous) */ |
| 1069 | if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq)) |
| 1070 | return 0; |
| 1071 | |
| 1072 | /* Nasty start_seq wrap-around check (see comments above) */ |
| 1073 | if (!before(start_seq, tp->snd_nxt)) |
| 1074 | return 0; |
| 1075 | |
| 1076 | /* In outstanding window? ...This is valid exit for DSACKs too. |
| 1077 | * start_seq == snd_una is non-sensical (see comments above) |
| 1078 | */ |
| 1079 | if (after(start_seq, tp->snd_una)) |
| 1080 | return 1; |
| 1081 | |
| 1082 | if (!is_dsack || !tp->undo_marker) |
| 1083 | return 0; |
| 1084 | |
| 1085 | /* ...Then it's D-SACK, and must reside below snd_una completely */ |
| 1086 | if (!after(end_seq, tp->snd_una)) |
| 1087 | return 0; |
| 1088 | |
| 1089 | if (!before(start_seq, tp->undo_marker)) |
| 1090 | return 1; |
| 1091 | |
| 1092 | /* Too old */ |
| 1093 | if (!after(end_seq, tp->undo_marker)) |
| 1094 | return 0; |
| 1095 | |
| 1096 | /* Undo_marker boundary crossing (overestimates a lot). Known already: |
| 1097 | * start_seq < undo_marker and end_seq >= undo_marker. |
| 1098 | */ |
| 1099 | return !before(start_seq, end_seq - tp->max_window); |
| 1100 | } |
| 1101 | |
| 1102 | |
David S. Miller | d06e021 | 2007-06-18 22:43:06 -0700 | [diff] [blame] | 1103 | static int tcp_check_dsack(struct tcp_sock *tp, struct sk_buff *ack_skb, |
| 1104 | struct tcp_sack_block_wire *sp, int num_sacks, |
| 1105 | u32 prior_snd_una) |
| 1106 | { |
| 1107 | u32 start_seq_0 = ntohl(get_unaligned(&sp[0].start_seq)); |
| 1108 | u32 end_seq_0 = ntohl(get_unaligned(&sp[0].end_seq)); |
| 1109 | int dup_sack = 0; |
| 1110 | |
| 1111 | if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) { |
| 1112 | dup_sack = 1; |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 1113 | tcp_dsack_seen(tp); |
David S. Miller | d06e021 | 2007-06-18 22:43:06 -0700 | [diff] [blame] | 1114 | NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV); |
| 1115 | } else if (num_sacks > 1) { |
| 1116 | u32 end_seq_1 = ntohl(get_unaligned(&sp[1].end_seq)); |
| 1117 | u32 start_seq_1 = ntohl(get_unaligned(&sp[1].start_seq)); |
| 1118 | |
| 1119 | if (!after(end_seq_0, end_seq_1) && |
| 1120 | !before(start_seq_0, start_seq_1)) { |
| 1121 | dup_sack = 1; |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 1122 | tcp_dsack_seen(tp); |
David S. Miller | d06e021 | 2007-06-18 22:43:06 -0700 | [diff] [blame] | 1123 | NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV); |
| 1124 | } |
| 1125 | } |
| 1126 | |
| 1127 | /* D-SACK for already forgotten data... Do dumb counting. */ |
| 1128 | if (dup_sack && |
| 1129 | !after(end_seq_0, prior_snd_una) && |
| 1130 | after(end_seq_0, tp->undo_marker)) |
| 1131 | tp->undo_retrans--; |
| 1132 | |
| 1133 | return dup_sack; |
| 1134 | } |
| 1135 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1136 | static int |
| 1137 | tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una) |
| 1138 | { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1139 | const struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1140 | struct tcp_sock *tp = tcp_sk(sk); |
Arnaldo Carvalho de Melo | 9c70220 | 2007-04-25 18:04:18 -0700 | [diff] [blame] | 1141 | unsigned char *ptr = (skb_transport_header(ack_skb) + |
| 1142 | TCP_SKB_CB(ack_skb)->sacked); |
Al Viro | 269bd27 | 2006-09-27 18:32:28 -0700 | [diff] [blame] | 1143 | struct tcp_sack_block_wire *sp = (struct tcp_sack_block_wire *)(ptr+2); |
Baruch Even | fda03fb | 2007-02-04 23:35:57 -0800 | [diff] [blame] | 1144 | struct sk_buff *cached_skb; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1145 | int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3; |
| 1146 | int reord = tp->packets_out; |
| 1147 | int prior_fackets; |
| 1148 | u32 lost_retrans = 0; |
| 1149 | int flag = 0; |
Ilpo Järvinen | 7769f40 | 2007-06-15 15:14:04 -0700 | [diff] [blame] | 1150 | int found_dup_sack = 0; |
Baruch Even | fda03fb | 2007-02-04 23:35:57 -0800 | [diff] [blame] | 1151 | int cached_fack_count; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1152 | int i; |
Baruch Even | fda03fb | 2007-02-04 23:35:57 -0800 | [diff] [blame] | 1153 | int first_sack_index; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1154 | |
Ilpo Järvinen | d738cd8 | 2007-03-24 21:03:23 -0700 | [diff] [blame] | 1155 | if (!tp->sacked_out) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1156 | tp->fackets_out = 0; |
Ilpo Järvinen | d738cd8 | 2007-03-24 21:03:23 -0700 | [diff] [blame] | 1157 | tp->highest_sack = tp->snd_una; |
| 1158 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1159 | prior_fackets = tp->fackets_out; |
| 1160 | |
David S. Miller | d06e021 | 2007-06-18 22:43:06 -0700 | [diff] [blame] | 1161 | found_dup_sack = tcp_check_dsack(tp, ack_skb, sp, |
| 1162 | num_sacks, prior_snd_una); |
| 1163 | if (found_dup_sack) |
Ilpo Järvinen | 49ff4bb | 2007-08-02 19:47:59 -0700 | [diff] [blame] | 1164 | flag |= FLAG_DSACKING_ACK; |
Baruch Even | 6f74651 | 2007-02-04 23:36:42 -0800 | [diff] [blame] | 1165 | |
| 1166 | /* Eliminate too old ACKs, but take into |
| 1167 | * account more or less fresh ones, they can |
| 1168 | * contain valid SACK info. |
| 1169 | */ |
| 1170 | if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window)) |
| 1171 | return 0; |
| 1172 | |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1173 | /* SACK fastpath: |
| 1174 | * if the only SACK change is the increase of the end_seq of |
| 1175 | * the first block then only apply that SACK block |
| 1176 | * and use retrans queue hinting otherwise slowpath */ |
| 1177 | flag = 1; |
Baruch Even | 6f74651 | 2007-02-04 23:36:42 -0800 | [diff] [blame] | 1178 | for (i = 0; i < num_sacks; i++) { |
| 1179 | __be32 start_seq = sp[i].start_seq; |
| 1180 | __be32 end_seq = sp[i].end_seq; |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1181 | |
Baruch Even | 6f74651 | 2007-02-04 23:36:42 -0800 | [diff] [blame] | 1182 | if (i == 0) { |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1183 | if (tp->recv_sack_cache[i].start_seq != start_seq) |
| 1184 | flag = 0; |
| 1185 | } else { |
| 1186 | if ((tp->recv_sack_cache[i].start_seq != start_seq) || |
| 1187 | (tp->recv_sack_cache[i].end_seq != end_seq)) |
| 1188 | flag = 0; |
| 1189 | } |
| 1190 | tp->recv_sack_cache[i].start_seq = start_seq; |
| 1191 | tp->recv_sack_cache[i].end_seq = end_seq; |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1192 | } |
Baruch Even | 8a3c3a9 | 2007-02-04 23:37:41 -0800 | [diff] [blame] | 1193 | /* Clear the rest of the cache sack blocks so they won't match mistakenly. */ |
| 1194 | for (; i < ARRAY_SIZE(tp->recv_sack_cache); i++) { |
| 1195 | tp->recv_sack_cache[i].start_seq = 0; |
| 1196 | tp->recv_sack_cache[i].end_seq = 0; |
| 1197 | } |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1198 | |
Baruch Even | fda03fb | 2007-02-04 23:35:57 -0800 | [diff] [blame] | 1199 | first_sack_index = 0; |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1200 | if (flag) |
| 1201 | num_sacks = 1; |
| 1202 | else { |
| 1203 | int j; |
| 1204 | tp->fastpath_skb_hint = NULL; |
| 1205 | |
| 1206 | /* order SACK blocks to allow in order walk of the retrans queue */ |
| 1207 | for (i = num_sacks-1; i > 0; i--) { |
| 1208 | for (j = 0; j < i; j++){ |
| 1209 | if (after(ntohl(sp[j].start_seq), |
| 1210 | ntohl(sp[j+1].start_seq))){ |
Baruch Even | db3ccda | 2007-01-25 13:35:06 -0800 | [diff] [blame] | 1211 | struct tcp_sack_block_wire tmp; |
| 1212 | |
| 1213 | tmp = sp[j]; |
| 1214 | sp[j] = sp[j+1]; |
| 1215 | sp[j+1] = tmp; |
Baruch Even | fda03fb | 2007-02-04 23:35:57 -0800 | [diff] [blame] | 1216 | |
| 1217 | /* Track where the first SACK block goes to */ |
| 1218 | if (j == first_sack_index) |
| 1219 | first_sack_index = j+1; |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1220 | } |
| 1221 | |
| 1222 | } |
| 1223 | } |
| 1224 | } |
| 1225 | |
| 1226 | /* clear flag as used for different purpose in following code */ |
| 1227 | flag = 0; |
| 1228 | |
Baruch Even | fda03fb | 2007-02-04 23:35:57 -0800 | [diff] [blame] | 1229 | /* Use SACK fastpath hint if valid */ |
| 1230 | cached_skb = tp->fastpath_skb_hint; |
| 1231 | cached_fack_count = tp->fastpath_cnt_hint; |
| 1232 | if (!cached_skb) { |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1233 | cached_skb = tcp_write_queue_head(sk); |
Baruch Even | fda03fb | 2007-02-04 23:35:57 -0800 | [diff] [blame] | 1234 | cached_fack_count = 0; |
| 1235 | } |
| 1236 | |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1237 | for (i=0; i<num_sacks; i++, sp++) { |
| 1238 | struct sk_buff *skb; |
| 1239 | __u32 start_seq = ntohl(sp->start_seq); |
| 1240 | __u32 end_seq = ntohl(sp->end_seq); |
| 1241 | int fack_count; |
Ilpo Järvinen | 7769f40 | 2007-06-15 15:14:04 -0700 | [diff] [blame] | 1242 | int dup_sack = (found_dup_sack && (i == first_sack_index)); |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1243 | |
Ilpo Järvinen | 18f0254 | 2007-08-24 22:55:52 -0700 | [diff] [blame] | 1244 | if (!tcp_is_sackblock_valid(tp, dup_sack, start_seq, end_seq)) { |
| 1245 | if (dup_sack) { |
| 1246 | if (!tp->undo_marker) |
| 1247 | NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDNOUNDO); |
| 1248 | else |
| 1249 | NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDOLD); |
Ilpo Järvinen | 93e6802 | 2007-09-25 22:46:50 -0700 | [diff] [blame] | 1250 | } else { |
| 1251 | /* Don't count olds caused by ACK reordering */ |
| 1252 | if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) && |
| 1253 | !after(end_seq, tp->snd_una)) |
| 1254 | continue; |
Ilpo Järvinen | 18f0254 | 2007-08-24 22:55:52 -0700 | [diff] [blame] | 1255 | NET_INC_STATS_BH(LINUX_MIB_TCPSACKDISCARD); |
Ilpo Järvinen | 93e6802 | 2007-09-25 22:46:50 -0700 | [diff] [blame] | 1256 | } |
Ilpo Järvinen | 5b3c988 | 2007-08-24 22:54:44 -0700 | [diff] [blame] | 1257 | continue; |
Ilpo Järvinen | 18f0254 | 2007-08-24 22:55:52 -0700 | [diff] [blame] | 1258 | } |
Ilpo Järvinen | 5b3c988 | 2007-08-24 22:54:44 -0700 | [diff] [blame] | 1259 | |
Baruch Even | fda03fb | 2007-02-04 23:35:57 -0800 | [diff] [blame] | 1260 | skb = cached_skb; |
| 1261 | fack_count = cached_fack_count; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1262 | |
| 1263 | /* Event "B" in the comment above. */ |
| 1264 | if (after(end_seq, tp->high_seq)) |
| 1265 | flag |= FLAG_DATA_LOST; |
| 1266 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1267 | tcp_for_write_queue_from(skb, sk) { |
David S. Miller | 6475be1 | 2005-09-01 22:47:01 -0700 | [diff] [blame] | 1268 | int in_sack, pcount; |
| 1269 | u8 sacked; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1270 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1271 | if (skb == tcp_send_head(sk)) |
| 1272 | break; |
| 1273 | |
Baruch Even | fda03fb | 2007-02-04 23:35:57 -0800 | [diff] [blame] | 1274 | cached_skb = skb; |
| 1275 | cached_fack_count = fack_count; |
| 1276 | if (i == first_sack_index) { |
| 1277 | tp->fastpath_skb_hint = skb; |
| 1278 | tp->fastpath_cnt_hint = fack_count; |
| 1279 | } |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1280 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1281 | /* The retransmission queue is always in order, so |
| 1282 | * we can short-circuit the walk early. |
| 1283 | */ |
David S. Miller | 6475be1 | 2005-09-01 22:47:01 -0700 | [diff] [blame] | 1284 | if (!before(TCP_SKB_CB(skb)->seq, end_seq)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1285 | break; |
| 1286 | |
Herbert Xu | 3c05d92 | 2005-09-14 20:50:35 -0700 | [diff] [blame] | 1287 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && |
| 1288 | !before(end_seq, TCP_SKB_CB(skb)->end_seq); |
| 1289 | |
David S. Miller | 6475be1 | 2005-09-01 22:47:01 -0700 | [diff] [blame] | 1290 | pcount = tcp_skb_pcount(skb); |
| 1291 | |
Herbert Xu | 3c05d92 | 2005-09-14 20:50:35 -0700 | [diff] [blame] | 1292 | if (pcount > 1 && !in_sack && |
| 1293 | after(TCP_SKB_CB(skb)->end_seq, start_seq)) { |
David S. Miller | 6475be1 | 2005-09-01 22:47:01 -0700 | [diff] [blame] | 1294 | unsigned int pkt_len; |
| 1295 | |
Herbert Xu | 3c05d92 | 2005-09-14 20:50:35 -0700 | [diff] [blame] | 1296 | in_sack = !after(start_seq, |
| 1297 | TCP_SKB_CB(skb)->seq); |
| 1298 | |
| 1299 | if (!in_sack) |
David S. Miller | 6475be1 | 2005-09-01 22:47:01 -0700 | [diff] [blame] | 1300 | pkt_len = (start_seq - |
| 1301 | TCP_SKB_CB(skb)->seq); |
| 1302 | else |
| 1303 | pkt_len = (end_seq - |
| 1304 | TCP_SKB_CB(skb)->seq); |
Herbert Xu | 7967168 | 2006-06-22 02:40:14 -0700 | [diff] [blame] | 1305 | if (tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size)) |
David S. Miller | 6475be1 | 2005-09-01 22:47:01 -0700 | [diff] [blame] | 1306 | break; |
| 1307 | pcount = tcp_skb_pcount(skb); |
| 1308 | } |
| 1309 | |
| 1310 | fack_count += pcount; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1311 | |
David S. Miller | 6475be1 | 2005-09-01 22:47:01 -0700 | [diff] [blame] | 1312 | sacked = TCP_SKB_CB(skb)->sacked; |
| 1313 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1314 | /* Account D-SACK for retransmitted packet. */ |
| 1315 | if ((dup_sack && in_sack) && |
| 1316 | (sacked & TCPCB_RETRANS) && |
| 1317 | after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker)) |
| 1318 | tp->undo_retrans--; |
| 1319 | |
| 1320 | /* The frame is ACKed. */ |
| 1321 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) { |
| 1322 | if (sacked&TCPCB_RETRANS) { |
| 1323 | if ((dup_sack && in_sack) && |
| 1324 | (sacked&TCPCB_SACKED_ACKED)) |
| 1325 | reord = min(fack_count, reord); |
| 1326 | } else { |
| 1327 | /* If it was in a hole, we detected reordering. */ |
| 1328 | if (fack_count < prior_fackets && |
| 1329 | !(sacked&TCPCB_SACKED_ACKED)) |
| 1330 | reord = min(fack_count, reord); |
| 1331 | } |
| 1332 | |
| 1333 | /* Nothing to do; acked frame is about to be dropped. */ |
| 1334 | continue; |
| 1335 | } |
| 1336 | |
| 1337 | if ((sacked&TCPCB_SACKED_RETRANS) && |
| 1338 | after(end_seq, TCP_SKB_CB(skb)->ack_seq) && |
| 1339 | (!lost_retrans || after(end_seq, lost_retrans))) |
| 1340 | lost_retrans = end_seq; |
| 1341 | |
| 1342 | if (!in_sack) |
| 1343 | continue; |
| 1344 | |
| 1345 | if (!(sacked&TCPCB_SACKED_ACKED)) { |
| 1346 | if (sacked & TCPCB_SACKED_RETRANS) { |
| 1347 | /* If the segment is not tagged as lost, |
| 1348 | * we do not clear RETRANS, believing |
| 1349 | * that retransmission is still in flight. |
| 1350 | */ |
| 1351 | if (sacked & TCPCB_LOST) { |
| 1352 | TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); |
| 1353 | tp->lost_out -= tcp_skb_pcount(skb); |
| 1354 | tp->retrans_out -= tcp_skb_pcount(skb); |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1355 | |
| 1356 | /* clear lost hint */ |
| 1357 | tp->retransmit_skb_hint = NULL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1358 | } |
| 1359 | } else { |
| 1360 | /* New sack for not retransmitted frame, |
| 1361 | * which was in hole. It is reordering. |
| 1362 | */ |
| 1363 | if (!(sacked & TCPCB_RETRANS) && |
| 1364 | fack_count < prior_fackets) |
| 1365 | reord = min(fack_count, reord); |
| 1366 | |
| 1367 | if (sacked & TCPCB_LOST) { |
| 1368 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; |
| 1369 | tp->lost_out -= tcp_skb_pcount(skb); |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1370 | |
| 1371 | /* clear lost hint */ |
| 1372 | tp->retransmit_skb_hint = NULL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1373 | } |
Ilpo Järvinen | 4dc2665 | 2007-02-21 23:16:11 -0800 | [diff] [blame] | 1374 | /* SACK enhanced F-RTO detection. |
| 1375 | * Set flag if and only if non-rexmitted |
| 1376 | * segments below frto_highmark are |
| 1377 | * SACKed (RFC4138; Appendix B). |
| 1378 | * Clearing correct due to in-order walk |
| 1379 | */ |
| 1380 | if (after(end_seq, tp->frto_highmark)) { |
| 1381 | flag &= ~FLAG_ONLY_ORIG_SACKED; |
| 1382 | } else { |
| 1383 | if (!(sacked & TCPCB_RETRANS)) |
| 1384 | flag |= FLAG_ONLY_ORIG_SACKED; |
| 1385 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1386 | } |
| 1387 | |
| 1388 | TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED; |
| 1389 | flag |= FLAG_DATA_SACKED; |
| 1390 | tp->sacked_out += tcp_skb_pcount(skb); |
| 1391 | |
| 1392 | if (fack_count > tp->fackets_out) |
| 1393 | tp->fackets_out = fack_count; |
Ilpo Järvinen | d738cd8 | 2007-03-24 21:03:23 -0700 | [diff] [blame] | 1394 | |
| 1395 | if (after(TCP_SKB_CB(skb)->seq, |
| 1396 | tp->highest_sack)) |
| 1397 | tp->highest_sack = TCP_SKB_CB(skb)->seq; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1398 | } else { |
| 1399 | if (dup_sack && (sacked&TCPCB_RETRANS)) |
| 1400 | reord = min(fack_count, reord); |
| 1401 | } |
| 1402 | |
| 1403 | /* D-SACK. We can detect redundant retransmission |
| 1404 | * in S|R and plain R frames and clear it. |
| 1405 | * undo_retrans is decreased above, L|R frames |
| 1406 | * are accounted above as well. |
| 1407 | */ |
| 1408 | if (dup_sack && |
| 1409 | (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) { |
| 1410 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; |
| 1411 | tp->retrans_out -= tcp_skb_pcount(skb); |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1412 | tp->retransmit_skb_hint = NULL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1413 | } |
| 1414 | } |
| 1415 | } |
| 1416 | |
| 1417 | /* Check for lost retransmit. This superb idea is |
| 1418 | * borrowed from "ratehalving". Event "C". |
| 1419 | * Later note: FACK people cheated me again 8), |
| 1420 | * we have to account for reordering! Ugly, |
| 1421 | * but should help. |
| 1422 | */ |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1423 | if (lost_retrans && icsk->icsk_ca_state == TCP_CA_Recovery) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1424 | struct sk_buff *skb; |
| 1425 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1426 | tcp_for_write_queue(skb, sk) { |
| 1427 | if (skb == tcp_send_head(sk)) |
| 1428 | break; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1429 | if (after(TCP_SKB_CB(skb)->seq, lost_retrans)) |
| 1430 | break; |
| 1431 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) |
| 1432 | continue; |
| 1433 | if ((TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) && |
| 1434 | after(lost_retrans, TCP_SKB_CB(skb)->ack_seq) && |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 1435 | (tcp_is_fack(tp) || |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1436 | !before(lost_retrans, |
| 1437 | TCP_SKB_CB(skb)->ack_seq + tp->reordering * |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1438 | tp->mss_cache))) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1439 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; |
| 1440 | tp->retrans_out -= tcp_skb_pcount(skb); |
| 1441 | |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1442 | /* clear lost hint */ |
| 1443 | tp->retransmit_skb_hint = NULL; |
| 1444 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1445 | if (!(TCP_SKB_CB(skb)->sacked&(TCPCB_LOST|TCPCB_SACKED_ACKED))) { |
| 1446 | tp->lost_out += tcp_skb_pcount(skb); |
| 1447 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; |
| 1448 | flag |= FLAG_DATA_SACKED; |
| 1449 | NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT); |
| 1450 | } |
| 1451 | } |
| 1452 | } |
| 1453 | } |
| 1454 | |
Ilpo Järvinen | 86426c2 | 2007-08-09 14:45:17 +0300 | [diff] [blame] | 1455 | tcp_verify_left_out(tp); |
| 1456 | |
Ilpo Järvinen | 288035f9 | 2007-02-21 23:14:42 -0800 | [diff] [blame] | 1457 | if ((reord < tp->fackets_out) && icsk->icsk_ca_state != TCP_CA_Loss && |
Ilpo Järvinen | c5e7af0 | 2007-02-23 16:22:06 -0800 | [diff] [blame] | 1458 | (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark))) |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1459 | tcp_update_reordering(sk, ((tp->fackets_out + 1) - reord), 0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1460 | |
| 1461 | #if FASTRETRANS_DEBUG > 0 |
| 1462 | BUG_TRAP((int)tp->sacked_out >= 0); |
| 1463 | BUG_TRAP((int)tp->lost_out >= 0); |
| 1464 | BUG_TRAP((int)tp->retrans_out >= 0); |
| 1465 | BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0); |
| 1466 | #endif |
| 1467 | return flag; |
| 1468 | } |
| 1469 | |
Ilpo Järvinen | 575ee71 | 2007-04-30 00:39:55 -0700 | [diff] [blame] | 1470 | /* F-RTO can only be used if TCP has never retransmitted anything other than |
| 1471 | * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here) |
Ilpo Järvinen | 30935cf | 2007-02-21 23:01:36 -0800 | [diff] [blame] | 1472 | */ |
Ilpo Järvinen | 4ddf667 | 2007-05-27 01:52:00 -0700 | [diff] [blame] | 1473 | static void tcp_check_reno_reordering(struct sock *sk, const int addend) |
| 1474 | { |
| 1475 | struct tcp_sock *tp = tcp_sk(sk); |
| 1476 | u32 holes; |
| 1477 | |
| 1478 | holes = max(tp->lost_out, 1U); |
| 1479 | holes = min(holes, tp->packets_out); |
| 1480 | |
| 1481 | if ((tp->sacked_out + holes) > tp->packets_out) { |
| 1482 | tp->sacked_out = tp->packets_out - holes; |
| 1483 | tcp_update_reordering(sk, tp->packets_out + addend, 0); |
| 1484 | } |
| 1485 | } |
| 1486 | |
| 1487 | /* Emulate SACKs for SACKless connection: account for a new dupack. */ |
| 1488 | |
| 1489 | static void tcp_add_reno_sack(struct sock *sk) |
| 1490 | { |
| 1491 | struct tcp_sock *tp = tcp_sk(sk); |
| 1492 | tp->sacked_out++; |
| 1493 | tcp_check_reno_reordering(sk, 0); |
Ilpo Järvinen | 005903b | 2007-08-09 14:44:16 +0300 | [diff] [blame] | 1494 | tcp_verify_left_out(tp); |
Ilpo Järvinen | 4ddf667 | 2007-05-27 01:52:00 -0700 | [diff] [blame] | 1495 | } |
| 1496 | |
| 1497 | /* Account for ACK, ACKing some data in Reno Recovery phase. */ |
| 1498 | |
| 1499 | static void tcp_remove_reno_sacks(struct sock *sk, int acked) |
| 1500 | { |
| 1501 | struct tcp_sock *tp = tcp_sk(sk); |
| 1502 | |
| 1503 | if (acked > 0) { |
| 1504 | /* One ACK acked hole. The rest eat duplicate ACKs. */ |
| 1505 | if (acked-1 >= tp->sacked_out) |
| 1506 | tp->sacked_out = 0; |
| 1507 | else |
| 1508 | tp->sacked_out -= acked-1; |
| 1509 | } |
| 1510 | tcp_check_reno_reordering(sk, acked); |
Ilpo Järvinen | 005903b | 2007-08-09 14:44:16 +0300 | [diff] [blame] | 1511 | tcp_verify_left_out(tp); |
Ilpo Järvinen | 4ddf667 | 2007-05-27 01:52:00 -0700 | [diff] [blame] | 1512 | } |
| 1513 | |
| 1514 | static inline void tcp_reset_reno_sack(struct tcp_sock *tp) |
| 1515 | { |
| 1516 | tp->sacked_out = 0; |
Ilpo Järvinen | 4ddf667 | 2007-05-27 01:52:00 -0700 | [diff] [blame] | 1517 | } |
| 1518 | |
Ilpo Järvinen | 46d0de4 | 2007-02-21 23:10:39 -0800 | [diff] [blame] | 1519 | int tcp_use_frto(struct sock *sk) |
Ilpo Järvinen | bdaae17 | 2007-02-21 22:59:58 -0800 | [diff] [blame] | 1520 | { |
| 1521 | const struct tcp_sock *tp = tcp_sk(sk); |
Ilpo Järvinen | 46d0de4 | 2007-02-21 23:10:39 -0800 | [diff] [blame] | 1522 | struct sk_buff *skb; |
Ilpo Järvinen | bdaae17 | 2007-02-21 22:59:58 -0800 | [diff] [blame] | 1523 | |
Ilpo Järvinen | 575ee71 | 2007-04-30 00:39:55 -0700 | [diff] [blame] | 1524 | if (!sysctl_tcp_frto) |
Ilpo Järvinen | 46d0de4 | 2007-02-21 23:10:39 -0800 | [diff] [blame] | 1525 | return 0; |
| 1526 | |
Ilpo Järvinen | 4dc2665 | 2007-02-21 23:16:11 -0800 | [diff] [blame] | 1527 | if (IsSackFrto()) |
| 1528 | return 1; |
| 1529 | |
Ilpo Järvinen | 46d0de4 | 2007-02-21 23:10:39 -0800 | [diff] [blame] | 1530 | /* Avoid expensive walking of rexmit queue if possible */ |
| 1531 | if (tp->retrans_out > 1) |
| 1532 | return 0; |
| 1533 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1534 | skb = tcp_write_queue_head(sk); |
| 1535 | skb = tcp_write_queue_next(sk, skb); /* Skips head */ |
| 1536 | tcp_for_write_queue_from(skb, sk) { |
| 1537 | if (skb == tcp_send_head(sk)) |
| 1538 | break; |
Ilpo Järvinen | 46d0de4 | 2007-02-21 23:10:39 -0800 | [diff] [blame] | 1539 | if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS) |
| 1540 | return 0; |
| 1541 | /* Short-circuit when first non-SACKed skb has been checked */ |
| 1542 | if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) |
| 1543 | break; |
| 1544 | } |
| 1545 | return 1; |
Ilpo Järvinen | bdaae17 | 2007-02-21 22:59:58 -0800 | [diff] [blame] | 1546 | } |
| 1547 | |
Ilpo Järvinen | 30935cf | 2007-02-21 23:01:36 -0800 | [diff] [blame] | 1548 | /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO |
| 1549 | * recovery a bit and use heuristics in tcp_process_frto() to detect if |
Ilpo Järvinen | d1a54c6 | 2007-02-21 23:11:57 -0800 | [diff] [blame] | 1550 | * the RTO was spurious. Only clear SACKED_RETRANS of the head here to |
| 1551 | * keep retrans_out counting accurate (with SACK F-RTO, other than head |
| 1552 | * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS |
| 1553 | * bits are handled if the Loss state is really to be entered (in |
| 1554 | * tcp_enter_frto_loss). |
Ilpo Järvinen | 7487c48 | 2007-02-21 23:02:30 -0800 | [diff] [blame] | 1555 | * |
| 1556 | * Do like tcp_enter_loss() would; when RTO expires the second time it |
| 1557 | * does: |
| 1558 | * "Reduce ssthresh if it has not yet been made inside this window." |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1559 | */ |
| 1560 | void tcp_enter_frto(struct sock *sk) |
| 1561 | { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1562 | const struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1563 | struct tcp_sock *tp = tcp_sk(sk); |
| 1564 | struct sk_buff *skb; |
| 1565 | |
Ilpo Järvinen | 7487c48 | 2007-02-21 23:02:30 -0800 | [diff] [blame] | 1566 | if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) || |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1567 | tp->snd_una == tp->high_seq || |
Ilpo Järvinen | 7487c48 | 2007-02-21 23:02:30 -0800 | [diff] [blame] | 1568 | ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) && |
| 1569 | !icsk->icsk_retransmits)) { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1570 | tp->prior_ssthresh = tcp_current_ssthresh(sk); |
Ilpo Järvinen | 66e93e4 | 2007-02-21 23:13:47 -0800 | [diff] [blame] | 1571 | /* Our state is too optimistic in ssthresh() call because cwnd |
| 1572 | * is not reduced until tcp_enter_frto_loss() when previous FRTO |
| 1573 | * recovery has not yet completed. Pattern would be this: RTO, |
| 1574 | * Cumulative ACK, RTO (2xRTO for the same segment does not end |
| 1575 | * up here twice). |
| 1576 | * RFC4138 should be more specific on what to do, even though |
| 1577 | * RTO is quite unlikely to occur after the first Cumulative ACK |
| 1578 | * due to back-off and complexity of triggering events ... |
| 1579 | */ |
| 1580 | if (tp->frto_counter) { |
| 1581 | u32 stored_cwnd; |
| 1582 | stored_cwnd = tp->snd_cwnd; |
| 1583 | tp->snd_cwnd = 2; |
| 1584 | tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); |
| 1585 | tp->snd_cwnd = stored_cwnd; |
| 1586 | } else { |
| 1587 | tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); |
| 1588 | } |
| 1589 | /* ... in theory, cong.control module could do "any tricks" in |
| 1590 | * ssthresh(), which means that ca_state, lost bits and lost_out |
| 1591 | * counter would have to be faked before the call occurs. We |
| 1592 | * consider that too expensive, unlikely and hacky, so modules |
| 1593 | * using these in ssthresh() must deal these incompatibility |
| 1594 | * issues if they receives CA_EVENT_FRTO and frto_counter != 0 |
| 1595 | */ |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1596 | tcp_ca_event(sk, CA_EVENT_FRTO); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1597 | } |
| 1598 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1599 | tp->undo_marker = tp->snd_una; |
| 1600 | tp->undo_retrans = 0; |
| 1601 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1602 | skb = tcp_write_queue_head(sk); |
Ilpo Järvinen | 009a2e3 | 2007-09-20 11:34:38 -0700 | [diff] [blame] | 1603 | if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) |
| 1604 | tp->undo_marker = 0; |
Ilpo Järvinen | d1a54c6 | 2007-02-21 23:11:57 -0800 | [diff] [blame] | 1605 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { |
Ilpo Järvinen | 522e754 | 2007-02-21 22:54:52 -0800 | [diff] [blame] | 1606 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; |
Ilpo Järvinen | d1a54c6 | 2007-02-21 23:11:57 -0800 | [diff] [blame] | 1607 | tp->retrans_out -= tcp_skb_pcount(skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1608 | } |
Ilpo Järvinen | 005903b | 2007-08-09 14:44:16 +0300 | [diff] [blame] | 1609 | tcp_verify_left_out(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1610 | |
Ilpo Järvinen | 4dc2665 | 2007-02-21 23:16:11 -0800 | [diff] [blame] | 1611 | /* Earlier loss recovery underway (see RFC4138; Appendix B). |
| 1612 | * The last condition is necessary at least in tp->frto_counter case. |
| 1613 | */ |
| 1614 | if (IsSackFrto() && (tp->frto_counter || |
| 1615 | ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) && |
| 1616 | after(tp->high_seq, tp->snd_una)) { |
| 1617 | tp->frto_highmark = tp->high_seq; |
| 1618 | } else { |
| 1619 | tp->frto_highmark = tp->snd_nxt; |
| 1620 | } |
Ilpo Järvinen | 7b0eb22 | 2007-02-21 23:03:35 -0800 | [diff] [blame] | 1621 | tcp_set_ca_state(sk, TCP_CA_Disorder); |
| 1622 | tp->high_seq = tp->snd_nxt; |
Ilpo Järvinen | 7487c48 | 2007-02-21 23:02:30 -0800 | [diff] [blame] | 1623 | tp->frto_counter = 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1624 | } |
| 1625 | |
| 1626 | /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO, |
| 1627 | * which indicates that we should follow the traditional RTO recovery, |
| 1628 | * i.e. mark everything lost and do go-back-N retransmission. |
| 1629 | */ |
Ilpo Järvinen | d1a54c6 | 2007-02-21 23:11:57 -0800 | [diff] [blame] | 1630 | static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1631 | { |
| 1632 | struct tcp_sock *tp = tcp_sk(sk); |
| 1633 | struct sk_buff *skb; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1634 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1635 | tp->lost_out = 0; |
Ilpo Järvinen | d1a54c6 | 2007-02-21 23:11:57 -0800 | [diff] [blame] | 1636 | tp->retrans_out = 0; |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 1637 | if (tcp_is_reno(tp)) |
Ilpo Järvinen | 9bff40f | 2007-05-27 01:53:49 -0700 | [diff] [blame] | 1638 | tcp_reset_reno_sack(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1639 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1640 | tcp_for_write_queue(skb, sk) { |
| 1641 | if (skb == tcp_send_head(sk)) |
| 1642 | break; |
Ilpo Järvinen | d1a54c6 | 2007-02-21 23:11:57 -0800 | [diff] [blame] | 1643 | /* |
| 1644 | * Count the retransmission made on RTO correctly (only when |
| 1645 | * waiting for the first ACK and did not get it)... |
| 1646 | */ |
| 1647 | if ((tp->frto_counter == 1) && !(flag&FLAG_DATA_ACKED)) { |
Ilpo Järvinen | 0a9f2a4 | 2007-07-15 00:19:29 -0700 | [diff] [blame] | 1648 | /* For some reason this R-bit might get cleared? */ |
| 1649 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) |
| 1650 | tp->retrans_out += tcp_skb_pcount(skb); |
Ilpo Järvinen | d1a54c6 | 2007-02-21 23:11:57 -0800 | [diff] [blame] | 1651 | /* ...enter this if branch just for the first segment */ |
| 1652 | flag |= FLAG_DATA_ACKED; |
| 1653 | } else { |
Ilpo Järvinen | 009a2e3 | 2007-09-20 11:34:38 -0700 | [diff] [blame] | 1654 | if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) |
| 1655 | tp->undo_marker = 0; |
Ilpo Järvinen | d1a54c6 | 2007-02-21 23:11:57 -0800 | [diff] [blame] | 1656 | TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); |
| 1657 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1658 | |
Ilpo Järvinen | 9bff40f | 2007-05-27 01:53:49 -0700 | [diff] [blame] | 1659 | /* Don't lost mark skbs that were fwd transmitted after RTO */ |
| 1660 | if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) && |
| 1661 | !after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark)) { |
| 1662 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; |
| 1663 | tp->lost_out += tcp_skb_pcount(skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1664 | } |
| 1665 | } |
Ilpo Järvinen | 005903b | 2007-08-09 14:44:16 +0300 | [diff] [blame] | 1666 | tcp_verify_left_out(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1667 | |
Ilpo Järvinen | 95c4922 | 2007-02-21 23:05:18 -0800 | [diff] [blame] | 1668 | tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1669 | tp->snd_cwnd_cnt = 0; |
| 1670 | tp->snd_cwnd_stamp = tcp_time_stamp; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1671 | tp->frto_counter = 0; |
| 1672 | |
| 1673 | tp->reordering = min_t(unsigned int, tp->reordering, |
| 1674 | sysctl_tcp_reordering); |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1675 | tcp_set_ca_state(sk, TCP_CA_Loss); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1676 | tp->high_seq = tp->frto_highmark; |
| 1677 | TCP_ECN_queue_cwr(tp); |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1678 | |
Ilpo Järvinen | b768920 | 2007-09-20 11:37:19 -0700 | [diff] [blame] | 1679 | tcp_clear_retrans_hints_partial(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1680 | } |
| 1681 | |
| 1682 | void tcp_clear_retrans(struct tcp_sock *tp) |
| 1683 | { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1684 | tp->retrans_out = 0; |
| 1685 | |
| 1686 | tp->fackets_out = 0; |
| 1687 | tp->sacked_out = 0; |
| 1688 | tp->lost_out = 0; |
| 1689 | |
| 1690 | tp->undo_marker = 0; |
| 1691 | tp->undo_retrans = 0; |
| 1692 | } |
| 1693 | |
| 1694 | /* Enter Loss state. If "how" is not zero, forget all SACK information |
| 1695 | * and reset tags completely, otherwise preserve SACKs. If receiver |
| 1696 | * dropped its ofo queue, we will know this due to reneging detection. |
| 1697 | */ |
| 1698 | void tcp_enter_loss(struct sock *sk, int how) |
| 1699 | { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1700 | const struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1701 | struct tcp_sock *tp = tcp_sk(sk); |
| 1702 | struct sk_buff *skb; |
| 1703 | int cnt = 0; |
| 1704 | |
| 1705 | /* Reduce ssthresh if it has not yet been made inside this window. */ |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1706 | if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq || |
| 1707 | (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) { |
| 1708 | tp->prior_ssthresh = tcp_current_ssthresh(sk); |
| 1709 | tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); |
| 1710 | tcp_ca_event(sk, CA_EVENT_LOSS); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1711 | } |
| 1712 | tp->snd_cwnd = 1; |
| 1713 | tp->snd_cwnd_cnt = 0; |
| 1714 | tp->snd_cwnd_stamp = tcp_time_stamp; |
| 1715 | |
Stephen Hemminger | 9772efb | 2005-11-10 17:09:53 -0800 | [diff] [blame] | 1716 | tp->bytes_acked = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1717 | tcp_clear_retrans(tp); |
| 1718 | |
Ilpo Järvinen | b768920 | 2007-09-20 11:37:19 -0700 | [diff] [blame] | 1719 | if (!how) { |
| 1720 | /* Push undo marker, if it was plain RTO and nothing |
| 1721 | * was retransmitted. */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1722 | tp->undo_marker = tp->snd_una; |
Ilpo Järvinen | b768920 | 2007-09-20 11:37:19 -0700 | [diff] [blame] | 1723 | tcp_clear_retrans_hints_partial(tp); |
| 1724 | } else { |
| 1725 | tcp_clear_all_retrans_hints(tp); |
| 1726 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1727 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1728 | tcp_for_write_queue(skb, sk) { |
| 1729 | if (skb == tcp_send_head(sk)) |
| 1730 | break; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1731 | cnt += tcp_skb_pcount(skb); |
| 1732 | if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS) |
| 1733 | tp->undo_marker = 0; |
| 1734 | TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED; |
| 1735 | if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) { |
| 1736 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED; |
| 1737 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; |
| 1738 | tp->lost_out += tcp_skb_pcount(skb); |
| 1739 | } else { |
| 1740 | tp->sacked_out += tcp_skb_pcount(skb); |
| 1741 | tp->fackets_out = cnt; |
| 1742 | } |
| 1743 | } |
Ilpo Järvinen | 005903b | 2007-08-09 14:44:16 +0300 | [diff] [blame] | 1744 | tcp_verify_left_out(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1745 | |
| 1746 | tp->reordering = min_t(unsigned int, tp->reordering, |
| 1747 | sysctl_tcp_reordering); |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1748 | tcp_set_ca_state(sk, TCP_CA_Loss); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1749 | tp->high_seq = tp->snd_nxt; |
| 1750 | TCP_ECN_queue_cwr(tp); |
Ilpo Järvinen | 580e572 | 2007-05-19 13:56:57 -0700 | [diff] [blame] | 1751 | /* Abort FRTO algorithm if one is in progress */ |
| 1752 | tp->frto_counter = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1753 | } |
| 1754 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 1755 | static int tcp_check_sack_reneging(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1756 | { |
| 1757 | struct sk_buff *skb; |
| 1758 | |
| 1759 | /* If ACK arrived pointing to a remembered SACK, |
| 1760 | * it means that our remembered SACKs do not reflect |
| 1761 | * real state of receiver i.e. |
| 1762 | * receiver _host_ is heavily congested (or buggy). |
| 1763 | * Do processing similar to RTO timeout. |
| 1764 | */ |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1765 | if ((skb = tcp_write_queue_head(sk)) != NULL && |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1766 | (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1767 | struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1768 | NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING); |
| 1769 | |
| 1770 | tcp_enter_loss(sk, 1); |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1771 | icsk->icsk_retransmits++; |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1772 | tcp_retransmit_skb(sk, tcp_write_queue_head(sk)); |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 1773 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1774 | icsk->icsk_rto, TCP_RTO_MAX); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1775 | return 1; |
| 1776 | } |
| 1777 | return 0; |
| 1778 | } |
| 1779 | |
| 1780 | static inline int tcp_fackets_out(struct tcp_sock *tp) |
| 1781 | { |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 1782 | return tcp_is_reno(tp) ? tp->sacked_out+1 : tp->fackets_out; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1783 | } |
| 1784 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 1785 | static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1786 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 1787 | return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1788 | } |
| 1789 | |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1790 | static inline int tcp_head_timedout(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1791 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1792 | struct tcp_sock *tp = tcp_sk(sk); |
| 1793 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1794 | return tp->packets_out && |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1795 | tcp_skb_timedout(sk, tcp_write_queue_head(sk)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1796 | } |
| 1797 | |
| 1798 | /* Linux NewReno/SACK/FACK/ECN state machine. |
| 1799 | * -------------------------------------- |
| 1800 | * |
| 1801 | * "Open" Normal state, no dubious events, fast path. |
| 1802 | * "Disorder" In all the respects it is "Open", |
| 1803 | * but requires a bit more attention. It is entered when |
| 1804 | * we see some SACKs or dupacks. It is split of "Open" |
| 1805 | * mainly to move some processing from fast path to slow one. |
| 1806 | * "CWR" CWND was reduced due to some Congestion Notification event. |
| 1807 | * It can be ECN, ICMP source quench, local device congestion. |
| 1808 | * "Recovery" CWND was reduced, we are fast-retransmitting. |
| 1809 | * "Loss" CWND was reduced due to RTO timeout or SACK reneging. |
| 1810 | * |
| 1811 | * tcp_fastretrans_alert() is entered: |
| 1812 | * - each incoming ACK, if state is not "Open" |
| 1813 | * - when arrived ACK is unusual, namely: |
| 1814 | * * SACK |
| 1815 | * * Duplicate ACK. |
| 1816 | * * ECN ECE. |
| 1817 | * |
| 1818 | * Counting packets in flight is pretty simple. |
| 1819 | * |
| 1820 | * in_flight = packets_out - left_out + retrans_out |
| 1821 | * |
| 1822 | * packets_out is SND.NXT-SND.UNA counted in packets. |
| 1823 | * |
| 1824 | * retrans_out is number of retransmitted segments. |
| 1825 | * |
| 1826 | * left_out is number of segments left network, but not ACKed yet. |
| 1827 | * |
| 1828 | * left_out = sacked_out + lost_out |
| 1829 | * |
| 1830 | * sacked_out: Packets, which arrived to receiver out of order |
| 1831 | * and hence not ACKed. With SACKs this number is simply |
| 1832 | * amount of SACKed data. Even without SACKs |
| 1833 | * it is easy to give pretty reliable estimate of this number, |
| 1834 | * counting duplicate ACKs. |
| 1835 | * |
| 1836 | * lost_out: Packets lost by network. TCP has no explicit |
| 1837 | * "loss notification" feedback from network (for now). |
| 1838 | * It means that this number can be only _guessed_. |
| 1839 | * Actually, it is the heuristics to predict lossage that |
| 1840 | * distinguishes different algorithms. |
| 1841 | * |
| 1842 | * F.e. after RTO, when all the queue is considered as lost, |
| 1843 | * lost_out = packets_out and in_flight = retrans_out. |
| 1844 | * |
| 1845 | * Essentially, we have now two algorithms counting |
| 1846 | * lost packets. |
| 1847 | * |
| 1848 | * FACK: It is the simplest heuristics. As soon as we decided |
| 1849 | * that something is lost, we decide that _all_ not SACKed |
| 1850 | * packets until the most forward SACK are lost. I.e. |
| 1851 | * lost_out = fackets_out - sacked_out and left_out = fackets_out. |
| 1852 | * It is absolutely correct estimate, if network does not reorder |
| 1853 | * packets. And it loses any connection to reality when reordering |
| 1854 | * takes place. We use FACK by default until reordering |
| 1855 | * is suspected on the path to this destination. |
| 1856 | * |
| 1857 | * NewReno: when Recovery is entered, we assume that one segment |
| 1858 | * is lost (classic Reno). While we are in Recovery and |
| 1859 | * a partial ACK arrives, we assume that one more packet |
| 1860 | * is lost (NewReno). This heuristics are the same in NewReno |
| 1861 | * and SACK. |
| 1862 | * |
| 1863 | * Imagine, that's all! Forget about all this shamanism about CWND inflation |
| 1864 | * deflation etc. CWND is real congestion window, never inflated, changes |
| 1865 | * only according to classic VJ rules. |
| 1866 | * |
| 1867 | * Really tricky (and requiring careful tuning) part of algorithm |
| 1868 | * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue(). |
| 1869 | * The first determines the moment _when_ we should reduce CWND and, |
| 1870 | * hence, slow down forward transmission. In fact, it determines the moment |
| 1871 | * when we decide that hole is caused by loss, rather than by a reorder. |
| 1872 | * |
| 1873 | * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill |
| 1874 | * holes, caused by lost packets. |
| 1875 | * |
| 1876 | * And the most logically complicated part of algorithm is undo |
| 1877 | * heuristics. We detect false retransmits due to both too early |
| 1878 | * fast retransmit (reordering) and underestimated RTO, analyzing |
| 1879 | * timestamps and D-SACKs. When we detect that some segments were |
| 1880 | * retransmitted by mistake and CWND reduction was wrong, we undo |
| 1881 | * window reduction and abort recovery phase. This logic is hidden |
| 1882 | * inside several functions named tcp_try_undo_<something>. |
| 1883 | */ |
| 1884 | |
| 1885 | /* This function decides, when we should leave Disordered state |
| 1886 | * and enter Recovery phase, reducing congestion window. |
| 1887 | * |
| 1888 | * Main question: may we further continue forward transmission |
| 1889 | * with the same cwnd? |
| 1890 | */ |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1891 | static int tcp_time_to_recover(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1892 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1893 | struct tcp_sock *tp = tcp_sk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1894 | __u32 packets_out; |
| 1895 | |
Ilpo Järvinen | 52c63f1 | 2007-02-21 23:06:52 -0800 | [diff] [blame] | 1896 | /* Do not perform any recovery during FRTO algorithm */ |
| 1897 | if (tp->frto_counter) |
| 1898 | return 0; |
| 1899 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1900 | /* Trick#1: The loss is proven. */ |
| 1901 | if (tp->lost_out) |
| 1902 | return 1; |
| 1903 | |
| 1904 | /* Not-A-Trick#2 : Classic rule... */ |
| 1905 | if (tcp_fackets_out(tp) > tp->reordering) |
| 1906 | return 1; |
| 1907 | |
| 1908 | /* Trick#3 : when we use RFC2988 timer restart, fast |
| 1909 | * retransmit can be triggered by timeout of queue head. |
| 1910 | */ |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1911 | if (tcp_head_timedout(sk)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1912 | return 1; |
| 1913 | |
| 1914 | /* Trick#4: It is still not OK... But will it be useful to delay |
| 1915 | * recovery more? |
| 1916 | */ |
| 1917 | packets_out = tp->packets_out; |
| 1918 | if (packets_out <= tp->reordering && |
| 1919 | tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) && |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1920 | !tcp_may_send_now(sk)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1921 | /* We have nothing to send. This connection is limited |
| 1922 | * either by receiver window or by application. |
| 1923 | */ |
| 1924 | return 1; |
| 1925 | } |
| 1926 | |
| 1927 | return 0; |
| 1928 | } |
| 1929 | |
Ilpo Järvinen | d8f4f22 | 2007-04-20 22:56:38 -0700 | [diff] [blame] | 1930 | /* RFC: This is from the original, I doubt that this is necessary at all: |
| 1931 | * clear xmit_retrans hint if seq of this skb is beyond hint. How could we |
| 1932 | * retransmitted past LOST markings in the first place? I'm not fully sure |
| 1933 | * about undo and end of connection cases, which can cause R without L? |
| 1934 | */ |
| 1935 | static void tcp_verify_retransmit_hint(struct tcp_sock *tp, |
| 1936 | struct sk_buff *skb) |
| 1937 | { |
| 1938 | if ((tp->retransmit_skb_hint != NULL) && |
| 1939 | before(TCP_SKB_CB(skb)->seq, |
| 1940 | TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) |
Ilpo Järvinen | 19b2b48 | 2007-03-28 12:06:37 -0700 | [diff] [blame] | 1941 | tp->retransmit_skb_hint = NULL; |
Ilpo Järvinen | d8f4f22 | 2007-04-20 22:56:38 -0700 | [diff] [blame] | 1942 | } |
| 1943 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1944 | /* Mark head of queue up as lost. */ |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1945 | static void tcp_mark_head_lost(struct sock *sk, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1946 | int packets, u32 high_seq) |
| 1947 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1948 | struct tcp_sock *tp = tcp_sk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1949 | struct sk_buff *skb; |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1950 | int cnt; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1951 | |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1952 | BUG_TRAP(packets <= tp->packets_out); |
| 1953 | if (tp->lost_skb_hint) { |
| 1954 | skb = tp->lost_skb_hint; |
| 1955 | cnt = tp->lost_cnt_hint; |
| 1956 | } else { |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1957 | skb = tcp_write_queue_head(sk); |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1958 | cnt = 0; |
| 1959 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1960 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1961 | tcp_for_write_queue_from(skb, sk) { |
| 1962 | if (skb == tcp_send_head(sk)) |
| 1963 | break; |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1964 | /* TODO: do this better */ |
| 1965 | /* this is not the most efficient way to do this... */ |
| 1966 | tp->lost_skb_hint = skb; |
| 1967 | tp->lost_cnt_hint = cnt; |
| 1968 | cnt += tcp_skb_pcount(skb); |
| 1969 | if (cnt > packets || after(TCP_SKB_CB(skb)->end_seq, high_seq)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1970 | break; |
| 1971 | if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) { |
| 1972 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; |
| 1973 | tp->lost_out += tcp_skb_pcount(skb); |
Ilpo Järvinen | d8f4f22 | 2007-04-20 22:56:38 -0700 | [diff] [blame] | 1974 | tcp_verify_retransmit_hint(tp, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1975 | } |
| 1976 | } |
Ilpo Järvinen | 005903b | 2007-08-09 14:44:16 +0300 | [diff] [blame] | 1977 | tcp_verify_left_out(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1978 | } |
| 1979 | |
| 1980 | /* Account newly detected lost packet(s) */ |
| 1981 | |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1982 | static void tcp_update_scoreboard(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1983 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1984 | struct tcp_sock *tp = tcp_sk(sk); |
| 1985 | |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 1986 | if (tcp_is_fack(tp)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1987 | int lost = tp->fackets_out - tp->reordering; |
| 1988 | if (lost <= 0) |
| 1989 | lost = 1; |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1990 | tcp_mark_head_lost(sk, lost, tp->high_seq); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1991 | } else { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1992 | tcp_mark_head_lost(sk, 1, tp->high_seq); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1993 | } |
| 1994 | |
| 1995 | /* New heuristics: it is possible only after we switched |
| 1996 | * to restart timer each time when something is ACKed. |
| 1997 | * Hence, we can detect timed out packets during fast |
| 1998 | * retransmit without falling to slow start. |
| 1999 | */ |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 2000 | if (!tcp_is_reno(tp) && tcp_head_timedout(sk)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2001 | struct sk_buff *skb; |
| 2002 | |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 2003 | skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2004 | : tcp_write_queue_head(sk); |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 2005 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2006 | tcp_for_write_queue_from(skb, sk) { |
| 2007 | if (skb == tcp_send_head(sk)) |
| 2008 | break; |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 2009 | if (!tcp_skb_timedout(sk, skb)) |
| 2010 | break; |
| 2011 | |
| 2012 | if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2013 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; |
| 2014 | tp->lost_out += tcp_skb_pcount(skb); |
Ilpo Järvinen | d8f4f22 | 2007-04-20 22:56:38 -0700 | [diff] [blame] | 2015 | tcp_verify_retransmit_hint(tp, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2016 | } |
| 2017 | } |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 2018 | |
| 2019 | tp->scoreboard_skb_hint = skb; |
| 2020 | |
Ilpo Järvinen | 005903b | 2007-08-09 14:44:16 +0300 | [diff] [blame] | 2021 | tcp_verify_left_out(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2022 | } |
| 2023 | } |
| 2024 | |
| 2025 | /* CWND moderation, preventing bursts due to too big ACKs |
| 2026 | * in dubious situations. |
| 2027 | */ |
| 2028 | static inline void tcp_moderate_cwnd(struct tcp_sock *tp) |
| 2029 | { |
| 2030 | tp->snd_cwnd = min(tp->snd_cwnd, |
| 2031 | tcp_packets_in_flight(tp)+tcp_max_burst(tp)); |
| 2032 | tp->snd_cwnd_stamp = tcp_time_stamp; |
| 2033 | } |
| 2034 | |
Stephen Hemminger | 72dc5b9 | 2006-06-05 17:30:08 -0700 | [diff] [blame] | 2035 | /* Lower bound on congestion window is slow start threshold |
| 2036 | * unless congestion avoidance choice decides to overide it. |
| 2037 | */ |
| 2038 | static inline u32 tcp_cwnd_min(const struct sock *sk) |
| 2039 | { |
| 2040 | const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; |
| 2041 | |
| 2042 | return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh; |
| 2043 | } |
| 2044 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2045 | /* Decrease cwnd each second ack. */ |
Ilpo Järvinen | 1e757f9 | 2007-07-30 19:48:37 -0700 | [diff] [blame] | 2046 | static void tcp_cwnd_down(struct sock *sk, int flag) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2047 | { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2048 | struct tcp_sock *tp = tcp_sk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2049 | int decr = tp->snd_cwnd_cnt + 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2050 | |
Ilpo Järvinen | 49ff4bb | 2007-08-02 19:47:59 -0700 | [diff] [blame] | 2051 | if ((flag&(FLAG_ANY_PROGRESS|FLAG_DSACKING_ACK)) || |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 2052 | (tcp_is_reno(tp) && !(flag&FLAG_NOT_DUP))) { |
Ilpo Järvinen | 1e757f9 | 2007-07-30 19:48:37 -0700 | [diff] [blame] | 2053 | tp->snd_cwnd_cnt = decr&1; |
| 2054 | decr >>= 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2055 | |
Ilpo Järvinen | 1e757f9 | 2007-07-30 19:48:37 -0700 | [diff] [blame] | 2056 | if (decr && tp->snd_cwnd > tcp_cwnd_min(sk)) |
| 2057 | tp->snd_cwnd -= decr; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2058 | |
Ilpo Järvinen | 1e757f9 | 2007-07-30 19:48:37 -0700 | [diff] [blame] | 2059 | tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1); |
| 2060 | tp->snd_cwnd_stamp = tcp_time_stamp; |
| 2061 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2062 | } |
| 2063 | |
| 2064 | /* Nothing was retransmitted or returned timestamp is less |
| 2065 | * than timestamp of the first retransmission. |
| 2066 | */ |
| 2067 | static inline int tcp_packet_delayed(struct tcp_sock *tp) |
| 2068 | { |
| 2069 | return !tp->retrans_stamp || |
| 2070 | (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && |
| 2071 | (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0); |
| 2072 | } |
| 2073 | |
| 2074 | /* Undo procedures. */ |
| 2075 | |
| 2076 | #if FASTRETRANS_DEBUG > 1 |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2077 | static void DBGUNDO(struct sock *sk, const char *msg) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2078 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2079 | struct tcp_sock *tp = tcp_sk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2080 | struct inet_sock *inet = inet_sk(sk); |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2081 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2082 | printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n", |
| 2083 | msg, |
| 2084 | NIPQUAD(inet->daddr), ntohs(inet->dport), |
Ilpo Järvinen | 83ae408 | 2007-08-09 14:37:30 +0300 | [diff] [blame] | 2085 | tp->snd_cwnd, tcp_left_out(tp), |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2086 | tp->snd_ssthresh, tp->prior_ssthresh, |
| 2087 | tp->packets_out); |
| 2088 | } |
| 2089 | #else |
| 2090 | #define DBGUNDO(x...) do { } while (0) |
| 2091 | #endif |
| 2092 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2093 | static void tcp_undo_cwr(struct sock *sk, const int undo) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2094 | { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2095 | struct tcp_sock *tp = tcp_sk(sk); |
| 2096 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2097 | if (tp->prior_ssthresh) { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2098 | const struct inet_connection_sock *icsk = inet_csk(sk); |
| 2099 | |
| 2100 | if (icsk->icsk_ca_ops->undo_cwnd) |
| 2101 | tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2102 | else |
| 2103 | tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1); |
| 2104 | |
| 2105 | if (undo && tp->prior_ssthresh > tp->snd_ssthresh) { |
| 2106 | tp->snd_ssthresh = tp->prior_ssthresh; |
| 2107 | TCP_ECN_withdraw_cwr(tp); |
| 2108 | } |
| 2109 | } else { |
| 2110 | tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh); |
| 2111 | } |
| 2112 | tcp_moderate_cwnd(tp); |
| 2113 | tp->snd_cwnd_stamp = tcp_time_stamp; |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 2114 | |
| 2115 | /* There is something screwy going on with the retrans hints after |
| 2116 | an undo */ |
Ilpo Järvinen | 5af4ec2 | 2007-09-20 11:30:48 -0700 | [diff] [blame] | 2117 | tcp_clear_all_retrans_hints(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2118 | } |
| 2119 | |
| 2120 | static inline int tcp_may_undo(struct tcp_sock *tp) |
| 2121 | { |
| 2122 | return tp->undo_marker && |
| 2123 | (!tp->undo_retrans || tcp_packet_delayed(tp)); |
| 2124 | } |
| 2125 | |
| 2126 | /* People celebrate: "We love our President!" */ |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2127 | static int tcp_try_undo_recovery(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2128 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2129 | struct tcp_sock *tp = tcp_sk(sk); |
| 2130 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2131 | if (tcp_may_undo(tp)) { |
| 2132 | /* Happy end! We did not retransmit anything |
| 2133 | * or our original transmission succeeded. |
| 2134 | */ |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2135 | DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans"); |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2136 | tcp_undo_cwr(sk, 1); |
| 2137 | if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2138 | NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO); |
| 2139 | else |
| 2140 | NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO); |
| 2141 | tp->undo_marker = 0; |
| 2142 | } |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 2143 | if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2144 | /* Hold old state until something *above* high_seq |
| 2145 | * is ACKed. For Reno it is MUST to prevent false |
| 2146 | * fast retransmits (RFC2582). SACK TCP is safe. */ |
| 2147 | tcp_moderate_cwnd(tp); |
| 2148 | return 1; |
| 2149 | } |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2150 | tcp_set_ca_state(sk, TCP_CA_Open); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2151 | return 0; |
| 2152 | } |
| 2153 | |
| 2154 | /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */ |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2155 | static void tcp_try_undo_dsack(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2156 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2157 | struct tcp_sock *tp = tcp_sk(sk); |
| 2158 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2159 | if (tp->undo_marker && !tp->undo_retrans) { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2160 | DBGUNDO(sk, "D-SACK"); |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2161 | tcp_undo_cwr(sk, 1); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2162 | tp->undo_marker = 0; |
| 2163 | NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO); |
| 2164 | } |
| 2165 | } |
| 2166 | |
| 2167 | /* Undo during fast recovery after partial ACK. */ |
| 2168 | |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2169 | static int tcp_try_undo_partial(struct sock *sk, int acked) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2170 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2171 | struct tcp_sock *tp = tcp_sk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2172 | /* Partial ACK arrived. Force Hoe's retransmit. */ |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 2173 | int failed = tcp_is_reno(tp) || tp->fackets_out>tp->reordering; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2174 | |
| 2175 | if (tcp_may_undo(tp)) { |
| 2176 | /* Plain luck! Hole if filled with delayed |
| 2177 | * packet, rather than with a retransmit. |
| 2178 | */ |
| 2179 | if (tp->retrans_out == 0) |
| 2180 | tp->retrans_stamp = 0; |
| 2181 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2182 | tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2183 | |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2184 | DBGUNDO(sk, "Hoe"); |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2185 | tcp_undo_cwr(sk, 0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2186 | NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO); |
| 2187 | |
| 2188 | /* So... Do not make Hoe's retransmit yet. |
| 2189 | * If the first packet was delayed, the rest |
| 2190 | * ones are most probably delayed as well. |
| 2191 | */ |
| 2192 | failed = 0; |
| 2193 | } |
| 2194 | return failed; |
| 2195 | } |
| 2196 | |
| 2197 | /* Undo during loss recovery after partial ACK. */ |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2198 | static int tcp_try_undo_loss(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2199 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2200 | struct tcp_sock *tp = tcp_sk(sk); |
| 2201 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2202 | if (tcp_may_undo(tp)) { |
| 2203 | struct sk_buff *skb; |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2204 | tcp_for_write_queue(skb, sk) { |
| 2205 | if (skb == tcp_send_head(sk)) |
| 2206 | break; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2207 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; |
| 2208 | } |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 2209 | |
Ilpo Järvinen | 5af4ec2 | 2007-09-20 11:30:48 -0700 | [diff] [blame] | 2210 | tcp_clear_all_retrans_hints(tp); |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 2211 | |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2212 | DBGUNDO(sk, "partial loss"); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2213 | tp->lost_out = 0; |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2214 | tcp_undo_cwr(sk, 1); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2215 | NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO); |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2216 | inet_csk(sk)->icsk_retransmits = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2217 | tp->undo_marker = 0; |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 2218 | if (tcp_is_sack(tp)) |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2219 | tcp_set_ca_state(sk, TCP_CA_Open); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2220 | return 1; |
| 2221 | } |
| 2222 | return 0; |
| 2223 | } |
| 2224 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2225 | static inline void tcp_complete_cwr(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2226 | { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2227 | struct tcp_sock *tp = tcp_sk(sk); |
Stephen Hemminger | 317a76f | 2005-06-23 12:19:55 -0700 | [diff] [blame] | 2228 | tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2229 | tp->snd_cwnd_stamp = tcp_time_stamp; |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2230 | tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2231 | } |
| 2232 | |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2233 | static void tcp_try_to_open(struct sock *sk, int flag) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2234 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2235 | struct tcp_sock *tp = tcp_sk(sk); |
| 2236 | |
Ilpo Järvinen | 86426c2 | 2007-08-09 14:45:17 +0300 | [diff] [blame] | 2237 | tcp_verify_left_out(tp); |
| 2238 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2239 | if (tp->retrans_out == 0) |
| 2240 | tp->retrans_stamp = 0; |
| 2241 | |
| 2242 | if (flag&FLAG_ECE) |
Ilpo Järvinen | 3cfe3ba | 2007-02-27 10:09:49 -0800 | [diff] [blame] | 2243 | tcp_enter_cwr(sk, 1); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2244 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2245 | if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2246 | int state = TCP_CA_Open; |
| 2247 | |
Ilpo Järvinen | d02596e | 2007-07-07 13:39:02 +0300 | [diff] [blame] | 2248 | if (tcp_left_out(tp) || tp->retrans_out || tp->undo_marker) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2249 | state = TCP_CA_Disorder; |
| 2250 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2251 | if (inet_csk(sk)->icsk_ca_state != state) { |
| 2252 | tcp_set_ca_state(sk, state); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2253 | tp->high_seq = tp->snd_nxt; |
| 2254 | } |
| 2255 | tcp_moderate_cwnd(tp); |
| 2256 | } else { |
Ilpo Järvinen | 1e757f9 | 2007-07-30 19:48:37 -0700 | [diff] [blame] | 2257 | tcp_cwnd_down(sk, flag); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2258 | } |
| 2259 | } |
| 2260 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 2261 | static void tcp_mtup_probe_failed(struct sock *sk) |
| 2262 | { |
| 2263 | struct inet_connection_sock *icsk = inet_csk(sk); |
| 2264 | |
| 2265 | icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1; |
| 2266 | icsk->icsk_mtup.probe_size = 0; |
| 2267 | } |
| 2268 | |
| 2269 | static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb) |
| 2270 | { |
| 2271 | struct tcp_sock *tp = tcp_sk(sk); |
| 2272 | struct inet_connection_sock *icsk = inet_csk(sk); |
| 2273 | |
| 2274 | /* FIXME: breaks with very large cwnd */ |
| 2275 | tp->prior_ssthresh = tcp_current_ssthresh(sk); |
| 2276 | tp->snd_cwnd = tp->snd_cwnd * |
| 2277 | tcp_mss_to_mtu(sk, tp->mss_cache) / |
| 2278 | icsk->icsk_mtup.probe_size; |
| 2279 | tp->snd_cwnd_cnt = 0; |
| 2280 | tp->snd_cwnd_stamp = tcp_time_stamp; |
| 2281 | tp->rcv_ssthresh = tcp_current_ssthresh(sk); |
| 2282 | |
| 2283 | icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size; |
| 2284 | icsk->icsk_mtup.probe_size = 0; |
| 2285 | tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); |
| 2286 | } |
| 2287 | |
| 2288 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2289 | /* Process an event, which can update packets-in-flight not trivially. |
| 2290 | * Main goal of this function is to calculate new estimate for left_out, |
| 2291 | * taking into account both packets sitting in receiver's buffer and |
| 2292 | * packets lost by network. |
| 2293 | * |
| 2294 | * Besides that it does CWND reduction, when packet loss is detected |
| 2295 | * and changes state of machine. |
| 2296 | * |
| 2297 | * It does _not_ decide what to send, it is made in function |
| 2298 | * tcp_xmit_retransmit_queue(). |
| 2299 | */ |
| 2300 | static void |
Ilpo Järvinen | 1b6d427b | 2007-08-09 14:53:36 +0300 | [diff] [blame] | 2301 | tcp_fastretrans_alert(struct sock *sk, int pkts_acked, int flag) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2302 | { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2303 | struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2304 | struct tcp_sock *tp = tcp_sk(sk); |
Ilpo Järvinen | 2e60529 | 2007-08-02 19:46:58 -0700 | [diff] [blame] | 2305 | int is_dupack = !(flag&(FLAG_SND_UNA_ADVANCED|FLAG_NOT_DUP)); |
| 2306 | int do_lost = is_dupack || ((flag&FLAG_DATA_SACKED) && |
| 2307 | (tp->fackets_out > tp->reordering)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2308 | |
| 2309 | /* Some technical things: |
| 2310 | * 1. Reno does not count dupacks (sacked_out) automatically. */ |
| 2311 | if (!tp->packets_out) |
| 2312 | tp->sacked_out = 0; |
Ilpo Järvinen | 91fed7a | 2007-10-09 01:24:15 -0700 | [diff] [blame] | 2313 | |
| 2314 | if (WARN_ON(!tp->sacked_out && tp->fackets_out)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2315 | tp->fackets_out = 0; |
| 2316 | |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2317 | /* Now state machine starts. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2318 | * A. ECE, hence prohibit cwnd undoing, the reduction is required. */ |
| 2319 | if (flag&FLAG_ECE) |
| 2320 | tp->prior_ssthresh = 0; |
| 2321 | |
| 2322 | /* B. In all the states check for reneging SACKs. */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2323 | if (tp->sacked_out && tcp_check_sack_reneging(sk)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2324 | return; |
| 2325 | |
| 2326 | /* C. Process data loss notification, provided it is valid. */ |
| 2327 | if ((flag&FLAG_DATA_LOST) && |
| 2328 | before(tp->snd_una, tp->high_seq) && |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2329 | icsk->icsk_ca_state != TCP_CA_Open && |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2330 | tp->fackets_out > tp->reordering) { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2331 | tcp_mark_head_lost(sk, tp->fackets_out-tp->reordering, tp->high_seq); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2332 | NET_INC_STATS_BH(LINUX_MIB_TCPLOSS); |
| 2333 | } |
| 2334 | |
Ilpo Järvinen | 005903b | 2007-08-09 14:44:16 +0300 | [diff] [blame] | 2335 | /* D. Check consistency of the current state. */ |
| 2336 | tcp_verify_left_out(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2337 | |
| 2338 | /* E. Check state exit conditions. State can be terminated |
| 2339 | * when high_seq is ACKed. */ |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2340 | if (icsk->icsk_ca_state == TCP_CA_Open) { |
Ilpo Järvinen | 7b0eb22 | 2007-02-21 23:03:35 -0800 | [diff] [blame] | 2341 | BUG_TRAP(tp->retrans_out == 0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2342 | tp->retrans_stamp = 0; |
| 2343 | } else if (!before(tp->snd_una, tp->high_seq)) { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2344 | switch (icsk->icsk_ca_state) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2345 | case TCP_CA_Loss: |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2346 | icsk->icsk_retransmits = 0; |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2347 | if (tcp_try_undo_recovery(sk)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2348 | return; |
| 2349 | break; |
| 2350 | |
| 2351 | case TCP_CA_CWR: |
| 2352 | /* CWR is to be held something *above* high_seq |
| 2353 | * is ACKed for CWR bit to reach receiver. */ |
| 2354 | if (tp->snd_una != tp->high_seq) { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2355 | tcp_complete_cwr(sk); |
| 2356 | tcp_set_ca_state(sk, TCP_CA_Open); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2357 | } |
| 2358 | break; |
| 2359 | |
| 2360 | case TCP_CA_Disorder: |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2361 | tcp_try_undo_dsack(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2362 | if (!tp->undo_marker || |
| 2363 | /* For SACK case do not Open to allow to undo |
| 2364 | * catching for all duplicate ACKs. */ |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 2365 | tcp_is_reno(tp) || tp->snd_una != tp->high_seq) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2366 | tp->undo_marker = 0; |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2367 | tcp_set_ca_state(sk, TCP_CA_Open); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2368 | } |
| 2369 | break; |
| 2370 | |
| 2371 | case TCP_CA_Recovery: |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 2372 | if (tcp_is_reno(tp)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2373 | tcp_reset_reno_sack(tp); |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2374 | if (tcp_try_undo_recovery(sk)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2375 | return; |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2376 | tcp_complete_cwr(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2377 | break; |
| 2378 | } |
| 2379 | } |
| 2380 | |
| 2381 | /* F. Process state. */ |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2382 | switch (icsk->icsk_ca_state) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2383 | case TCP_CA_Recovery: |
Ilpo Järvinen | 2e60529 | 2007-08-02 19:46:58 -0700 | [diff] [blame] | 2384 | if (!(flag & FLAG_SND_UNA_ADVANCED)) { |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 2385 | if (tcp_is_reno(tp) && is_dupack) |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2386 | tcp_add_reno_sack(sk); |
Ilpo Järvinen | 1b6d427b | 2007-08-09 14:53:36 +0300 | [diff] [blame] | 2387 | } else |
| 2388 | do_lost = tcp_try_undo_partial(sk, pkts_acked); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2389 | break; |
| 2390 | case TCP_CA_Loss: |
| 2391 | if (flag&FLAG_DATA_ACKED) |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2392 | icsk->icsk_retransmits = 0; |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2393 | if (!tcp_try_undo_loss(sk)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2394 | tcp_moderate_cwnd(tp); |
| 2395 | tcp_xmit_retransmit_queue(sk); |
| 2396 | return; |
| 2397 | } |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2398 | if (icsk->icsk_ca_state != TCP_CA_Open) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2399 | return; |
| 2400 | /* Loss is undone; fall through to processing in Open state. */ |
| 2401 | default: |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 2402 | if (tcp_is_reno(tp)) { |
Ilpo Järvinen | 2e60529 | 2007-08-02 19:46:58 -0700 | [diff] [blame] | 2403 | if (flag & FLAG_SND_UNA_ADVANCED) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2404 | tcp_reset_reno_sack(tp); |
| 2405 | if (is_dupack) |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2406 | tcp_add_reno_sack(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2407 | } |
| 2408 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2409 | if (icsk->icsk_ca_state == TCP_CA_Disorder) |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2410 | tcp_try_undo_dsack(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2411 | |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2412 | if (!tcp_time_to_recover(sk)) { |
| 2413 | tcp_try_to_open(sk, flag); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2414 | return; |
| 2415 | } |
| 2416 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 2417 | /* MTU probe failure: don't reduce cwnd */ |
| 2418 | if (icsk->icsk_ca_state < TCP_CA_CWR && |
| 2419 | icsk->icsk_mtup.probe_size && |
John Heffner | 0e7b136 | 2006-03-20 21:32:58 -0800 | [diff] [blame] | 2420 | tp->snd_una == tp->mtu_probe.probe_seq_start) { |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 2421 | tcp_mtup_probe_failed(sk); |
| 2422 | /* Restores the reduction we did in tcp_mtup_probe() */ |
| 2423 | tp->snd_cwnd++; |
| 2424 | tcp_simple_retransmit(sk); |
| 2425 | return; |
| 2426 | } |
| 2427 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2428 | /* Otherwise enter Recovery state */ |
| 2429 | |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 2430 | if (tcp_is_reno(tp)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2431 | NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY); |
| 2432 | else |
| 2433 | NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY); |
| 2434 | |
| 2435 | tp->high_seq = tp->snd_nxt; |
| 2436 | tp->prior_ssthresh = 0; |
| 2437 | tp->undo_marker = tp->snd_una; |
| 2438 | tp->undo_retrans = tp->retrans_out; |
| 2439 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2440 | if (icsk->icsk_ca_state < TCP_CA_CWR) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2441 | if (!(flag&FLAG_ECE)) |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2442 | tp->prior_ssthresh = tcp_current_ssthresh(sk); |
| 2443 | tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2444 | TCP_ECN_queue_cwr(tp); |
| 2445 | } |
| 2446 | |
Stephen Hemminger | 9772efb | 2005-11-10 17:09:53 -0800 | [diff] [blame] | 2447 | tp->bytes_acked = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2448 | tp->snd_cwnd_cnt = 0; |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2449 | tcp_set_ca_state(sk, TCP_CA_Recovery); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2450 | } |
| 2451 | |
Ilpo Järvinen | 2e60529 | 2007-08-02 19:46:58 -0700 | [diff] [blame] | 2452 | if (do_lost || tcp_head_timedout(sk)) |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2453 | tcp_update_scoreboard(sk); |
Ilpo Järvinen | 1e757f9 | 2007-07-30 19:48:37 -0700 | [diff] [blame] | 2454 | tcp_cwnd_down(sk, flag); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2455 | tcp_xmit_retransmit_queue(sk); |
| 2456 | } |
| 2457 | |
| 2458 | /* Read draft-ietf-tcplw-high-performance before mucking |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 2459 | * with this code. (Supersedes RFC1323) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2460 | */ |
Stephen Hemminger | 2d2abba | 2005-11-10 16:56:12 -0800 | [diff] [blame] | 2461 | static void tcp_ack_saw_tstamp(struct sock *sk, int flag) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2462 | { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2463 | /* RTTM Rule: A TSecr value received in a segment is used to |
| 2464 | * update the averaged RTT measurement only if the segment |
| 2465 | * acknowledges some new data, i.e., only if it advances the |
| 2466 | * left edge of the send window. |
| 2467 | * |
| 2468 | * See draft-ietf-tcplw-high-performance-00, section 3.3. |
| 2469 | * 1998/04/10 Andrey V. Savochkin <saw@msu.ru> |
| 2470 | * |
| 2471 | * Changed: reset backoff as soon as we see the first valid sample. |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 2472 | * If we do not, we get strongly overestimated rto. With timestamps |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2473 | * samples are accepted even from very old segments: f.e., when rtt=1 |
| 2474 | * increases to 8, we retransmit 5 times and after 8 seconds delayed |
| 2475 | * answer arrives rto becomes 120 seconds! If at least one of segments |
| 2476 | * in window is lost... Voila. --ANK (010210) |
| 2477 | */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2478 | struct tcp_sock *tp = tcp_sk(sk); |
| 2479 | const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr; |
Stephen Hemminger | 2d2abba | 2005-11-10 16:56:12 -0800 | [diff] [blame] | 2480 | tcp_rtt_estimator(sk, seq_rtt); |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2481 | tcp_set_rto(sk); |
| 2482 | inet_csk(sk)->icsk_backoff = 0; |
| 2483 | tcp_bound_rto(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2484 | } |
| 2485 | |
Stephen Hemminger | 2d2abba | 2005-11-10 16:56:12 -0800 | [diff] [blame] | 2486 | static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2487 | { |
| 2488 | /* We don't have a timestamp. Can only use |
| 2489 | * packets that are not retransmitted to determine |
| 2490 | * rtt estimates. Also, we must not reset the |
| 2491 | * backoff for rto until we get a non-retransmitted |
| 2492 | * packet. This allows us to deal with a situation |
| 2493 | * where the network delay has increased suddenly. |
| 2494 | * I.e. Karn's algorithm. (SIGCOMM '87, p5.) |
| 2495 | */ |
| 2496 | |
| 2497 | if (flag & FLAG_RETRANS_DATA_ACKED) |
| 2498 | return; |
| 2499 | |
Stephen Hemminger | 2d2abba | 2005-11-10 16:56:12 -0800 | [diff] [blame] | 2500 | tcp_rtt_estimator(sk, seq_rtt); |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2501 | tcp_set_rto(sk); |
| 2502 | inet_csk(sk)->icsk_backoff = 0; |
| 2503 | tcp_bound_rto(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2504 | } |
| 2505 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2506 | static inline void tcp_ack_update_rtt(struct sock *sk, const int flag, |
Stephen Hemminger | 2d2abba | 2005-11-10 16:56:12 -0800 | [diff] [blame] | 2507 | const s32 seq_rtt) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2508 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2509 | const struct tcp_sock *tp = tcp_sk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2510 | /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */ |
| 2511 | if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) |
Stephen Hemminger | 2d2abba | 2005-11-10 16:56:12 -0800 | [diff] [blame] | 2512 | tcp_ack_saw_tstamp(sk, flag); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2513 | else if (seq_rtt >= 0) |
Stephen Hemminger | 2d2abba | 2005-11-10 16:56:12 -0800 | [diff] [blame] | 2514 | tcp_ack_no_tstamp(sk, seq_rtt, flag); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2515 | } |
| 2516 | |
Stephen Hemminger | 1675134 | 2007-07-16 18:35:52 -0700 | [diff] [blame] | 2517 | static void tcp_cong_avoid(struct sock *sk, u32 ack, |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 2518 | u32 in_flight, int good) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2519 | { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2520 | const struct inet_connection_sock *icsk = inet_csk(sk); |
Stephen Hemminger | 1675134 | 2007-07-16 18:35:52 -0700 | [diff] [blame] | 2521 | icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight, good); |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2522 | tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2523 | } |
| 2524 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2525 | /* Restart timer after forward progress on connection. |
| 2526 | * RFC2988 recommends to restart timer to now+rto. |
| 2527 | */ |
Ilpo Järvinen | 6728e7d | 2007-08-24 22:53:26 -0700 | [diff] [blame] | 2528 | static void tcp_rearm_rto(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2529 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2530 | struct tcp_sock *tp = tcp_sk(sk); |
| 2531 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2532 | if (!tp->packets_out) { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2533 | inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2534 | } else { |
Arnaldo Carvalho de Melo | 3f421ba | 2005-08-09 20:11:08 -0700 | [diff] [blame] | 2535 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, inet_csk(sk)->icsk_rto, TCP_RTO_MAX); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2536 | } |
| 2537 | } |
| 2538 | |
Ilpo Järvinen | 7c46a03 | 2007-09-20 11:33:43 -0700 | [diff] [blame] | 2539 | /* If we get here, the whole TSO packet has not been acked. */ |
Ilpo Järvinen | 13fcf85 | 2007-10-09 01:28:45 -0700 | [diff] [blame] | 2540 | static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2541 | { |
| 2542 | struct tcp_sock *tp = tcp_sk(sk); |
Ilpo Järvinen | 7c46a03 | 2007-09-20 11:33:43 -0700 | [diff] [blame] | 2543 | u32 packets_acked; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2544 | |
Ilpo Järvinen | 7c46a03 | 2007-09-20 11:33:43 -0700 | [diff] [blame] | 2545 | BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2546 | |
| 2547 | packets_acked = tcp_skb_pcount(skb); |
Ilpo Järvinen | 7c46a03 | 2007-09-20 11:33:43 -0700 | [diff] [blame] | 2548 | if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2549 | return 0; |
| 2550 | packets_acked -= tcp_skb_pcount(skb); |
| 2551 | |
| 2552 | if (packets_acked) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2553 | BUG_ON(tcp_skb_pcount(skb) == 0); |
Ilpo Järvinen | 7c46a03 | 2007-09-20 11:33:43 -0700 | [diff] [blame] | 2554 | BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2555 | } |
| 2556 | |
Ilpo Järvinen | 13fcf85 | 2007-10-09 01:28:45 -0700 | [diff] [blame] | 2557 | return packets_acked; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2558 | } |
| 2559 | |
Ilpo Järvinen | 7c46a03 | 2007-09-20 11:33:43 -0700 | [diff] [blame] | 2560 | /* Remove acknowledged frames from the retransmission queue. If our packet |
| 2561 | * is before the ack sequence we can discard it as it's confirmed to have |
| 2562 | * arrived at the other end. |
| 2563 | */ |
| 2564 | static int tcp_clean_rtx_queue(struct sock *sk, s32 *seq_rtt_p) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2565 | { |
| 2566 | struct tcp_sock *tp = tcp_sk(sk); |
Stephen Hemminger | 2d2abba | 2005-11-10 16:56:12 -0800 | [diff] [blame] | 2567 | const struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2568 | struct sk_buff *skb; |
Ilpo Järvinen | 7c46a03 | 2007-09-20 11:33:43 -0700 | [diff] [blame] | 2569 | u32 now = tcp_time_stamp; |
Ilpo Järvinen | 13fcf85 | 2007-10-09 01:28:45 -0700 | [diff] [blame] | 2570 | int fully_acked = 1; |
Ilpo Järvinen | 7c46a03 | 2007-09-20 11:33:43 -0700 | [diff] [blame] | 2571 | int flag = 0; |
Ilpo Järvinen | 6418204 | 2007-05-31 21:37:55 -0700 | [diff] [blame] | 2572 | int prior_packets = tp->packets_out; |
Ilpo Järvinen | 7c46a03 | 2007-09-20 11:33:43 -0700 | [diff] [blame] | 2573 | s32 seq_rtt = -1; |
Ilpo Järvinen | b9ce204 | 2007-06-15 15:08:43 -0700 | [diff] [blame] | 2574 | ktime_t last_ackt = net_invalid_timestamp(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2575 | |
Ilpo Järvinen | 7c46a03 | 2007-09-20 11:33:43 -0700 | [diff] [blame] | 2576 | while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) { |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2577 | struct tcp_skb_cb *scb = TCP_SKB_CB(skb); |
Ilpo Järvinen | 13fcf85 | 2007-10-09 01:28:45 -0700 | [diff] [blame] | 2578 | u32 end_seq; |
| 2579 | u32 packets_acked; |
Ilpo Järvinen | 7c46a03 | 2007-09-20 11:33:43 -0700 | [diff] [blame] | 2580 | u8 sacked = scb->sacked; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2581 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2582 | if (after(scb->end_seq, tp->snd_una)) { |
Ilpo Järvinen | 13fcf85 | 2007-10-09 01:28:45 -0700 | [diff] [blame] | 2583 | if (tcp_skb_pcount(skb) == 1 || |
| 2584 | !after(tp->snd_una, scb->seq)) |
| 2585 | break; |
| 2586 | |
| 2587 | packets_acked = tcp_tso_acked(sk, skb); |
| 2588 | if (!packets_acked) |
| 2589 | break; |
| 2590 | |
| 2591 | fully_acked = 0; |
| 2592 | end_seq = tp->snd_una; |
| 2593 | } else { |
| 2594 | packets_acked = tcp_skb_pcount(skb); |
| 2595 | end_seq = scb->end_seq; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2596 | } |
| 2597 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 2598 | /* MTU probing checks */ |
Ilpo Järvinen | 7c46a03 | 2007-09-20 11:33:43 -0700 | [diff] [blame] | 2599 | if (fully_acked && icsk->icsk_mtup.probe_size && |
| 2600 | !after(tp->mtu_probe.probe_seq_end, scb->end_seq)) { |
| 2601 | tcp_mtup_probe_success(sk, skb); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 2602 | } |
| 2603 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2604 | if (sacked) { |
| 2605 | if (sacked & TCPCB_RETRANS) { |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 2606 | if (sacked & TCPCB_SACKED_RETRANS) |
Ilpo Järvinen | 13fcf85 | 2007-10-09 01:28:45 -0700 | [diff] [blame] | 2607 | tp->retrans_out -= packets_acked; |
Ilpo Järvinen | 7c46a03 | 2007-09-20 11:33:43 -0700 | [diff] [blame] | 2608 | flag |= FLAG_RETRANS_DATA_ACKED; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2609 | seq_rtt = -1; |
Ilpo Järvinen | 009a2e3 | 2007-09-20 11:34:38 -0700 | [diff] [blame] | 2610 | if ((flag & FLAG_DATA_ACKED) || |
| 2611 | (packets_acked > 1)) |
| 2612 | flag |= FLAG_NONHEAD_RETRANS_ACKED; |
Stephen Hemminger | 2d2abba | 2005-11-10 16:56:12 -0800 | [diff] [blame] | 2613 | } else if (seq_rtt < 0) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2614 | seq_rtt = now - scb->when; |
Ilpo Järvinen | 13fcf85 | 2007-10-09 01:28:45 -0700 | [diff] [blame] | 2615 | if (fully_acked) |
| 2616 | last_ackt = skb->tstamp; |
Patrick McHardy | a61bbcf | 2005-08-14 17:24:31 -0700 | [diff] [blame] | 2617 | } |
Ilpo Järvinen | 7c46a03 | 2007-09-20 11:33:43 -0700 | [diff] [blame] | 2618 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2619 | if (sacked & TCPCB_SACKED_ACKED) |
Ilpo Järvinen | 13fcf85 | 2007-10-09 01:28:45 -0700 | [diff] [blame] | 2620 | tp->sacked_out -= packets_acked; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2621 | if (sacked & TCPCB_LOST) |
Ilpo Järvinen | 13fcf85 | 2007-10-09 01:28:45 -0700 | [diff] [blame] | 2622 | tp->lost_out -= packets_acked; |
Ilpo Järvinen | 7c46a03 | 2007-09-20 11:33:43 -0700 | [diff] [blame] | 2623 | |
| 2624 | if ((sacked & TCPCB_URG) && tp->urg_mode && |
| 2625 | !before(end_seq, tp->snd_up)) |
| 2626 | tp->urg_mode = 0; |
Stephen Hemminger | 2d2abba | 2005-11-10 16:56:12 -0800 | [diff] [blame] | 2627 | } else if (seq_rtt < 0) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2628 | seq_rtt = now - scb->when; |
Ilpo Järvinen | 13fcf85 | 2007-10-09 01:28:45 -0700 | [diff] [blame] | 2629 | if (fully_acked) |
| 2630 | last_ackt = skb->tstamp; |
Stephen Hemminger | 2d2abba | 2005-11-10 16:56:12 -0800 | [diff] [blame] | 2631 | } |
Ilpo Järvinen | 13fcf85 | 2007-10-09 01:28:45 -0700 | [diff] [blame] | 2632 | tp->packets_out -= packets_acked; |
| 2633 | |
Ilpo Järvinen | 009a2e3 | 2007-09-20 11:34:38 -0700 | [diff] [blame] | 2634 | /* Initial outgoing SYN's get put onto the write_queue |
| 2635 | * just like anything else we transmit. It is not |
| 2636 | * true data, and if we misinform our callers that |
| 2637 | * this ACK acks real data, we will erroneously exit |
| 2638 | * connection startup slow start one packet too |
| 2639 | * quickly. This is severely frowned upon behavior. |
| 2640 | */ |
| 2641 | if (!(scb->flags & TCPCB_FLAG_SYN)) { |
| 2642 | flag |= FLAG_DATA_ACKED; |
| 2643 | } else { |
| 2644 | flag |= FLAG_SYN_ACKED; |
| 2645 | tp->retrans_stamp = 0; |
| 2646 | } |
| 2647 | |
Ilpo Järvinen | 13fcf85 | 2007-10-09 01:28:45 -0700 | [diff] [blame] | 2648 | if (!fully_acked) |
| 2649 | break; |
| 2650 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2651 | tcp_unlink_write_queue(skb, sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2652 | sk_stream_free_skb(sk, skb); |
Ilpo Järvinen | 5af4ec2 | 2007-09-20 11:30:48 -0700 | [diff] [blame] | 2653 | tcp_clear_all_retrans_hints(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2654 | } |
| 2655 | |
Ilpo Järvinen | 7c46a03 | 2007-09-20 11:33:43 -0700 | [diff] [blame] | 2656 | if (flag & FLAG_ACKED) { |
Ilpo Järvinen | 6418204 | 2007-05-31 21:37:55 -0700 | [diff] [blame] | 2657 | u32 pkts_acked = prior_packets - tp->packets_out; |
Stephen Hemminger | 164891a | 2007-04-23 22:26:16 -0700 | [diff] [blame] | 2658 | const struct tcp_congestion_ops *ca_ops |
| 2659 | = inet_csk(sk)->icsk_ca_ops; |
| 2660 | |
Ilpo Järvinen | 7c46a03 | 2007-09-20 11:33:43 -0700 | [diff] [blame] | 2661 | tcp_ack_update_rtt(sk, flag, seq_rtt); |
Ilpo Järvinen | 6728e7d | 2007-08-24 22:53:26 -0700 | [diff] [blame] | 2662 | tcp_rearm_rto(sk); |
Stephen Hemminger | 317a76f | 2005-06-23 12:19:55 -0700 | [diff] [blame] | 2663 | |
Ilpo Järvinen | 91fed7a | 2007-10-09 01:24:15 -0700 | [diff] [blame] | 2664 | tp->fackets_out -= min(pkts_acked, tp->fackets_out); |
Ilpo Järvinen | 13fcf85 | 2007-10-09 01:28:45 -0700 | [diff] [blame] | 2665 | /* hint's skb might be NULL but we don't need to care */ |
| 2666 | tp->fastpath_cnt_hint -= min_t(u32, pkts_acked, |
| 2667 | tp->fastpath_cnt_hint); |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 2668 | if (tcp_is_reno(tp)) |
Ilpo Järvinen | 1b6d427b | 2007-08-09 14:53:36 +0300 | [diff] [blame] | 2669 | tcp_remove_reno_sacks(sk, pkts_acked); |
| 2670 | |
Stephen Hemminger | 30cfd0b | 2007-07-25 23:49:34 -0700 | [diff] [blame] | 2671 | if (ca_ops->pkts_acked) { |
| 2672 | s32 rtt_us = -1; |
Ilpo Järvinen | b9ce204 | 2007-06-15 15:08:43 -0700 | [diff] [blame] | 2673 | |
Stephen Hemminger | 30cfd0b | 2007-07-25 23:49:34 -0700 | [diff] [blame] | 2674 | /* Is the ACK triggering packet unambiguous? */ |
Ilpo Järvinen | 7c46a03 | 2007-09-20 11:33:43 -0700 | [diff] [blame] | 2675 | if (!(flag & FLAG_RETRANS_DATA_ACKED)) { |
Stephen Hemminger | 30cfd0b | 2007-07-25 23:49:34 -0700 | [diff] [blame] | 2676 | /* High resolution needed and available? */ |
| 2677 | if (ca_ops->flags & TCP_CONG_RTT_STAMP && |
| 2678 | !ktime_equal(last_ackt, |
| 2679 | net_invalid_timestamp())) |
| 2680 | rtt_us = ktime_us_delta(ktime_get_real(), |
| 2681 | last_ackt); |
| 2682 | else if (seq_rtt > 0) |
| 2683 | rtt_us = jiffies_to_usecs(seq_rtt); |
| 2684 | } |
| 2685 | |
| 2686 | ca_ops->pkts_acked(sk, pkts_acked, rtt_us); |
| 2687 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2688 | } |
| 2689 | |
| 2690 | #if FASTRETRANS_DEBUG > 0 |
| 2691 | BUG_TRAP((int)tp->sacked_out >= 0); |
| 2692 | BUG_TRAP((int)tp->lost_out >= 0); |
| 2693 | BUG_TRAP((int)tp->retrans_out >= 0); |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 2694 | if (!tp->packets_out && tcp_is_sack(tp)) { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2695 | const struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2696 | if (tp->lost_out) { |
| 2697 | printk(KERN_DEBUG "Leak l=%u %d\n", |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2698 | tp->lost_out, icsk->icsk_ca_state); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2699 | tp->lost_out = 0; |
| 2700 | } |
| 2701 | if (tp->sacked_out) { |
| 2702 | printk(KERN_DEBUG "Leak s=%u %d\n", |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2703 | tp->sacked_out, icsk->icsk_ca_state); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2704 | tp->sacked_out = 0; |
| 2705 | } |
| 2706 | if (tp->retrans_out) { |
| 2707 | printk(KERN_DEBUG "Leak r=%u %d\n", |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2708 | tp->retrans_out, icsk->icsk_ca_state); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2709 | tp->retrans_out = 0; |
| 2710 | } |
| 2711 | } |
| 2712 | #endif |
| 2713 | *seq_rtt_p = seq_rtt; |
Ilpo Järvinen | 7c46a03 | 2007-09-20 11:33:43 -0700 | [diff] [blame] | 2714 | return flag; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2715 | } |
| 2716 | |
| 2717 | static void tcp_ack_probe(struct sock *sk) |
| 2718 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2719 | const struct tcp_sock *tp = tcp_sk(sk); |
| 2720 | struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2721 | |
| 2722 | /* Was it a usable window open? */ |
| 2723 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2724 | if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2725 | tp->snd_una + tp->snd_wnd)) { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2726 | icsk->icsk_backoff = 0; |
| 2727 | inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2728 | /* Socket must be waked up by subsequent tcp_data_snd_check(). |
| 2729 | * This function is not for random using! |
| 2730 | */ |
| 2731 | } else { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2732 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, |
Arnaldo Carvalho de Melo | 3f421ba | 2005-08-09 20:11:08 -0700 | [diff] [blame] | 2733 | min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), |
| 2734 | TCP_RTO_MAX); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2735 | } |
| 2736 | } |
| 2737 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2738 | static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2739 | { |
| 2740 | return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) || |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2741 | inet_csk(sk)->icsk_ca_state != TCP_CA_Open); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2742 | } |
| 2743 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2744 | static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2745 | { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2746 | const struct tcp_sock *tp = tcp_sk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2747 | return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) && |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2748 | !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2749 | } |
| 2750 | |
| 2751 | /* Check that window update is acceptable. |
| 2752 | * The function assumes that snd_una<=ack<=snd_next. |
| 2753 | */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2754 | static inline int tcp_may_update_window(const struct tcp_sock *tp, const u32 ack, |
| 2755 | const u32 ack_seq, const u32 nwin) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2756 | { |
| 2757 | return (after(ack, tp->snd_una) || |
| 2758 | after(ack_seq, tp->snd_wl1) || |
| 2759 | (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd)); |
| 2760 | } |
| 2761 | |
| 2762 | /* Update our send window. |
| 2763 | * |
| 2764 | * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2 |
| 2765 | * and in FreeBSD. NetBSD's one is even worse.) is wrong. |
| 2766 | */ |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2767 | static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack, |
| 2768 | u32 ack_seq) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2769 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2770 | struct tcp_sock *tp = tcp_sk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2771 | int flag = 0; |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 2772 | u32 nwin = ntohs(tcp_hdr(skb)->window); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2773 | |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 2774 | if (likely(!tcp_hdr(skb)->syn)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2775 | nwin <<= tp->rx_opt.snd_wscale; |
| 2776 | |
| 2777 | if (tcp_may_update_window(tp, ack, ack_seq, nwin)) { |
| 2778 | flag |= FLAG_WIN_UPDATE; |
| 2779 | tcp_update_wl(tp, ack, ack_seq); |
| 2780 | |
| 2781 | if (tp->snd_wnd != nwin) { |
| 2782 | tp->snd_wnd = nwin; |
| 2783 | |
| 2784 | /* Note, it is the only place, where |
| 2785 | * fast path is recovered for sending TCP. |
| 2786 | */ |
Herbert Xu | 2ad4106 | 2005-10-27 18:47:46 +1000 | [diff] [blame] | 2787 | tp->pred_flags = 0; |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2788 | tcp_fast_path_check(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2789 | |
| 2790 | if (nwin > tp->max_window) { |
| 2791 | tp->max_window = nwin; |
Arnaldo Carvalho de Melo | d83d846 | 2005-12-13 23:26:10 -0800 | [diff] [blame] | 2792 | tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2793 | } |
| 2794 | } |
| 2795 | } |
| 2796 | |
| 2797 | tp->snd_una = ack; |
| 2798 | |
| 2799 | return flag; |
| 2800 | } |
| 2801 | |
Ilpo Järvinen | 9ead9a1 | 2007-02-21 22:56:19 -0800 | [diff] [blame] | 2802 | /* A very conservative spurious RTO response algorithm: reduce cwnd and |
| 2803 | * continue in congestion avoidance. |
| 2804 | */ |
| 2805 | static void tcp_conservative_spur_to_response(struct tcp_sock *tp) |
| 2806 | { |
| 2807 | tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); |
Ilpo Järvinen | aa8b6a7 | 2007-02-21 23:06:03 -0800 | [diff] [blame] | 2808 | tp->snd_cwnd_cnt = 0; |
Ilpo Järvinen | 4632365 | 2007-05-19 13:56:23 -0700 | [diff] [blame] | 2809 | TCP_ECN_queue_cwr(tp); |
Ilpo Järvinen | 9ead9a1 | 2007-02-21 22:56:19 -0800 | [diff] [blame] | 2810 | tcp_moderate_cwnd(tp); |
| 2811 | } |
| 2812 | |
Ilpo Järvinen | 3cfe3ba | 2007-02-27 10:09:49 -0800 | [diff] [blame] | 2813 | /* A conservative spurious RTO response algorithm: reduce cwnd using |
| 2814 | * rate halving and continue in congestion avoidance. |
| 2815 | */ |
| 2816 | static void tcp_ratehalving_spur_to_response(struct sock *sk) |
| 2817 | { |
Ilpo Järvinen | 3cfe3ba | 2007-02-27 10:09:49 -0800 | [diff] [blame] | 2818 | tcp_enter_cwr(sk, 0); |
Ilpo Järvinen | 3cfe3ba | 2007-02-27 10:09:49 -0800 | [diff] [blame] | 2819 | } |
| 2820 | |
Ilpo Järvinen | e317f6f | 2007-03-02 13:34:19 -0800 | [diff] [blame] | 2821 | static void tcp_undo_spur_to_response(struct sock *sk, int flag) |
Ilpo Järvinen | 3cfe3ba | 2007-02-27 10:09:49 -0800 | [diff] [blame] | 2822 | { |
Ilpo Järvinen | e317f6f | 2007-03-02 13:34:19 -0800 | [diff] [blame] | 2823 | if (flag&FLAG_ECE) |
| 2824 | tcp_ratehalving_spur_to_response(sk); |
| 2825 | else |
| 2826 | tcp_undo_cwr(sk, 1); |
Ilpo Järvinen | 3cfe3ba | 2007-02-27 10:09:49 -0800 | [diff] [blame] | 2827 | } |
| 2828 | |
Ilpo Järvinen | 30935cf | 2007-02-21 23:01:36 -0800 | [diff] [blame] | 2829 | /* F-RTO spurious RTO detection algorithm (RFC4138) |
| 2830 | * |
Ilpo Järvinen | 6408d20 | 2007-02-21 23:04:11 -0800 | [diff] [blame] | 2831 | * F-RTO affects during two new ACKs following RTO (well, almost, see inline |
| 2832 | * comments). State (ACK number) is kept in frto_counter. When ACK advances |
| 2833 | * window (but not to or beyond highest sequence sent before RTO): |
Ilpo Järvinen | 30935cf | 2007-02-21 23:01:36 -0800 | [diff] [blame] | 2834 | * On First ACK, send two new segments out. |
| 2835 | * On Second ACK, RTO was likely spurious. Do spurious response (response |
| 2836 | * algorithm is not part of the F-RTO detection algorithm |
| 2837 | * given in RFC4138 but can be selected separately). |
| 2838 | * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss |
Ilpo Järvinen | d551e45 | 2007-04-30 00:42:20 -0700 | [diff] [blame] | 2839 | * and TCP falls back to conventional RTO recovery. F-RTO allows overriding |
| 2840 | * of Nagle, this is done using frto_counter states 2 and 3, when a new data |
| 2841 | * segment of any size sent during F-RTO, state 2 is upgraded to 3. |
Ilpo Järvinen | 30935cf | 2007-02-21 23:01:36 -0800 | [diff] [blame] | 2842 | * |
| 2843 | * Rationale: if the RTO was spurious, new ACKs should arrive from the |
| 2844 | * original window even after we transmit two new data segments. |
| 2845 | * |
Ilpo Järvinen | 4dc2665 | 2007-02-21 23:16:11 -0800 | [diff] [blame] | 2846 | * SACK version: |
| 2847 | * on first step, wait until first cumulative ACK arrives, then move to |
| 2848 | * the second step. In second step, the next ACK decides. |
| 2849 | * |
Ilpo Järvinen | 30935cf | 2007-02-21 23:01:36 -0800 | [diff] [blame] | 2850 | * F-RTO is implemented (mainly) in four functions: |
| 2851 | * - tcp_use_frto() is used to determine if TCP is can use F-RTO |
| 2852 | * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is |
| 2853 | * called when tcp_use_frto() showed green light |
| 2854 | * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm |
| 2855 | * - tcp_enter_frto_loss() is called if there is not enough evidence |
| 2856 | * to prove that the RTO is indeed spurious. It transfers the control |
| 2857 | * from F-RTO to the conventional RTO recovery |
| 2858 | */ |
Ilpo Järvinen | 2e60529 | 2007-08-02 19:46:58 -0700 | [diff] [blame] | 2859 | static int tcp_process_frto(struct sock *sk, int flag) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2860 | { |
| 2861 | struct tcp_sock *tp = tcp_sk(sk); |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2862 | |
Ilpo Järvinen | 005903b | 2007-08-09 14:44:16 +0300 | [diff] [blame] | 2863 | tcp_verify_left_out(tp); |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2864 | |
Ilpo Järvinen | 7487c48 | 2007-02-21 23:02:30 -0800 | [diff] [blame] | 2865 | /* Duplicate the behavior from Loss state (fastretrans_alert) */ |
| 2866 | if (flag&FLAG_DATA_ACKED) |
| 2867 | inet_csk(sk)->icsk_retransmits = 0; |
| 2868 | |
Ilpo Järvinen | 009a2e3 | 2007-09-20 11:34:38 -0700 | [diff] [blame] | 2869 | if ((flag & FLAG_NONHEAD_RETRANS_ACKED) || |
| 2870 | ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED))) |
| 2871 | tp->undo_marker = 0; |
| 2872 | |
Ilpo Järvinen | 95c4922 | 2007-02-21 23:05:18 -0800 | [diff] [blame] | 2873 | if (!before(tp->snd_una, tp->frto_highmark)) { |
Ilpo Järvinen | d551e45 | 2007-04-30 00:42:20 -0700 | [diff] [blame] | 2874 | tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag); |
Ilpo Järvinen | 7c9a4a5 | 2007-02-21 23:08:34 -0800 | [diff] [blame] | 2875 | return 1; |
Ilpo Järvinen | 95c4922 | 2007-02-21 23:05:18 -0800 | [diff] [blame] | 2876 | } |
| 2877 | |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 2878 | if (!IsSackFrto() || tcp_is_reno(tp)) { |
Ilpo Järvinen | 4dc2665 | 2007-02-21 23:16:11 -0800 | [diff] [blame] | 2879 | /* RFC4138 shortcoming in step 2; should also have case c): |
| 2880 | * ACK isn't duplicate nor advances window, e.g., opposite dir |
| 2881 | * data, winupdate |
| 2882 | */ |
Ilpo Järvinen | 2e60529 | 2007-08-02 19:46:58 -0700 | [diff] [blame] | 2883 | if (!(flag&FLAG_ANY_PROGRESS) && (flag&FLAG_NOT_DUP)) |
Ilpo Järvinen | 4dc2665 | 2007-02-21 23:16:11 -0800 | [diff] [blame] | 2884 | return 1; |
Ilpo Järvinen | 6408d20 | 2007-02-21 23:04:11 -0800 | [diff] [blame] | 2885 | |
Ilpo Järvinen | 4dc2665 | 2007-02-21 23:16:11 -0800 | [diff] [blame] | 2886 | if (!(flag&FLAG_DATA_ACKED)) { |
| 2887 | tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3), |
| 2888 | flag); |
| 2889 | return 1; |
| 2890 | } |
| 2891 | } else { |
| 2892 | if (!(flag&FLAG_DATA_ACKED) && (tp->frto_counter == 1)) { |
| 2893 | /* Prevent sending of new data. */ |
| 2894 | tp->snd_cwnd = min(tp->snd_cwnd, |
| 2895 | tcp_packets_in_flight(tp)); |
| 2896 | return 1; |
| 2897 | } |
| 2898 | |
Ilpo Järvinen | d551e45 | 2007-04-30 00:42:20 -0700 | [diff] [blame] | 2899 | if ((tp->frto_counter >= 2) && |
Ilpo Järvinen | 4dc2665 | 2007-02-21 23:16:11 -0800 | [diff] [blame] | 2900 | (!(flag&FLAG_FORWARD_PROGRESS) || |
| 2901 | ((flag&FLAG_DATA_SACKED) && !(flag&FLAG_ONLY_ORIG_SACKED)))) { |
| 2902 | /* RFC4138 shortcoming (see comment above) */ |
| 2903 | if (!(flag&FLAG_FORWARD_PROGRESS) && (flag&FLAG_NOT_DUP)) |
| 2904 | return 1; |
| 2905 | |
| 2906 | tcp_enter_frto_loss(sk, 3, flag); |
| 2907 | return 1; |
| 2908 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2909 | } |
| 2910 | |
| 2911 | if (tp->frto_counter == 1) { |
Ilpo Järvinen | 575ee71 | 2007-04-30 00:39:55 -0700 | [diff] [blame] | 2912 | /* Sending of the next skb must be allowed or no FRTO */ |
| 2913 | if (!tcp_send_head(sk) || |
| 2914 | after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, |
| 2915 | tp->snd_una + tp->snd_wnd)) { |
Ilpo Järvinen | d551e45 | 2007-04-30 00:42:20 -0700 | [diff] [blame] | 2916 | tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), |
| 2917 | flag); |
Ilpo Järvinen | 575ee71 | 2007-04-30 00:39:55 -0700 | [diff] [blame] | 2918 | return 1; |
| 2919 | } |
| 2920 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2921 | tp->snd_cwnd = tcp_packets_in_flight(tp) + 2; |
Ilpo Järvinen | 94d0ea7 | 2007-02-21 23:07:27 -0800 | [diff] [blame] | 2922 | tp->frto_counter = 2; |
Ilpo Järvinen | 7c9a4a5 | 2007-02-21 23:08:34 -0800 | [diff] [blame] | 2923 | return 1; |
Ilpo Järvinen | d551e45 | 2007-04-30 00:42:20 -0700 | [diff] [blame] | 2924 | } else { |
Ilpo Järvinen | 3cfe3ba | 2007-02-27 10:09:49 -0800 | [diff] [blame] | 2925 | switch (sysctl_tcp_frto_response) { |
| 2926 | case 2: |
Ilpo Järvinen | e317f6f | 2007-03-02 13:34:19 -0800 | [diff] [blame] | 2927 | tcp_undo_spur_to_response(sk, flag); |
Ilpo Järvinen | 3cfe3ba | 2007-02-27 10:09:49 -0800 | [diff] [blame] | 2928 | break; |
| 2929 | case 1: |
| 2930 | tcp_conservative_spur_to_response(tp); |
| 2931 | break; |
| 2932 | default: |
| 2933 | tcp_ratehalving_spur_to_response(sk); |
| 2934 | break; |
Stephen Hemminger | 3ff50b7 | 2007-04-20 17:09:22 -0700 | [diff] [blame] | 2935 | } |
Ilpo Järvinen | 94d0ea7 | 2007-02-21 23:07:27 -0800 | [diff] [blame] | 2936 | tp->frto_counter = 0; |
Ilpo Järvinen | 009a2e3 | 2007-09-20 11:34:38 -0700 | [diff] [blame] | 2937 | tp->undo_marker = 0; |
Ilpo Järvinen | 912d8f0 | 2007-09-25 22:47:31 -0700 | [diff] [blame^] | 2938 | NET_INC_STATS_BH(LINUX_MIB_TCPSPURIOUSRTOS); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2939 | } |
Ilpo Järvinen | 7c9a4a5 | 2007-02-21 23:08:34 -0800 | [diff] [blame] | 2940 | return 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2941 | } |
| 2942 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2943 | /* This routine deals with incoming acks, but not outgoing ones. */ |
| 2944 | static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag) |
| 2945 | { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2946 | struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2947 | struct tcp_sock *tp = tcp_sk(sk); |
| 2948 | u32 prior_snd_una = tp->snd_una; |
| 2949 | u32 ack_seq = TCP_SKB_CB(skb)->seq; |
| 2950 | u32 ack = TCP_SKB_CB(skb)->ack_seq; |
| 2951 | u32 prior_in_flight; |
| 2952 | s32 seq_rtt; |
| 2953 | int prior_packets; |
Ilpo Järvinen | 7c9a4a5 | 2007-02-21 23:08:34 -0800 | [diff] [blame] | 2954 | int frto_cwnd = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2955 | |
| 2956 | /* If the ack is newer than sent or older than previous acks |
| 2957 | * then we can probably ignore it. |
| 2958 | */ |
| 2959 | if (after(ack, tp->snd_nxt)) |
| 2960 | goto uninteresting_ack; |
| 2961 | |
| 2962 | if (before(ack, prior_snd_una)) |
| 2963 | goto old_ack; |
| 2964 | |
Ilpo Järvinen | 2e60529 | 2007-08-02 19:46:58 -0700 | [diff] [blame] | 2965 | if (after(ack, prior_snd_una)) |
| 2966 | flag |= FLAG_SND_UNA_ADVANCED; |
| 2967 | |
Daikichi Osuga | 3fdf3f0 | 2006-08-29 02:01:44 -0700 | [diff] [blame] | 2968 | if (sysctl_tcp_abc) { |
| 2969 | if (icsk->icsk_ca_state < TCP_CA_CWR) |
| 2970 | tp->bytes_acked += ack - prior_snd_una; |
| 2971 | else if (icsk->icsk_ca_state == TCP_CA_Loss) |
| 2972 | /* we assume just one segment left network */ |
| 2973 | tp->bytes_acked += min(ack - prior_snd_una, tp->mss_cache); |
| 2974 | } |
Stephen Hemminger | 9772efb | 2005-11-10 17:09:53 -0800 | [diff] [blame] | 2975 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2976 | if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) { |
| 2977 | /* Window is constant, pure forward advance. |
| 2978 | * No more checks are required. |
| 2979 | * Note, we use the fact that SND.UNA>=SND.WL2. |
| 2980 | */ |
| 2981 | tcp_update_wl(tp, ack, ack_seq); |
| 2982 | tp->snd_una = ack; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2983 | flag |= FLAG_WIN_UPDATE; |
| 2984 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2985 | tcp_ca_event(sk, CA_EVENT_FAST_ACK); |
Stephen Hemminger | 317a76f | 2005-06-23 12:19:55 -0700 | [diff] [blame] | 2986 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2987 | NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS); |
| 2988 | } else { |
| 2989 | if (ack_seq != TCP_SKB_CB(skb)->end_seq) |
| 2990 | flag |= FLAG_DATA; |
| 2991 | else |
| 2992 | NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS); |
| 2993 | |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2994 | flag |= tcp_ack_update_window(sk, skb, ack, ack_seq); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2995 | |
| 2996 | if (TCP_SKB_CB(skb)->sacked) |
| 2997 | flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una); |
| 2998 | |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 2999 | if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb))) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3000 | flag |= FLAG_ECE; |
| 3001 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 3002 | tcp_ca_event(sk, CA_EVENT_SLOW_ACK); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3003 | } |
| 3004 | |
| 3005 | /* We passed data and got it acked, remove any soft error |
| 3006 | * log. Something worked... |
| 3007 | */ |
| 3008 | sk->sk_err_soft = 0; |
| 3009 | tp->rcv_tstamp = tcp_time_stamp; |
| 3010 | prior_packets = tp->packets_out; |
| 3011 | if (!prior_packets) |
| 3012 | goto no_queue; |
| 3013 | |
| 3014 | prior_in_flight = tcp_packets_in_flight(tp); |
| 3015 | |
| 3016 | /* See if we can take anything off of the retransmit queue. */ |
Stephen Hemminger | 2d2abba | 2005-11-10 16:56:12 -0800 | [diff] [blame] | 3017 | flag |= tcp_clean_rtx_queue(sk, &seq_rtt); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3018 | |
| 3019 | if (tp->frto_counter) |
Ilpo Järvinen | 2e60529 | 2007-08-02 19:46:58 -0700 | [diff] [blame] | 3020 | frto_cwnd = tcp_process_frto(sk, flag); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3021 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 3022 | if (tcp_ack_is_dubious(sk, flag)) { |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 3023 | /* Advance CWND, if state allows this. */ |
Ilpo Järvinen | 7c9a4a5 | 2007-02-21 23:08:34 -0800 | [diff] [blame] | 3024 | if ((flag & FLAG_DATA_ACKED) && !frto_cwnd && |
| 3025 | tcp_may_raise_cwnd(sk, flag)) |
Stephen Hemminger | 1675134 | 2007-07-16 18:35:52 -0700 | [diff] [blame] | 3026 | tcp_cong_avoid(sk, ack, prior_in_flight, 0); |
Ilpo Järvinen | 1b6d427b | 2007-08-09 14:53:36 +0300 | [diff] [blame] | 3027 | tcp_fastretrans_alert(sk, prior_packets - tp->packets_out, flag); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3028 | } else { |
Ilpo Järvinen | 7c9a4a5 | 2007-02-21 23:08:34 -0800 | [diff] [blame] | 3029 | if ((flag & FLAG_DATA_ACKED) && !frto_cwnd) |
Stephen Hemminger | 1675134 | 2007-07-16 18:35:52 -0700 | [diff] [blame] | 3030 | tcp_cong_avoid(sk, ack, prior_in_flight, 1); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3031 | } |
| 3032 | |
| 3033 | if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP)) |
| 3034 | dst_confirm(sk->sk_dst_cache); |
| 3035 | |
| 3036 | return 1; |
| 3037 | |
| 3038 | no_queue: |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 3039 | icsk->icsk_probes_out = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3040 | |
| 3041 | /* If this ack opens up a zero window, clear backoff. It was |
| 3042 | * being used to time the probes, and is probably far higher than |
| 3043 | * it needs to be for normal retransmission. |
| 3044 | */ |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 3045 | if (tcp_send_head(sk)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3046 | tcp_ack_probe(sk); |
| 3047 | return 1; |
| 3048 | |
| 3049 | old_ack: |
| 3050 | if (TCP_SKB_CB(skb)->sacked) |
| 3051 | tcp_sacktag_write_queue(sk, skb, prior_snd_una); |
| 3052 | |
| 3053 | uninteresting_ack: |
| 3054 | SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt); |
| 3055 | return 0; |
| 3056 | } |
| 3057 | |
| 3058 | |
| 3059 | /* Look for tcp options. Normally only called on SYN and SYNACK packets. |
| 3060 | * But, this can also be called on packets in the established flow when |
| 3061 | * the fast version below fails. |
| 3062 | */ |
| 3063 | void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, int estab) |
| 3064 | { |
| 3065 | unsigned char *ptr; |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 3066 | struct tcphdr *th = tcp_hdr(skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3067 | int length=(th->doff*4)-sizeof(struct tcphdr); |
| 3068 | |
| 3069 | ptr = (unsigned char *)(th + 1); |
| 3070 | opt_rx->saw_tstamp = 0; |
| 3071 | |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 3072 | while (length > 0) { |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 3073 | int opcode=*ptr++; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3074 | int opsize; |
| 3075 | |
| 3076 | switch (opcode) { |
| 3077 | case TCPOPT_EOL: |
| 3078 | return; |
| 3079 | case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ |
| 3080 | length--; |
| 3081 | continue; |
| 3082 | default: |
| 3083 | opsize=*ptr++; |
| 3084 | if (opsize < 2) /* "silly options" */ |
| 3085 | return; |
| 3086 | if (opsize > length) |
| 3087 | return; /* don't parse partial options */ |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 3088 | switch (opcode) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3089 | case TCPOPT_MSS: |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 3090 | if (opsize==TCPOLEN_MSS && th->syn && !estab) { |
Al Viro | 4f3608b | 2006-09-27 18:35:09 -0700 | [diff] [blame] | 3091 | u16 in_mss = ntohs(get_unaligned((__be16 *)ptr)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3092 | if (in_mss) { |
| 3093 | if (opt_rx->user_mss && opt_rx->user_mss < in_mss) |
| 3094 | in_mss = opt_rx->user_mss; |
| 3095 | opt_rx->mss_clamp = in_mss; |
| 3096 | } |
| 3097 | } |
| 3098 | break; |
| 3099 | case TCPOPT_WINDOW: |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 3100 | if (opsize==TCPOLEN_WINDOW && th->syn && !estab) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3101 | if (sysctl_tcp_window_scaling) { |
| 3102 | __u8 snd_wscale = *(__u8 *) ptr; |
| 3103 | opt_rx->wscale_ok = 1; |
| 3104 | if (snd_wscale > 14) { |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 3105 | if (net_ratelimit()) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3106 | printk(KERN_INFO "tcp_parse_options: Illegal window " |
| 3107 | "scaling value %d >14 received.\n", |
| 3108 | snd_wscale); |
| 3109 | snd_wscale = 14; |
| 3110 | } |
| 3111 | opt_rx->snd_wscale = snd_wscale; |
| 3112 | } |
| 3113 | break; |
| 3114 | case TCPOPT_TIMESTAMP: |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 3115 | if (opsize==TCPOLEN_TIMESTAMP) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3116 | if ((estab && opt_rx->tstamp_ok) || |
| 3117 | (!estab && sysctl_tcp_timestamps)) { |
| 3118 | opt_rx->saw_tstamp = 1; |
Al Viro | 4f3608b | 2006-09-27 18:35:09 -0700 | [diff] [blame] | 3119 | opt_rx->rcv_tsval = ntohl(get_unaligned((__be32 *)ptr)); |
| 3120 | opt_rx->rcv_tsecr = ntohl(get_unaligned((__be32 *)(ptr+4))); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3121 | } |
| 3122 | } |
| 3123 | break; |
| 3124 | case TCPOPT_SACK_PERM: |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 3125 | if (opsize==TCPOLEN_SACK_PERM && th->syn && !estab) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3126 | if (sysctl_tcp_sack) { |
| 3127 | opt_rx->sack_ok = 1; |
| 3128 | tcp_sack_reset(opt_rx); |
| 3129 | } |
| 3130 | } |
| 3131 | break; |
| 3132 | |
| 3133 | case TCPOPT_SACK: |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 3134 | if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) && |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3135 | !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) && |
| 3136 | opt_rx->sack_ok) { |
| 3137 | TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th; |
| 3138 | } |
Ilpo Järvinen | d7ea5b9 | 2007-06-14 12:58:26 -0700 | [diff] [blame] | 3139 | break; |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 3140 | #ifdef CONFIG_TCP_MD5SIG |
| 3141 | case TCPOPT_MD5SIG: |
| 3142 | /* |
| 3143 | * The MD5 Hash has already been |
| 3144 | * checked (see tcp_v{4,6}_do_rcv()). |
| 3145 | */ |
| 3146 | break; |
| 3147 | #endif |
Stephen Hemminger | 3ff50b7 | 2007-04-20 17:09:22 -0700 | [diff] [blame] | 3148 | } |
| 3149 | |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 3150 | ptr+=opsize-2; |
| 3151 | length-=opsize; |
Stephen Hemminger | 3ff50b7 | 2007-04-20 17:09:22 -0700 | [diff] [blame] | 3152 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3153 | } |
| 3154 | } |
| 3155 | |
| 3156 | /* Fast parse options. This hopes to only see timestamps. |
| 3157 | * If it is wrong it falls back on tcp_parse_options(). |
| 3158 | */ |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 3159 | static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th, |
| 3160 | struct tcp_sock *tp) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3161 | { |
| 3162 | if (th->doff == sizeof(struct tcphdr)>>2) { |
| 3163 | tp->rx_opt.saw_tstamp = 0; |
| 3164 | return 0; |
| 3165 | } else if (tp->rx_opt.tstamp_ok && |
| 3166 | th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) { |
Al Viro | 4f3608b | 2006-09-27 18:35:09 -0700 | [diff] [blame] | 3167 | __be32 *ptr = (__be32 *)(th + 1); |
| 3168 | if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3169 | | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) { |
| 3170 | tp->rx_opt.saw_tstamp = 1; |
| 3171 | ++ptr; |
| 3172 | tp->rx_opt.rcv_tsval = ntohl(*ptr); |
| 3173 | ++ptr; |
| 3174 | tp->rx_opt.rcv_tsecr = ntohl(*ptr); |
| 3175 | return 1; |
| 3176 | } |
| 3177 | } |
| 3178 | tcp_parse_options(skb, &tp->rx_opt, 1); |
| 3179 | return 1; |
| 3180 | } |
| 3181 | |
| 3182 | static inline void tcp_store_ts_recent(struct tcp_sock *tp) |
| 3183 | { |
| 3184 | tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval; |
James Morris | 9d729f7 | 2007-03-04 16:12:44 -0800 | [diff] [blame] | 3185 | tp->rx_opt.ts_recent_stamp = get_seconds(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3186 | } |
| 3187 | |
| 3188 | static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq) |
| 3189 | { |
| 3190 | if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) { |
| 3191 | /* PAWS bug workaround wrt. ACK frames, the PAWS discard |
| 3192 | * extra check below makes sure this can only happen |
| 3193 | * for pure ACK frames. -DaveM |
| 3194 | * |
| 3195 | * Not only, also it occurs for expired timestamps. |
| 3196 | */ |
| 3197 | |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 3198 | if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 || |
James Morris | 9d729f7 | 2007-03-04 16:12:44 -0800 | [diff] [blame] | 3199 | get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3200 | tcp_store_ts_recent(tp); |
| 3201 | } |
| 3202 | } |
| 3203 | |
| 3204 | /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM |
| 3205 | * |
| 3206 | * It is not fatal. If this ACK does _not_ change critical state (seqs, window) |
| 3207 | * it can pass through stack. So, the following predicate verifies that |
| 3208 | * this segment is not used for anything but congestion avoidance or |
| 3209 | * fast retransmit. Moreover, we even are able to eliminate most of such |
| 3210 | * second order effects, if we apply some small "replay" window (~RTO) |
| 3211 | * to timestamp space. |
| 3212 | * |
| 3213 | * All these measures still do not guarantee that we reject wrapped ACKs |
| 3214 | * on networks with high bandwidth, when sequence space is recycled fastly, |
| 3215 | * but it guarantees that such events will be very rare and do not affect |
| 3216 | * connection seriously. This doesn't look nice, but alas, PAWS is really |
| 3217 | * buggy extension. |
| 3218 | * |
| 3219 | * [ Later note. Even worse! It is buggy for segments _with_ data. RFC |
| 3220 | * states that events when retransmit arrives after original data are rare. |
| 3221 | * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is |
| 3222 | * the biggest problem on large power networks even with minor reordering. |
| 3223 | * OK, let's give it small replay window. If peer clock is even 1hz, it is safe |
| 3224 | * up to bandwidth of 18Gigabit/sec. 8) ] |
| 3225 | */ |
| 3226 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3227 | static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3228 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3229 | struct tcp_sock *tp = tcp_sk(sk); |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 3230 | struct tcphdr *th = tcp_hdr(skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3231 | u32 seq = TCP_SKB_CB(skb)->seq; |
| 3232 | u32 ack = TCP_SKB_CB(skb)->ack_seq; |
| 3233 | |
| 3234 | return (/* 1. Pure ACK with correct sequence number. */ |
| 3235 | (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) && |
| 3236 | |
| 3237 | /* 2. ... and duplicate ACK. */ |
| 3238 | ack == tp->snd_una && |
| 3239 | |
| 3240 | /* 3. ... and does not update window. */ |
| 3241 | !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) && |
| 3242 | |
| 3243 | /* 4. ... and sits in replay window. */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3244 | (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3245 | } |
| 3246 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3247 | static inline int tcp_paws_discard(const struct sock *sk, const struct sk_buff *skb) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3248 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3249 | const struct tcp_sock *tp = tcp_sk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3250 | return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW && |
James Morris | 9d729f7 | 2007-03-04 16:12:44 -0800 | [diff] [blame] | 3251 | get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS && |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3252 | !tcp_disordered_ack(sk, skb)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3253 | } |
| 3254 | |
| 3255 | /* Check segment sequence number for validity. |
| 3256 | * |
| 3257 | * Segment controls are considered valid, if the segment |
| 3258 | * fits to the window after truncation to the window. Acceptability |
| 3259 | * of data (and SYN, FIN, of course) is checked separately. |
| 3260 | * See tcp_data_queue(), for example. |
| 3261 | * |
| 3262 | * Also, controls (RST is main one) are accepted using RCV.WUP instead |
| 3263 | * of RCV.NXT. Peer still did not advance his SND.UNA when we |
| 3264 | * delayed ACK, so that hisSND.UNA<=ourRCV.WUP. |
| 3265 | * (borrowed from freebsd) |
| 3266 | */ |
| 3267 | |
| 3268 | static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq) |
| 3269 | { |
| 3270 | return !before(end_seq, tp->rcv_wup) && |
| 3271 | !after(seq, tp->rcv_nxt + tcp_receive_window(tp)); |
| 3272 | } |
| 3273 | |
| 3274 | /* When we get a reset we do this. */ |
| 3275 | static void tcp_reset(struct sock *sk) |
| 3276 | { |
| 3277 | /* We want the right error as BSD sees it (and indeed as we do). */ |
| 3278 | switch (sk->sk_state) { |
| 3279 | case TCP_SYN_SENT: |
| 3280 | sk->sk_err = ECONNREFUSED; |
| 3281 | break; |
| 3282 | case TCP_CLOSE_WAIT: |
| 3283 | sk->sk_err = EPIPE; |
| 3284 | break; |
| 3285 | case TCP_CLOSE: |
| 3286 | return; |
| 3287 | default: |
| 3288 | sk->sk_err = ECONNRESET; |
| 3289 | } |
| 3290 | |
| 3291 | if (!sock_flag(sk, SOCK_DEAD)) |
| 3292 | sk->sk_error_report(sk); |
| 3293 | |
| 3294 | tcp_done(sk); |
| 3295 | } |
| 3296 | |
| 3297 | /* |
| 3298 | * Process the FIN bit. This now behaves as it is supposed to work |
| 3299 | * and the FIN takes effect when it is validly part of sequence |
| 3300 | * space. Not before when we get holes. |
| 3301 | * |
| 3302 | * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT |
| 3303 | * (and thence onto LAST-ACK and finally, CLOSE, we never enter |
| 3304 | * TIME-WAIT) |
| 3305 | * |
| 3306 | * If we are in FINWAIT-1, a received FIN indicates simultaneous |
| 3307 | * close and we go into CLOSING (and later onto TIME-WAIT) |
| 3308 | * |
| 3309 | * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT. |
| 3310 | */ |
| 3311 | static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th) |
| 3312 | { |
| 3313 | struct tcp_sock *tp = tcp_sk(sk); |
| 3314 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3315 | inet_csk_schedule_ack(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3316 | |
| 3317 | sk->sk_shutdown |= RCV_SHUTDOWN; |
| 3318 | sock_set_flag(sk, SOCK_DONE); |
| 3319 | |
| 3320 | switch (sk->sk_state) { |
| 3321 | case TCP_SYN_RECV: |
| 3322 | case TCP_ESTABLISHED: |
| 3323 | /* Move to CLOSE_WAIT */ |
| 3324 | tcp_set_state(sk, TCP_CLOSE_WAIT); |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3325 | inet_csk(sk)->icsk_ack.pingpong = 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3326 | break; |
| 3327 | |
| 3328 | case TCP_CLOSE_WAIT: |
| 3329 | case TCP_CLOSING: |
| 3330 | /* Received a retransmission of the FIN, do |
| 3331 | * nothing. |
| 3332 | */ |
| 3333 | break; |
| 3334 | case TCP_LAST_ACK: |
| 3335 | /* RFC793: Remain in the LAST-ACK state. */ |
| 3336 | break; |
| 3337 | |
| 3338 | case TCP_FIN_WAIT1: |
| 3339 | /* This case occurs when a simultaneous close |
| 3340 | * happens, we must ack the received FIN and |
| 3341 | * enter the CLOSING state. |
| 3342 | */ |
| 3343 | tcp_send_ack(sk); |
| 3344 | tcp_set_state(sk, TCP_CLOSING); |
| 3345 | break; |
| 3346 | case TCP_FIN_WAIT2: |
| 3347 | /* Received a FIN -- send ACK and enter TIME_WAIT. */ |
| 3348 | tcp_send_ack(sk); |
| 3349 | tcp_time_wait(sk, TCP_TIME_WAIT, 0); |
| 3350 | break; |
| 3351 | default: |
| 3352 | /* Only TCP_LISTEN and TCP_CLOSE are left, in these |
| 3353 | * cases we should never reach this piece of code. |
| 3354 | */ |
| 3355 | printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n", |
| 3356 | __FUNCTION__, sk->sk_state); |
| 3357 | break; |
Stephen Hemminger | 3ff50b7 | 2007-04-20 17:09:22 -0700 | [diff] [blame] | 3358 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3359 | |
| 3360 | /* It _is_ possible, that we have something out-of-order _after_ FIN. |
| 3361 | * Probably, we should reset in this case. For now drop them. |
| 3362 | */ |
| 3363 | __skb_queue_purge(&tp->out_of_order_queue); |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 3364 | if (tcp_is_sack(tp)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3365 | tcp_sack_reset(&tp->rx_opt); |
| 3366 | sk_stream_mem_reclaim(sk); |
| 3367 | |
| 3368 | if (!sock_flag(sk, SOCK_DEAD)) { |
| 3369 | sk->sk_state_change(sk); |
| 3370 | |
| 3371 | /* Do not send POLL_HUP for half duplex close. */ |
| 3372 | if (sk->sk_shutdown == SHUTDOWN_MASK || |
| 3373 | sk->sk_state == TCP_CLOSE) |
| 3374 | sk_wake_async(sk, 1, POLL_HUP); |
| 3375 | else |
| 3376 | sk_wake_async(sk, 1, POLL_IN); |
| 3377 | } |
| 3378 | } |
| 3379 | |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 3380 | static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3381 | { |
| 3382 | if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) { |
| 3383 | if (before(seq, sp->start_seq)) |
| 3384 | sp->start_seq = seq; |
| 3385 | if (after(end_seq, sp->end_seq)) |
| 3386 | sp->end_seq = end_seq; |
| 3387 | return 1; |
| 3388 | } |
| 3389 | return 0; |
| 3390 | } |
| 3391 | |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 3392 | static void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3393 | { |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 3394 | if (tcp_is_sack(tp) && sysctl_tcp_dsack) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3395 | if (before(seq, tp->rcv_nxt)) |
| 3396 | NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT); |
| 3397 | else |
| 3398 | NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT); |
| 3399 | |
| 3400 | tp->rx_opt.dsack = 1; |
| 3401 | tp->duplicate_sack[0].start_seq = seq; |
| 3402 | tp->duplicate_sack[0].end_seq = end_seq; |
| 3403 | tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 4 - tp->rx_opt.tstamp_ok); |
| 3404 | } |
| 3405 | } |
| 3406 | |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 3407 | static void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3408 | { |
| 3409 | if (!tp->rx_opt.dsack) |
| 3410 | tcp_dsack_set(tp, seq, end_seq); |
| 3411 | else |
| 3412 | tcp_sack_extend(tp->duplicate_sack, seq, end_seq); |
| 3413 | } |
| 3414 | |
| 3415 | static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb) |
| 3416 | { |
| 3417 | struct tcp_sock *tp = tcp_sk(sk); |
| 3418 | |
| 3419 | if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && |
| 3420 | before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { |
| 3421 | NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST); |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3422 | tcp_enter_quickack_mode(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3423 | |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 3424 | if (tcp_is_sack(tp) && sysctl_tcp_dsack) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3425 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; |
| 3426 | |
| 3427 | if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) |
| 3428 | end_seq = tp->rcv_nxt; |
| 3429 | tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq); |
| 3430 | } |
| 3431 | } |
| 3432 | |
| 3433 | tcp_send_ack(sk); |
| 3434 | } |
| 3435 | |
| 3436 | /* These routines update the SACK block as out-of-order packets arrive or |
| 3437 | * in-order packets close up the sequence space. |
| 3438 | */ |
| 3439 | static void tcp_sack_maybe_coalesce(struct tcp_sock *tp) |
| 3440 | { |
| 3441 | int this_sack; |
| 3442 | struct tcp_sack_block *sp = &tp->selective_acks[0]; |
| 3443 | struct tcp_sack_block *swalk = sp+1; |
| 3444 | |
| 3445 | /* See if the recent change to the first SACK eats into |
| 3446 | * or hits the sequence space of other SACK blocks, if so coalesce. |
| 3447 | */ |
| 3448 | for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; ) { |
| 3449 | if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) { |
| 3450 | int i; |
| 3451 | |
| 3452 | /* Zap SWALK, by moving every further SACK up by one slot. |
| 3453 | * Decrease num_sacks. |
| 3454 | */ |
| 3455 | tp->rx_opt.num_sacks--; |
| 3456 | tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok); |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 3457 | for (i=this_sack; i < tp->rx_opt.num_sacks; i++) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3458 | sp[i] = sp[i+1]; |
| 3459 | continue; |
| 3460 | } |
| 3461 | this_sack++, swalk++; |
| 3462 | } |
| 3463 | } |
| 3464 | |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 3465 | static inline void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3466 | { |
| 3467 | __u32 tmp; |
| 3468 | |
| 3469 | tmp = sack1->start_seq; |
| 3470 | sack1->start_seq = sack2->start_seq; |
| 3471 | sack2->start_seq = tmp; |
| 3472 | |
| 3473 | tmp = sack1->end_seq; |
| 3474 | sack1->end_seq = sack2->end_seq; |
| 3475 | sack2->end_seq = tmp; |
| 3476 | } |
| 3477 | |
| 3478 | static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq) |
| 3479 | { |
| 3480 | struct tcp_sock *tp = tcp_sk(sk); |
| 3481 | struct tcp_sack_block *sp = &tp->selective_acks[0]; |
| 3482 | int cur_sacks = tp->rx_opt.num_sacks; |
| 3483 | int this_sack; |
| 3484 | |
| 3485 | if (!cur_sacks) |
| 3486 | goto new_sack; |
| 3487 | |
| 3488 | for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) { |
| 3489 | if (tcp_sack_extend(sp, seq, end_seq)) { |
| 3490 | /* Rotate this_sack to the first one. */ |
| 3491 | for (; this_sack>0; this_sack--, sp--) |
| 3492 | tcp_sack_swap(sp, sp-1); |
| 3493 | if (cur_sacks > 1) |
| 3494 | tcp_sack_maybe_coalesce(tp); |
| 3495 | return; |
| 3496 | } |
| 3497 | } |
| 3498 | |
| 3499 | /* Could not find an adjacent existing SACK, build a new one, |
| 3500 | * put it at the front, and shift everyone else down. We |
| 3501 | * always know there is at least one SACK present already here. |
| 3502 | * |
| 3503 | * If the sack array is full, forget about the last one. |
| 3504 | */ |
| 3505 | if (this_sack >= 4) { |
| 3506 | this_sack--; |
| 3507 | tp->rx_opt.num_sacks--; |
| 3508 | sp--; |
| 3509 | } |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 3510 | for (; this_sack > 0; this_sack--, sp--) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3511 | *sp = *(sp-1); |
| 3512 | |
| 3513 | new_sack: |
| 3514 | /* Build the new head SACK, and we're done. */ |
| 3515 | sp->start_seq = seq; |
| 3516 | sp->end_seq = end_seq; |
| 3517 | tp->rx_opt.num_sacks++; |
| 3518 | tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok); |
| 3519 | } |
| 3520 | |
| 3521 | /* RCV.NXT advances, some SACKs should be eaten. */ |
| 3522 | |
| 3523 | static void tcp_sack_remove(struct tcp_sock *tp) |
| 3524 | { |
| 3525 | struct tcp_sack_block *sp = &tp->selective_acks[0]; |
| 3526 | int num_sacks = tp->rx_opt.num_sacks; |
| 3527 | int this_sack; |
| 3528 | |
| 3529 | /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */ |
David S. Miller | b03efcf | 2005-07-08 14:57:23 -0700 | [diff] [blame] | 3530 | if (skb_queue_empty(&tp->out_of_order_queue)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3531 | tp->rx_opt.num_sacks = 0; |
| 3532 | tp->rx_opt.eff_sacks = tp->rx_opt.dsack; |
| 3533 | return; |
| 3534 | } |
| 3535 | |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 3536 | for (this_sack = 0; this_sack < num_sacks; ) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3537 | /* Check if the start of the sack is covered by RCV.NXT. */ |
| 3538 | if (!before(tp->rcv_nxt, sp->start_seq)) { |
| 3539 | int i; |
| 3540 | |
| 3541 | /* RCV.NXT must cover all the block! */ |
| 3542 | BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq)); |
| 3543 | |
| 3544 | /* Zap this SACK, by moving forward any other SACKS. */ |
| 3545 | for (i=this_sack+1; i < num_sacks; i++) |
| 3546 | tp->selective_acks[i-1] = tp->selective_acks[i]; |
| 3547 | num_sacks--; |
| 3548 | continue; |
| 3549 | } |
| 3550 | this_sack++; |
| 3551 | sp++; |
| 3552 | } |
| 3553 | if (num_sacks != tp->rx_opt.num_sacks) { |
| 3554 | tp->rx_opt.num_sacks = num_sacks; |
| 3555 | tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok); |
| 3556 | } |
| 3557 | } |
| 3558 | |
| 3559 | /* This one checks to see if we can put data from the |
| 3560 | * out_of_order queue into the receive_queue. |
| 3561 | */ |
| 3562 | static void tcp_ofo_queue(struct sock *sk) |
| 3563 | { |
| 3564 | struct tcp_sock *tp = tcp_sk(sk); |
| 3565 | __u32 dsack_high = tp->rcv_nxt; |
| 3566 | struct sk_buff *skb; |
| 3567 | |
| 3568 | while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) { |
| 3569 | if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) |
| 3570 | break; |
| 3571 | |
| 3572 | if (before(TCP_SKB_CB(skb)->seq, dsack_high)) { |
| 3573 | __u32 dsack = dsack_high; |
| 3574 | if (before(TCP_SKB_CB(skb)->end_seq, dsack_high)) |
| 3575 | dsack_high = TCP_SKB_CB(skb)->end_seq; |
| 3576 | tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack); |
| 3577 | } |
| 3578 | |
| 3579 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { |
| 3580 | SOCK_DEBUG(sk, "ofo packet was already received \n"); |
David S. Miller | 8728b83 | 2005-08-09 19:25:21 -0700 | [diff] [blame] | 3581 | __skb_unlink(skb, &tp->out_of_order_queue); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3582 | __kfree_skb(skb); |
| 3583 | continue; |
| 3584 | } |
| 3585 | SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n", |
| 3586 | tp->rcv_nxt, TCP_SKB_CB(skb)->seq, |
| 3587 | TCP_SKB_CB(skb)->end_seq); |
| 3588 | |
David S. Miller | 8728b83 | 2005-08-09 19:25:21 -0700 | [diff] [blame] | 3589 | __skb_unlink(skb, &tp->out_of_order_queue); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3590 | __skb_queue_tail(&sk->sk_receive_queue, skb); |
| 3591 | tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 3592 | if (tcp_hdr(skb)->fin) |
| 3593 | tcp_fin(skb, sk, tcp_hdr(skb)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3594 | } |
| 3595 | } |
| 3596 | |
| 3597 | static int tcp_prune_queue(struct sock *sk); |
| 3598 | |
| 3599 | static void tcp_data_queue(struct sock *sk, struct sk_buff *skb) |
| 3600 | { |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 3601 | struct tcphdr *th = tcp_hdr(skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3602 | struct tcp_sock *tp = tcp_sk(sk); |
| 3603 | int eaten = -1; |
| 3604 | |
| 3605 | if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) |
| 3606 | goto drop; |
| 3607 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3608 | __skb_pull(skb, th->doff*4); |
| 3609 | |
| 3610 | TCP_ECN_accept_cwr(tp, skb); |
| 3611 | |
| 3612 | if (tp->rx_opt.dsack) { |
| 3613 | tp->rx_opt.dsack = 0; |
| 3614 | tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks, |
| 3615 | 4 - tp->rx_opt.tstamp_ok); |
| 3616 | } |
| 3617 | |
| 3618 | /* Queue data for delivery to the user. |
| 3619 | * Packets in sequence go to the receive queue. |
| 3620 | * Out of sequence packets to the out_of_order_queue. |
| 3621 | */ |
| 3622 | if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { |
| 3623 | if (tcp_receive_window(tp) == 0) |
| 3624 | goto out_of_window; |
| 3625 | |
| 3626 | /* Ok. In sequence. In window. */ |
| 3627 | if (tp->ucopy.task == current && |
| 3628 | tp->copied_seq == tp->rcv_nxt && tp->ucopy.len && |
| 3629 | sock_owned_by_user(sk) && !tp->urg_data) { |
| 3630 | int chunk = min_t(unsigned int, skb->len, |
| 3631 | tp->ucopy.len); |
| 3632 | |
| 3633 | __set_current_state(TASK_RUNNING); |
| 3634 | |
| 3635 | local_bh_enable(); |
| 3636 | if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) { |
| 3637 | tp->ucopy.len -= chunk; |
| 3638 | tp->copied_seq += chunk; |
| 3639 | eaten = (chunk == skb->len && !th->fin); |
| 3640 | tcp_rcv_space_adjust(sk); |
| 3641 | } |
| 3642 | local_bh_disable(); |
| 3643 | } |
| 3644 | |
| 3645 | if (eaten <= 0) { |
| 3646 | queue_and_out: |
| 3647 | if (eaten < 0 && |
| 3648 | (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || |
| 3649 | !sk_stream_rmem_schedule(sk, skb))) { |
| 3650 | if (tcp_prune_queue(sk) < 0 || |
| 3651 | !sk_stream_rmem_schedule(sk, skb)) |
| 3652 | goto drop; |
| 3653 | } |
| 3654 | sk_stream_set_owner_r(skb, sk); |
| 3655 | __skb_queue_tail(&sk->sk_receive_queue, skb); |
| 3656 | } |
| 3657 | tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 3658 | if (skb->len) |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 3659 | tcp_event_data_recv(sk, skb); |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 3660 | if (th->fin) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3661 | tcp_fin(skb, sk, th); |
| 3662 | |
David S. Miller | b03efcf | 2005-07-08 14:57:23 -0700 | [diff] [blame] | 3663 | if (!skb_queue_empty(&tp->out_of_order_queue)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3664 | tcp_ofo_queue(sk); |
| 3665 | |
| 3666 | /* RFC2581. 4.2. SHOULD send immediate ACK, when |
| 3667 | * gap in queue is filled. |
| 3668 | */ |
David S. Miller | b03efcf | 2005-07-08 14:57:23 -0700 | [diff] [blame] | 3669 | if (skb_queue_empty(&tp->out_of_order_queue)) |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3670 | inet_csk(sk)->icsk_ack.pingpong = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3671 | } |
| 3672 | |
| 3673 | if (tp->rx_opt.num_sacks) |
| 3674 | tcp_sack_remove(tp); |
| 3675 | |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 3676 | tcp_fast_path_check(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3677 | |
| 3678 | if (eaten > 0) |
| 3679 | __kfree_skb(skb); |
| 3680 | else if (!sock_flag(sk, SOCK_DEAD)) |
| 3681 | sk->sk_data_ready(sk, 0); |
| 3682 | return; |
| 3683 | } |
| 3684 | |
| 3685 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { |
| 3686 | /* A retransmit, 2nd most common case. Force an immediate ack. */ |
| 3687 | NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST); |
| 3688 | tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); |
| 3689 | |
| 3690 | out_of_window: |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3691 | tcp_enter_quickack_mode(sk); |
| 3692 | inet_csk_schedule_ack(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3693 | drop: |
| 3694 | __kfree_skb(skb); |
| 3695 | return; |
| 3696 | } |
| 3697 | |
| 3698 | /* Out of window. F.e. zero window probe. */ |
| 3699 | if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp))) |
| 3700 | goto out_of_window; |
| 3701 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3702 | tcp_enter_quickack_mode(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3703 | |
| 3704 | if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { |
| 3705 | /* Partial packet, seq < rcv_next < end_seq */ |
| 3706 | SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n", |
| 3707 | tp->rcv_nxt, TCP_SKB_CB(skb)->seq, |
| 3708 | TCP_SKB_CB(skb)->end_seq); |
| 3709 | |
| 3710 | tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt); |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 3711 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3712 | /* If window is closed, drop tail of packet. But after |
| 3713 | * remembering D-SACK for its head made in previous line. |
| 3714 | */ |
| 3715 | if (!tcp_receive_window(tp)) |
| 3716 | goto out_of_window; |
| 3717 | goto queue_and_out; |
| 3718 | } |
| 3719 | |
| 3720 | TCP_ECN_check_ce(tp, skb); |
| 3721 | |
| 3722 | if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || |
| 3723 | !sk_stream_rmem_schedule(sk, skb)) { |
| 3724 | if (tcp_prune_queue(sk) < 0 || |
| 3725 | !sk_stream_rmem_schedule(sk, skb)) |
| 3726 | goto drop; |
| 3727 | } |
| 3728 | |
| 3729 | /* Disable header prediction. */ |
| 3730 | tp->pred_flags = 0; |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3731 | inet_csk_schedule_ack(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3732 | |
| 3733 | SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n", |
| 3734 | tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); |
| 3735 | |
| 3736 | sk_stream_set_owner_r(skb, sk); |
| 3737 | |
| 3738 | if (!skb_peek(&tp->out_of_order_queue)) { |
| 3739 | /* Initial out of order segment, build 1 SACK. */ |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 3740 | if (tcp_is_sack(tp)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3741 | tp->rx_opt.num_sacks = 1; |
| 3742 | tp->rx_opt.dsack = 0; |
| 3743 | tp->rx_opt.eff_sacks = 1; |
| 3744 | tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq; |
| 3745 | tp->selective_acks[0].end_seq = |
| 3746 | TCP_SKB_CB(skb)->end_seq; |
| 3747 | } |
| 3748 | __skb_queue_head(&tp->out_of_order_queue,skb); |
| 3749 | } else { |
| 3750 | struct sk_buff *skb1 = tp->out_of_order_queue.prev; |
| 3751 | u32 seq = TCP_SKB_CB(skb)->seq; |
| 3752 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; |
| 3753 | |
| 3754 | if (seq == TCP_SKB_CB(skb1)->end_seq) { |
David S. Miller | 8728b83 | 2005-08-09 19:25:21 -0700 | [diff] [blame] | 3755 | __skb_append(skb1, skb, &tp->out_of_order_queue); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3756 | |
| 3757 | if (!tp->rx_opt.num_sacks || |
| 3758 | tp->selective_acks[0].end_seq != seq) |
| 3759 | goto add_sack; |
| 3760 | |
| 3761 | /* Common case: data arrive in order after hole. */ |
| 3762 | tp->selective_acks[0].end_seq = end_seq; |
| 3763 | return; |
| 3764 | } |
| 3765 | |
| 3766 | /* Find place to insert this segment. */ |
| 3767 | do { |
| 3768 | if (!after(TCP_SKB_CB(skb1)->seq, seq)) |
| 3769 | break; |
| 3770 | } while ((skb1 = skb1->prev) != |
| 3771 | (struct sk_buff*)&tp->out_of_order_queue); |
| 3772 | |
| 3773 | /* Do skb overlap to previous one? */ |
| 3774 | if (skb1 != (struct sk_buff*)&tp->out_of_order_queue && |
| 3775 | before(seq, TCP_SKB_CB(skb1)->end_seq)) { |
| 3776 | if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) { |
| 3777 | /* All the bits are present. Drop. */ |
| 3778 | __kfree_skb(skb); |
| 3779 | tcp_dsack_set(tp, seq, end_seq); |
| 3780 | goto add_sack; |
| 3781 | } |
| 3782 | if (after(seq, TCP_SKB_CB(skb1)->seq)) { |
| 3783 | /* Partial overlap. */ |
| 3784 | tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq); |
| 3785 | } else { |
| 3786 | skb1 = skb1->prev; |
| 3787 | } |
| 3788 | } |
| 3789 | __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue); |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 3790 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3791 | /* And clean segments covered by new one as whole. */ |
| 3792 | while ((skb1 = skb->next) != |
| 3793 | (struct sk_buff*)&tp->out_of_order_queue && |
| 3794 | after(end_seq, TCP_SKB_CB(skb1)->seq)) { |
| 3795 | if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) { |
| 3796 | tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq); |
| 3797 | break; |
| 3798 | } |
David S. Miller | 8728b83 | 2005-08-09 19:25:21 -0700 | [diff] [blame] | 3799 | __skb_unlink(skb1, &tp->out_of_order_queue); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3800 | tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq); |
| 3801 | __kfree_skb(skb1); |
| 3802 | } |
| 3803 | |
| 3804 | add_sack: |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 3805 | if (tcp_is_sack(tp)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3806 | tcp_sack_new_ofo_skb(sk, seq, end_seq); |
| 3807 | } |
| 3808 | } |
| 3809 | |
| 3810 | /* Collapse contiguous sequence of skbs head..tail with |
| 3811 | * sequence numbers start..end. |
| 3812 | * Segments with FIN/SYN are not collapsed (only because this |
| 3813 | * simplifies code) |
| 3814 | */ |
| 3815 | static void |
David S. Miller | 8728b83 | 2005-08-09 19:25:21 -0700 | [diff] [blame] | 3816 | tcp_collapse(struct sock *sk, struct sk_buff_head *list, |
| 3817 | struct sk_buff *head, struct sk_buff *tail, |
| 3818 | u32 start, u32 end) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3819 | { |
| 3820 | struct sk_buff *skb; |
| 3821 | |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 3822 | /* First, check that queue is collapsible and find |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3823 | * the point where collapsing can be useful. */ |
| 3824 | for (skb = head; skb != tail; ) { |
| 3825 | /* No new bits? It is possible on ofo queue. */ |
| 3826 | if (!before(start, TCP_SKB_CB(skb)->end_seq)) { |
| 3827 | struct sk_buff *next = skb->next; |
David S. Miller | 8728b83 | 2005-08-09 19:25:21 -0700 | [diff] [blame] | 3828 | __skb_unlink(skb, list); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3829 | __kfree_skb(skb); |
| 3830 | NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED); |
| 3831 | skb = next; |
| 3832 | continue; |
| 3833 | } |
| 3834 | |
| 3835 | /* The first skb to collapse is: |
| 3836 | * - not SYN/FIN and |
| 3837 | * - bloated or contains data before "start" or |
| 3838 | * overlaps to the next one. |
| 3839 | */ |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 3840 | if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin && |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3841 | (tcp_win_from_space(skb->truesize) > skb->len || |
| 3842 | before(TCP_SKB_CB(skb)->seq, start) || |
| 3843 | (skb->next != tail && |
| 3844 | TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq))) |
| 3845 | break; |
| 3846 | |
| 3847 | /* Decided to skip this, advance start seq. */ |
| 3848 | start = TCP_SKB_CB(skb)->end_seq; |
| 3849 | skb = skb->next; |
| 3850 | } |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 3851 | if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3852 | return; |
| 3853 | |
| 3854 | while (before(start, end)) { |
| 3855 | struct sk_buff *nskb; |
| 3856 | int header = skb_headroom(skb); |
| 3857 | int copy = SKB_MAX_ORDER(header, 0); |
| 3858 | |
| 3859 | /* Too big header? This can happen with IPv6. */ |
| 3860 | if (copy < 0) |
| 3861 | return; |
| 3862 | if (end-start < copy) |
| 3863 | copy = end-start; |
| 3864 | nskb = alloc_skb(copy+header, GFP_ATOMIC); |
| 3865 | if (!nskb) |
| 3866 | return; |
Arnaldo Carvalho de Melo | c51957d | 2007-03-10 12:47:22 -0300 | [diff] [blame] | 3867 | |
Arnaldo Carvalho de Melo | 98e399f | 2007-03-19 15:33:04 -0700 | [diff] [blame] | 3868 | skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head); |
Arnaldo Carvalho de Melo | 9c70220 | 2007-04-25 18:04:18 -0700 | [diff] [blame] | 3869 | skb_set_network_header(nskb, (skb_network_header(skb) - |
| 3870 | skb->head)); |
| 3871 | skb_set_transport_header(nskb, (skb_transport_header(skb) - |
| 3872 | skb->head)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3873 | skb_reserve(nskb, header); |
| 3874 | memcpy(nskb->head, skb->head, header); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3875 | memcpy(nskb->cb, skb->cb, sizeof(skb->cb)); |
| 3876 | TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start; |
David S. Miller | 8728b83 | 2005-08-09 19:25:21 -0700 | [diff] [blame] | 3877 | __skb_insert(nskb, skb->prev, skb, list); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3878 | sk_stream_set_owner_r(nskb, sk); |
| 3879 | |
| 3880 | /* Copy data, releasing collapsed skbs. */ |
| 3881 | while (copy > 0) { |
| 3882 | int offset = start - TCP_SKB_CB(skb)->seq; |
| 3883 | int size = TCP_SKB_CB(skb)->end_seq - start; |
| 3884 | |
Kris Katterjohn | 09a6266 | 2006-01-08 22:24:28 -0800 | [diff] [blame] | 3885 | BUG_ON(offset < 0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3886 | if (size > 0) { |
| 3887 | size = min(copy, size); |
| 3888 | if (skb_copy_bits(skb, offset, skb_put(nskb, size), size)) |
| 3889 | BUG(); |
| 3890 | TCP_SKB_CB(nskb)->end_seq += size; |
| 3891 | copy -= size; |
| 3892 | start += size; |
| 3893 | } |
| 3894 | if (!before(start, TCP_SKB_CB(skb)->end_seq)) { |
| 3895 | struct sk_buff *next = skb->next; |
David S. Miller | 8728b83 | 2005-08-09 19:25:21 -0700 | [diff] [blame] | 3896 | __skb_unlink(skb, list); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3897 | __kfree_skb(skb); |
| 3898 | NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED); |
| 3899 | skb = next; |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 3900 | if (skb == tail || |
| 3901 | tcp_hdr(skb)->syn || |
| 3902 | tcp_hdr(skb)->fin) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3903 | return; |
| 3904 | } |
| 3905 | } |
| 3906 | } |
| 3907 | } |
| 3908 | |
| 3909 | /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs |
| 3910 | * and tcp_collapse() them until all the queue is collapsed. |
| 3911 | */ |
| 3912 | static void tcp_collapse_ofo_queue(struct sock *sk) |
| 3913 | { |
| 3914 | struct tcp_sock *tp = tcp_sk(sk); |
| 3915 | struct sk_buff *skb = skb_peek(&tp->out_of_order_queue); |
| 3916 | struct sk_buff *head; |
| 3917 | u32 start, end; |
| 3918 | |
| 3919 | if (skb == NULL) |
| 3920 | return; |
| 3921 | |
| 3922 | start = TCP_SKB_CB(skb)->seq; |
| 3923 | end = TCP_SKB_CB(skb)->end_seq; |
| 3924 | head = skb; |
| 3925 | |
| 3926 | for (;;) { |
| 3927 | skb = skb->next; |
| 3928 | |
| 3929 | /* Segment is terminated when we see gap or when |
| 3930 | * we are at the end of all the queue. */ |
| 3931 | if (skb == (struct sk_buff *)&tp->out_of_order_queue || |
| 3932 | after(TCP_SKB_CB(skb)->seq, end) || |
| 3933 | before(TCP_SKB_CB(skb)->end_seq, start)) { |
David S. Miller | 8728b83 | 2005-08-09 19:25:21 -0700 | [diff] [blame] | 3934 | tcp_collapse(sk, &tp->out_of_order_queue, |
| 3935 | head, skb, start, end); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3936 | head = skb; |
| 3937 | if (skb == (struct sk_buff *)&tp->out_of_order_queue) |
| 3938 | break; |
| 3939 | /* Start new segment */ |
| 3940 | start = TCP_SKB_CB(skb)->seq; |
| 3941 | end = TCP_SKB_CB(skb)->end_seq; |
| 3942 | } else { |
| 3943 | if (before(TCP_SKB_CB(skb)->seq, start)) |
| 3944 | start = TCP_SKB_CB(skb)->seq; |
| 3945 | if (after(TCP_SKB_CB(skb)->end_seq, end)) |
| 3946 | end = TCP_SKB_CB(skb)->end_seq; |
| 3947 | } |
| 3948 | } |
| 3949 | } |
| 3950 | |
| 3951 | /* Reduce allocated memory if we can, trying to get |
| 3952 | * the socket within its memory limits again. |
| 3953 | * |
| 3954 | * Return less than zero if we should start dropping frames |
| 3955 | * until the socket owning process reads some of the data |
| 3956 | * to stabilize the situation. |
| 3957 | */ |
| 3958 | static int tcp_prune_queue(struct sock *sk) |
| 3959 | { |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 3960 | struct tcp_sock *tp = tcp_sk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3961 | |
| 3962 | SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq); |
| 3963 | |
| 3964 | NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED); |
| 3965 | |
| 3966 | if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf) |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 3967 | tcp_clamp_window(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3968 | else if (tcp_memory_pressure) |
| 3969 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss); |
| 3970 | |
| 3971 | tcp_collapse_ofo_queue(sk); |
David S. Miller | 8728b83 | 2005-08-09 19:25:21 -0700 | [diff] [blame] | 3972 | tcp_collapse(sk, &sk->sk_receive_queue, |
| 3973 | sk->sk_receive_queue.next, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3974 | (struct sk_buff*)&sk->sk_receive_queue, |
| 3975 | tp->copied_seq, tp->rcv_nxt); |
| 3976 | sk_stream_mem_reclaim(sk); |
| 3977 | |
| 3978 | if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) |
| 3979 | return 0; |
| 3980 | |
| 3981 | /* Collapsing did not help, destructive actions follow. |
| 3982 | * This must not ever occur. */ |
| 3983 | |
| 3984 | /* First, purge the out_of_order queue. */ |
David S. Miller | b03efcf | 2005-07-08 14:57:23 -0700 | [diff] [blame] | 3985 | if (!skb_queue_empty(&tp->out_of_order_queue)) { |
| 3986 | NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3987 | __skb_queue_purge(&tp->out_of_order_queue); |
| 3988 | |
| 3989 | /* Reset SACK state. A conforming SACK implementation will |
| 3990 | * do the same at a timeout based retransmit. When a connection |
| 3991 | * is in a sad state like this, we care only about integrity |
| 3992 | * of the connection not performance. |
| 3993 | */ |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 3994 | if (tcp_is_sack(tp)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3995 | tcp_sack_reset(&tp->rx_opt); |
| 3996 | sk_stream_mem_reclaim(sk); |
| 3997 | } |
| 3998 | |
| 3999 | if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) |
| 4000 | return 0; |
| 4001 | |
| 4002 | /* If we are really being abused, tell the caller to silently |
| 4003 | * drop receive data on the floor. It will get retransmitted |
| 4004 | * and hopefully then we'll have sufficient space. |
| 4005 | */ |
| 4006 | NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED); |
| 4007 | |
| 4008 | /* Massive buffer overcommit. */ |
| 4009 | tp->pred_flags = 0; |
| 4010 | return -1; |
| 4011 | } |
| 4012 | |
| 4013 | |
| 4014 | /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. |
| 4015 | * As additional protections, we do not touch cwnd in retransmission phases, |
| 4016 | * and if application hit its sndbuf limit recently. |
| 4017 | */ |
| 4018 | void tcp_cwnd_application_limited(struct sock *sk) |
| 4019 | { |
| 4020 | struct tcp_sock *tp = tcp_sk(sk); |
| 4021 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 4022 | if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4023 | sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { |
| 4024 | /* Limited by application or receiver window. */ |
Ilpo Järvinen | d254bcd | 2006-08-04 16:57:42 -0700 | [diff] [blame] | 4025 | u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); |
| 4026 | u32 win_used = max(tp->snd_cwnd_used, init_win); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4027 | if (win_used < tp->snd_cwnd) { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 4028 | tp->snd_ssthresh = tcp_current_ssthresh(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4029 | tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; |
| 4030 | } |
| 4031 | tp->snd_cwnd_used = 0; |
| 4032 | } |
| 4033 | tp->snd_cwnd_stamp = tcp_time_stamp; |
| 4034 | } |
| 4035 | |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 4036 | static int tcp_should_expand_sndbuf(struct sock *sk) |
David S. Miller | 0d9901d | 2005-07-05 15:21:10 -0700 | [diff] [blame] | 4037 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 4038 | struct tcp_sock *tp = tcp_sk(sk); |
| 4039 | |
David S. Miller | 0d9901d | 2005-07-05 15:21:10 -0700 | [diff] [blame] | 4040 | /* If the user specified a specific send buffer setting, do |
| 4041 | * not modify it. |
| 4042 | */ |
| 4043 | if (sk->sk_userlocks & SOCK_SNDBUF_LOCK) |
| 4044 | return 0; |
| 4045 | |
| 4046 | /* If we are under global TCP memory pressure, do not expand. */ |
| 4047 | if (tcp_memory_pressure) |
| 4048 | return 0; |
| 4049 | |
| 4050 | /* If we are under soft global TCP memory pressure, do not expand. */ |
| 4051 | if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0]) |
| 4052 | return 0; |
| 4053 | |
| 4054 | /* If we filled the congestion window, do not expand. */ |
| 4055 | if (tp->packets_out >= tp->snd_cwnd) |
| 4056 | return 0; |
| 4057 | |
| 4058 | return 1; |
| 4059 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4060 | |
| 4061 | /* When incoming ACK allowed to free some skb from write_queue, |
| 4062 | * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket |
| 4063 | * on the exit from tcp input handler. |
| 4064 | * |
| 4065 | * PROBLEM: sndbuf expansion does not work well with largesend. |
| 4066 | */ |
| 4067 | static void tcp_new_space(struct sock *sk) |
| 4068 | { |
| 4069 | struct tcp_sock *tp = tcp_sk(sk); |
| 4070 | |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 4071 | if (tcp_should_expand_sndbuf(sk)) { |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 4072 | int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) + |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4073 | MAX_TCP_HEADER + 16 + sizeof(struct sk_buff), |
| 4074 | demanded = max_t(unsigned int, tp->snd_cwnd, |
| 4075 | tp->reordering + 1); |
| 4076 | sndmem *= 2*demanded; |
| 4077 | if (sndmem > sk->sk_sndbuf) |
| 4078 | sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]); |
| 4079 | tp->snd_cwnd_stamp = tcp_time_stamp; |
| 4080 | } |
| 4081 | |
| 4082 | sk->sk_write_space(sk); |
| 4083 | } |
| 4084 | |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 4085 | static void tcp_check_space(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4086 | { |
| 4087 | if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) { |
| 4088 | sock_reset_flag(sk, SOCK_QUEUE_SHRUNK); |
| 4089 | if (sk->sk_socket && |
| 4090 | test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) |
| 4091 | tcp_new_space(sk); |
| 4092 | } |
| 4093 | } |
| 4094 | |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 4095 | static inline void tcp_data_snd_check(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4096 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 4097 | tcp_push_pending_frames(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4098 | tcp_check_space(sk); |
| 4099 | } |
| 4100 | |
| 4101 | /* |
| 4102 | * Check if sending an ack is needed. |
| 4103 | */ |
| 4104 | static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible) |
| 4105 | { |
| 4106 | struct tcp_sock *tp = tcp_sk(sk); |
| 4107 | |
| 4108 | /* More than one full frame received... */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 4109 | if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4110 | /* ... and right edge of window advances far enough. |
| 4111 | * (tcp_recvmsg() will send ACK otherwise). Or... |
| 4112 | */ |
| 4113 | && __tcp_select_window(sk) >= tp->rcv_wnd) || |
| 4114 | /* We ACK each frame or... */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 4115 | tcp_in_quickack_mode(sk) || |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4116 | /* We have out of order data. */ |
| 4117 | (ofo_possible && |
| 4118 | skb_peek(&tp->out_of_order_queue))) { |
| 4119 | /* Then ack it now */ |
| 4120 | tcp_send_ack(sk); |
| 4121 | } else { |
| 4122 | /* Else, send delayed ack. */ |
| 4123 | tcp_send_delayed_ack(sk); |
| 4124 | } |
| 4125 | } |
| 4126 | |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 4127 | static inline void tcp_ack_snd_check(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4128 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 4129 | if (!inet_csk_ack_scheduled(sk)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4130 | /* We sent a data segment already. */ |
| 4131 | return; |
| 4132 | } |
| 4133 | __tcp_ack_snd_check(sk, 1); |
| 4134 | } |
| 4135 | |
| 4136 | /* |
| 4137 | * This routine is only called when we have urgent data |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 4138 | * signaled. Its the 'slow' part of tcp_urg. It could be |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4139 | * moved inline now as tcp_urg is only called from one |
| 4140 | * place. We handle URGent data wrong. We have to - as |
| 4141 | * BSD still doesn't use the correction from RFC961. |
| 4142 | * For 1003.1g we should support a new option TCP_STDURG to permit |
| 4143 | * either form (or just set the sysctl tcp_stdurg). |
| 4144 | */ |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 4145 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4146 | static void tcp_check_urg(struct sock * sk, struct tcphdr * th) |
| 4147 | { |
| 4148 | struct tcp_sock *tp = tcp_sk(sk); |
| 4149 | u32 ptr = ntohs(th->urg_ptr); |
| 4150 | |
| 4151 | if (ptr && !sysctl_tcp_stdurg) |
| 4152 | ptr--; |
| 4153 | ptr += ntohl(th->seq); |
| 4154 | |
| 4155 | /* Ignore urgent data that we've already seen and read. */ |
| 4156 | if (after(tp->copied_seq, ptr)) |
| 4157 | return; |
| 4158 | |
| 4159 | /* Do not replay urg ptr. |
| 4160 | * |
| 4161 | * NOTE: interesting situation not covered by specs. |
| 4162 | * Misbehaving sender may send urg ptr, pointing to segment, |
| 4163 | * which we already have in ofo queue. We are not able to fetch |
| 4164 | * such data and will stay in TCP_URG_NOTYET until will be eaten |
| 4165 | * by recvmsg(). Seems, we are not obliged to handle such wicked |
| 4166 | * situations. But it is worth to think about possibility of some |
| 4167 | * DoSes using some hypothetical application level deadlock. |
| 4168 | */ |
| 4169 | if (before(ptr, tp->rcv_nxt)) |
| 4170 | return; |
| 4171 | |
| 4172 | /* Do we already have a newer (or duplicate) urgent pointer? */ |
| 4173 | if (tp->urg_data && !after(ptr, tp->urg_seq)) |
| 4174 | return; |
| 4175 | |
| 4176 | /* Tell the world about our new urgent pointer. */ |
| 4177 | sk_send_sigurg(sk); |
| 4178 | |
| 4179 | /* We may be adding urgent data when the last byte read was |
| 4180 | * urgent. To do this requires some care. We cannot just ignore |
| 4181 | * tp->copied_seq since we would read the last urgent byte again |
| 4182 | * as data, nor can we alter copied_seq until this data arrives |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 4183 | * or we break the semantics of SIOCATMARK (and thus sockatmark()) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4184 | * |
| 4185 | * NOTE. Double Dutch. Rendering to plain English: author of comment |
| 4186 | * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB); |
| 4187 | * and expect that both A and B disappear from stream. This is _wrong_. |
| 4188 | * Though this happens in BSD with high probability, this is occasional. |
| 4189 | * Any application relying on this is buggy. Note also, that fix "works" |
| 4190 | * only in this artificial test. Insert some normal data between A and B and we will |
| 4191 | * decline of BSD again. Verdict: it is better to remove to trap |
| 4192 | * buggy users. |
| 4193 | */ |
| 4194 | if (tp->urg_seq == tp->copied_seq && tp->urg_data && |
| 4195 | !sock_flag(sk, SOCK_URGINLINE) && |
| 4196 | tp->copied_seq != tp->rcv_nxt) { |
| 4197 | struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); |
| 4198 | tp->copied_seq++; |
| 4199 | if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) { |
David S. Miller | 8728b83 | 2005-08-09 19:25:21 -0700 | [diff] [blame] | 4200 | __skb_unlink(skb, &sk->sk_receive_queue); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4201 | __kfree_skb(skb); |
| 4202 | } |
| 4203 | } |
| 4204 | |
| 4205 | tp->urg_data = TCP_URG_NOTYET; |
| 4206 | tp->urg_seq = ptr; |
| 4207 | |
| 4208 | /* Disable header prediction. */ |
| 4209 | tp->pred_flags = 0; |
| 4210 | } |
| 4211 | |
| 4212 | /* This is the 'fast' part of urgent handling. */ |
| 4213 | static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th) |
| 4214 | { |
| 4215 | struct tcp_sock *tp = tcp_sk(sk); |
| 4216 | |
| 4217 | /* Check if we get a new urgent pointer - normally not. */ |
| 4218 | if (th->urg) |
| 4219 | tcp_check_urg(sk,th); |
| 4220 | |
| 4221 | /* Do we wait for any urgent data? - normally not... */ |
| 4222 | if (tp->urg_data == TCP_URG_NOTYET) { |
| 4223 | u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) - |
| 4224 | th->syn; |
| 4225 | |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 4226 | /* Is the urgent pointer pointing into this packet? */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4227 | if (ptr < skb->len) { |
| 4228 | u8 tmp; |
| 4229 | if (skb_copy_bits(skb, ptr, &tmp, 1)) |
| 4230 | BUG(); |
| 4231 | tp->urg_data = TCP_URG_VALID | tmp; |
| 4232 | if (!sock_flag(sk, SOCK_DEAD)) |
| 4233 | sk->sk_data_ready(sk, 0); |
| 4234 | } |
| 4235 | } |
| 4236 | } |
| 4237 | |
| 4238 | static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen) |
| 4239 | { |
| 4240 | struct tcp_sock *tp = tcp_sk(sk); |
| 4241 | int chunk = skb->len - hlen; |
| 4242 | int err; |
| 4243 | |
| 4244 | local_bh_enable(); |
Herbert Xu | 6047637 | 2007-04-09 11:59:39 -0700 | [diff] [blame] | 4245 | if (skb_csum_unnecessary(skb)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4246 | err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk); |
| 4247 | else |
| 4248 | err = skb_copy_and_csum_datagram_iovec(skb, hlen, |
| 4249 | tp->ucopy.iov); |
| 4250 | |
| 4251 | if (!err) { |
| 4252 | tp->ucopy.len -= chunk; |
| 4253 | tp->copied_seq += chunk; |
| 4254 | tcp_rcv_space_adjust(sk); |
| 4255 | } |
| 4256 | |
| 4257 | local_bh_disable(); |
| 4258 | return err; |
| 4259 | } |
| 4260 | |
Al Viro | b51655b | 2006-11-14 21:40:42 -0800 | [diff] [blame] | 4261 | static __sum16 __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4262 | { |
Al Viro | b51655b | 2006-11-14 21:40:42 -0800 | [diff] [blame] | 4263 | __sum16 result; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4264 | |
| 4265 | if (sock_owned_by_user(sk)) { |
| 4266 | local_bh_enable(); |
| 4267 | result = __tcp_checksum_complete(skb); |
| 4268 | local_bh_disable(); |
| 4269 | } else { |
| 4270 | result = __tcp_checksum_complete(skb); |
| 4271 | } |
| 4272 | return result; |
| 4273 | } |
| 4274 | |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 4275 | static inline int tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4276 | { |
Herbert Xu | 6047637 | 2007-04-09 11:59:39 -0700 | [diff] [blame] | 4277 | return !skb_csum_unnecessary(skb) && |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4278 | __tcp_checksum_complete_user(sk, skb); |
| 4279 | } |
| 4280 | |
Chris Leech | 1a2449a | 2006-05-23 18:05:53 -0700 | [diff] [blame] | 4281 | #ifdef CONFIG_NET_DMA |
| 4282 | static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, int hlen) |
| 4283 | { |
| 4284 | struct tcp_sock *tp = tcp_sk(sk); |
| 4285 | int chunk = skb->len - hlen; |
| 4286 | int dma_cookie; |
| 4287 | int copied_early = 0; |
| 4288 | |
| 4289 | if (tp->ucopy.wakeup) |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 4290 | return 0; |
Chris Leech | 1a2449a | 2006-05-23 18:05:53 -0700 | [diff] [blame] | 4291 | |
| 4292 | if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list) |
| 4293 | tp->ucopy.dma_chan = get_softnet_dma(); |
| 4294 | |
Herbert Xu | 6047637 | 2007-04-09 11:59:39 -0700 | [diff] [blame] | 4295 | if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) { |
Chris Leech | 1a2449a | 2006-05-23 18:05:53 -0700 | [diff] [blame] | 4296 | |
| 4297 | dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan, |
| 4298 | skb, hlen, tp->ucopy.iov, chunk, tp->ucopy.pinned_list); |
| 4299 | |
| 4300 | if (dma_cookie < 0) |
| 4301 | goto out; |
| 4302 | |
| 4303 | tp->ucopy.dma_cookie = dma_cookie; |
| 4304 | copied_early = 1; |
| 4305 | |
| 4306 | tp->ucopy.len -= chunk; |
| 4307 | tp->copied_seq += chunk; |
| 4308 | tcp_rcv_space_adjust(sk); |
| 4309 | |
| 4310 | if ((tp->ucopy.len == 0) || |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 4311 | (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) || |
Chris Leech | 1a2449a | 2006-05-23 18:05:53 -0700 | [diff] [blame] | 4312 | (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) { |
| 4313 | tp->ucopy.wakeup = 1; |
| 4314 | sk->sk_data_ready(sk, 0); |
| 4315 | } |
| 4316 | } else if (chunk > 0) { |
| 4317 | tp->ucopy.wakeup = 1; |
| 4318 | sk->sk_data_ready(sk, 0); |
| 4319 | } |
| 4320 | out: |
| 4321 | return copied_early; |
| 4322 | } |
| 4323 | #endif /* CONFIG_NET_DMA */ |
| 4324 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4325 | /* |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 4326 | * TCP receive function for the ESTABLISHED state. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4327 | * |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 4328 | * It is split into a fast path and a slow path. The fast path is |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4329 | * disabled when: |
| 4330 | * - A zero window was announced from us - zero window probing |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 4331 | * is only handled properly in the slow path. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4332 | * - Out of order segments arrived. |
| 4333 | * - Urgent data is expected. |
| 4334 | * - There is no buffer space left |
| 4335 | * - Unexpected TCP flags/window values/header lengths are received |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 4336 | * (detected by checking the TCP header against pred_flags) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4337 | * - Data is sent in both directions. Fast path only supports pure senders |
| 4338 | * or pure receivers (this means either the sequence number or the ack |
| 4339 | * value must stay constant) |
| 4340 | * - Unexpected TCP option. |
| 4341 | * |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 4342 | * When these conditions are not satisfied it drops into a standard |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4343 | * receive procedure patterned after RFC793 to handle all cases. |
| 4344 | * The first three cases are guaranteed by proper pred_flags setting, |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 4345 | * the rest is checked inline. Fast processing is turned on in |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4346 | * tcp_data_queue when everything is OK. |
| 4347 | */ |
| 4348 | int tcp_rcv_established(struct sock *sk, struct sk_buff *skb, |
| 4349 | struct tcphdr *th, unsigned len) |
| 4350 | { |
| 4351 | struct tcp_sock *tp = tcp_sk(sk); |
| 4352 | |
| 4353 | /* |
| 4354 | * Header prediction. |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 4355 | * The code loosely follows the one in the famous |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4356 | * "30 instruction TCP receive" Van Jacobson mail. |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 4357 | * |
| 4358 | * Van's trick is to deposit buffers into socket queue |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4359 | * on a device interrupt, to call tcp_recv function |
| 4360 | * on the receive process context and checksum and copy |
| 4361 | * the buffer to user space. smart... |
| 4362 | * |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 4363 | * Our current scheme is not silly either but we take the |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4364 | * extra cost of the net_bh soft interrupt processing... |
| 4365 | * We do checksum and copy also but from device to kernel. |
| 4366 | */ |
| 4367 | |
| 4368 | tp->rx_opt.saw_tstamp = 0; |
| 4369 | |
| 4370 | /* pred_flags is 0xS?10 << 16 + snd_wnd |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 4371 | * if header_prediction is to be made |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4372 | * 'S' will always be tp->tcp_header_len >> 2 |
| 4373 | * '?' will be 0 for the fast path, otherwise pred_flags is 0 to |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 4374 | * turn it off (when there are holes in the receive |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4375 | * space for instance) |
| 4376 | * PSH flag is ignored. |
| 4377 | */ |
| 4378 | |
| 4379 | if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags && |
| 4380 | TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { |
| 4381 | int tcp_header_len = tp->tcp_header_len; |
| 4382 | |
| 4383 | /* Timestamp header prediction: tcp_header_len |
| 4384 | * is automatically equal to th->doff*4 due to pred_flags |
| 4385 | * match. |
| 4386 | */ |
| 4387 | |
| 4388 | /* Check timestamp */ |
| 4389 | if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) { |
Al Viro | 4f3608b | 2006-09-27 18:35:09 -0700 | [diff] [blame] | 4390 | __be32 *ptr = (__be32 *)(th + 1); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4391 | |
| 4392 | /* No? Slow path! */ |
Al Viro | 4f3608b | 2006-09-27 18:35:09 -0700 | [diff] [blame] | 4393 | if (*ptr != htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4394 | | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) |
| 4395 | goto slow_path; |
| 4396 | |
| 4397 | tp->rx_opt.saw_tstamp = 1; |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 4398 | ++ptr; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4399 | tp->rx_opt.rcv_tsval = ntohl(*ptr); |
| 4400 | ++ptr; |
| 4401 | tp->rx_opt.rcv_tsecr = ntohl(*ptr); |
| 4402 | |
| 4403 | /* If PAWS failed, check it more carefully in slow path */ |
| 4404 | if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0) |
| 4405 | goto slow_path; |
| 4406 | |
| 4407 | /* DO NOT update ts_recent here, if checksum fails |
| 4408 | * and timestamp was corrupted part, it will result |
| 4409 | * in a hung connection since we will drop all |
| 4410 | * future packets due to the PAWS test. |
| 4411 | */ |
| 4412 | } |
| 4413 | |
| 4414 | if (len <= tcp_header_len) { |
| 4415 | /* Bulk data transfer: sender */ |
| 4416 | if (len == tcp_header_len) { |
| 4417 | /* Predicted packet is in window by definition. |
| 4418 | * seq == rcv_nxt and rcv_wup <= rcv_nxt. |
| 4419 | * Hence, check seq<=rcv_wup reduces to: |
| 4420 | */ |
| 4421 | if (tcp_header_len == |
| 4422 | (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && |
| 4423 | tp->rcv_nxt == tp->rcv_wup) |
| 4424 | tcp_store_ts_recent(tp); |
| 4425 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4426 | /* We know that such packets are checksummed |
| 4427 | * on entry. |
| 4428 | */ |
| 4429 | tcp_ack(sk, skb, 0); |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 4430 | __kfree_skb(skb); |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 4431 | tcp_data_snd_check(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4432 | return 0; |
| 4433 | } else { /* Header too small */ |
| 4434 | TCP_INC_STATS_BH(TCP_MIB_INERRS); |
| 4435 | goto discard; |
| 4436 | } |
| 4437 | } else { |
| 4438 | int eaten = 0; |
Chris Leech | 1a2449a | 2006-05-23 18:05:53 -0700 | [diff] [blame] | 4439 | int copied_early = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4440 | |
Chris Leech | 1a2449a | 2006-05-23 18:05:53 -0700 | [diff] [blame] | 4441 | if (tp->copied_seq == tp->rcv_nxt && |
| 4442 | len - tcp_header_len <= tp->ucopy.len) { |
| 4443 | #ifdef CONFIG_NET_DMA |
| 4444 | if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) { |
| 4445 | copied_early = 1; |
| 4446 | eaten = 1; |
| 4447 | } |
| 4448 | #endif |
| 4449 | if (tp->ucopy.task == current && sock_owned_by_user(sk) && !copied_early) { |
| 4450 | __set_current_state(TASK_RUNNING); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4451 | |
Chris Leech | 1a2449a | 2006-05-23 18:05:53 -0700 | [diff] [blame] | 4452 | if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) |
| 4453 | eaten = 1; |
| 4454 | } |
| 4455 | if (eaten) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4456 | /* Predicted packet is in window by definition. |
| 4457 | * seq == rcv_nxt and rcv_wup <= rcv_nxt. |
| 4458 | * Hence, check seq<=rcv_wup reduces to: |
| 4459 | */ |
| 4460 | if (tcp_header_len == |
| 4461 | (sizeof(struct tcphdr) + |
| 4462 | TCPOLEN_TSTAMP_ALIGNED) && |
| 4463 | tp->rcv_nxt == tp->rcv_wup) |
| 4464 | tcp_store_ts_recent(tp); |
| 4465 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 4466 | tcp_rcv_rtt_measure_ts(sk, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4467 | |
| 4468 | __skb_pull(skb, tcp_header_len); |
| 4469 | tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; |
| 4470 | NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4471 | } |
Chris Leech | 1a2449a | 2006-05-23 18:05:53 -0700 | [diff] [blame] | 4472 | if (copied_early) |
| 4473 | tcp_cleanup_rbuf(sk, skb->len); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4474 | } |
| 4475 | if (!eaten) { |
| 4476 | if (tcp_checksum_complete_user(sk, skb)) |
| 4477 | goto csum_error; |
| 4478 | |
| 4479 | /* Predicted packet is in window by definition. |
| 4480 | * seq == rcv_nxt and rcv_wup <= rcv_nxt. |
| 4481 | * Hence, check seq<=rcv_wup reduces to: |
| 4482 | */ |
| 4483 | if (tcp_header_len == |
| 4484 | (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && |
| 4485 | tp->rcv_nxt == tp->rcv_wup) |
| 4486 | tcp_store_ts_recent(tp); |
| 4487 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 4488 | tcp_rcv_rtt_measure_ts(sk, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4489 | |
| 4490 | if ((int)skb->truesize > sk->sk_forward_alloc) |
| 4491 | goto step5; |
| 4492 | |
| 4493 | NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS); |
| 4494 | |
| 4495 | /* Bulk data transfer: receiver */ |
| 4496 | __skb_pull(skb,tcp_header_len); |
| 4497 | __skb_queue_tail(&sk->sk_receive_queue, skb); |
| 4498 | sk_stream_set_owner_r(skb, sk); |
| 4499 | tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; |
| 4500 | } |
| 4501 | |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 4502 | tcp_event_data_recv(sk, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4503 | |
| 4504 | if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) { |
| 4505 | /* Well, only one small jumplet in fast path... */ |
| 4506 | tcp_ack(sk, skb, FLAG_DATA); |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 4507 | tcp_data_snd_check(sk); |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 4508 | if (!inet_csk_ack_scheduled(sk)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4509 | goto no_ack; |
| 4510 | } |
| 4511 | |
David S. Miller | 3143241 | 2005-05-23 12:03:06 -0700 | [diff] [blame] | 4512 | __tcp_ack_snd_check(sk, 0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4513 | no_ack: |
Chris Leech | 1a2449a | 2006-05-23 18:05:53 -0700 | [diff] [blame] | 4514 | #ifdef CONFIG_NET_DMA |
| 4515 | if (copied_early) |
| 4516 | __skb_queue_tail(&sk->sk_async_wait_queue, skb); |
| 4517 | else |
| 4518 | #endif |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4519 | if (eaten) |
| 4520 | __kfree_skb(skb); |
| 4521 | else |
| 4522 | sk->sk_data_ready(sk, 0); |
| 4523 | return 0; |
| 4524 | } |
| 4525 | } |
| 4526 | |
| 4527 | slow_path: |
| 4528 | if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb)) |
| 4529 | goto csum_error; |
| 4530 | |
| 4531 | /* |
| 4532 | * RFC1323: H1. Apply PAWS check first. |
| 4533 | */ |
| 4534 | if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp && |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 4535 | tcp_paws_discard(sk, skb)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4536 | if (!th->rst) { |
| 4537 | NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED); |
| 4538 | tcp_send_dupack(sk, skb); |
| 4539 | goto discard; |
| 4540 | } |
| 4541 | /* Resets are accepted even if PAWS failed. |
| 4542 | |
| 4543 | ts_recent update must be made after we are sure |
| 4544 | that the packet is in window. |
| 4545 | */ |
| 4546 | } |
| 4547 | |
| 4548 | /* |
| 4549 | * Standard slow path. |
| 4550 | */ |
| 4551 | |
| 4552 | if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) { |
| 4553 | /* RFC793, page 37: "In all states except SYN-SENT, all reset |
| 4554 | * (RST) segments are validated by checking their SEQ-fields." |
| 4555 | * And page 69: "If an incoming segment is not acceptable, |
| 4556 | * an acknowledgment should be sent in reply (unless the RST bit |
| 4557 | * is set, if so drop the segment and return)". |
| 4558 | */ |
| 4559 | if (!th->rst) |
| 4560 | tcp_send_dupack(sk, skb); |
| 4561 | goto discard; |
| 4562 | } |
| 4563 | |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 4564 | if (th->rst) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4565 | tcp_reset(sk); |
| 4566 | goto discard; |
| 4567 | } |
| 4568 | |
| 4569 | tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); |
| 4570 | |
| 4571 | if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { |
| 4572 | TCP_INC_STATS_BH(TCP_MIB_INERRS); |
| 4573 | NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN); |
| 4574 | tcp_reset(sk); |
| 4575 | return 1; |
| 4576 | } |
| 4577 | |
| 4578 | step5: |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 4579 | if (th->ack) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4580 | tcp_ack(sk, skb, FLAG_SLOWPATH); |
| 4581 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 4582 | tcp_rcv_rtt_measure_ts(sk, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4583 | |
| 4584 | /* Process urgent data. */ |
| 4585 | tcp_urg(sk, skb, th); |
| 4586 | |
| 4587 | /* step 7: process the segment text */ |
| 4588 | tcp_data_queue(sk, skb); |
| 4589 | |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 4590 | tcp_data_snd_check(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4591 | tcp_ack_snd_check(sk); |
| 4592 | return 0; |
| 4593 | |
| 4594 | csum_error: |
| 4595 | TCP_INC_STATS_BH(TCP_MIB_INERRS); |
| 4596 | |
| 4597 | discard: |
| 4598 | __kfree_skb(skb); |
| 4599 | return 0; |
| 4600 | } |
| 4601 | |
| 4602 | static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb, |
| 4603 | struct tcphdr *th, unsigned len) |
| 4604 | { |
| 4605 | struct tcp_sock *tp = tcp_sk(sk); |
Arnaldo Carvalho de Melo | d83d846 | 2005-12-13 23:26:10 -0800 | [diff] [blame] | 4606 | struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4607 | int saved_clamp = tp->rx_opt.mss_clamp; |
| 4608 | |
| 4609 | tcp_parse_options(skb, &tp->rx_opt, 0); |
| 4610 | |
| 4611 | if (th->ack) { |
| 4612 | /* rfc793: |
| 4613 | * "If the state is SYN-SENT then |
| 4614 | * first check the ACK bit |
| 4615 | * If the ACK bit is set |
| 4616 | * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send |
| 4617 | * a reset (unless the RST bit is set, if so drop |
| 4618 | * the segment and return)" |
| 4619 | * |
| 4620 | * We do not send data with SYN, so that RFC-correct |
| 4621 | * test reduces to: |
| 4622 | */ |
| 4623 | if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt) |
| 4624 | goto reset_and_undo; |
| 4625 | |
| 4626 | if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && |
| 4627 | !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp, |
| 4628 | tcp_time_stamp)) { |
| 4629 | NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED); |
| 4630 | goto reset_and_undo; |
| 4631 | } |
| 4632 | |
| 4633 | /* Now ACK is acceptable. |
| 4634 | * |
| 4635 | * "If the RST bit is set |
| 4636 | * If the ACK was acceptable then signal the user "error: |
| 4637 | * connection reset", drop the segment, enter CLOSED state, |
| 4638 | * delete TCB, and return." |
| 4639 | */ |
| 4640 | |
| 4641 | if (th->rst) { |
| 4642 | tcp_reset(sk); |
| 4643 | goto discard; |
| 4644 | } |
| 4645 | |
| 4646 | /* rfc793: |
| 4647 | * "fifth, if neither of the SYN or RST bits is set then |
| 4648 | * drop the segment and return." |
| 4649 | * |
| 4650 | * See note below! |
| 4651 | * --ANK(990513) |
| 4652 | */ |
| 4653 | if (!th->syn) |
| 4654 | goto discard_and_undo; |
| 4655 | |
| 4656 | /* rfc793: |
| 4657 | * "If the SYN bit is on ... |
| 4658 | * are acceptable then ... |
| 4659 | * (our SYN has been ACKed), change the connection |
| 4660 | * state to ESTABLISHED..." |
| 4661 | */ |
| 4662 | |
| 4663 | TCP_ECN_rcv_synack(tp, th); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4664 | |
| 4665 | tp->snd_wl1 = TCP_SKB_CB(skb)->seq; |
| 4666 | tcp_ack(sk, skb, FLAG_SLOWPATH); |
| 4667 | |
| 4668 | /* Ok.. it's good. Set up sequence numbers and |
| 4669 | * move to established. |
| 4670 | */ |
| 4671 | tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; |
| 4672 | tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; |
| 4673 | |
| 4674 | /* RFC1323: The window in SYN & SYN/ACK segments is |
| 4675 | * never scaled. |
| 4676 | */ |
| 4677 | tp->snd_wnd = ntohs(th->window); |
| 4678 | tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq); |
| 4679 | |
| 4680 | if (!tp->rx_opt.wscale_ok) { |
| 4681 | tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0; |
| 4682 | tp->window_clamp = min(tp->window_clamp, 65535U); |
| 4683 | } |
| 4684 | |
| 4685 | if (tp->rx_opt.saw_tstamp) { |
| 4686 | tp->rx_opt.tstamp_ok = 1; |
| 4687 | tp->tcp_header_len = |
| 4688 | sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; |
| 4689 | tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; |
| 4690 | tcp_store_ts_recent(tp); |
| 4691 | } else { |
| 4692 | tp->tcp_header_len = sizeof(struct tcphdr); |
| 4693 | } |
| 4694 | |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 4695 | if (tcp_is_sack(tp) && sysctl_tcp_fack) |
| 4696 | tcp_enable_fack(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4697 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 4698 | tcp_mtup_init(sk); |
Arnaldo Carvalho de Melo | d83d846 | 2005-12-13 23:26:10 -0800 | [diff] [blame] | 4699 | tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4700 | tcp_initialize_rcv_mss(sk); |
| 4701 | |
| 4702 | /* Remember, tcp_poll() does not lock socket! |
| 4703 | * Change state from SYN-SENT only after copied_seq |
| 4704 | * is initialized. */ |
| 4705 | tp->copied_seq = tp->rcv_nxt; |
Ralf Baechle | e16aa20 | 2006-12-07 00:11:33 -0800 | [diff] [blame] | 4706 | smp_mb(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4707 | tcp_set_state(sk, TCP_ESTABLISHED); |
| 4708 | |
Venkat Yekkirala | 6b87769 | 2006-11-08 17:04:09 -0600 | [diff] [blame] | 4709 | security_inet_conn_established(sk, skb); |
| 4710 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4711 | /* Make sure socket is routed, for correct metrics. */ |
Arnaldo Carvalho de Melo | 8292a17 | 2005-12-13 23:15:52 -0800 | [diff] [blame] | 4712 | icsk->icsk_af_ops->rebuild_header(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4713 | |
| 4714 | tcp_init_metrics(sk); |
| 4715 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 4716 | tcp_init_congestion_control(sk); |
Stephen Hemminger | 317a76f | 2005-06-23 12:19:55 -0700 | [diff] [blame] | 4717 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4718 | /* Prevent spurious tcp_cwnd_restart() on first data |
| 4719 | * packet. |
| 4720 | */ |
| 4721 | tp->lsndtime = tcp_time_stamp; |
| 4722 | |
| 4723 | tcp_init_buffer_space(sk); |
| 4724 | |
| 4725 | if (sock_flag(sk, SOCK_KEEPOPEN)) |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 4726 | inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4727 | |
| 4728 | if (!tp->rx_opt.snd_wscale) |
| 4729 | __tcp_fast_path_on(tp, tp->snd_wnd); |
| 4730 | else |
| 4731 | tp->pred_flags = 0; |
| 4732 | |
| 4733 | if (!sock_flag(sk, SOCK_DEAD)) { |
| 4734 | sk->sk_state_change(sk); |
| 4735 | sk_wake_async(sk, 0, POLL_OUT); |
| 4736 | } |
| 4737 | |
Arnaldo Carvalho de Melo | 295f732 | 2005-08-09 20:11:56 -0700 | [diff] [blame] | 4738 | if (sk->sk_write_pending || |
| 4739 | icsk->icsk_accept_queue.rskq_defer_accept || |
| 4740 | icsk->icsk_ack.pingpong) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4741 | /* Save one ACK. Data will be ready after |
| 4742 | * several ticks, if write_pending is set. |
| 4743 | * |
| 4744 | * It may be deleted, but with this feature tcpdumps |
| 4745 | * look so _wonderfully_ clever, that I was not able |
| 4746 | * to stand against the temptation 8) --ANK |
| 4747 | */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 4748 | inet_csk_schedule_ack(sk); |
Arnaldo Carvalho de Melo | 295f732 | 2005-08-09 20:11:56 -0700 | [diff] [blame] | 4749 | icsk->icsk_ack.lrcvtime = tcp_time_stamp; |
| 4750 | icsk->icsk_ack.ato = TCP_ATO_MIN; |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 4751 | tcp_incr_quickack(sk); |
| 4752 | tcp_enter_quickack_mode(sk); |
Arnaldo Carvalho de Melo | 3f421ba | 2005-08-09 20:11:08 -0700 | [diff] [blame] | 4753 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, |
| 4754 | TCP_DELACK_MAX, TCP_RTO_MAX); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4755 | |
| 4756 | discard: |
| 4757 | __kfree_skb(skb); |
| 4758 | return 0; |
| 4759 | } else { |
| 4760 | tcp_send_ack(sk); |
| 4761 | } |
| 4762 | return -1; |
| 4763 | } |
| 4764 | |
| 4765 | /* No ACK in the segment */ |
| 4766 | |
| 4767 | if (th->rst) { |
| 4768 | /* rfc793: |
| 4769 | * "If the RST bit is set |
| 4770 | * |
| 4771 | * Otherwise (no ACK) drop the segment and return." |
| 4772 | */ |
| 4773 | |
| 4774 | goto discard_and_undo; |
| 4775 | } |
| 4776 | |
| 4777 | /* PAWS check. */ |
| 4778 | if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && tcp_paws_check(&tp->rx_opt, 0)) |
| 4779 | goto discard_and_undo; |
| 4780 | |
| 4781 | if (th->syn) { |
| 4782 | /* We see SYN without ACK. It is attempt of |
| 4783 | * simultaneous connect with crossed SYNs. |
| 4784 | * Particularly, it can be connect to self. |
| 4785 | */ |
| 4786 | tcp_set_state(sk, TCP_SYN_RECV); |
| 4787 | |
| 4788 | if (tp->rx_opt.saw_tstamp) { |
| 4789 | tp->rx_opt.tstamp_ok = 1; |
| 4790 | tcp_store_ts_recent(tp); |
| 4791 | tp->tcp_header_len = |
| 4792 | sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; |
| 4793 | } else { |
| 4794 | tp->tcp_header_len = sizeof(struct tcphdr); |
| 4795 | } |
| 4796 | |
| 4797 | tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; |
| 4798 | tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; |
| 4799 | |
| 4800 | /* RFC1323: The window in SYN & SYN/ACK segments is |
| 4801 | * never scaled. |
| 4802 | */ |
| 4803 | tp->snd_wnd = ntohs(th->window); |
| 4804 | tp->snd_wl1 = TCP_SKB_CB(skb)->seq; |
| 4805 | tp->max_window = tp->snd_wnd; |
| 4806 | |
| 4807 | TCP_ECN_rcv_syn(tp, th); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4808 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 4809 | tcp_mtup_init(sk); |
Arnaldo Carvalho de Melo | d83d846 | 2005-12-13 23:26:10 -0800 | [diff] [blame] | 4810 | tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4811 | tcp_initialize_rcv_mss(sk); |
| 4812 | |
| 4813 | |
| 4814 | tcp_send_synack(sk); |
| 4815 | #if 0 |
| 4816 | /* Note, we could accept data and URG from this segment. |
| 4817 | * There are no obstacles to make this. |
| 4818 | * |
| 4819 | * However, if we ignore data in ACKless segments sometimes, |
| 4820 | * we have no reasons to accept it sometimes. |
| 4821 | * Also, seems the code doing it in step6 of tcp_rcv_state_process |
| 4822 | * is not flawless. So, discard packet for sanity. |
| 4823 | * Uncomment this return to process the data. |
| 4824 | */ |
| 4825 | return -1; |
| 4826 | #else |
| 4827 | goto discard; |
| 4828 | #endif |
| 4829 | } |
| 4830 | /* "fifth, if neither of the SYN or RST bits is set then |
| 4831 | * drop the segment and return." |
| 4832 | */ |
| 4833 | |
| 4834 | discard_and_undo: |
| 4835 | tcp_clear_options(&tp->rx_opt); |
| 4836 | tp->rx_opt.mss_clamp = saved_clamp; |
| 4837 | goto discard; |
| 4838 | |
| 4839 | reset_and_undo: |
| 4840 | tcp_clear_options(&tp->rx_opt); |
| 4841 | tp->rx_opt.mss_clamp = saved_clamp; |
| 4842 | return 1; |
| 4843 | } |
| 4844 | |
| 4845 | |
| 4846 | /* |
| 4847 | * This function implements the receiving procedure of RFC 793 for |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 4848 | * all states except ESTABLISHED and TIME_WAIT. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4849 | * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be |
| 4850 | * address independent. |
| 4851 | */ |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 4852 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4853 | int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb, |
| 4854 | struct tcphdr *th, unsigned len) |
| 4855 | { |
| 4856 | struct tcp_sock *tp = tcp_sk(sk); |
Arnaldo Carvalho de Melo | 8292a17 | 2005-12-13 23:15:52 -0800 | [diff] [blame] | 4857 | struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4858 | int queued = 0; |
| 4859 | |
| 4860 | tp->rx_opt.saw_tstamp = 0; |
| 4861 | |
| 4862 | switch (sk->sk_state) { |
| 4863 | case TCP_CLOSE: |
| 4864 | goto discard; |
| 4865 | |
| 4866 | case TCP_LISTEN: |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 4867 | if (th->ack) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4868 | return 1; |
| 4869 | |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 4870 | if (th->rst) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4871 | goto discard; |
| 4872 | |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 4873 | if (th->syn) { |
Arnaldo Carvalho de Melo | 8292a17 | 2005-12-13 23:15:52 -0800 | [diff] [blame] | 4874 | if (icsk->icsk_af_ops->conn_request(sk, skb) < 0) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4875 | return 1; |
| 4876 | |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 4877 | /* Now we have several options: In theory there is |
| 4878 | * nothing else in the frame. KA9Q has an option to |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4879 | * send data with the syn, BSD accepts data with the |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 4880 | * syn up to the [to be] advertised window and |
| 4881 | * Solaris 2.1 gives you a protocol error. For now |
| 4882 | * we just ignore it, that fits the spec precisely |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4883 | * and avoids incompatibilities. It would be nice in |
| 4884 | * future to drop through and process the data. |
| 4885 | * |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 4886 | * Now that TTCP is starting to be used we ought to |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4887 | * queue this data. |
| 4888 | * But, this leaves one open to an easy denial of |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 4889 | * service attack, and SYN cookies can't defend |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4890 | * against this problem. So, we drop the data |
Masayuki Nakagawa | fb7e239 | 2007-01-23 20:15:06 -0800 | [diff] [blame] | 4891 | * in the interest of security over speed unless |
| 4892 | * it's still in use. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4893 | */ |
Masayuki Nakagawa | fb7e239 | 2007-01-23 20:15:06 -0800 | [diff] [blame] | 4894 | kfree_skb(skb); |
| 4895 | return 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4896 | } |
| 4897 | goto discard; |
| 4898 | |
| 4899 | case TCP_SYN_SENT: |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4900 | queued = tcp_rcv_synsent_state_process(sk, skb, th, len); |
| 4901 | if (queued >= 0) |
| 4902 | return queued; |
| 4903 | |
| 4904 | /* Do step6 onward by hand. */ |
| 4905 | tcp_urg(sk, skb, th); |
| 4906 | __kfree_skb(skb); |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 4907 | tcp_data_snd_check(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4908 | return 0; |
| 4909 | } |
| 4910 | |
| 4911 | if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp && |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 4912 | tcp_paws_discard(sk, skb)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4913 | if (!th->rst) { |
| 4914 | NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED); |
| 4915 | tcp_send_dupack(sk, skb); |
| 4916 | goto discard; |
| 4917 | } |
| 4918 | /* Reset is accepted even if it did not pass PAWS. */ |
| 4919 | } |
| 4920 | |
| 4921 | /* step 1: check sequence number */ |
| 4922 | if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) { |
| 4923 | if (!th->rst) |
| 4924 | tcp_send_dupack(sk, skb); |
| 4925 | goto discard; |
| 4926 | } |
| 4927 | |
| 4928 | /* step 2: check RST bit */ |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 4929 | if (th->rst) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4930 | tcp_reset(sk); |
| 4931 | goto discard; |
| 4932 | } |
| 4933 | |
| 4934 | tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); |
| 4935 | |
| 4936 | /* step 3: check security and precedence [ignored] */ |
| 4937 | |
| 4938 | /* step 4: |
| 4939 | * |
| 4940 | * Check for a SYN in window. |
| 4941 | */ |
| 4942 | if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { |
| 4943 | NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN); |
| 4944 | tcp_reset(sk); |
| 4945 | return 1; |
| 4946 | } |
| 4947 | |
| 4948 | /* step 5: check the ACK field */ |
| 4949 | if (th->ack) { |
| 4950 | int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH); |
| 4951 | |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 4952 | switch (sk->sk_state) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4953 | case TCP_SYN_RECV: |
| 4954 | if (acceptable) { |
| 4955 | tp->copied_seq = tp->rcv_nxt; |
Ralf Baechle | e16aa20 | 2006-12-07 00:11:33 -0800 | [diff] [blame] | 4956 | smp_mb(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4957 | tcp_set_state(sk, TCP_ESTABLISHED); |
| 4958 | sk->sk_state_change(sk); |
| 4959 | |
| 4960 | /* Note, that this wakeup is only for marginal |
| 4961 | * crossed SYN case. Passively open sockets |
| 4962 | * are not waked up, because sk->sk_sleep == |
| 4963 | * NULL and sk->sk_socket == NULL. |
| 4964 | */ |
| 4965 | if (sk->sk_socket) { |
| 4966 | sk_wake_async(sk,0,POLL_OUT); |
| 4967 | } |
| 4968 | |
| 4969 | tp->snd_una = TCP_SKB_CB(skb)->ack_seq; |
| 4970 | tp->snd_wnd = ntohs(th->window) << |
| 4971 | tp->rx_opt.snd_wscale; |
| 4972 | tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, |
| 4973 | TCP_SKB_CB(skb)->seq); |
| 4974 | |
| 4975 | /* tcp_ack considers this ACK as duplicate |
| 4976 | * and does not calculate rtt. |
| 4977 | * Fix it at least with timestamps. |
| 4978 | */ |
| 4979 | if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && |
| 4980 | !tp->srtt) |
Stephen Hemminger | 2d2abba | 2005-11-10 16:56:12 -0800 | [diff] [blame] | 4981 | tcp_ack_saw_tstamp(sk, 0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4982 | |
| 4983 | if (tp->rx_opt.tstamp_ok) |
| 4984 | tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; |
| 4985 | |
| 4986 | /* Make sure socket is routed, for |
| 4987 | * correct metrics. |
| 4988 | */ |
Arnaldo Carvalho de Melo | 8292a17 | 2005-12-13 23:15:52 -0800 | [diff] [blame] | 4989 | icsk->icsk_af_ops->rebuild_header(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4990 | |
| 4991 | tcp_init_metrics(sk); |
| 4992 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 4993 | tcp_init_congestion_control(sk); |
Stephen Hemminger | 317a76f | 2005-06-23 12:19:55 -0700 | [diff] [blame] | 4994 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4995 | /* Prevent spurious tcp_cwnd_restart() on |
| 4996 | * first data packet. |
| 4997 | */ |
| 4998 | tp->lsndtime = tcp_time_stamp; |
| 4999 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 5000 | tcp_mtup_init(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5001 | tcp_initialize_rcv_mss(sk); |
| 5002 | tcp_init_buffer_space(sk); |
| 5003 | tcp_fast_path_on(tp); |
| 5004 | } else { |
| 5005 | return 1; |
| 5006 | } |
| 5007 | break; |
| 5008 | |
| 5009 | case TCP_FIN_WAIT1: |
| 5010 | if (tp->snd_una == tp->write_seq) { |
| 5011 | tcp_set_state(sk, TCP_FIN_WAIT2); |
| 5012 | sk->sk_shutdown |= SEND_SHUTDOWN; |
| 5013 | dst_confirm(sk->sk_dst_cache); |
| 5014 | |
| 5015 | if (!sock_flag(sk, SOCK_DEAD)) |
| 5016 | /* Wake up lingering close() */ |
| 5017 | sk->sk_state_change(sk); |
| 5018 | else { |
| 5019 | int tmo; |
| 5020 | |
| 5021 | if (tp->linger2 < 0 || |
| 5022 | (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && |
| 5023 | after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) { |
| 5024 | tcp_done(sk); |
| 5025 | NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA); |
| 5026 | return 1; |
| 5027 | } |
| 5028 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 5029 | tmo = tcp_fin_time(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5030 | if (tmo > TCP_TIMEWAIT_LEN) { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 5031 | inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5032 | } else if (th->fin || sock_owned_by_user(sk)) { |
| 5033 | /* Bad case. We could lose such FIN otherwise. |
| 5034 | * It is not a big problem, but it looks confusing |
| 5035 | * and not so rare event. We still can lose it now, |
| 5036 | * if it spins in bh_lock_sock(), but it is really |
| 5037 | * marginal case. |
| 5038 | */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 5039 | inet_csk_reset_keepalive_timer(sk, tmo); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5040 | } else { |
| 5041 | tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); |
| 5042 | goto discard; |
| 5043 | } |
| 5044 | } |
| 5045 | } |
| 5046 | break; |
| 5047 | |
| 5048 | case TCP_CLOSING: |
| 5049 | if (tp->snd_una == tp->write_seq) { |
| 5050 | tcp_time_wait(sk, TCP_TIME_WAIT, 0); |
| 5051 | goto discard; |
| 5052 | } |
| 5053 | break; |
| 5054 | |
| 5055 | case TCP_LAST_ACK: |
| 5056 | if (tp->snd_una == tp->write_seq) { |
| 5057 | tcp_update_metrics(sk); |
| 5058 | tcp_done(sk); |
| 5059 | goto discard; |
| 5060 | } |
| 5061 | break; |
| 5062 | } |
| 5063 | } else |
| 5064 | goto discard; |
| 5065 | |
| 5066 | /* step 6: check the URG bit */ |
| 5067 | tcp_urg(sk, skb, th); |
| 5068 | |
| 5069 | /* step 7: process the segment text */ |
| 5070 | switch (sk->sk_state) { |
| 5071 | case TCP_CLOSE_WAIT: |
| 5072 | case TCP_CLOSING: |
| 5073 | case TCP_LAST_ACK: |
| 5074 | if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) |
| 5075 | break; |
| 5076 | case TCP_FIN_WAIT1: |
| 5077 | case TCP_FIN_WAIT2: |
| 5078 | /* RFC 793 says to queue data in these states, |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 5079 | * RFC 1122 says we MUST send a reset. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5080 | * BSD 4.4 also does reset. |
| 5081 | */ |
| 5082 | if (sk->sk_shutdown & RCV_SHUTDOWN) { |
| 5083 | if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && |
| 5084 | after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { |
| 5085 | NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA); |
| 5086 | tcp_reset(sk); |
| 5087 | return 1; |
| 5088 | } |
| 5089 | } |
| 5090 | /* Fall through */ |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 5091 | case TCP_ESTABLISHED: |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5092 | tcp_data_queue(sk, skb); |
| 5093 | queued = 1; |
| 5094 | break; |
| 5095 | } |
| 5096 | |
| 5097 | /* tcp_data could move socket to TIME-WAIT */ |
| 5098 | if (sk->sk_state != TCP_CLOSE) { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 5099 | tcp_data_snd_check(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5100 | tcp_ack_snd_check(sk); |
| 5101 | } |
| 5102 | |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 5103 | if (!queued) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5104 | discard: |
| 5105 | __kfree_skb(skb); |
| 5106 | } |
| 5107 | return 0; |
| 5108 | } |
| 5109 | |
| 5110 | EXPORT_SYMBOL(sysctl_tcp_ecn); |
| 5111 | EXPORT_SYMBOL(sysctl_tcp_reordering); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5112 | EXPORT_SYMBOL(tcp_parse_options); |
| 5113 | EXPORT_SYMBOL(tcp_rcv_established); |
| 5114 | EXPORT_SYMBOL(tcp_rcv_state_process); |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 5115 | EXPORT_SYMBOL(tcp_initialize_rcv_mss); |