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 | * |
Jesper Juhl | 02c30a8 | 2005-05-05 16:16:16 -0700 | [diff] [blame] | 8 | * Authors: Ross Biro |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 9 | * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> |
| 10 | * Mark Evans, <evansmp@uhura.aston.ac.uk> |
| 11 | * Corey Minyard <wf-rch!minyard@relay.EU.net> |
| 12 | * Florian La Roche, <flla@stud.uni-sb.de> |
| 13 | * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> |
| 14 | * Linus Torvalds, <torvalds@cs.helsinki.fi> |
| 15 | * Alan Cox, <gw4pts@gw4pts.ampr.org> |
| 16 | * Matthew Dillon, <dillon@apollo.west.oic.com> |
| 17 | * Arnt Gulbrandsen, <agulbra@nvg.unit.no> |
| 18 | * Jorge Cwik, <jorge@laser.satlink.net> |
| 19 | */ |
| 20 | |
| 21 | /* |
| 22 | * Changes: |
| 23 | * Pedro Roque : Fast Retransmit/Recovery. |
| 24 | * Two receive queues. |
| 25 | * Retransmit queue handled by TCP. |
| 26 | * Better retransmit timer handling. |
| 27 | * New congestion avoidance. |
| 28 | * Header prediction. |
| 29 | * Variable renaming. |
| 30 | * |
| 31 | * Eric : Fast Retransmit. |
| 32 | * Randy Scott : MSS option defines. |
| 33 | * Eric Schenk : Fixes to slow start algorithm. |
| 34 | * Eric Schenk : Yet another double ACK bug. |
| 35 | * Eric Schenk : Delayed ACK bug fixes. |
| 36 | * Eric Schenk : Floyd style fast retrans war avoidance. |
| 37 | * David S. Miller : Don't allow zero congestion window. |
| 38 | * Eric Schenk : Fix retransmitter so that it sends |
| 39 | * next packet on ack of previous packet. |
| 40 | * Andi Kleen : Moved open_request checking here |
| 41 | * and process RSTs for open_requests. |
| 42 | * Andi Kleen : Better prune_queue, and other fixes. |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 43 | * Andrey Savochkin: Fix RTT measurements in the presence of |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 44 | * timestamps. |
| 45 | * Andrey Savochkin: Check sequence numbers correctly when |
| 46 | * removing SACKs due to in sequence incoming |
| 47 | * data segments. |
| 48 | * Andi Kleen: Make sure we never ack data there is not |
| 49 | * enough room for. Also make this condition |
| 50 | * a fatal error if it might still happen. |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 51 | * Andi Kleen: Add tcp_measure_rcv_mss to make |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 52 | * connections with MSS<min(MTU,ann. MSS) |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 53 | * work without delayed acks. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 54 | * Andi Kleen: Process packets with PSH set in the |
| 55 | * fast path. |
| 56 | * J Hadi Salim: ECN support |
| 57 | * Andrei Gurtov, |
| 58 | * Pasi Sarolahti, |
| 59 | * Panu Kuhlberg: Experimental audit of TCP (re)transmission |
| 60 | * engine. Lots of bugs are found. |
| 61 | * Pasi Sarolahti: F-RTO for dealing with spurious RTOs |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 62 | */ |
| 63 | |
Joe Perches | afd46503 | 2012-03-12 07:03:32 +0000 | [diff] [blame] | 64 | #define pr_fmt(fmt) "TCP: " fmt |
| 65 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 66 | #include <linux/mm.h> |
Tejun Heo | 5a0e3ad | 2010-03-24 17:04:11 +0900 | [diff] [blame] | 67 | #include <linux/slab.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 68 | #include <linux/module.h> |
| 69 | #include <linux/sysctl.h> |
Ilpo Järvinen | a0bffff | 2009-03-21 13:36:17 -0700 | [diff] [blame] | 70 | #include <linux/kernel.h> |
Eric Dumazet | ad971f6 | 2014-10-11 15:17:29 -0700 | [diff] [blame] | 71 | #include <linux/prefetch.h> |
Satoru SATOH | 5ffc02a | 2008-05-04 22:14:42 -0700 | [diff] [blame] | 72 | #include <net/dst.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 73 | #include <net/tcp.h> |
| 74 | #include <net/inet_common.h> |
| 75 | #include <linux/ipsec.h> |
| 76 | #include <asm/unaligned.h> |
Willem de Bruijn | e1c8a60 | 2014-08-04 22:11:50 -0400 | [diff] [blame] | 77 | #include <linux/errqueue.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 78 | |
Brian Haley | ab32ea5 | 2006-09-22 14:15:41 -0700 | [diff] [blame] | 79 | int sysctl_tcp_timestamps __read_mostly = 1; |
| 80 | int sysctl_tcp_window_scaling __read_mostly = 1; |
| 81 | int sysctl_tcp_sack __read_mostly = 1; |
| 82 | int sysctl_tcp_fack __read_mostly = 1; |
Eric Dumazet | dca145f | 2014-10-27 21:45:24 -0700 | [diff] [blame] | 83 | int sysctl_tcp_max_reordering __read_mostly = 300; |
Brian Haley | ab32ea5 | 2006-09-22 14:15:41 -0700 | [diff] [blame] | 84 | int sysctl_tcp_dsack __read_mostly = 1; |
| 85 | int sysctl_tcp_app_win __read_mostly = 31; |
Eric Dumazet | b49960a | 2012-05-02 02:28:41 +0000 | [diff] [blame] | 86 | int sysctl_tcp_adv_win_scale __read_mostly = 1; |
Eric Dumazet | 4bc2f18 | 2010-07-09 21:22:10 +0000 | [diff] [blame] | 87 | EXPORT_SYMBOL(sysctl_tcp_adv_win_scale); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 88 | |
Eric Dumazet | 282f23c | 2012-07-17 10:13:05 +0200 | [diff] [blame] | 89 | /* rfc5961 challenge ack rate limiting */ |
Eric Dumazet | 75ff39c | 2016-07-10 10:04:02 +0200 | [diff] [blame] | 90 | int sysctl_tcp_challenge_ack_limit = 1000; |
Eric Dumazet | 282f23c | 2012-07-17 10:13:05 +0200 | [diff] [blame] | 91 | |
Brian Haley | ab32ea5 | 2006-09-22 14:15:41 -0700 | [diff] [blame] | 92 | int sysctl_tcp_stdurg __read_mostly; |
| 93 | int sysctl_tcp_rfc1337 __read_mostly; |
| 94 | int sysctl_tcp_max_orphans __read_mostly = NR_FILE; |
Ilpo Järvinen | c96fd3d | 2007-09-20 11:36:37 -0700 | [diff] [blame] | 95 | int sysctl_tcp_frto __read_mostly = 2; |
Yuchung Cheng | f672258 | 2015-10-16 21:57:42 -0700 | [diff] [blame] | 96 | int sysctl_tcp_min_rtt_wlen __read_mostly = 300; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 97 | |
Andreas Petlund | 7e38017 | 2010-02-18 04:48:19 +0000 | [diff] [blame] | 98 | int sysctl_tcp_thin_dupack __read_mostly; |
| 99 | |
Brian Haley | ab32ea5 | 2006-09-22 14:15:41 -0700 | [diff] [blame] | 100 | int sysctl_tcp_moderate_rcvbuf __read_mostly = 1; |
Nandita Dukkipati | 6ba8a3b | 2013-03-11 10:00:43 +0000 | [diff] [blame] | 101 | int sysctl_tcp_early_retrans __read_mostly = 3; |
Neal Cardwell | 032ee42 | 2015-02-06 16:04:38 -0500 | [diff] [blame] | 102 | int sysctl_tcp_invalid_ratelimit __read_mostly = HZ/2; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 103 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 104 | #define FLAG_DATA 0x01 /* Incoming frame contained data. */ |
| 105 | #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */ |
| 106 | #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */ |
| 107 | #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */ |
| 108 | #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */ |
| 109 | #define FLAG_DATA_SACKED 0x20 /* New SACK. */ |
| 110 | #define FLAG_ECE 0x40 /* ECE in this ACK */ |
Yuchung Cheng | 291a00d | 2015-07-01 14:11:14 -0700 | [diff] [blame] | 111 | #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 112 | #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/ |
Yuchung Cheng | e33099f | 2013-03-20 13:33:00 +0000 | [diff] [blame] | 113 | #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */ |
Ilpo Järvinen | 2e60529 | 2007-08-02 19:46:58 -0700 | [diff] [blame] | 114 | #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */ |
Ryousei Takano | 564262c | 2007-10-25 23:03:52 -0700 | [diff] [blame] | 115 | #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */ |
Ilpo Järvinen | cadbd03 | 2007-12-31 04:49:21 -0800 | [diff] [blame] | 116 | #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */ |
Eric Dumazet | 12fb3dd | 2013-04-19 07:19:48 +0000 | [diff] [blame] | 117 | #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 118 | |
| 119 | #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED) |
| 120 | #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED) |
| 121 | #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE) |
| 122 | #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED) |
| 123 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 124 | #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] | 125 | #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 126 | |
Yuchung Cheng | e662ca4 | 2016-02-02 10:33:04 -0800 | [diff] [blame] | 127 | #define REXMIT_NONE 0 /* no loss recovery to do */ |
| 128 | #define REXMIT_LOST 1 /* retransmit packets marked lost */ |
| 129 | #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */ |
| 130 | |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 131 | /* Adapt the MSS value used to make delayed ack decision to the |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 132 | * real world. |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 133 | */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 134 | static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 135 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 136 | struct inet_connection_sock *icsk = inet_csk(sk); |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 137 | const unsigned int lss = icsk->icsk_ack.last_seg_size; |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 138 | unsigned int len; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 139 | |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 140 | icsk->icsk_ack.last_seg_size = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 141 | |
| 142 | /* skb->len may jitter because of SACKs, even if peer |
| 143 | * sends good full-sized frames. |
| 144 | */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 145 | len = skb_shinfo(skb)->gso_size ? : skb->len; |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 146 | if (len >= icsk->icsk_ack.rcv_mss) { |
| 147 | icsk->icsk_ack.rcv_mss = len; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 148 | } else { |
| 149 | /* Otherwise, we make more careful check taking into account, |
| 150 | * that SACKs block is variable. |
| 151 | * |
| 152 | * "len" is invariant segment length, including TCP header. |
| 153 | */ |
Arnaldo Carvalho de Melo | 9c70220 | 2007-04-25 18:04:18 -0700 | [diff] [blame] | 154 | len += skb->data - skb_transport_header(skb); |
William Allen Simpson | bee7ca9 | 2009-11-10 09:51:18 +0000 | [diff] [blame] | 155 | if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) || |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 156 | /* If PSH is not set, packet should be |
| 157 | * full sized, provided peer TCP is not badly broken. |
| 158 | * This observation (if it is correct 8)) allows |
| 159 | * to handle super-low mtu links fairly. |
| 160 | */ |
| 161 | (len >= TCP_MIN_MSS + sizeof(struct tcphdr) && |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 162 | !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 163 | /* Subtract also invariant (if peer is RFC compliant), |
| 164 | * tcp header plus fixed timestamp option length. |
| 165 | * Resulting "len" is MSS free of SACK jitter. |
| 166 | */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 167 | len -= tcp_sk(sk)->tcp_header_len; |
| 168 | icsk->icsk_ack.last_seg_size = len; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 169 | if (len == lss) { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 170 | icsk->icsk_ack.rcv_mss = len; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 171 | return; |
| 172 | } |
| 173 | } |
Alexey Kuznetsov | 1ef9696 | 2006-09-19 12:52:50 -0700 | [diff] [blame] | 174 | if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) |
| 175 | icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2; |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 176 | icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 177 | } |
| 178 | } |
| 179 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 180 | static void tcp_incr_quickack(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 181 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 182 | struct inet_connection_sock *icsk = inet_csk(sk); |
Eric Dumazet | 95c9617 | 2012-04-15 05:58:06 +0000 | [diff] [blame] | 183 | unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 184 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 185 | if (quickacks == 0) |
| 186 | quickacks = 2; |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 187 | if (quickacks > icsk->icsk_ack.quick) |
| 188 | icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 189 | } |
| 190 | |
stephen hemminger | 1b9f409 | 2010-09-28 19:30:14 +0000 | [diff] [blame] | 191 | static void tcp_enter_quickack_mode(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 192 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 193 | struct inet_connection_sock *icsk = inet_csk(sk); |
| 194 | tcp_incr_quickack(sk); |
| 195 | icsk->icsk_ack.pingpong = 0; |
| 196 | icsk->icsk_ack.ato = TCP_ATO_MIN; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 197 | } |
| 198 | |
| 199 | /* Send ACKs quickly, if "quick" count is not exhausted |
| 200 | * and the session is not interactive. |
| 201 | */ |
| 202 | |
Jon Maxwell | 2251ae4 | 2015-07-08 10:12:28 +1000 | [diff] [blame] | 203 | static bool tcp_in_quickack_mode(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 204 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 205 | const struct inet_connection_sock *icsk = inet_csk(sk); |
Jon Maxwell | 2251ae4 | 2015-07-08 10:12:28 +1000 | [diff] [blame] | 206 | const struct dst_entry *dst = __sk_dst_get(sk); |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 207 | |
Jon Maxwell | 2251ae4 | 2015-07-08 10:12:28 +1000 | [diff] [blame] | 208 | return (dst && dst_metric(dst, RTAX_QUICKACK)) || |
| 209 | (icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 210 | } |
| 211 | |
Florian Westphal | 735d383 | 2014-09-29 13:08:30 +0200 | [diff] [blame] | 212 | static void tcp_ecn_queue_cwr(struct tcp_sock *tp) |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 213 | { |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 214 | if (tp->ecn_flags & TCP_ECN_OK) |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 215 | tp->ecn_flags |= TCP_ECN_QUEUE_CWR; |
| 216 | } |
| 217 | |
Florian Westphal | 735d383 | 2014-09-29 13:08:30 +0200 | [diff] [blame] | 218 | static void tcp_ecn_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb) |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 219 | { |
| 220 | if (tcp_hdr(skb)->cwr) |
| 221 | tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; |
| 222 | } |
| 223 | |
Florian Westphal | 735d383 | 2014-09-29 13:08:30 +0200 | [diff] [blame] | 224 | static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp) |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 225 | { |
| 226 | tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; |
| 227 | } |
| 228 | |
Florian Westphal | 735d383 | 2014-09-29 13:08:30 +0200 | [diff] [blame] | 229 | static void __tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb) |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 230 | { |
Eric Dumazet | b82d1bb | 2011-09-27 02:20:08 -0400 | [diff] [blame] | 231 | switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) { |
Eric Dumazet | 7a269ff | 2011-09-22 20:02:19 +0000 | [diff] [blame] | 232 | case INET_ECN_NOT_ECT: |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 233 | /* Funny extension: if ECT is not set on a segment, |
Eric Dumazet | 7a269ff | 2011-09-22 20:02:19 +0000 | [diff] [blame] | 234 | * and we already seen ECT on a previous segment, |
| 235 | * it is probably a retransmit. |
| 236 | */ |
| 237 | if (tp->ecn_flags & TCP_ECN_SEEN) |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 238 | tcp_enter_quickack_mode((struct sock *)tp); |
Eric Dumazet | 7a269ff | 2011-09-22 20:02:19 +0000 | [diff] [blame] | 239 | break; |
| 240 | case INET_ECN_CE: |
Florian Westphal | 9890092 | 2014-09-26 22:37:35 +0200 | [diff] [blame] | 241 | if (tcp_ca_needs_ecn((struct sock *)tp)) |
| 242 | tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_IS_CE); |
| 243 | |
Eric Dumazet | aae06bf | 2012-08-06 11:04:43 +0000 | [diff] [blame] | 244 | if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) { |
| 245 | /* Better not delay acks, sender can have a very low cwnd */ |
| 246 | tcp_enter_quickack_mode((struct sock *)tp); |
| 247 | tp->ecn_flags |= TCP_ECN_DEMAND_CWR; |
| 248 | } |
Eric Dumazet | 7a269ff | 2011-09-22 20:02:19 +0000 | [diff] [blame] | 249 | tp->ecn_flags |= TCP_ECN_SEEN; |
Florian Westphal | 9890092 | 2014-09-26 22:37:35 +0200 | [diff] [blame] | 250 | break; |
| 251 | default: |
| 252 | if (tcp_ca_needs_ecn((struct sock *)tp)) |
| 253 | tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_NO_CE); |
| 254 | tp->ecn_flags |= TCP_ECN_SEEN; |
| 255 | break; |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 256 | } |
| 257 | } |
| 258 | |
Florian Westphal | 735d383 | 2014-09-29 13:08:30 +0200 | [diff] [blame] | 259 | static void tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb) |
| 260 | { |
| 261 | if (tp->ecn_flags & TCP_ECN_OK) |
| 262 | __tcp_ecn_check_ce(tp, skb); |
| 263 | } |
| 264 | |
| 265 | static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th) |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 266 | { |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 267 | if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr)) |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 268 | tp->ecn_flags &= ~TCP_ECN_OK; |
| 269 | } |
| 270 | |
Florian Westphal | 735d383 | 2014-09-29 13:08:30 +0200 | [diff] [blame] | 271 | static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th) |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 272 | { |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 273 | if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr)) |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 274 | tp->ecn_flags &= ~TCP_ECN_OK; |
| 275 | } |
| 276 | |
Florian Westphal | 735d383 | 2014-09-29 13:08:30 +0200 | [diff] [blame] | 277 | static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th) |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 278 | { |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 279 | if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK)) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 280 | return true; |
| 281 | return false; |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 282 | } |
| 283 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 284 | /* Buffer size and advertised window tuning. |
| 285 | * |
| 286 | * 1. Tuning sk->sk_sndbuf, when connection enters established state. |
| 287 | */ |
| 288 | |
Eric Dumazet | 6ae7053 | 2013-10-01 10:23:44 -0700 | [diff] [blame] | 289 | static void tcp_sndbuf_expand(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 290 | { |
Eric Dumazet | 6ae7053 | 2013-10-01 10:23:44 -0700 | [diff] [blame] | 291 | const struct tcp_sock *tp = tcp_sk(sk); |
Yuchung Cheng | 77bfc17 | 2016-09-19 23:39:20 -0400 | [diff] [blame] | 292 | const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; |
Eric Dumazet | 6ae7053 | 2013-10-01 10:23:44 -0700 | [diff] [blame] | 293 | int sndmem, per_mss; |
| 294 | u32 nr_segs; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 295 | |
Eric Dumazet | 6ae7053 | 2013-10-01 10:23:44 -0700 | [diff] [blame] | 296 | /* Worst case is non GSO/TSO : each frame consumes one skb |
| 297 | * and skb->head is kmalloced using power of two area of memory |
| 298 | */ |
| 299 | per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) + |
| 300 | MAX_TCP_HEADER + |
| 301 | SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); |
| 302 | |
| 303 | per_mss = roundup_pow_of_two(per_mss) + |
| 304 | SKB_DATA_ALIGN(sizeof(struct sk_buff)); |
| 305 | |
| 306 | nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd); |
| 307 | nr_segs = max_t(u32, nr_segs, tp->reordering + 1); |
| 308 | |
| 309 | /* Fast Recovery (RFC 5681 3.2) : |
| 310 | * Cubic needs 1.7 factor, rounded to 2 to include |
| 311 | * extra cushion (application might react slowly to POLLOUT) |
| 312 | */ |
Yuchung Cheng | 77bfc17 | 2016-09-19 23:39:20 -0400 | [diff] [blame] | 313 | sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2; |
| 314 | sndmem *= nr_segs * per_mss; |
Eric Dumazet | 6ae7053 | 2013-10-01 10:23:44 -0700 | [diff] [blame] | 315 | |
Eric Dumazet | 06a59ec | 2011-10-13 18:24:42 +0000 | [diff] [blame] | 316 | if (sk->sk_sndbuf < sndmem) |
| 317 | sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 318 | } |
| 319 | |
| 320 | /* 2. Tuning advertised window (window_clamp, rcv_ssthresh) |
| 321 | * |
| 322 | * All tcp_full_space() is split to two parts: "network" buffer, allocated |
| 323 | * forward and advertised in receiver window (tp->rcv_wnd) and |
| 324 | * "application buffer", required to isolate scheduling/application |
| 325 | * latencies from network. |
| 326 | * window_clamp is maximal advertised window. It can be less than |
| 327 | * tcp_full_space(), in this case tcp_full_space() - window_clamp |
| 328 | * is reserved for "application" buffer. The less window_clamp is |
| 329 | * the smoother our behaviour from viewpoint of network, but the lower |
| 330 | * throughput and the higher sensitivity of the connection to losses. 8) |
| 331 | * |
| 332 | * rcv_ssthresh is more strict window_clamp used at "slow start" |
| 333 | * phase to predict further behaviour of this connection. |
| 334 | * It is used for two goals: |
| 335 | * - to enforce header prediction at sender, even when application |
| 336 | * requires some significant "application buffer". It is check #1. |
| 337 | * - to prevent pruning of receive queue because of misprediction |
| 338 | * of receiver window. Check #2. |
| 339 | * |
| 340 | * The scheme does not work when sender sends good segments opening |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 341 | * window and then starts to feed us spaghetti. But it should work |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 342 | * in common situations. Otherwise, we have to rely on queue collapsing. |
| 343 | */ |
| 344 | |
| 345 | /* Slow part of check#2. */ |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 346 | 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] | 347 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 348 | struct tcp_sock *tp = tcp_sk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 349 | /* Optimize this! */ |
Eric Dumazet | dfd4f0a | 2007-12-21 06:07:53 -0800 | [diff] [blame] | 350 | int truesize = tcp_win_from_space(skb->truesize) >> 1; |
| 351 | int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 352 | |
| 353 | while (tp->rcv_ssthresh <= window) { |
| 354 | if (truesize <= skb->len) |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 355 | return 2 * inet_csk(sk)->icsk_ack.rcv_mss; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 356 | |
| 357 | truesize >>= 1; |
| 358 | window >>= 1; |
| 359 | } |
| 360 | return 0; |
| 361 | } |
| 362 | |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 363 | static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 364 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 365 | struct tcp_sock *tp = tcp_sk(sk); |
| 366 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 367 | /* Check #1 */ |
| 368 | if (tp->rcv_ssthresh < tp->window_clamp && |
| 369 | (int)tp->rcv_ssthresh < tcp_space(sk) && |
Eric Dumazet | b8da51e | 2015-05-15 12:39:27 -0700 | [diff] [blame] | 370 | !tcp_under_memory_pressure(sk)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 371 | int incr; |
| 372 | |
| 373 | /* Check #2. Increase window, if skb with such overhead |
| 374 | * will fit to rcvbuf in future. |
| 375 | */ |
| 376 | if (tcp_win_from_space(skb->truesize) <= skb->len) |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 377 | incr = 2 * tp->advmss; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 378 | else |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 379 | incr = __tcp_grow_window(sk, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 380 | |
| 381 | if (incr) { |
Eric Dumazet | 4d846f0 | 2012-04-16 23:28:07 +0000 | [diff] [blame] | 382 | incr = max_t(int, incr, 2 * skb->len); |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 383 | tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, |
| 384 | tp->window_clamp); |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 385 | inet_csk(sk)->icsk_ack.quick |= 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 386 | } |
| 387 | } |
| 388 | } |
| 389 | |
| 390 | /* 3. Tuning rcvbuf, when connection enters established state. */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 391 | static void tcp_fixup_rcvbuf(struct sock *sk) |
| 392 | { |
Eric Dumazet | e9266a0 | 2011-10-20 16:53:56 -0400 | [diff] [blame] | 393 | u32 mss = tcp_sk(sk)->advmss; |
Eric Dumazet | e9266a0 | 2011-10-20 16:53:56 -0400 | [diff] [blame] | 394 | int rcvmem; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 395 | |
Yuchung Cheng | 85f1652 | 2013-06-11 15:35:32 -0700 | [diff] [blame] | 396 | rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) * |
| 397 | tcp_default_init_rwnd(mss); |
Eric Dumazet | e9266a0 | 2011-10-20 16:53:56 -0400 | [diff] [blame] | 398 | |
Eric Dumazet | b0983d3 | 2013-09-20 13:56:58 -0700 | [diff] [blame] | 399 | /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency |
| 400 | * Allow enough cushion so that sender is not limited by our window |
| 401 | */ |
| 402 | if (sysctl_tcp_moderate_rcvbuf) |
| 403 | rcvmem <<= 2; |
| 404 | |
Eric Dumazet | e9266a0 | 2011-10-20 16:53:56 -0400 | [diff] [blame] | 405 | if (sk->sk_rcvbuf < rcvmem) |
| 406 | sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 407 | } |
| 408 | |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 409 | /* 4. Try to fixup all. It is made immediately after connection enters |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 410 | * established state. |
| 411 | */ |
Jerry Chu | 1046716 | 2012-08-31 12:29:11 +0000 | [diff] [blame] | 412 | void tcp_init_buffer_space(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 413 | { |
| 414 | struct tcp_sock *tp = tcp_sk(sk); |
| 415 | int maxwin; |
| 416 | |
| 417 | if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) |
| 418 | tcp_fixup_rcvbuf(sk); |
| 419 | if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK)) |
Eric Dumazet | 6ae7053 | 2013-10-01 10:23:44 -0700 | [diff] [blame] | 420 | tcp_sndbuf_expand(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 421 | |
| 422 | tp->rcvq_space.space = tp->rcv_wnd; |
Eric Dumazet | b0983d3 | 2013-09-20 13:56:58 -0700 | [diff] [blame] | 423 | tp->rcvq_space.time = tcp_time_stamp; |
| 424 | tp->rcvq_space.seq = tp->copied_seq; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 425 | |
| 426 | maxwin = tcp_full_space(sk); |
| 427 | |
| 428 | if (tp->window_clamp >= maxwin) { |
| 429 | tp->window_clamp = maxwin; |
| 430 | |
| 431 | if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss) |
| 432 | tp->window_clamp = max(maxwin - |
| 433 | (maxwin >> sysctl_tcp_app_win), |
| 434 | 4 * tp->advmss); |
| 435 | } |
| 436 | |
| 437 | /* Force reservation of one segment. */ |
| 438 | if (sysctl_tcp_app_win && |
| 439 | tp->window_clamp > 2 * tp->advmss && |
| 440 | tp->window_clamp + tp->advmss > maxwin) |
| 441 | tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss); |
| 442 | |
| 443 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp); |
| 444 | tp->snd_cwnd_stamp = tcp_time_stamp; |
| 445 | } |
| 446 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 447 | /* 5. Recalculate window clamp after socket hit its memory bounds. */ |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 448 | static void tcp_clamp_window(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 449 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 450 | struct tcp_sock *tp = tcp_sk(sk); |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 451 | struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 452 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 453 | icsk->icsk_ack.quick = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 454 | |
John Heffner | 326f36e | 2005-11-10 17:11:48 -0800 | [diff] [blame] | 455 | if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] && |
| 456 | !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) && |
Eric Dumazet | b8da51e | 2015-05-15 12:39:27 -0700 | [diff] [blame] | 457 | !tcp_under_memory_pressure(sk) && |
Glauber Costa | 180d8cd | 2011-12-11 21:47:02 +0000 | [diff] [blame] | 458 | sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) { |
John Heffner | 326f36e | 2005-11-10 17:11:48 -0800 | [diff] [blame] | 459 | sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc), |
| 460 | sysctl_tcp_rmem[2]); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 461 | } |
John Heffner | 326f36e | 2005-11-10 17:11:48 -0800 | [diff] [blame] | 462 | if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 463 | tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 464 | } |
| 465 | |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 466 | /* Initialize RCV_MSS value. |
| 467 | * RCV_MSS is an our guess about MSS used by the peer. |
| 468 | * We haven't any direct information about the MSS. |
| 469 | * It's better to underestimate the RCV_MSS rather than overestimate. |
| 470 | * Overestimations make us ACKing less frequently than needed. |
| 471 | * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss(). |
| 472 | */ |
| 473 | void tcp_initialize_rcv_mss(struct sock *sk) |
| 474 | { |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 475 | const struct tcp_sock *tp = tcp_sk(sk); |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 476 | unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache); |
| 477 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 478 | hint = min(hint, tp->rcv_wnd / 2); |
William Allen Simpson | bee7ca9 | 2009-11-10 09:51:18 +0000 | [diff] [blame] | 479 | hint = min(hint, TCP_MSS_DEFAULT); |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 480 | hint = max(hint, TCP_MIN_MSS); |
| 481 | |
| 482 | inet_csk(sk)->icsk_ack.rcv_mss = hint; |
| 483 | } |
Eric Dumazet | 4bc2f18 | 2010-07-09 21:22:10 +0000 | [diff] [blame] | 484 | EXPORT_SYMBOL(tcp_initialize_rcv_mss); |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 485 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 486 | /* Receiver "autotuning" code. |
| 487 | * |
| 488 | * The algorithm for RTT estimation w/o timestamps is based on |
| 489 | * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL. |
Justin P. Mattock | 631dd1a | 2010-10-18 11:03:14 +0200 | [diff] [blame] | 490 | * <http://public.lanl.gov/radiant/pubs.html#DRS> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 491 | * |
| 492 | * More detail on this code can be found at |
Justin P. Mattock | 631dd1a | 2010-10-18 11:03:14 +0200 | [diff] [blame] | 493 | * <http://staff.psc.edu/jheffner/>, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 494 | * though this reference is out of date. A new paper |
| 495 | * is pending. |
| 496 | */ |
| 497 | static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep) |
| 498 | { |
| 499 | u32 new_sample = tp->rcv_rtt_est.rtt; |
| 500 | long m = sample; |
| 501 | |
| 502 | if (m == 0) |
| 503 | m = 1; |
| 504 | |
| 505 | if (new_sample != 0) { |
| 506 | /* If we sample in larger samples in the non-timestamp |
| 507 | * case, we could grossly overestimate the RTT especially |
| 508 | * with chatty applications or bulk transfer apps which |
| 509 | * are stalled on filesystem I/O. |
| 510 | * |
| 511 | * Also, since we are only going for a minimum in the |
Stephen Hemminger | 31f3426 | 2005-11-15 15:17:10 -0800 | [diff] [blame] | 512 | * non-timestamp case, we do not smooth things out |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 513 | * else with timestamps disabled convergence takes too |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 514 | * long. |
| 515 | */ |
| 516 | if (!win_dep) { |
| 517 | m -= (new_sample >> 3); |
| 518 | new_sample += m; |
Neal Cardwell | 18a223e | 2012-04-10 07:59:20 +0000 | [diff] [blame] | 519 | } else { |
| 520 | m <<= 3; |
| 521 | if (m < new_sample) |
| 522 | new_sample = m; |
| 523 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 524 | } else { |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 525 | /* No previous measure. */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 526 | new_sample = m << 3; |
| 527 | } |
| 528 | |
| 529 | if (tp->rcv_rtt_est.rtt != new_sample) |
| 530 | tp->rcv_rtt_est.rtt = new_sample; |
| 531 | } |
| 532 | |
| 533 | static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp) |
| 534 | { |
| 535 | if (tp->rcv_rtt_est.time == 0) |
| 536 | goto new_measure; |
| 537 | if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq)) |
| 538 | return; |
Neal Cardwell | 651913c | 2012-04-27 11:29:37 -0400 | [diff] [blame] | 539 | tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 540 | |
| 541 | new_measure: |
| 542 | tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; |
| 543 | tp->rcv_rtt_est.time = tcp_time_stamp; |
| 544 | } |
| 545 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 546 | static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, |
| 547 | const struct sk_buff *skb) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 548 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 549 | struct tcp_sock *tp = tcp_sk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 550 | if (tp->rx_opt.rcv_tsecr && |
| 551 | (TCP_SKB_CB(skb)->end_seq - |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 552 | TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 553 | tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0); |
| 554 | } |
| 555 | |
| 556 | /* |
| 557 | * This function should be called every time data is copied to user space. |
| 558 | * It calculates the appropriate TCP receive buffer space. |
| 559 | */ |
| 560 | void tcp_rcv_space_adjust(struct sock *sk) |
| 561 | { |
| 562 | struct tcp_sock *tp = tcp_sk(sk); |
| 563 | int time; |
Eric Dumazet | b0983d3 | 2013-09-20 13:56:58 -0700 | [diff] [blame] | 564 | int copied; |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 565 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 566 | time = tcp_time_stamp - tp->rcvq_space.time; |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 567 | if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 568 | return; |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 569 | |
Eric Dumazet | b0983d3 | 2013-09-20 13:56:58 -0700 | [diff] [blame] | 570 | /* Number of bytes copied to user in last RTT */ |
| 571 | copied = tp->copied_seq - tp->rcvq_space.seq; |
| 572 | if (copied <= tp->rcvq_space.space) |
| 573 | goto new_measure; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 574 | |
Eric Dumazet | b0983d3 | 2013-09-20 13:56:58 -0700 | [diff] [blame] | 575 | /* A bit of theory : |
| 576 | * copied = bytes received in previous RTT, our base window |
| 577 | * To cope with packet losses, we need a 2x factor |
| 578 | * To cope with slow start, and sender growing its cwin by 100 % |
| 579 | * every RTT, we need a 4x factor, because the ACK we are sending |
| 580 | * now is for the next RTT, not the current one : |
| 581 | * <prev RTT . ><current RTT .. ><next RTT .... > |
| 582 | */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 583 | |
Eric Dumazet | b0983d3 | 2013-09-20 13:56:58 -0700 | [diff] [blame] | 584 | if (sysctl_tcp_moderate_rcvbuf && |
| 585 | !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { |
| 586 | int rcvwin, rcvmem, rcvbuf; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 587 | |
Eric Dumazet | b0983d3 | 2013-09-20 13:56:58 -0700 | [diff] [blame] | 588 | /* minimal window to cope with packet losses, assuming |
| 589 | * steady state. Add some cushion because of small variations. |
| 590 | */ |
| 591 | rcvwin = (copied << 1) + 16 * tp->advmss; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 592 | |
Eric Dumazet | b0983d3 | 2013-09-20 13:56:58 -0700 | [diff] [blame] | 593 | /* If rate increased by 25%, |
| 594 | * assume slow start, rcvwin = 3 * copied |
| 595 | * If rate increased by 50%, |
| 596 | * assume sender can use 2x growth, rcvwin = 4 * copied |
| 597 | */ |
| 598 | if (copied >= |
| 599 | tp->rcvq_space.space + (tp->rcvq_space.space >> 2)) { |
| 600 | if (copied >= |
| 601 | tp->rcvq_space.space + (tp->rcvq_space.space >> 1)) |
| 602 | rcvwin <<= 1; |
| 603 | else |
| 604 | rcvwin += (rcvwin >> 1); |
| 605 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 606 | |
Eric Dumazet | b0983d3 | 2013-09-20 13:56:58 -0700 | [diff] [blame] | 607 | rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER); |
| 608 | while (tcp_win_from_space(rcvmem) < tp->advmss) |
| 609 | rcvmem += 128; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 610 | |
Eric Dumazet | b0983d3 | 2013-09-20 13:56:58 -0700 | [diff] [blame] | 611 | rcvbuf = min(rcvwin / tp->advmss * rcvmem, sysctl_tcp_rmem[2]); |
| 612 | if (rcvbuf > sk->sk_rcvbuf) { |
| 613 | sk->sk_rcvbuf = rcvbuf; |
| 614 | |
| 615 | /* Make the window clamp follow along. */ |
| 616 | tp->window_clamp = rcvwin; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 617 | } |
| 618 | } |
Eric Dumazet | b0983d3 | 2013-09-20 13:56:58 -0700 | [diff] [blame] | 619 | tp->rcvq_space.space = copied; |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 620 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 621 | new_measure: |
| 622 | tp->rcvq_space.seq = tp->copied_seq; |
| 623 | tp->rcvq_space.time = tcp_time_stamp; |
| 624 | } |
| 625 | |
| 626 | /* There is something which you must keep in mind when you analyze the |
| 627 | * behavior of the tp->ato delayed ack timeout interval. When a |
| 628 | * connection starts up, we want to ack as quickly as possible. The |
| 629 | * problem is that "good" TCP's do slow start at the beginning of data |
| 630 | * transmission. The means that until we send the first few ACK's the |
| 631 | * sender will sit on his end and only queue most of his data, because |
| 632 | * he can only send snd_cwnd unacked packets at any given time. For |
| 633 | * each ACK we send, he increments snd_cwnd and transmits more of his |
| 634 | * queue. -DaveM |
| 635 | */ |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 636 | 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] | 637 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 638 | struct tcp_sock *tp = tcp_sk(sk); |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 639 | struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 640 | u32 now; |
| 641 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 642 | inet_csk_schedule_ack(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 643 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 644 | tcp_measure_rcv_mss(sk, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 645 | |
| 646 | tcp_rcv_rtt_measure(tp); |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 647 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 648 | now = tcp_time_stamp; |
| 649 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 650 | if (!icsk->icsk_ack.ato) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 651 | /* The _first_ data packet received, initialize |
| 652 | * delayed ACK engine. |
| 653 | */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 654 | tcp_incr_quickack(sk); |
| 655 | icsk->icsk_ack.ato = TCP_ATO_MIN; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 656 | } else { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 657 | int m = now - icsk->icsk_ack.lrcvtime; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 658 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 659 | if (m <= TCP_ATO_MIN / 2) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 660 | /* The fastest case is the first. */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 661 | icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2; |
| 662 | } else if (m < icsk->icsk_ack.ato) { |
| 663 | icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m; |
| 664 | if (icsk->icsk_ack.ato > icsk->icsk_rto) |
| 665 | icsk->icsk_ack.ato = icsk->icsk_rto; |
| 666 | } else if (m > icsk->icsk_rto) { |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 667 | /* Too long gap. Apparently sender failed to |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 668 | * restart window, so that we send ACKs quickly. |
| 669 | */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 670 | tcp_incr_quickack(sk); |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 671 | sk_mem_reclaim(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 672 | } |
| 673 | } |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 674 | icsk->icsk_ack.lrcvtime = now; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 675 | |
Florian Westphal | 735d383 | 2014-09-29 13:08:30 +0200 | [diff] [blame] | 676 | tcp_ecn_check_ce(tp, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 677 | |
| 678 | if (skb->len >= 128) |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 679 | tcp_grow_window(sk, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 680 | } |
| 681 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 682 | /* Called to compute a smoothed rtt estimate. The data fed to this |
| 683 | * routine either comes from timestamps, or from segments that were |
| 684 | * known _not_ to have been retransmitted [see Karn/Partridge |
| 685 | * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88 |
| 686 | * piece by Van Jacobson. |
| 687 | * NOTE: the next three routines used to be one big routine. |
| 688 | * To save cycles in the RFC 1323 implementation it was better to break |
| 689 | * it up into three procedures. -- erics |
| 690 | */ |
Eric Dumazet | 740b0f1 | 2014-02-26 14:02:48 -0800 | [diff] [blame] | 691 | static void tcp_rtt_estimator(struct sock *sk, long mrtt_us) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 692 | { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 693 | struct tcp_sock *tp = tcp_sk(sk); |
Eric Dumazet | 740b0f1 | 2014-02-26 14:02:48 -0800 | [diff] [blame] | 694 | long m = mrtt_us; /* RTT */ |
| 695 | u32 srtt = tp->srtt_us; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 696 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 697 | /* The following amusing code comes from Jacobson's |
| 698 | * article in SIGCOMM '88. Note that rtt and mdev |
| 699 | * are scaled versions of rtt and mean deviation. |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 700 | * This is designed to be as fast as possible |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 701 | * m stands for "measurement". |
| 702 | * |
| 703 | * On a 1990 paper the rto value is changed to: |
| 704 | * RTO = rtt + 4 * mdev |
| 705 | * |
| 706 | * Funny. This algorithm seems to be very broken. |
| 707 | * These formulae increase RTO, when it should be decreased, increase |
Stephen Hemminger | 31f3426 | 2005-11-15 15:17:10 -0800 | [diff] [blame] | 708 | * too slowly, when it should be increased quickly, decrease too quickly |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 709 | * etc. I guess in BSD RTO takes ONE value, so that it is absolutely |
| 710 | * does not matter how to _calculate_ it. Seems, it was trap |
| 711 | * that VJ failed to avoid. 8) |
| 712 | */ |
Eric Dumazet | 4a5ab4e | 2014-02-06 15:57:10 -0800 | [diff] [blame] | 713 | if (srtt != 0) { |
| 714 | m -= (srtt >> 3); /* m is now error in rtt est */ |
| 715 | srtt += m; /* rtt = 7/8 rtt + 1/8 new */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 716 | if (m < 0) { |
| 717 | m = -m; /* m is now abs(error) */ |
Eric Dumazet | 740b0f1 | 2014-02-26 14:02:48 -0800 | [diff] [blame] | 718 | m -= (tp->mdev_us >> 2); /* similar update on mdev */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 719 | /* This is similar to one of Eifel findings. |
| 720 | * Eifel blocks mdev updates when rtt decreases. |
| 721 | * This solution is a bit different: we use finer gain |
| 722 | * for mdev in this case (alpha*beta). |
| 723 | * Like Eifel it also prevents growth of rto, |
| 724 | * but also it limits too fast rto decreases, |
| 725 | * happening in pure Eifel. |
| 726 | */ |
| 727 | if (m > 0) |
| 728 | m >>= 3; |
| 729 | } else { |
Eric Dumazet | 740b0f1 | 2014-02-26 14:02:48 -0800 | [diff] [blame] | 730 | m -= (tp->mdev_us >> 2); /* similar update on mdev */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 731 | } |
Eric Dumazet | 740b0f1 | 2014-02-26 14:02:48 -0800 | [diff] [blame] | 732 | tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */ |
| 733 | if (tp->mdev_us > tp->mdev_max_us) { |
| 734 | tp->mdev_max_us = tp->mdev_us; |
| 735 | if (tp->mdev_max_us > tp->rttvar_us) |
| 736 | tp->rttvar_us = tp->mdev_max_us; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 737 | } |
| 738 | if (after(tp->snd_una, tp->rtt_seq)) { |
Eric Dumazet | 740b0f1 | 2014-02-26 14:02:48 -0800 | [diff] [blame] | 739 | if (tp->mdev_max_us < tp->rttvar_us) |
| 740 | tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 741 | tp->rtt_seq = tp->snd_nxt; |
Eric Dumazet | 740b0f1 | 2014-02-26 14:02:48 -0800 | [diff] [blame] | 742 | tp->mdev_max_us = tcp_rto_min_us(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 743 | } |
| 744 | } else { |
| 745 | /* no previous measure. */ |
Eric Dumazet | 4a5ab4e | 2014-02-06 15:57:10 -0800 | [diff] [blame] | 746 | srtt = m << 3; /* take the measured time to be rtt */ |
Eric Dumazet | 740b0f1 | 2014-02-26 14:02:48 -0800 | [diff] [blame] | 747 | tp->mdev_us = m << 1; /* make sure rto = 3*rtt */ |
| 748 | tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk)); |
| 749 | tp->mdev_max_us = tp->rttvar_us; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 750 | tp->rtt_seq = tp->snd_nxt; |
| 751 | } |
Eric Dumazet | 740b0f1 | 2014-02-26 14:02:48 -0800 | [diff] [blame] | 752 | tp->srtt_us = max(1U, srtt); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 753 | } |
| 754 | |
Eric Dumazet | 95bd09e | 2013-08-27 05:46:32 -0700 | [diff] [blame] | 755 | /* Set the sk_pacing_rate to allow proper sizing of TSO packets. |
| 756 | * Note: TCP stack does not yet implement pacing. |
| 757 | * FQ packet scheduler can be used to implement cheap but effective |
| 758 | * TCP pacing, to smooth the burst on large writes when packets |
| 759 | * in flight is significantly lower than cwnd (or rwin) |
| 760 | */ |
Eric Dumazet | 43e122b | 2015-08-21 17:38:02 -0700 | [diff] [blame] | 761 | int sysctl_tcp_pacing_ss_ratio __read_mostly = 200; |
| 762 | int sysctl_tcp_pacing_ca_ratio __read_mostly = 120; |
| 763 | |
Eric Dumazet | 95bd09e | 2013-08-27 05:46:32 -0700 | [diff] [blame] | 764 | static void tcp_update_pacing_rate(struct sock *sk) |
| 765 | { |
| 766 | const struct tcp_sock *tp = tcp_sk(sk); |
| 767 | u64 rate; |
| 768 | |
| 769 | /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */ |
Eric Dumazet | 43e122b | 2015-08-21 17:38:02 -0700 | [diff] [blame] | 770 | rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3); |
| 771 | |
| 772 | /* current rate is (cwnd * mss) / srtt |
| 773 | * In Slow Start [1], set sk_pacing_rate to 200 % the current rate. |
| 774 | * In Congestion Avoidance phase, set it to 120 % the current rate. |
| 775 | * |
| 776 | * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh) |
| 777 | * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching |
| 778 | * end of slow start and should slow down. |
| 779 | */ |
| 780 | if (tp->snd_cwnd < tp->snd_ssthresh / 2) |
| 781 | rate *= sysctl_tcp_pacing_ss_ratio; |
| 782 | else |
| 783 | rate *= sysctl_tcp_pacing_ca_ratio; |
Eric Dumazet | 95bd09e | 2013-08-27 05:46:32 -0700 | [diff] [blame] | 784 | |
| 785 | rate *= max(tp->snd_cwnd, tp->packets_out); |
| 786 | |
Eric Dumazet | 740b0f1 | 2014-02-26 14:02:48 -0800 | [diff] [blame] | 787 | if (likely(tp->srtt_us)) |
| 788 | do_div(rate, tp->srtt_us); |
Eric Dumazet | 95bd09e | 2013-08-27 05:46:32 -0700 | [diff] [blame] | 789 | |
Eric Dumazet | ba53742 | 2013-10-09 17:14:52 -0700 | [diff] [blame] | 790 | /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate |
| 791 | * without any lock. We want to make sure compiler wont store |
| 792 | * intermediate values in this location. |
| 793 | */ |
| 794 | ACCESS_ONCE(sk->sk_pacing_rate) = min_t(u64, rate, |
| 795 | sk->sk_max_pacing_rate); |
Eric Dumazet | 95bd09e | 2013-08-27 05:46:32 -0700 | [diff] [blame] | 796 | } |
| 797 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 798 | /* Calculate rto without backoff. This is the second half of Van Jacobson's |
| 799 | * routine referred to above. |
| 800 | */ |
stephen hemminger | f7e56a7 | 2013-12-29 11:39:51 -0800 | [diff] [blame] | 801 | static void tcp_set_rto(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 802 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 803 | const struct tcp_sock *tp = tcp_sk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 804 | /* Old crap is replaced with new one. 8) |
| 805 | * |
| 806 | * More seriously: |
| 807 | * 1. If rtt variance happened to be less 50msec, it is hallucination. |
| 808 | * It cannot be less due to utterly erratic ACK generation made |
| 809 | * at least by solaris and freebsd. "Erratic ACKs" has _nothing_ |
| 810 | * to do with delayed acks, because at cwnd>2 true delack timeout |
| 811 | * is invisible. Actually, Linux-2.4 also generates erratic |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 812 | * ACKs in some circumstances. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 813 | */ |
Damian Lukowski | f1ecd5d | 2009-08-26 00:16:31 +0000 | [diff] [blame] | 814 | inet_csk(sk)->icsk_rto = __tcp_set_rto(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 815 | |
| 816 | /* 2. Fixups made earlier cannot be right. |
| 817 | * If we do not estimate RTO correctly without them, |
| 818 | * all the algo is pure shit and should be replaced |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 819 | * with correct one. It is exactly, which we pretend to do. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 820 | */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 821 | |
Ilpo Järvinen | ee6aac5 | 2008-12-05 22:43:08 -0800 | [diff] [blame] | 822 | /* NOTE: clamping at TCP_RTO_MIN is not required, current algo |
| 823 | * guarantees that rto is higher. |
| 824 | */ |
Damian Lukowski | f1ecd5d | 2009-08-26 00:16:31 +0000 | [diff] [blame] | 825 | tcp_bound_rto(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 826 | } |
| 827 | |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 828 | __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 829 | { |
| 830 | __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0); |
| 831 | |
Gerrit Renker | 22b71c8 | 2010-08-29 19:23:12 +0000 | [diff] [blame] | 832 | if (!cwnd) |
David S. Miller | 442b963 | 2011-02-02 17:05:11 -0800 | [diff] [blame] | 833 | cwnd = TCP_INIT_CWND; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 834 | return min_t(__u32, cwnd, tp->snd_cwnd_clamp); |
| 835 | } |
| 836 | |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 837 | /* |
| 838 | * Packet counting of FACK is based on in-order assumptions, therefore TCP |
| 839 | * disables it when reordering is detected |
| 840 | */ |
David S. Miller | 4aabd8e | 2012-07-09 16:07:30 -0700 | [diff] [blame] | 841 | void tcp_disable_fack(struct tcp_sock *tp) |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 842 | { |
Ilpo Järvinen | 85cc391 | 2007-11-15 19:39:31 -0800 | [diff] [blame] | 843 | /* RFC3517 uses different metric in lost marker => reset on change */ |
| 844 | if (tcp_is_fack(tp)) |
| 845 | tp->lost_skb_hint = NULL; |
Vijay Subramanian | ab56222 | 2011-12-20 13:23:24 +0000 | [diff] [blame] | 846 | tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED; |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 847 | } |
| 848 | |
Ryousei Takano | 564262c | 2007-10-25 23:03:52 -0700 | [diff] [blame] | 849 | /* Take a notice that peer is sending D-SACKs */ |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 850 | static void tcp_dsack_seen(struct tcp_sock *tp) |
| 851 | { |
Vijay Subramanian | ab56222 | 2011-12-20 13:23:24 +0000 | [diff] [blame] | 852 | tp->rx_opt.sack_ok |= TCP_DSACK_SEEN; |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 853 | } |
| 854 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 855 | static void tcp_update_reordering(struct sock *sk, const int metric, |
| 856 | const int ts) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 857 | { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 858 | struct tcp_sock *tp = tcp_sk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 859 | if (metric > tp->reordering) { |
Pavel Emelyanov | 40b215e | 2008-07-03 01:05:41 -0700 | [diff] [blame] | 860 | int mib_idx; |
| 861 | |
Eric Dumazet | dca145f | 2014-10-27 21:45:24 -0700 | [diff] [blame] | 862 | tp->reordering = min(sysctl_tcp_max_reordering, metric); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 863 | |
| 864 | /* This exciting event is worth to be remembered. 8) */ |
| 865 | if (ts) |
Pavel Emelyanov | 40b215e | 2008-07-03 01:05:41 -0700 | [diff] [blame] | 866 | mib_idx = LINUX_MIB_TCPTSREORDER; |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 867 | else if (tcp_is_reno(tp)) |
Pavel Emelyanov | 40b215e | 2008-07-03 01:05:41 -0700 | [diff] [blame] | 868 | mib_idx = LINUX_MIB_TCPRENOREORDER; |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 869 | else if (tcp_is_fack(tp)) |
Pavel Emelyanov | 40b215e | 2008-07-03 01:05:41 -0700 | [diff] [blame] | 870 | mib_idx = LINUX_MIB_TCPFACKREORDER; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 871 | else |
Pavel Emelyanov | 40b215e | 2008-07-03 01:05:41 -0700 | [diff] [blame] | 872 | mib_idx = LINUX_MIB_TCPSACKREORDER; |
| 873 | |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 874 | NET_INC_STATS(sock_net(sk), mib_idx); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 875 | #if FASTRETRANS_DEBUG > 1 |
Joe Perches | 91df42b | 2012-05-15 14:11:54 +0000 | [diff] [blame] | 876 | pr_debug("Disorder%d %d %u f%u s%u rr%d\n", |
| 877 | tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state, |
| 878 | tp->reordering, |
| 879 | tp->fackets_out, |
| 880 | tp->sacked_out, |
| 881 | tp->undo_marker ? tp->undo_retrans : 0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 882 | #endif |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 883 | tcp_disable_fack(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 884 | } |
Yuchung Cheng | eed530b | 2012-05-02 13:30:03 +0000 | [diff] [blame] | 885 | |
| 886 | if (metric > 0) |
| 887 | tcp_disable_early_retrans(tp); |
Yuchung Cheng | 4f41b1c | 2015-10-16 21:57:47 -0700 | [diff] [blame] | 888 | tp->rack.reord = 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 889 | } |
| 890 | |
Ilpo Järvinen | 006f582 | 2008-09-20 21:20:20 -0700 | [diff] [blame] | 891 | /* This must be called before lost_out is incremented */ |
Ilpo Järvinen | c8c213f | 2008-09-20 21:18:55 -0700 | [diff] [blame] | 892 | static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb) |
| 893 | { |
Ian Morris | 51456b2 | 2015-04-03 09:17:26 +0100 | [diff] [blame] | 894 | if (!tp->retransmit_skb_hint || |
Ilpo Järvinen | c8c213f | 2008-09-20 21:18:55 -0700 | [diff] [blame] | 895 | before(TCP_SKB_CB(skb)->seq, |
| 896 | TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) |
Ilpo Järvinen | 006f582 | 2008-09-20 21:20:20 -0700 | [diff] [blame] | 897 | tp->retransmit_skb_hint = skb; |
| 898 | |
| 899 | if (!tp->lost_out || |
| 900 | after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high)) |
| 901 | tp->retransmit_high = TCP_SKB_CB(skb)->end_seq; |
Ilpo Järvinen | c8c213f | 2008-09-20 21:18:55 -0700 | [diff] [blame] | 902 | } |
| 903 | |
Neal Cardwell | 0682e69 | 2016-09-19 23:39:13 -0400 | [diff] [blame] | 904 | /* Sum the number of packets on the wire we have marked as lost. |
| 905 | * There are two cases we care about here: |
| 906 | * a) Packet hasn't been marked lost (nor retransmitted), |
| 907 | * and this is the first loss. |
| 908 | * b) Packet has been marked both lost and retransmitted, |
| 909 | * and this means we think it was lost again. |
| 910 | */ |
| 911 | static void tcp_sum_lost(struct tcp_sock *tp, struct sk_buff *skb) |
| 912 | { |
| 913 | __u8 sacked = TCP_SKB_CB(skb)->sacked; |
| 914 | |
| 915 | if (!(sacked & TCPCB_LOST) || |
| 916 | ((sacked & TCPCB_LOST) && (sacked & TCPCB_SACKED_RETRANS))) |
| 917 | tp->lost += tcp_skb_pcount(skb); |
| 918 | } |
| 919 | |
Ilpo Järvinen | 41ea36e | 2008-09-20 21:19:22 -0700 | [diff] [blame] | 920 | static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb) |
| 921 | { |
| 922 | if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { |
| 923 | tcp_verify_retransmit_hint(tp, skb); |
| 924 | |
| 925 | tp->lost_out += tcp_skb_pcount(skb); |
Neal Cardwell | 0682e69 | 2016-09-19 23:39:13 -0400 | [diff] [blame] | 926 | tcp_sum_lost(tp, skb); |
Ilpo Järvinen | 41ea36e | 2008-09-20 21:19:22 -0700 | [diff] [blame] | 927 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; |
| 928 | } |
| 929 | } |
| 930 | |
Yuchung Cheng | 4f41b1c | 2015-10-16 21:57:47 -0700 | [diff] [blame] | 931 | void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb) |
Ilpo Järvinen | 006f582 | 2008-09-20 21:20:20 -0700 | [diff] [blame] | 932 | { |
| 933 | tcp_verify_retransmit_hint(tp, skb); |
| 934 | |
Neal Cardwell | 0682e69 | 2016-09-19 23:39:13 -0400 | [diff] [blame] | 935 | tcp_sum_lost(tp, skb); |
Ilpo Järvinen | 006f582 | 2008-09-20 21:20:20 -0700 | [diff] [blame] | 936 | if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { |
| 937 | tp->lost_out += tcp_skb_pcount(skb); |
| 938 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; |
| 939 | } |
| 940 | } |
| 941 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 942 | /* This procedure tags the retransmission queue when SACKs arrive. |
| 943 | * |
| 944 | * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L). |
| 945 | * Packets in queue with these bits set are counted in variables |
| 946 | * sacked_out, retrans_out and lost_out, correspondingly. |
| 947 | * |
| 948 | * Valid combinations are: |
| 949 | * Tag InFlight Description |
| 950 | * 0 1 - orig segment is in flight. |
| 951 | * S 0 - nothing flies, orig reached receiver. |
| 952 | * L 0 - nothing flies, orig lost by net. |
| 953 | * R 2 - both orig and retransmit are in flight. |
| 954 | * L|R 1 - orig is lost, retransmit is in flight. |
| 955 | * S|R 1 - orig reached receiver, retrans is still in flight. |
| 956 | * (L|S|R is logically valid, it could occur when L|R is sacked, |
| 957 | * but it is equivalent to plain S and code short-curcuits it to S. |
| 958 | * L|S is logically invalid, it would mean -1 packet in flight 8)) |
| 959 | * |
| 960 | * These 6 states form finite state machine, controlled by the following events: |
| 961 | * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue()) |
| 962 | * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue()) |
Yuchung Cheng | 974c123 | 2012-01-19 14:42:21 +0000 | [diff] [blame] | 963 | * 3. Loss detection event of two flavors: |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 964 | * A. Scoreboard estimator decided the packet is lost. |
| 965 | * A'. Reno "three dupacks" marks head of queue lost. |
Yuchung Cheng | 974c123 | 2012-01-19 14:42:21 +0000 | [diff] [blame] | 966 | * A''. Its FACK modification, head until snd.fack is lost. |
| 967 | * B. SACK arrives sacking SND.NXT at the moment, when the |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 968 | * segment was retransmitted. |
| 969 | * 4. D-SACK added new rule: D-SACK changes any tag to S. |
| 970 | * |
| 971 | * It is pleasant to note, that state diagram turns out to be commutative, |
| 972 | * so that we are allowed not to be bothered by order of our actions, |
| 973 | * when multiple events arrive simultaneously. (see the function below). |
| 974 | * |
| 975 | * Reordering detection. |
| 976 | * -------------------- |
| 977 | * Reordering metric is maximal distance, which a packet can be displaced |
| 978 | * in packet stream. With SACKs we can estimate it: |
| 979 | * |
| 980 | * 1. SACK fills old hole and the corresponding segment was not |
| 981 | * ever retransmitted -> reordering. Alas, we cannot use it |
| 982 | * when segment was retransmitted. |
| 983 | * 2. The last flaw is solved with D-SACK. D-SACK arrives |
| 984 | * for retransmitted and already SACKed segment -> reordering.. |
| 985 | * Both of these heuristics are not used in Loss state, when we cannot |
| 986 | * account for retransmits accurately. |
Ilpo Järvinen | 5b3c988 | 2007-08-24 22:54:44 -0700 | [diff] [blame] | 987 | * |
| 988 | * SACK block validation. |
| 989 | * ---------------------- |
| 990 | * |
| 991 | * SACK block range validation checks that the received SACK block fits to |
| 992 | * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT. |
| 993 | * Note that SND.UNA is not included to the range though being valid because |
Ilpo Järvinen | 0e83533 | 2007-10-01 15:28:17 -0700 | [diff] [blame] | 994 | * it means that the receiver is rather inconsistent with itself reporting |
| 995 | * SACK reneging when it should advance SND.UNA. Such SACK block this is |
| 996 | * perfectly valid, however, in light of RFC2018 which explicitly states |
| 997 | * that "SACK block MUST reflect the newest segment. Even if the newest |
| 998 | * segment is going to be discarded ...", not that it looks very clever |
| 999 | * in case of head skb. Due to potentional receiver driven attacks, we |
| 1000 | * choose to avoid immediate execution of a walk in write queue due to |
| 1001 | * reneging and defer head skb's loss recovery to standard loss recovery |
| 1002 | * procedure that will eventually trigger (nothing forbids us doing this). |
Ilpo Järvinen | 5b3c988 | 2007-08-24 22:54:44 -0700 | [diff] [blame] | 1003 | * |
| 1004 | * Implements also blockage to start_seq wrap-around. Problem lies in the |
| 1005 | * fact that though start_seq (s) is before end_seq (i.e., not reversed), |
| 1006 | * there's no guarantee that it will be before snd_nxt (n). The problem |
| 1007 | * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt |
| 1008 | * wrap (s_w): |
| 1009 | * |
| 1010 | * <- outs wnd -> <- wrapzone -> |
| 1011 | * u e n u_w e_w s n_w |
| 1012 | * | | | | | | | |
| 1013 | * |<------------+------+----- TCP seqno space --------------+---------->| |
| 1014 | * ...-- <2^31 ->| |<--------... |
| 1015 | * ...---- >2^31 ------>| |<--------... |
| 1016 | * |
| 1017 | * Current code wouldn't be vulnerable but it's better still to discard such |
| 1018 | * crazy SACK blocks. Doing this check for start_seq alone closes somewhat |
| 1019 | * similar case (end_seq after snd_nxt wrap) as earlier reversed check in |
| 1020 | * snd_nxt wrap -> snd_una region will then become "well defined", i.e., |
| 1021 | * equal to the ideal case (infinite seqno space without wrap caused issues). |
| 1022 | * |
| 1023 | * With D-SACK the lower bound is extended to cover sequence space below |
| 1024 | * SND.UNA down to undo_marker, which is the last point of interest. Yet |
Ryousei Takano | 564262c | 2007-10-25 23:03:52 -0700 | [diff] [blame] | 1025 | * again, D-SACK block must not to go across snd_una (for the same reason as |
Ilpo Järvinen | 5b3c988 | 2007-08-24 22:54:44 -0700 | [diff] [blame] | 1026 | * for the normal SACK blocks, explained above). But there all simplicity |
| 1027 | * ends, TCP might receive valid D-SACKs below that. As long as they reside |
| 1028 | * fully below undo_marker they do not affect behavior in anyway and can |
| 1029 | * therefore be safely ignored. In rare cases (which are more or less |
| 1030 | * theoretical ones), the D-SACK will nicely cross that boundary due to skb |
| 1031 | * fragmentation and packet reordering past skb's retransmission. To consider |
| 1032 | * them correctly, the acceptable range must be extended even more though |
| 1033 | * the exact amount is rather hard to quantify. However, tp->max_window can |
| 1034 | * be used as an exaggerated estimate. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1035 | */ |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1036 | static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack, |
| 1037 | u32 start_seq, u32 end_seq) |
Ilpo Järvinen | 5b3c988 | 2007-08-24 22:54:44 -0700 | [diff] [blame] | 1038 | { |
| 1039 | /* Too far in future, or reversed (interpretation is ambiguous) */ |
| 1040 | if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq)) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1041 | return false; |
Ilpo Järvinen | 5b3c988 | 2007-08-24 22:54:44 -0700 | [diff] [blame] | 1042 | |
| 1043 | /* Nasty start_seq wrap-around check (see comments above) */ |
| 1044 | if (!before(start_seq, tp->snd_nxt)) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1045 | return false; |
Ilpo Järvinen | 5b3c988 | 2007-08-24 22:54:44 -0700 | [diff] [blame] | 1046 | |
Ryousei Takano | 564262c | 2007-10-25 23:03:52 -0700 | [diff] [blame] | 1047 | /* In outstanding window? ...This is valid exit for D-SACKs too. |
Ilpo Järvinen | 5b3c988 | 2007-08-24 22:54:44 -0700 | [diff] [blame] | 1048 | * start_seq == snd_una is non-sensical (see comments above) |
| 1049 | */ |
| 1050 | if (after(start_seq, tp->snd_una)) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1051 | return true; |
Ilpo Järvinen | 5b3c988 | 2007-08-24 22:54:44 -0700 | [diff] [blame] | 1052 | |
| 1053 | if (!is_dsack || !tp->undo_marker) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1054 | return false; |
Ilpo Järvinen | 5b3c988 | 2007-08-24 22:54:44 -0700 | [diff] [blame] | 1055 | |
| 1056 | /* ...Then it's D-SACK, and must reside below snd_una completely */ |
Zheng Yan | f779b2d | 2011-09-18 22:37:34 -0400 | [diff] [blame] | 1057 | if (after(end_seq, tp->snd_una)) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1058 | return false; |
Ilpo Järvinen | 5b3c988 | 2007-08-24 22:54:44 -0700 | [diff] [blame] | 1059 | |
| 1060 | if (!before(start_seq, tp->undo_marker)) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1061 | return true; |
Ilpo Järvinen | 5b3c988 | 2007-08-24 22:54:44 -0700 | [diff] [blame] | 1062 | |
| 1063 | /* Too old */ |
| 1064 | if (!after(end_seq, tp->undo_marker)) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1065 | return false; |
Ilpo Järvinen | 5b3c988 | 2007-08-24 22:54:44 -0700 | [diff] [blame] | 1066 | |
| 1067 | /* Undo_marker boundary crossing (overestimates a lot). Known already: |
| 1068 | * start_seq < undo_marker and end_seq >= undo_marker. |
| 1069 | */ |
| 1070 | return !before(start_seq, end_seq - tp->max_window); |
| 1071 | } |
| 1072 | |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1073 | static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb, |
| 1074 | struct tcp_sack_block_wire *sp, int num_sacks, |
| 1075 | u32 prior_snd_una) |
David S. Miller | d06e021 | 2007-06-18 22:43:06 -0700 | [diff] [blame] | 1076 | { |
Pavel Emelyanov | 1ed8346 | 2008-07-16 20:29:51 -0700 | [diff] [blame] | 1077 | struct tcp_sock *tp = tcp_sk(sk); |
Harvey Harrison | d3e2ce3 | 2008-05-02 16:26:16 -0700 | [diff] [blame] | 1078 | u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq); |
| 1079 | u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq); |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1080 | bool dup_sack = false; |
David S. Miller | d06e021 | 2007-06-18 22:43:06 -0700 | [diff] [blame] | 1081 | |
| 1082 | if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) { |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1083 | dup_sack = true; |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 1084 | tcp_dsack_seen(tp); |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 1085 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV); |
David S. Miller | d06e021 | 2007-06-18 22:43:06 -0700 | [diff] [blame] | 1086 | } else if (num_sacks > 1) { |
Harvey Harrison | d3e2ce3 | 2008-05-02 16:26:16 -0700 | [diff] [blame] | 1087 | u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq); |
| 1088 | u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq); |
David S. Miller | d06e021 | 2007-06-18 22:43:06 -0700 | [diff] [blame] | 1089 | |
| 1090 | if (!after(end_seq_0, end_seq_1) && |
| 1091 | !before(start_seq_0, start_seq_1)) { |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1092 | dup_sack = true; |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 1093 | tcp_dsack_seen(tp); |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 1094 | NET_INC_STATS(sock_net(sk), |
Pavel Emelyanov | de0744a | 2008-07-16 20:31:16 -0700 | [diff] [blame] | 1095 | LINUX_MIB_TCPDSACKOFORECV); |
David S. Miller | d06e021 | 2007-06-18 22:43:06 -0700 | [diff] [blame] | 1096 | } |
| 1097 | } |
| 1098 | |
| 1099 | /* D-SACK for already forgotten data... Do dumb counting. */ |
Yuchung Cheng | 6e08d5e | 2014-07-02 12:07:16 -0700 | [diff] [blame] | 1100 | if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 && |
David S. Miller | d06e021 | 2007-06-18 22:43:06 -0700 | [diff] [blame] | 1101 | !after(end_seq_0, prior_snd_una) && |
| 1102 | after(end_seq_0, tp->undo_marker)) |
| 1103 | tp->undo_retrans--; |
| 1104 | |
| 1105 | return dup_sack; |
| 1106 | } |
| 1107 | |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1108 | struct tcp_sacktag_state { |
Eric Dumazet | 740b0f1 | 2014-02-26 14:02:48 -0800 | [diff] [blame] | 1109 | int reord; |
| 1110 | int fack_count; |
Kenneth Klette Jonassen | 31231a8 | 2015-05-01 01:10:58 +0200 | [diff] [blame] | 1111 | /* Timestamps for earliest and latest never-retransmitted segment |
| 1112 | * that was SACKed. RTO needs the earliest RTT to stay conservative, |
| 1113 | * but congestion control should still get an accurate delay signal. |
| 1114 | */ |
| 1115 | struct skb_mstamp first_sackt; |
| 1116 | struct skb_mstamp last_sackt; |
Yuchung Cheng | b9f6482 | 2016-09-19 23:39:14 -0400 | [diff] [blame] | 1117 | struct rate_sample *rate; |
Eric Dumazet | 740b0f1 | 2014-02-26 14:02:48 -0800 | [diff] [blame] | 1118 | int flag; |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1119 | }; |
| 1120 | |
Ilpo Järvinen | d193594 | 2007-10-11 17:34:25 -0700 | [diff] [blame] | 1121 | /* Check if skb is fully within the SACK block. In presence of GSO skbs, |
| 1122 | * the incoming SACK may not exactly match but we can find smaller MSS |
| 1123 | * aligned portion of it that matches. Therefore we might need to fragment |
| 1124 | * which may fail and creates some hassle (caller must handle error case |
| 1125 | * returns). |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1126 | * |
| 1127 | * FIXME: this could be merged to shift decision code |
Ilpo Järvinen | d193594 | 2007-10-11 17:34:25 -0700 | [diff] [blame] | 1128 | */ |
Adrian Bunk | 0f79efd | 2007-10-26 03:57:36 -0700 | [diff] [blame] | 1129 | static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb, |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1130 | u32 start_seq, u32 end_seq) |
Ilpo Järvinen | d193594 | 2007-10-11 17:34:25 -0700 | [diff] [blame] | 1131 | { |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1132 | int err; |
| 1133 | bool in_sack; |
Ilpo Järvinen | d193594 | 2007-10-11 17:34:25 -0700 | [diff] [blame] | 1134 | unsigned int pkt_len; |
Ilpo Järvinen | adb92db | 2008-11-24 21:13:50 -0800 | [diff] [blame] | 1135 | unsigned int mss; |
Ilpo Järvinen | d193594 | 2007-10-11 17:34:25 -0700 | [diff] [blame] | 1136 | |
| 1137 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && |
| 1138 | !before(end_seq, TCP_SKB_CB(skb)->end_seq); |
| 1139 | |
| 1140 | if (tcp_skb_pcount(skb) > 1 && !in_sack && |
| 1141 | after(TCP_SKB_CB(skb)->end_seq, start_seq)) { |
Ilpo Järvinen | adb92db | 2008-11-24 21:13:50 -0800 | [diff] [blame] | 1142 | mss = tcp_skb_mss(skb); |
Ilpo Järvinen | d193594 | 2007-10-11 17:34:25 -0700 | [diff] [blame] | 1143 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); |
| 1144 | |
Ilpo Järvinen | adb92db | 2008-11-24 21:13:50 -0800 | [diff] [blame] | 1145 | if (!in_sack) { |
Ilpo Järvinen | d193594 | 2007-10-11 17:34:25 -0700 | [diff] [blame] | 1146 | pkt_len = start_seq - TCP_SKB_CB(skb)->seq; |
Ilpo Järvinen | adb92db | 2008-11-24 21:13:50 -0800 | [diff] [blame] | 1147 | if (pkt_len < mss) |
| 1148 | pkt_len = mss; |
| 1149 | } else { |
Ilpo Järvinen | d193594 | 2007-10-11 17:34:25 -0700 | [diff] [blame] | 1150 | pkt_len = end_seq - TCP_SKB_CB(skb)->seq; |
Ilpo Järvinen | adb92db | 2008-11-24 21:13:50 -0800 | [diff] [blame] | 1151 | if (pkt_len < mss) |
| 1152 | return -EINVAL; |
| 1153 | } |
| 1154 | |
| 1155 | /* Round if necessary so that SACKs cover only full MSSes |
| 1156 | * and/or the remaining small portion (if present) |
| 1157 | */ |
| 1158 | if (pkt_len > mss) { |
| 1159 | unsigned int new_len = (pkt_len / mss) * mss; |
| 1160 | if (!in_sack && new_len < pkt_len) { |
| 1161 | new_len += mss; |
Neal Cardwell | 2cd0d74 | 2014-06-18 21:15:03 -0400 | [diff] [blame] | 1162 | if (new_len >= skb->len) |
Ilpo Järvinen | adb92db | 2008-11-24 21:13:50 -0800 | [diff] [blame] | 1163 | return 0; |
| 1164 | } |
| 1165 | pkt_len = new_len; |
| 1166 | } |
Octavian Purdila | 6cc55e0 | 2014-06-06 17:32:37 +0300 | [diff] [blame] | 1167 | err = tcp_fragment(sk, skb, pkt_len, mss, GFP_ATOMIC); |
Ilpo Järvinen | d193594 | 2007-10-11 17:34:25 -0700 | [diff] [blame] | 1168 | if (err < 0) |
| 1169 | return err; |
| 1170 | } |
| 1171 | |
| 1172 | return in_sack; |
| 1173 | } |
| 1174 | |
Neal Cardwell | cc9a672 | 2012-02-12 18:37:09 +0000 | [diff] [blame] | 1175 | /* Mark the given newly-SACKed range as such, adjusting counters and hints. */ |
| 1176 | static u8 tcp_sacktag_one(struct sock *sk, |
| 1177 | struct tcp_sacktag_state *state, u8 sacked, |
| 1178 | u32 start_seq, u32 end_seq, |
Eric Dumazet | 740b0f1 | 2014-02-26 14:02:48 -0800 | [diff] [blame] | 1179 | int dup_sack, int pcount, |
| 1180 | const struct skb_mstamp *xmit_time) |
Ilpo Järvinen | 9e10c47c | 2007-11-15 19:44:56 -0800 | [diff] [blame] | 1181 | { |
Ilpo Järvinen | 6859d49 | 2007-12-02 00:48:06 +0200 | [diff] [blame] | 1182 | struct tcp_sock *tp = tcp_sk(sk); |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1183 | int fack_count = state->fack_count; |
Ilpo Järvinen | 9e10c47c | 2007-11-15 19:44:56 -0800 | [diff] [blame] | 1184 | |
| 1185 | /* Account D-SACK for retransmitted packet. */ |
| 1186 | if (dup_sack && (sacked & TCPCB_RETRANS)) { |
Yuchung Cheng | 6e08d5e | 2014-07-02 12:07:16 -0700 | [diff] [blame] | 1187 | if (tp->undo_marker && tp->undo_retrans > 0 && |
Neal Cardwell | cc9a672 | 2012-02-12 18:37:09 +0000 | [diff] [blame] | 1188 | after(end_seq, tp->undo_marker)) |
Ilpo Järvinen | 9e10c47c | 2007-11-15 19:44:56 -0800 | [diff] [blame] | 1189 | tp->undo_retrans--; |
Ilpo Järvinen | ede9f3b | 2007-12-02 00:47:58 +0200 | [diff] [blame] | 1190 | if (sacked & TCPCB_SACKED_ACKED) |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1191 | state->reord = min(fack_count, state->reord); |
Ilpo Järvinen | 9e10c47c | 2007-11-15 19:44:56 -0800 | [diff] [blame] | 1192 | } |
| 1193 | |
| 1194 | /* Nothing to do; acked frame is about to be dropped (was ACKed). */ |
Neal Cardwell | cc9a672 | 2012-02-12 18:37:09 +0000 | [diff] [blame] | 1195 | if (!after(end_seq, tp->snd_una)) |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1196 | return sacked; |
Ilpo Järvinen | 9e10c47c | 2007-11-15 19:44:56 -0800 | [diff] [blame] | 1197 | |
| 1198 | if (!(sacked & TCPCB_SACKED_ACKED)) { |
Yuchung Cheng | 659a8ad | 2015-10-16 21:57:46 -0700 | [diff] [blame] | 1199 | tcp_rack_advance(tp, xmit_time, sacked); |
| 1200 | |
Ilpo Järvinen | 9e10c47c | 2007-11-15 19:44:56 -0800 | [diff] [blame] | 1201 | if (sacked & TCPCB_SACKED_RETRANS) { |
| 1202 | /* If the segment is not tagged as lost, |
| 1203 | * we do not clear RETRANS, believing |
| 1204 | * that retransmission is still in flight. |
| 1205 | */ |
| 1206 | if (sacked & TCPCB_LOST) { |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1207 | sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); |
Ilpo Järvinen | f58b22f | 2008-11-24 21:14:43 -0800 | [diff] [blame] | 1208 | tp->lost_out -= pcount; |
| 1209 | tp->retrans_out -= pcount; |
Ilpo Järvinen | 9e10c47c | 2007-11-15 19:44:56 -0800 | [diff] [blame] | 1210 | } |
| 1211 | } else { |
| 1212 | if (!(sacked & TCPCB_RETRANS)) { |
| 1213 | /* New sack for not retransmitted frame, |
| 1214 | * which was in hole. It is reordering. |
| 1215 | */ |
Neal Cardwell | cc9a672 | 2012-02-12 18:37:09 +0000 | [diff] [blame] | 1216 | if (before(start_seq, |
Ilpo Järvinen | 9e10c47c | 2007-11-15 19:44:56 -0800 | [diff] [blame] | 1217 | tcp_highest_sack_seq(tp))) |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1218 | state->reord = min(fack_count, |
| 1219 | state->reord); |
Yuchung Cheng | e33099f | 2013-03-20 13:33:00 +0000 | [diff] [blame] | 1220 | if (!after(end_seq, tp->high_seq)) |
| 1221 | state->flag |= FLAG_ORIG_SACK_ACKED; |
Kenneth Klette Jonassen | 31231a8 | 2015-05-01 01:10:58 +0200 | [diff] [blame] | 1222 | if (state->first_sackt.v64 == 0) |
| 1223 | state->first_sackt = *xmit_time; |
| 1224 | state->last_sackt = *xmit_time; |
Ilpo Järvinen | 9e10c47c | 2007-11-15 19:44:56 -0800 | [diff] [blame] | 1225 | } |
| 1226 | |
| 1227 | if (sacked & TCPCB_LOST) { |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1228 | sacked &= ~TCPCB_LOST; |
Ilpo Järvinen | f58b22f | 2008-11-24 21:14:43 -0800 | [diff] [blame] | 1229 | tp->lost_out -= pcount; |
Ilpo Järvinen | 9e10c47c | 2007-11-15 19:44:56 -0800 | [diff] [blame] | 1230 | } |
| 1231 | } |
| 1232 | |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1233 | sacked |= TCPCB_SACKED_ACKED; |
| 1234 | state->flag |= FLAG_DATA_SACKED; |
Ilpo Järvinen | f58b22f | 2008-11-24 21:14:43 -0800 | [diff] [blame] | 1235 | tp->sacked_out += pcount; |
Yuchung Cheng | ddf1af6 | 2016-02-02 10:33:06 -0800 | [diff] [blame] | 1236 | tp->delivered += pcount; /* Out-of-order packets delivered */ |
Ilpo Järvinen | 9e10c47c | 2007-11-15 19:44:56 -0800 | [diff] [blame] | 1237 | |
Ilpo Järvinen | f58b22f | 2008-11-24 21:14:43 -0800 | [diff] [blame] | 1238 | fack_count += pcount; |
Ilpo Järvinen | 9e10c47c | 2007-11-15 19:44:56 -0800 | [diff] [blame] | 1239 | |
| 1240 | /* Lost marker hint past SACKed? Tweak RFC3517 cnt */ |
Ian Morris | 00db412 | 2015-04-03 09:17:27 +0100 | [diff] [blame] | 1241 | if (!tcp_is_fack(tp) && tp->lost_skb_hint && |
Neal Cardwell | cc9a672 | 2012-02-12 18:37:09 +0000 | [diff] [blame] | 1242 | before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq)) |
Ilpo Järvinen | f58b22f | 2008-11-24 21:14:43 -0800 | [diff] [blame] | 1243 | tp->lost_cnt_hint += pcount; |
Ilpo Järvinen | 9e10c47c | 2007-11-15 19:44:56 -0800 | [diff] [blame] | 1244 | |
| 1245 | if (fack_count > tp->fackets_out) |
| 1246 | tp->fackets_out = fack_count; |
Ilpo Järvinen | 9e10c47c | 2007-11-15 19:44:56 -0800 | [diff] [blame] | 1247 | } |
| 1248 | |
| 1249 | /* D-SACK. We can detect redundant retransmission in S|R and plain R |
| 1250 | * frames and clear it. undo_retrans is decreased above, L|R frames |
| 1251 | * are accounted above as well. |
| 1252 | */ |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1253 | if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) { |
| 1254 | sacked &= ~TCPCB_SACKED_RETRANS; |
Ilpo Järvinen | f58b22f | 2008-11-24 21:14:43 -0800 | [diff] [blame] | 1255 | tp->retrans_out -= pcount; |
Ilpo Järvinen | 9e10c47c | 2007-11-15 19:44:56 -0800 | [diff] [blame] | 1256 | } |
| 1257 | |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1258 | return sacked; |
Ilpo Järvinen | 9e10c47c | 2007-11-15 19:44:56 -0800 | [diff] [blame] | 1259 | } |
| 1260 | |
Neal Cardwell | daef52b | 2012-02-12 18:37:10 +0000 | [diff] [blame] | 1261 | /* Shift newly-SACKed bytes from this skb to the immediately previous |
| 1262 | * already-SACKed sk_buff. Mark the newly-SACKed bytes as such. |
| 1263 | */ |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1264 | static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb, |
| 1265 | struct tcp_sacktag_state *state, |
| 1266 | unsigned int pcount, int shifted, int mss, |
| 1267 | bool dup_sack) |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1268 | { |
| 1269 | struct tcp_sock *tp = tcp_sk(sk); |
Ilpo Järvinen | 5013316 | 2008-12-05 22:42:41 -0800 | [diff] [blame] | 1270 | struct sk_buff *prev = tcp_write_queue_prev(sk, skb); |
Neal Cardwell | daef52b | 2012-02-12 18:37:10 +0000 | [diff] [blame] | 1271 | u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */ |
| 1272 | u32 end_seq = start_seq + shifted; /* end of newly-SACKed */ |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1273 | |
| 1274 | BUG_ON(!pcount); |
| 1275 | |
Neal Cardwell | 4c90d3b | 2012-02-26 10:06:19 +0000 | [diff] [blame] | 1276 | /* Adjust counters and hints for the newly sacked sequence |
| 1277 | * range but discard the return value since prev is already |
| 1278 | * marked. We must tag the range first because the seq |
| 1279 | * advancement below implicitly advances |
| 1280 | * tcp_highest_sack_seq() when skb is highest_sack. |
| 1281 | */ |
| 1282 | tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked, |
Yuchung Cheng | 59c9af4 | 2013-07-22 16:20:47 -0700 | [diff] [blame] | 1283 | start_seq, end_seq, dup_sack, pcount, |
Eric Dumazet | 740b0f1 | 2014-02-26 14:02:48 -0800 | [diff] [blame] | 1284 | &skb->skb_mstamp); |
Yuchung Cheng | b9f6482 | 2016-09-19 23:39:14 -0400 | [diff] [blame] | 1285 | tcp_rate_skb_delivered(sk, skb, state->rate); |
Neal Cardwell | 4c90d3b | 2012-02-26 10:06:19 +0000 | [diff] [blame] | 1286 | |
| 1287 | if (skb == tp->lost_skb_hint) |
Neal Cardwell | 0af2a0d | 2012-02-13 20:22:08 +0000 | [diff] [blame] | 1288 | tp->lost_cnt_hint += pcount; |
| 1289 | |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1290 | TCP_SKB_CB(prev)->end_seq += shifted; |
| 1291 | TCP_SKB_CB(skb)->seq += shifted; |
| 1292 | |
Eric Dumazet | cd7d849 | 2014-09-24 04:11:22 -0700 | [diff] [blame] | 1293 | tcp_skb_pcount_add(prev, pcount); |
| 1294 | BUG_ON(tcp_skb_pcount(skb) < pcount); |
| 1295 | tcp_skb_pcount_add(skb, -pcount); |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1296 | |
| 1297 | /* When we're adding to gso_segs == 1, gso_size will be zero, |
| 1298 | * in theory this shouldn't be necessary but as long as DSACK |
| 1299 | * code can come after this skb later on it's better to keep |
| 1300 | * setting gso_size to something. |
| 1301 | */ |
Eric Dumazet | f69ad29 | 2015-06-11 09:15:18 -0700 | [diff] [blame] | 1302 | if (!TCP_SKB_CB(prev)->tcp_gso_size) |
| 1303 | TCP_SKB_CB(prev)->tcp_gso_size = mss; |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1304 | |
| 1305 | /* CHECKME: To clear or not to clear? Mimics normal skb currently */ |
Eric Dumazet | 51466a7 | 2015-06-11 09:15:16 -0700 | [diff] [blame] | 1306 | if (tcp_skb_pcount(skb) <= 1) |
Eric Dumazet | f69ad29 | 2015-06-11 09:15:18 -0700 | [diff] [blame] | 1307 | TCP_SKB_CB(skb)->tcp_gso_size = 0; |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1308 | |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1309 | /* Difference in this won't matter, both ACKed by the same cumul. ACK */ |
| 1310 | TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS); |
| 1311 | |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1312 | if (skb->len > 0) { |
| 1313 | BUG_ON(!tcp_skb_pcount(skb)); |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 1314 | NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED); |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1315 | return false; |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1316 | } |
| 1317 | |
| 1318 | /* Whole SKB was eaten :-) */ |
| 1319 | |
Ilpo Järvinen | 92ee76b | 2008-11-24 21:26:56 -0800 | [diff] [blame] | 1320 | if (skb == tp->retransmit_skb_hint) |
| 1321 | tp->retransmit_skb_hint = prev; |
Ilpo Järvinen | 92ee76b | 2008-11-24 21:26:56 -0800 | [diff] [blame] | 1322 | if (skb == tp->lost_skb_hint) { |
| 1323 | tp->lost_skb_hint = prev; |
| 1324 | tp->lost_cnt_hint -= tcp_skb_pcount(prev); |
| 1325 | } |
| 1326 | |
Eric Dumazet | 5e8a402f | 2013-10-04 10:31:41 -0700 | [diff] [blame] | 1327 | TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; |
Martin KaFai Lau | a643b5d | 2016-04-25 14:44:49 -0700 | [diff] [blame] | 1328 | TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor; |
Eric Dumazet | 5e8a402f | 2013-10-04 10:31:41 -0700 | [diff] [blame] | 1329 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) |
| 1330 | TCP_SKB_CB(prev)->end_seq++; |
| 1331 | |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1332 | if (skb == tcp_highest_sack(sk)) |
| 1333 | tcp_advance_highest_sack(sk, skb); |
| 1334 | |
Martin KaFai Lau | cfea5a6 | 2016-04-19 22:39:29 -0700 | [diff] [blame] | 1335 | tcp_skb_collapse_tstamp(prev, skb); |
Yuchung Cheng | b9f6482 | 2016-09-19 23:39:14 -0400 | [diff] [blame] | 1336 | if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp.v64)) |
| 1337 | TCP_SKB_CB(prev)->tx.delivered_mstamp.v64 = 0; |
| 1338 | |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1339 | tcp_unlink_write_queue(skb, sk); |
| 1340 | sk_wmem_free_skb(sk, skb); |
| 1341 | |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 1342 | NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED); |
Ilpo Järvinen | 111cc8b | 2008-11-24 21:27:22 -0800 | [diff] [blame] | 1343 | |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1344 | return true; |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1345 | } |
| 1346 | |
| 1347 | /* I wish gso_size would have a bit more sane initialization than |
| 1348 | * something-or-zero which complicates things |
| 1349 | */ |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 1350 | static int tcp_skb_seglen(const struct sk_buff *skb) |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1351 | { |
Ilpo Järvinen | 775ffab | 2008-12-05 22:41:26 -0800 | [diff] [blame] | 1352 | return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb); |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1353 | } |
| 1354 | |
| 1355 | /* Shifting pages past head area doesn't work */ |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 1356 | static int skb_can_shift(const struct sk_buff *skb) |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1357 | { |
| 1358 | return !skb_headlen(skb) && skb_is_nonlinear(skb); |
| 1359 | } |
| 1360 | |
| 1361 | /* Try collapsing SACK blocks spanning across multiple skbs to a single |
| 1362 | * skb. |
| 1363 | */ |
| 1364 | static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb, |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1365 | struct tcp_sacktag_state *state, |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1366 | u32 start_seq, u32 end_seq, |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1367 | bool dup_sack) |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1368 | { |
| 1369 | struct tcp_sock *tp = tcp_sk(sk); |
| 1370 | struct sk_buff *prev; |
| 1371 | int mss; |
| 1372 | int pcount = 0; |
| 1373 | int len; |
| 1374 | int in_sack; |
| 1375 | |
| 1376 | if (!sk_can_gso(sk)) |
| 1377 | goto fallback; |
| 1378 | |
| 1379 | /* Normally R but no L won't result in plain S */ |
| 1380 | if (!dup_sack && |
Ilpo Järvinen | 9969ca5 | 2008-12-05 22:41:06 -0800 | [diff] [blame] | 1381 | (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS) |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1382 | goto fallback; |
| 1383 | if (!skb_can_shift(skb)) |
| 1384 | goto fallback; |
| 1385 | /* This frame is about to be dropped (was ACKed). */ |
| 1386 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) |
| 1387 | goto fallback; |
| 1388 | |
| 1389 | /* Can only happen with delayed DSACK + discard craziness */ |
| 1390 | if (unlikely(skb == tcp_write_queue_head(sk))) |
| 1391 | goto fallback; |
| 1392 | prev = tcp_write_queue_prev(sk, skb); |
| 1393 | |
| 1394 | if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) |
| 1395 | goto fallback; |
| 1396 | |
Martin KaFai Lau | a643b5d | 2016-04-25 14:44:49 -0700 | [diff] [blame] | 1397 | if (!tcp_skb_can_collapse_to(prev)) |
| 1398 | goto fallback; |
| 1399 | |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1400 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && |
| 1401 | !before(end_seq, TCP_SKB_CB(skb)->end_seq); |
| 1402 | |
| 1403 | if (in_sack) { |
| 1404 | len = skb->len; |
| 1405 | pcount = tcp_skb_pcount(skb); |
Ilpo Järvinen | 775ffab | 2008-12-05 22:41:26 -0800 | [diff] [blame] | 1406 | mss = tcp_skb_seglen(skb); |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1407 | |
| 1408 | /* TODO: Fix DSACKs to not fragment already SACKed and we can |
| 1409 | * drop this restriction as unnecessary |
| 1410 | */ |
Ilpo Järvinen | 775ffab | 2008-12-05 22:41:26 -0800 | [diff] [blame] | 1411 | if (mss != tcp_skb_seglen(prev)) |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1412 | goto fallback; |
| 1413 | } else { |
| 1414 | if (!after(TCP_SKB_CB(skb)->end_seq, start_seq)) |
| 1415 | goto noop; |
| 1416 | /* CHECKME: This is non-MSS split case only?, this will |
| 1417 | * cause skipped skbs due to advancing loop btw, original |
| 1418 | * has that feature too |
| 1419 | */ |
| 1420 | if (tcp_skb_pcount(skb) <= 1) |
| 1421 | goto noop; |
| 1422 | |
| 1423 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); |
| 1424 | if (!in_sack) { |
| 1425 | /* TODO: head merge to next could be attempted here |
| 1426 | * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)), |
| 1427 | * though it might not be worth of the additional hassle |
| 1428 | * |
| 1429 | * ...we can probably just fallback to what was done |
| 1430 | * previously. We could try merging non-SACKed ones |
| 1431 | * as well but it probably isn't going to buy off |
| 1432 | * because later SACKs might again split them, and |
| 1433 | * it would make skb timestamp tracking considerably |
| 1434 | * harder problem. |
| 1435 | */ |
| 1436 | goto fallback; |
| 1437 | } |
| 1438 | |
| 1439 | len = end_seq - TCP_SKB_CB(skb)->seq; |
| 1440 | BUG_ON(len < 0); |
| 1441 | BUG_ON(len > skb->len); |
| 1442 | |
| 1443 | /* MSS boundaries should be honoured or else pcount will |
| 1444 | * severely break even though it makes things bit trickier. |
| 1445 | * Optimize common case to avoid most of the divides |
| 1446 | */ |
| 1447 | mss = tcp_skb_mss(skb); |
| 1448 | |
| 1449 | /* TODO: Fix DSACKs to not fragment already SACKed and we can |
| 1450 | * drop this restriction as unnecessary |
| 1451 | */ |
Ilpo Järvinen | 775ffab | 2008-12-05 22:41:26 -0800 | [diff] [blame] | 1452 | if (mss != tcp_skb_seglen(prev)) |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1453 | goto fallback; |
| 1454 | |
| 1455 | if (len == mss) { |
| 1456 | pcount = 1; |
| 1457 | } else if (len < mss) { |
| 1458 | goto noop; |
| 1459 | } else { |
| 1460 | pcount = len / mss; |
| 1461 | len = pcount * mss; |
| 1462 | } |
| 1463 | } |
| 1464 | |
Neal Cardwell | 4648dc9 | 2012-03-05 19:35:04 +0000 | [diff] [blame] | 1465 | /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */ |
| 1466 | if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una)) |
| 1467 | goto fallback; |
| 1468 | |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1469 | if (!skb_shift(prev, skb, len)) |
| 1470 | goto fallback; |
Ilpo Järvinen | 9ec06ff | 2009-03-01 00:21:36 -0800 | [diff] [blame] | 1471 | if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack)) |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1472 | goto out; |
| 1473 | |
| 1474 | /* Hole filled allows collapsing with the next as well, this is very |
| 1475 | * useful when hole on every nth skb pattern happens |
| 1476 | */ |
| 1477 | if (prev == tcp_write_queue_tail(sk)) |
| 1478 | goto out; |
| 1479 | skb = tcp_write_queue_next(sk, prev); |
| 1480 | |
Ilpo Järvinen | f0bc52f | 2008-12-05 22:40:47 -0800 | [diff] [blame] | 1481 | if (!skb_can_shift(skb) || |
| 1482 | (skb == tcp_send_head(sk)) || |
| 1483 | ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) || |
Ilpo Järvinen | 775ffab | 2008-12-05 22:41:26 -0800 | [diff] [blame] | 1484 | (mss != tcp_skb_seglen(skb))) |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1485 | goto out; |
| 1486 | |
| 1487 | len = skb->len; |
| 1488 | if (skb_shift(prev, skb, len)) { |
| 1489 | pcount += tcp_skb_pcount(skb); |
Ilpo Järvinen | 9ec06ff | 2009-03-01 00:21:36 -0800 | [diff] [blame] | 1490 | tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0); |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1491 | } |
| 1492 | |
| 1493 | out: |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1494 | state->fack_count += pcount; |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1495 | return prev; |
| 1496 | |
| 1497 | noop: |
| 1498 | return skb; |
| 1499 | |
| 1500 | fallback: |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 1501 | NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK); |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1502 | return NULL; |
| 1503 | } |
| 1504 | |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1505 | static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk, |
| 1506 | struct tcp_sack_block *next_dup, |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1507 | struct tcp_sacktag_state *state, |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1508 | u32 start_seq, u32 end_seq, |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1509 | bool dup_sack_in) |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1510 | { |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1511 | struct tcp_sock *tp = tcp_sk(sk); |
| 1512 | struct sk_buff *tmp; |
| 1513 | |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1514 | tcp_for_write_queue_from(skb, sk) { |
| 1515 | int in_sack = 0; |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1516 | bool dup_sack = dup_sack_in; |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1517 | |
| 1518 | if (skb == tcp_send_head(sk)) |
| 1519 | break; |
| 1520 | |
| 1521 | /* queue is in-order => we can short-circuit the walk early */ |
| 1522 | if (!before(TCP_SKB_CB(skb)->seq, end_seq)) |
| 1523 | break; |
| 1524 | |
Ian Morris | 00db412 | 2015-04-03 09:17:27 +0100 | [diff] [blame] | 1525 | if (next_dup && |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1526 | before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) { |
| 1527 | in_sack = tcp_match_skb_to_sack(sk, skb, |
| 1528 | next_dup->start_seq, |
| 1529 | next_dup->end_seq); |
| 1530 | if (in_sack > 0) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1531 | dup_sack = true; |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1532 | } |
| 1533 | |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1534 | /* skb reference here is a bit tricky to get right, since |
| 1535 | * shifting can eat and free both this skb and the next, |
| 1536 | * so not even _safe variant of the loop is enough. |
| 1537 | */ |
| 1538 | if (in_sack <= 0) { |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1539 | tmp = tcp_shift_skb_data(sk, skb, state, |
| 1540 | start_seq, end_seq, dup_sack); |
Ian Morris | 00db412 | 2015-04-03 09:17:27 +0100 | [diff] [blame] | 1541 | if (tmp) { |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1542 | if (tmp != skb) { |
| 1543 | skb = tmp; |
| 1544 | continue; |
| 1545 | } |
| 1546 | |
| 1547 | in_sack = 0; |
| 1548 | } else { |
| 1549 | in_sack = tcp_match_skb_to_sack(sk, skb, |
| 1550 | start_seq, |
| 1551 | end_seq); |
| 1552 | } |
| 1553 | } |
| 1554 | |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1555 | if (unlikely(in_sack < 0)) |
| 1556 | break; |
| 1557 | |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1558 | if (in_sack) { |
Neal Cardwell | cc9a672 | 2012-02-12 18:37:09 +0000 | [diff] [blame] | 1559 | TCP_SKB_CB(skb)->sacked = |
| 1560 | tcp_sacktag_one(sk, |
| 1561 | state, |
| 1562 | TCP_SKB_CB(skb)->sacked, |
| 1563 | TCP_SKB_CB(skb)->seq, |
| 1564 | TCP_SKB_CB(skb)->end_seq, |
| 1565 | dup_sack, |
Yuchung Cheng | 59c9af4 | 2013-07-22 16:20:47 -0700 | [diff] [blame] | 1566 | tcp_skb_pcount(skb), |
Eric Dumazet | 740b0f1 | 2014-02-26 14:02:48 -0800 | [diff] [blame] | 1567 | &skb->skb_mstamp); |
Yuchung Cheng | b9f6482 | 2016-09-19 23:39:14 -0400 | [diff] [blame] | 1568 | tcp_rate_skb_delivered(sk, skb, state->rate); |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1569 | |
Ilpo Järvinen | 832d11c | 2008-11-24 21:20:15 -0800 | [diff] [blame] | 1570 | if (!before(TCP_SKB_CB(skb)->seq, |
| 1571 | tcp_highest_sack_seq(tp))) |
| 1572 | tcp_advance_highest_sack(sk, skb); |
| 1573 | } |
| 1574 | |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1575 | state->fack_count += tcp_skb_pcount(skb); |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1576 | } |
| 1577 | return skb; |
| 1578 | } |
| 1579 | |
| 1580 | /* Avoid all extra work that is being done by sacktag while walking in |
| 1581 | * a normal way |
| 1582 | */ |
| 1583 | static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk, |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1584 | struct tcp_sacktag_state *state, |
| 1585 | u32 skip_to_seq) |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1586 | { |
| 1587 | tcp_for_write_queue_from(skb, sk) { |
| 1588 | if (skb == tcp_send_head(sk)) |
| 1589 | break; |
| 1590 | |
Ilpo Järvinen | e8bae27 | 2008-11-24 21:12:28 -0800 | [diff] [blame] | 1591 | if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq)) |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1592 | break; |
Ilpo Järvinen | d152a7d | 2008-03-03 12:10:16 -0800 | [diff] [blame] | 1593 | |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1594 | state->fack_count += tcp_skb_pcount(skb); |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1595 | } |
| 1596 | return skb; |
| 1597 | } |
| 1598 | |
| 1599 | static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb, |
| 1600 | struct sock *sk, |
| 1601 | struct tcp_sack_block *next_dup, |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1602 | struct tcp_sacktag_state *state, |
| 1603 | u32 skip_to_seq) |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1604 | { |
Ian Morris | 51456b2 | 2015-04-03 09:17:26 +0100 | [diff] [blame] | 1605 | if (!next_dup) |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1606 | return skb; |
| 1607 | |
| 1608 | if (before(next_dup->start_seq, skip_to_seq)) { |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1609 | skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq); |
| 1610 | skb = tcp_sacktag_walk(skb, sk, NULL, state, |
| 1611 | next_dup->start_seq, next_dup->end_seq, |
| 1612 | 1); |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1613 | } |
| 1614 | |
| 1615 | return skb; |
| 1616 | } |
| 1617 | |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 1618 | static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache) |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1619 | { |
| 1620 | return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); |
| 1621 | } |
| 1622 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1623 | static int |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 1624 | tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb, |
Kenneth Klette Jonassen | 196da97 | 2015-05-01 01:10:57 +0200 | [diff] [blame] | 1625 | u32 prior_snd_una, struct tcp_sacktag_state *state) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1626 | { |
| 1627 | struct tcp_sock *tp = tcp_sk(sk); |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 1628 | const unsigned char *ptr = (skb_transport_header(ack_skb) + |
| 1629 | TCP_SKB_CB(ack_skb)->sacked); |
Ilpo Järvinen | fd6dad6 | 2007-11-15 19:49:47 -0800 | [diff] [blame] | 1630 | struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2); |
Adam Langley | 4389dde | 2008-07-19 00:07:02 -0700 | [diff] [blame] | 1631 | struct tcp_sack_block sp[TCP_NUM_SACKS]; |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1632 | struct tcp_sack_block *cache; |
| 1633 | struct sk_buff *skb; |
Adam Langley | 4389dde | 2008-07-19 00:07:02 -0700 | [diff] [blame] | 1634 | int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3); |
Ilpo Järvinen | fd6dad6 | 2007-11-15 19:49:47 -0800 | [diff] [blame] | 1635 | int used_sacks; |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1636 | bool found_dup_sack = false; |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1637 | int i, j; |
Baruch Even | fda03fb | 2007-02-04 23:35:57 -0800 | [diff] [blame] | 1638 | int first_sack_index; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1639 | |
Kenneth Klette Jonassen | 196da97 | 2015-05-01 01:10:57 +0200 | [diff] [blame] | 1640 | state->flag = 0; |
| 1641 | state->reord = tp->packets_out; |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1642 | |
Ilpo Järvinen | d738cd8 | 2007-03-24 21:03:23 -0700 | [diff] [blame] | 1643 | if (!tp->sacked_out) { |
Ilpo Järvinen | de83c05 | 2007-10-07 23:37:55 -0700 | [diff] [blame] | 1644 | if (WARN_ON(tp->fackets_out)) |
| 1645 | tp->fackets_out = 0; |
Ilpo Järvinen | 6859d49 | 2007-12-02 00:48:06 +0200 | [diff] [blame] | 1646 | tcp_highest_sack_reset(sk); |
Ilpo Järvinen | d738cd8 | 2007-03-24 21:03:23 -0700 | [diff] [blame] | 1647 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1648 | |
Pavel Emelyanov | 1ed8346 | 2008-07-16 20:29:51 -0700 | [diff] [blame] | 1649 | found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire, |
David S. Miller | d06e021 | 2007-06-18 22:43:06 -0700 | [diff] [blame] | 1650 | num_sacks, prior_snd_una); |
Yuchung Cheng | b9f6482 | 2016-09-19 23:39:14 -0400 | [diff] [blame] | 1651 | if (found_dup_sack) { |
Kenneth Klette Jonassen | 196da97 | 2015-05-01 01:10:57 +0200 | [diff] [blame] | 1652 | state->flag |= FLAG_DSACKING_ACK; |
Yuchung Cheng | b9f6482 | 2016-09-19 23:39:14 -0400 | [diff] [blame] | 1653 | tp->delivered++; /* A spurious retransmission is delivered */ |
| 1654 | } |
Baruch Even | 6f74651 | 2007-02-04 23:36:42 -0800 | [diff] [blame] | 1655 | |
| 1656 | /* Eliminate too old ACKs, but take into |
| 1657 | * account more or less fresh ones, they can |
| 1658 | * contain valid SACK info. |
| 1659 | */ |
| 1660 | if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window)) |
| 1661 | return 0; |
| 1662 | |
Ilpo Järvinen | 96a2d41 | 2007-11-14 15:47:18 -0800 | [diff] [blame] | 1663 | if (!tp->packets_out) |
| 1664 | goto out; |
| 1665 | |
Ilpo Järvinen | fd6dad6 | 2007-11-15 19:49:47 -0800 | [diff] [blame] | 1666 | used_sacks = 0; |
| 1667 | first_sack_index = 0; |
| 1668 | for (i = 0; i < num_sacks; i++) { |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1669 | bool dup_sack = !i && found_dup_sack; |
Ilpo Järvinen | fd6dad6 | 2007-11-15 19:49:47 -0800 | [diff] [blame] | 1670 | |
Harvey Harrison | d3e2ce3 | 2008-05-02 16:26:16 -0700 | [diff] [blame] | 1671 | sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq); |
| 1672 | sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq); |
Ilpo Järvinen | fd6dad6 | 2007-11-15 19:49:47 -0800 | [diff] [blame] | 1673 | |
| 1674 | if (!tcp_is_sackblock_valid(tp, dup_sack, |
| 1675 | sp[used_sacks].start_seq, |
| 1676 | sp[used_sacks].end_seq)) { |
Pavel Emelyanov | 40b215e | 2008-07-03 01:05:41 -0700 | [diff] [blame] | 1677 | int mib_idx; |
| 1678 | |
Ilpo Järvinen | fd6dad6 | 2007-11-15 19:49:47 -0800 | [diff] [blame] | 1679 | if (dup_sack) { |
| 1680 | if (!tp->undo_marker) |
Pavel Emelyanov | 40b215e | 2008-07-03 01:05:41 -0700 | [diff] [blame] | 1681 | mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO; |
Ilpo Järvinen | fd6dad6 | 2007-11-15 19:49:47 -0800 | [diff] [blame] | 1682 | else |
Pavel Emelyanov | 40b215e | 2008-07-03 01:05:41 -0700 | [diff] [blame] | 1683 | mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD; |
Ilpo Järvinen | fd6dad6 | 2007-11-15 19:49:47 -0800 | [diff] [blame] | 1684 | } else { |
| 1685 | /* Don't count olds caused by ACK reordering */ |
| 1686 | if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) && |
| 1687 | !after(sp[used_sacks].end_seq, tp->snd_una)) |
| 1688 | continue; |
Pavel Emelyanov | 40b215e | 2008-07-03 01:05:41 -0700 | [diff] [blame] | 1689 | mib_idx = LINUX_MIB_TCPSACKDISCARD; |
Ilpo Järvinen | fd6dad6 | 2007-11-15 19:49:47 -0800 | [diff] [blame] | 1690 | } |
Pavel Emelyanov | 40b215e | 2008-07-03 01:05:41 -0700 | [diff] [blame] | 1691 | |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 1692 | NET_INC_STATS(sock_net(sk), mib_idx); |
Ilpo Järvinen | fd6dad6 | 2007-11-15 19:49:47 -0800 | [diff] [blame] | 1693 | if (i == 0) |
| 1694 | first_sack_index = -1; |
| 1695 | continue; |
| 1696 | } |
| 1697 | |
| 1698 | /* Ignore very old stuff early */ |
| 1699 | if (!after(sp[used_sacks].end_seq, prior_snd_una)) |
| 1700 | continue; |
| 1701 | |
| 1702 | used_sacks++; |
| 1703 | } |
| 1704 | |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1705 | /* order SACK blocks to allow in order walk of the retrans queue */ |
| 1706 | for (i = used_sacks - 1; i > 0; i--) { |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1707 | for (j = 0; j < i; j++) { |
| 1708 | if (after(sp[j].start_seq, sp[j + 1].start_seq)) { |
Ilpo Järvinen | a0bffff | 2009-03-21 13:36:17 -0700 | [diff] [blame] | 1709 | swap(sp[j], sp[j + 1]); |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1710 | |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1711 | /* Track where the first SACK block goes to */ |
| 1712 | if (j == first_sack_index) |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1713 | first_sack_index = j + 1; |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1714 | } |
| 1715 | } |
| 1716 | } |
| 1717 | |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1718 | skb = tcp_write_queue_head(sk); |
Kenneth Klette Jonassen | 196da97 | 2015-05-01 01:10:57 +0200 | [diff] [blame] | 1719 | state->fack_count = 0; |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1720 | i = 0; |
| 1721 | |
| 1722 | if (!tp->sacked_out) { |
| 1723 | /* It's already past, so skip checking against it */ |
| 1724 | cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); |
| 1725 | } else { |
| 1726 | cache = tp->recv_sack_cache; |
| 1727 | /* Skip empty blocks in at head of the cache */ |
| 1728 | while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq && |
| 1729 | !cache->end_seq) |
| 1730 | cache++; |
Baruch Even | fda03fb | 2007-02-04 23:35:57 -0800 | [diff] [blame] | 1731 | } |
| 1732 | |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1733 | while (i < used_sacks) { |
Ilpo Järvinen | fd6dad6 | 2007-11-15 19:49:47 -0800 | [diff] [blame] | 1734 | u32 start_seq = sp[i].start_seq; |
| 1735 | u32 end_seq = sp[i].end_seq; |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1736 | bool dup_sack = (found_dup_sack && (i == first_sack_index)); |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1737 | struct tcp_sack_block *next_dup = NULL; |
Ilpo Järvinen | e56d6cd | 2007-11-01 00:09:37 -0700 | [diff] [blame] | 1738 | |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1739 | if (found_dup_sack && ((i + 1) == first_sack_index)) |
| 1740 | next_dup = &sp[i + 1]; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1741 | |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1742 | /* Skip too early cached blocks */ |
| 1743 | while (tcp_sack_cache_ok(tp, cache) && |
| 1744 | !before(start_seq, cache->end_seq)) |
| 1745 | cache++; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1746 | |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1747 | /* Can skip some work by looking recv_sack_cache? */ |
| 1748 | if (tcp_sack_cache_ok(tp, cache) && !dup_sack && |
| 1749 | after(end_seq, cache->start_seq)) { |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1750 | |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1751 | /* Head todo? */ |
| 1752 | if (before(start_seq, cache->start_seq)) { |
Kenneth Klette Jonassen | 196da97 | 2015-05-01 01:10:57 +0200 | [diff] [blame] | 1753 | skb = tcp_sacktag_skip(skb, sk, state, |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1754 | start_seq); |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1755 | skb = tcp_sacktag_walk(skb, sk, next_dup, |
Kenneth Klette Jonassen | 196da97 | 2015-05-01 01:10:57 +0200 | [diff] [blame] | 1756 | state, |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1757 | start_seq, |
| 1758 | cache->start_seq, |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1759 | dup_sack); |
Baruch Even | fda03fb | 2007-02-04 23:35:57 -0800 | [diff] [blame] | 1760 | } |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1761 | |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1762 | /* Rest of the block already fully processed? */ |
Ilpo Järvinen | 20de20b | 2007-11-16 16:17:05 -0800 | [diff] [blame] | 1763 | if (!after(end_seq, cache->end_seq)) |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1764 | goto advance_sp; |
Ilpo Järvinen | 20de20b | 2007-11-16 16:17:05 -0800 | [diff] [blame] | 1765 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1766 | skb = tcp_maybe_skipping_dsack(skb, sk, next_dup, |
Kenneth Klette Jonassen | 196da97 | 2015-05-01 01:10:57 +0200 | [diff] [blame] | 1767 | state, |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1768 | cache->end_seq); |
Ilpo Järvinen | e56d6cd | 2007-11-01 00:09:37 -0700 | [diff] [blame] | 1769 | |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1770 | /* ...tail remains todo... */ |
Ilpo Järvinen | 6859d49 | 2007-12-02 00:48:06 +0200 | [diff] [blame] | 1771 | if (tcp_highest_sack_seq(tp) == cache->end_seq) { |
Ilpo Järvinen | 20de20b | 2007-11-16 16:17:05 -0800 | [diff] [blame] | 1772 | /* ...but better entrypoint exists! */ |
Ilpo Järvinen | 6859d49 | 2007-12-02 00:48:06 +0200 | [diff] [blame] | 1773 | skb = tcp_highest_sack(sk); |
Ian Morris | 51456b2 | 2015-04-03 09:17:26 +0100 | [diff] [blame] | 1774 | if (!skb) |
Ilpo Järvinen | 6859d49 | 2007-12-02 00:48:06 +0200 | [diff] [blame] | 1775 | break; |
Kenneth Klette Jonassen | 196da97 | 2015-05-01 01:10:57 +0200 | [diff] [blame] | 1776 | state->fack_count = tp->fackets_out; |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1777 | cache++; |
| 1778 | goto walk; |
| 1779 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1780 | |
Kenneth Klette Jonassen | 196da97 | 2015-05-01 01:10:57 +0200 | [diff] [blame] | 1781 | skb = tcp_sacktag_skip(skb, sk, state, cache->end_seq); |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1782 | /* Check overlap against next cached too (past this one already) */ |
| 1783 | cache++; |
| 1784 | continue; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1785 | } |
Ilpo Järvinen | fbd52eb | 2007-11-10 21:24:19 -0800 | [diff] [blame] | 1786 | |
Ilpo Järvinen | 6859d49 | 2007-12-02 00:48:06 +0200 | [diff] [blame] | 1787 | if (!before(start_seq, tcp_highest_sack_seq(tp))) { |
| 1788 | skb = tcp_highest_sack(sk); |
Ian Morris | 51456b2 | 2015-04-03 09:17:26 +0100 | [diff] [blame] | 1789 | if (!skb) |
Ilpo Järvinen | 6859d49 | 2007-12-02 00:48:06 +0200 | [diff] [blame] | 1790 | break; |
Kenneth Klette Jonassen | 196da97 | 2015-05-01 01:10:57 +0200 | [diff] [blame] | 1791 | state->fack_count = tp->fackets_out; |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1792 | } |
Kenneth Klette Jonassen | 196da97 | 2015-05-01 01:10:57 +0200 | [diff] [blame] | 1793 | skb = tcp_sacktag_skip(skb, sk, state, start_seq); |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1794 | |
| 1795 | walk: |
Kenneth Klette Jonassen | 196da97 | 2015-05-01 01:10:57 +0200 | [diff] [blame] | 1796 | skb = tcp_sacktag_walk(skb, sk, next_dup, state, |
Ilpo Järvinen | a1197f5 | 2008-12-05 22:42:22 -0800 | [diff] [blame] | 1797 | start_seq, end_seq, dup_sack); |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1798 | |
| 1799 | advance_sp: |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1800 | i++; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1801 | } |
| 1802 | |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1803 | /* Clear the head of the cache sack blocks so we can skip it next time */ |
| 1804 | for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) { |
| 1805 | tp->recv_sack_cache[i].start_seq = 0; |
| 1806 | tp->recv_sack_cache[i].end_seq = 0; |
| 1807 | } |
| 1808 | for (j = 0; j < used_sacks; j++) |
| 1809 | tp->recv_sack_cache[i++] = sp[j]; |
| 1810 | |
Kenneth Klette Jonassen | 196da97 | 2015-05-01 01:10:57 +0200 | [diff] [blame] | 1811 | if ((state->reord < tp->fackets_out) && |
Yuchung Cheng | 9b44190 | 2013-03-20 13:32:58 +0000 | [diff] [blame] | 1812 | ((inet_csk(sk)->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker)) |
Kenneth Klette Jonassen | 196da97 | 2015-05-01 01:10:57 +0200 | [diff] [blame] | 1813 | tcp_update_reordering(sk, tp->fackets_out - state->reord, 0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1814 | |
Yuchung Cheng | 9dac883 | 2015-04-29 11:28:30 -0700 | [diff] [blame] | 1815 | tcp_verify_left_out(tp); |
Ilpo Järvinen | 96a2d41 | 2007-11-14 15:47:18 -0800 | [diff] [blame] | 1816 | out: |
| 1817 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1818 | #if FASTRETRANS_DEBUG > 0 |
Ilpo Järvinen | 547b792 | 2008-07-25 21:43:18 -0700 | [diff] [blame] | 1819 | WARN_ON((int)tp->sacked_out < 0); |
| 1820 | WARN_ON((int)tp->lost_out < 0); |
| 1821 | WARN_ON((int)tp->retrans_out < 0); |
| 1822 | WARN_ON((int)tcp_packets_in_flight(tp) < 0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1823 | #endif |
Kenneth Klette Jonassen | 196da97 | 2015-05-01 01:10:57 +0200 | [diff] [blame] | 1824 | return state->flag; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1825 | } |
| 1826 | |
Ilpo Järvinen | 882beba | 2008-04-07 22:33:07 -0700 | [diff] [blame] | 1827 | /* Limits sacked_out so that sum with lost_out isn't ever larger than |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1828 | * packets_out. Returns false if sacked_out adjustement wasn't necessary. |
Ilpo Järvinen | 30935cf | 2007-02-21 23:01:36 -0800 | [diff] [blame] | 1829 | */ |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1830 | static bool tcp_limit_reno_sacked(struct tcp_sock *tp) |
Ilpo Järvinen | 4ddf667 | 2007-05-27 01:52:00 -0700 | [diff] [blame] | 1831 | { |
Ilpo Järvinen | 4ddf667 | 2007-05-27 01:52:00 -0700 | [diff] [blame] | 1832 | u32 holes; |
| 1833 | |
| 1834 | holes = max(tp->lost_out, 1U); |
| 1835 | holes = min(holes, tp->packets_out); |
| 1836 | |
| 1837 | if ((tp->sacked_out + holes) > tp->packets_out) { |
| 1838 | tp->sacked_out = tp->packets_out - holes; |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1839 | return true; |
Ilpo Järvinen | 4ddf667 | 2007-05-27 01:52:00 -0700 | [diff] [blame] | 1840 | } |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1841 | return false; |
Ilpo Järvinen | 882beba | 2008-04-07 22:33:07 -0700 | [diff] [blame] | 1842 | } |
| 1843 | |
| 1844 | /* If we receive more dupacks than we expected counting segments |
| 1845 | * in assumption of absent reordering, interpret this as reordering. |
| 1846 | * The only another reason could be bug in receiver TCP. |
| 1847 | */ |
| 1848 | static void tcp_check_reno_reordering(struct sock *sk, const int addend) |
| 1849 | { |
| 1850 | struct tcp_sock *tp = tcp_sk(sk); |
| 1851 | if (tcp_limit_reno_sacked(tp)) |
| 1852 | tcp_update_reordering(sk, tp->packets_out + addend, 0); |
Ilpo Järvinen | 4ddf667 | 2007-05-27 01:52:00 -0700 | [diff] [blame] | 1853 | } |
| 1854 | |
| 1855 | /* Emulate SACKs for SACKless connection: account for a new dupack. */ |
| 1856 | |
| 1857 | static void tcp_add_reno_sack(struct sock *sk) |
| 1858 | { |
| 1859 | struct tcp_sock *tp = tcp_sk(sk); |
Yuchung Cheng | ddf1af6 | 2016-02-02 10:33:06 -0800 | [diff] [blame] | 1860 | u32 prior_sacked = tp->sacked_out; |
| 1861 | |
Ilpo Järvinen | 4ddf667 | 2007-05-27 01:52:00 -0700 | [diff] [blame] | 1862 | tp->sacked_out++; |
| 1863 | tcp_check_reno_reordering(sk, 0); |
Yuchung Cheng | ddf1af6 | 2016-02-02 10:33:06 -0800 | [diff] [blame] | 1864 | if (tp->sacked_out > prior_sacked) |
| 1865 | tp->delivered++; /* Some out-of-order packet is delivered */ |
Ilpo Järvinen | 005903b | 2007-08-09 14:44:16 +0300 | [diff] [blame] | 1866 | tcp_verify_left_out(tp); |
Ilpo Järvinen | 4ddf667 | 2007-05-27 01:52:00 -0700 | [diff] [blame] | 1867 | } |
| 1868 | |
| 1869 | /* Account for ACK, ACKing some data in Reno Recovery phase. */ |
| 1870 | |
| 1871 | static void tcp_remove_reno_sacks(struct sock *sk, int acked) |
| 1872 | { |
| 1873 | struct tcp_sock *tp = tcp_sk(sk); |
| 1874 | |
| 1875 | if (acked > 0) { |
| 1876 | /* One ACK acked hole. The rest eat duplicate ACKs. */ |
Yuchung Cheng | ddf1af6 | 2016-02-02 10:33:06 -0800 | [diff] [blame] | 1877 | tp->delivered += max_t(int, acked - tp->sacked_out, 1); |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1878 | if (acked - 1 >= tp->sacked_out) |
Ilpo Järvinen | 4ddf667 | 2007-05-27 01:52:00 -0700 | [diff] [blame] | 1879 | tp->sacked_out = 0; |
| 1880 | else |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1881 | tp->sacked_out -= acked - 1; |
Ilpo Järvinen | 4ddf667 | 2007-05-27 01:52:00 -0700 | [diff] [blame] | 1882 | } |
| 1883 | tcp_check_reno_reordering(sk, acked); |
Ilpo Järvinen | 005903b | 2007-08-09 14:44:16 +0300 | [diff] [blame] | 1884 | tcp_verify_left_out(tp); |
Ilpo Järvinen | 4ddf667 | 2007-05-27 01:52:00 -0700 | [diff] [blame] | 1885 | } |
| 1886 | |
| 1887 | static inline void tcp_reset_reno_sack(struct tcp_sock *tp) |
| 1888 | { |
| 1889 | tp->sacked_out = 0; |
Ilpo Järvinen | 4ddf667 | 2007-05-27 01:52:00 -0700 | [diff] [blame] | 1890 | } |
| 1891 | |
Yuchung Cheng | 989e04c | 2014-08-22 14:15:22 -0700 | [diff] [blame] | 1892 | void tcp_clear_retrans(struct tcp_sock *tp) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1893 | { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1894 | tp->retrans_out = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1895 | tp->lost_out = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1896 | tp->undo_marker = 0; |
Yuchung Cheng | 6e08d5e | 2014-07-02 12:07:16 -0700 | [diff] [blame] | 1897 | tp->undo_retrans = -1; |
Ilpo Järvinen | 4cd8299 | 2007-10-11 17:34:57 -0700 | [diff] [blame] | 1898 | tp->fackets_out = 0; |
| 1899 | tp->sacked_out = 0; |
| 1900 | } |
| 1901 | |
Yuchung Cheng | 989e04c | 2014-08-22 14:15:22 -0700 | [diff] [blame] | 1902 | static inline void tcp_init_undo(struct tcp_sock *tp) |
| 1903 | { |
| 1904 | tp->undo_marker = tp->snd_una; |
| 1905 | /* Retransmission still in flight may cause DSACKs later. */ |
| 1906 | tp->undo_retrans = tp->retrans_out ? : -1; |
| 1907 | } |
| 1908 | |
Neal Cardwell | 5ae344c | 2014-08-04 19:12:29 -0400 | [diff] [blame] | 1909 | /* Enter Loss state. If we detect SACK reneging, forget all SACK information |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1910 | * and reset tags completely, otherwise preserve SACKs. If receiver |
| 1911 | * dropped its ofo queue, we will know this due to reneging detection. |
| 1912 | */ |
Neal Cardwell | 5ae344c | 2014-08-04 19:12:29 -0400 | [diff] [blame] | 1913 | void tcp_enter_loss(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1914 | { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1915 | const struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1916 | struct tcp_sock *tp = tcp_sk(sk); |
Nikolay Borisov | 1043e25 | 2016-02-03 09:46:52 +0200 | [diff] [blame] | 1917 | struct net *net = sock_net(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1918 | struct sk_buff *skb; |
Yuchung Cheng | f82b681 | 2015-07-13 12:10:20 -0700 | [diff] [blame] | 1919 | bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery; |
Neal Cardwell | 5ae344c | 2014-08-04 19:12:29 -0400 | [diff] [blame] | 1920 | bool is_reneg; /* is receiver reneging on SACKs? */ |
Neal Cardwell | 0682e69 | 2016-09-19 23:39:13 -0400 | [diff] [blame] | 1921 | bool mark_lost; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1922 | |
| 1923 | /* Reduce ssthresh if it has not yet been made inside this window. */ |
Yuchung Cheng | e33099f | 2013-03-20 13:33:00 +0000 | [diff] [blame] | 1924 | if (icsk->icsk_ca_state <= TCP_CA_Disorder || |
| 1925 | !after(tp->high_seq, tp->snd_una) || |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1926 | (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) { |
| 1927 | tp->prior_ssthresh = tcp_current_ssthresh(sk); |
| 1928 | tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); |
| 1929 | tcp_ca_event(sk, CA_EVENT_LOSS); |
Yuchung Cheng | 989e04c | 2014-08-22 14:15:22 -0700 | [diff] [blame] | 1930 | tcp_init_undo(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1931 | } |
| 1932 | tp->snd_cwnd = 1; |
| 1933 | tp->snd_cwnd_cnt = 0; |
| 1934 | tp->snd_cwnd_stamp = tcp_time_stamp; |
| 1935 | |
Yuchung Cheng | 989e04c | 2014-08-22 14:15:22 -0700 | [diff] [blame] | 1936 | tp->retrans_out = 0; |
| 1937 | tp->lost_out = 0; |
Ilpo Järvinen | 4cd8299 | 2007-10-11 17:34:57 -0700 | [diff] [blame] | 1938 | |
| 1939 | if (tcp_is_reno(tp)) |
| 1940 | tcp_reset_reno_sack(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1941 | |
Neal Cardwell | 5ae344c | 2014-08-04 19:12:29 -0400 | [diff] [blame] | 1942 | skb = tcp_write_queue_head(sk); |
| 1943 | is_reneg = skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED); |
| 1944 | if (is_reneg) { |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 1945 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING); |
Ilpo Järvinen | 4cd8299 | 2007-10-11 17:34:57 -0700 | [diff] [blame] | 1946 | tp->sacked_out = 0; |
| 1947 | tp->fackets_out = 0; |
Ilpo Järvinen | b768920 | 2007-09-20 11:37:19 -0700 | [diff] [blame] | 1948 | } |
Ilpo Järvinen | 64edc27 | 2008-09-20 21:18:32 -0700 | [diff] [blame] | 1949 | tcp_clear_all_retrans_hints(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1950 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1951 | tcp_for_write_queue(skb, sk) { |
| 1952 | if (skb == tcp_send_head(sk)) |
| 1953 | break; |
Ilpo Järvinen | 4cd8299 | 2007-10-11 17:34:57 -0700 | [diff] [blame] | 1954 | |
Neal Cardwell | 0682e69 | 2016-09-19 23:39:13 -0400 | [diff] [blame] | 1955 | mark_lost = (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) || |
| 1956 | is_reneg); |
| 1957 | if (mark_lost) |
| 1958 | tcp_sum_lost(tp, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1959 | TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED; |
Neal Cardwell | 0682e69 | 2016-09-19 23:39:13 -0400 | [diff] [blame] | 1960 | if (mark_lost) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1961 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED; |
| 1962 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; |
| 1963 | tp->lost_out += tcp_skb_pcount(skb); |
Ilpo Järvinen | 006f582 | 2008-09-20 21:20:20 -0700 | [diff] [blame] | 1964 | tp->retransmit_high = TCP_SKB_CB(skb)->end_seq; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1965 | } |
| 1966 | } |
Ilpo Järvinen | 005903b | 2007-08-09 14:44:16 +0300 | [diff] [blame] | 1967 | tcp_verify_left_out(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1968 | |
Yuchung Cheng | 74c181d | 2013-08-12 16:41:25 -0700 | [diff] [blame] | 1969 | /* Timeout in disordered state after receiving substantial DUPACKs |
| 1970 | * suggests that the degree of reordering is over-estimated. |
| 1971 | */ |
| 1972 | if (icsk->icsk_ca_state <= TCP_CA_Disorder && |
Nikolay Borisov | 1043e25 | 2016-02-03 09:46:52 +0200 | [diff] [blame] | 1973 | tp->sacked_out >= net->ipv4.sysctl_tcp_reordering) |
Yuchung Cheng | 74c181d | 2013-08-12 16:41:25 -0700 | [diff] [blame] | 1974 | tp->reordering = min_t(unsigned int, tp->reordering, |
Nikolay Borisov | 1043e25 | 2016-02-03 09:46:52 +0200 | [diff] [blame] | 1975 | net->ipv4.sysctl_tcp_reordering); |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1976 | tcp_set_ca_state(sk, TCP_CA_Loss); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1977 | tp->high_seq = tp->snd_nxt; |
Florian Westphal | 735d383 | 2014-09-29 13:08:30 +0200 | [diff] [blame] | 1978 | tcp_ecn_queue_cwr(tp); |
Yuchung Cheng | e33099f | 2013-03-20 13:33:00 +0000 | [diff] [blame] | 1979 | |
| 1980 | /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous |
| 1981 | * loss recovery is underway except recurring timeout(s) on |
| 1982 | * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing |
| 1983 | */ |
| 1984 | tp->frto = sysctl_tcp_frto && |
| 1985 | (new_recovery || icsk->icsk_retransmits) && |
| 1986 | !inet_csk(sk)->icsk_mtup.probe_size; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1987 | } |
| 1988 | |
Ilpo Järvinen | cadbd03 | 2007-12-31 04:49:21 -0800 | [diff] [blame] | 1989 | /* If ACK arrived pointing to a remembered SACK, it means that our |
| 1990 | * remembered SACKs do not reflect real state of receiver i.e. |
| 1991 | * receiver _host_ is heavily congested (or buggy). |
| 1992 | * |
Neal Cardwell | 5ae344c | 2014-08-04 19:12:29 -0400 | [diff] [blame] | 1993 | * To avoid big spurious retransmission bursts due to transient SACK |
| 1994 | * scoreboard oddities that look like reneging, we give the receiver a |
| 1995 | * little time (max(RTT/2, 10ms)) to send us some more ACKs that will |
| 1996 | * restore sanity to the SACK scoreboard. If the apparent reneging |
| 1997 | * persists until this RTO then we'll clear the SACK scoreboard. |
Ilpo Järvinen | cadbd03 | 2007-12-31 04:49:21 -0800 | [diff] [blame] | 1998 | */ |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1999 | static bool tcp_check_sack_reneging(struct sock *sk, int flag) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2000 | { |
Ilpo Järvinen | cadbd03 | 2007-12-31 04:49:21 -0800 | [diff] [blame] | 2001 | if (flag & FLAG_SACK_RENEGING) { |
Neal Cardwell | 5ae344c | 2014-08-04 19:12:29 -0400 | [diff] [blame] | 2002 | struct tcp_sock *tp = tcp_sk(sk); |
| 2003 | unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4), |
| 2004 | msecs_to_jiffies(10)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2005 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2006 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
Neal Cardwell | 5ae344c | 2014-08-04 19:12:29 -0400 | [diff] [blame] | 2007 | delay, TCP_RTO_MAX); |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2008 | return true; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2009 | } |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2010 | return false; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2011 | } |
| 2012 | |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 2013 | static inline int tcp_fackets_out(const struct tcp_sock *tp) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2014 | { |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2015 | return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2016 | } |
| 2017 | |
Ilpo Järvinen | 85cc391 | 2007-11-15 19:39:31 -0800 | [diff] [blame] | 2018 | /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs |
| 2019 | * counter when SACK is enabled (without SACK, sacked_out is used for |
| 2020 | * that purpose). |
| 2021 | * |
| 2022 | * Instead, with FACK TCP uses fackets_out that includes both SACKed |
| 2023 | * segments up to the highest received SACK block so far and holes in |
| 2024 | * between them. |
| 2025 | * |
| 2026 | * With reordering, holes may still be in flight, so RFC3517 recovery |
| 2027 | * uses pure sacked_out (total number of SACKed segments) even though |
| 2028 | * it violates the RFC that uses duplicate ACKs, often these are equal |
| 2029 | * but when e.g. out-of-window ACKs or packet duplication occurs, |
| 2030 | * they differ. Since neither occurs due to loss, TCP should really |
| 2031 | * ignore them. |
| 2032 | */ |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 2033 | static inline int tcp_dupack_heuristics(const struct tcp_sock *tp) |
Ilpo Järvinen | 85cc391 | 2007-11-15 19:39:31 -0800 | [diff] [blame] | 2034 | { |
| 2035 | return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1; |
| 2036 | } |
| 2037 | |
Yuchung Cheng | 750ea2b | 2012-05-02 13:30:04 +0000 | [diff] [blame] | 2038 | static bool tcp_pause_early_retransmit(struct sock *sk, int flag) |
| 2039 | { |
| 2040 | struct tcp_sock *tp = tcp_sk(sk); |
| 2041 | unsigned long delay; |
| 2042 | |
| 2043 | /* Delay early retransmit and entering fast recovery for |
| 2044 | * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples |
| 2045 | * available, or RTO is scheduled to fire first. |
| 2046 | */ |
Nandita Dukkipati | 6ba8a3b | 2013-03-11 10:00:43 +0000 | [diff] [blame] | 2047 | if (sysctl_tcp_early_retrans < 2 || sysctl_tcp_early_retrans > 3 || |
Eric Dumazet | 740b0f1 | 2014-02-26 14:02:48 -0800 | [diff] [blame] | 2048 | (flag & FLAG_ECE) || !tp->srtt_us) |
Yuchung Cheng | 750ea2b | 2012-05-02 13:30:04 +0000 | [diff] [blame] | 2049 | return false; |
| 2050 | |
Eric Dumazet | 740b0f1 | 2014-02-26 14:02:48 -0800 | [diff] [blame] | 2051 | delay = max(usecs_to_jiffies(tp->srtt_us >> 5), |
| 2052 | msecs_to_jiffies(2)); |
| 2053 | |
Yuchung Cheng | 750ea2b | 2012-05-02 13:30:04 +0000 | [diff] [blame] | 2054 | if (!time_after(inet_csk(sk)->icsk_timeout, (jiffies + delay))) |
| 2055 | return false; |
| 2056 | |
Nandita Dukkipati | 6ba8a3b | 2013-03-11 10:00:43 +0000 | [diff] [blame] | 2057 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_EARLY_RETRANS, delay, |
| 2058 | TCP_RTO_MAX); |
Yuchung Cheng | 750ea2b | 2012-05-02 13:30:04 +0000 | [diff] [blame] | 2059 | return true; |
| 2060 | } |
| 2061 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2062 | /* Linux NewReno/SACK/FACK/ECN state machine. |
| 2063 | * -------------------------------------- |
| 2064 | * |
| 2065 | * "Open" Normal state, no dubious events, fast path. |
| 2066 | * "Disorder" In all the respects it is "Open", |
| 2067 | * but requires a bit more attention. It is entered when |
| 2068 | * we see some SACKs or dupacks. It is split of "Open" |
| 2069 | * mainly to move some processing from fast path to slow one. |
| 2070 | * "CWR" CWND was reduced due to some Congestion Notification event. |
| 2071 | * It can be ECN, ICMP source quench, local device congestion. |
| 2072 | * "Recovery" CWND was reduced, we are fast-retransmitting. |
| 2073 | * "Loss" CWND was reduced due to RTO timeout or SACK reneging. |
| 2074 | * |
| 2075 | * tcp_fastretrans_alert() is entered: |
| 2076 | * - each incoming ACK, if state is not "Open" |
| 2077 | * - when arrived ACK is unusual, namely: |
| 2078 | * * SACK |
| 2079 | * * Duplicate ACK. |
| 2080 | * * ECN ECE. |
| 2081 | * |
| 2082 | * Counting packets in flight is pretty simple. |
| 2083 | * |
| 2084 | * in_flight = packets_out - left_out + retrans_out |
| 2085 | * |
| 2086 | * packets_out is SND.NXT-SND.UNA counted in packets. |
| 2087 | * |
| 2088 | * retrans_out is number of retransmitted segments. |
| 2089 | * |
| 2090 | * left_out is number of segments left network, but not ACKed yet. |
| 2091 | * |
| 2092 | * left_out = sacked_out + lost_out |
| 2093 | * |
| 2094 | * sacked_out: Packets, which arrived to receiver out of order |
| 2095 | * and hence not ACKed. With SACKs this number is simply |
| 2096 | * amount of SACKed data. Even without SACKs |
| 2097 | * it is easy to give pretty reliable estimate of this number, |
| 2098 | * counting duplicate ACKs. |
| 2099 | * |
| 2100 | * lost_out: Packets lost by network. TCP has no explicit |
| 2101 | * "loss notification" feedback from network (for now). |
| 2102 | * It means that this number can be only _guessed_. |
| 2103 | * Actually, it is the heuristics to predict lossage that |
| 2104 | * distinguishes different algorithms. |
| 2105 | * |
| 2106 | * F.e. after RTO, when all the queue is considered as lost, |
| 2107 | * lost_out = packets_out and in_flight = retrans_out. |
| 2108 | * |
| 2109 | * Essentially, we have now two algorithms counting |
| 2110 | * lost packets. |
| 2111 | * |
| 2112 | * FACK: It is the simplest heuristics. As soon as we decided |
| 2113 | * that something is lost, we decide that _all_ not SACKed |
| 2114 | * packets until the most forward SACK are lost. I.e. |
| 2115 | * lost_out = fackets_out - sacked_out and left_out = fackets_out. |
| 2116 | * It is absolutely correct estimate, if network does not reorder |
| 2117 | * packets. And it loses any connection to reality when reordering |
| 2118 | * takes place. We use FACK by default until reordering |
| 2119 | * is suspected on the path to this destination. |
| 2120 | * |
| 2121 | * NewReno: when Recovery is entered, we assume that one segment |
| 2122 | * is lost (classic Reno). While we are in Recovery and |
| 2123 | * a partial ACK arrives, we assume that one more packet |
| 2124 | * is lost (NewReno). This heuristics are the same in NewReno |
| 2125 | * and SACK. |
| 2126 | * |
| 2127 | * Imagine, that's all! Forget about all this shamanism about CWND inflation |
| 2128 | * deflation etc. CWND is real congestion window, never inflated, changes |
| 2129 | * only according to classic VJ rules. |
| 2130 | * |
| 2131 | * Really tricky (and requiring careful tuning) part of algorithm |
| 2132 | * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue(). |
| 2133 | * The first determines the moment _when_ we should reduce CWND and, |
| 2134 | * hence, slow down forward transmission. In fact, it determines the moment |
| 2135 | * when we decide that hole is caused by loss, rather than by a reorder. |
| 2136 | * |
| 2137 | * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill |
| 2138 | * holes, caused by lost packets. |
| 2139 | * |
| 2140 | * And the most logically complicated part of algorithm is undo |
| 2141 | * heuristics. We detect false retransmits due to both too early |
| 2142 | * fast retransmit (reordering) and underestimated RTO, analyzing |
| 2143 | * timestamps and D-SACKs. When we detect that some segments were |
| 2144 | * retransmitted by mistake and CWND reduction was wrong, we undo |
| 2145 | * window reduction and abort recovery phase. This logic is hidden |
| 2146 | * inside several functions named tcp_try_undo_<something>. |
| 2147 | */ |
| 2148 | |
| 2149 | /* This function decides, when we should leave Disordered state |
| 2150 | * and enter Recovery phase, reducing congestion window. |
| 2151 | * |
| 2152 | * Main question: may we further continue forward transmission |
| 2153 | * with the same cwnd? |
| 2154 | */ |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2155 | static bool tcp_time_to_recover(struct sock *sk, int flag) |
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); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2158 | __u32 packets_out; |
Nikolay Borisov | 1043e25 | 2016-02-03 09:46:52 +0200 | [diff] [blame] | 2159 | int tcp_reordering = sock_net(sk)->ipv4.sysctl_tcp_reordering; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2160 | |
| 2161 | /* Trick#1: The loss is proven. */ |
| 2162 | if (tp->lost_out) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2163 | return true; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2164 | |
| 2165 | /* Not-A-Trick#2 : Classic rule... */ |
Andreas Petlund | ea84e55 | 2009-10-27 03:27:21 +0000 | [diff] [blame] | 2166 | if (tcp_dupack_heuristics(tp) > tp->reordering) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2167 | return true; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2168 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2169 | /* Trick#4: It is still not OK... But will it be useful to delay |
| 2170 | * recovery more? |
| 2171 | */ |
| 2172 | packets_out = tp->packets_out; |
| 2173 | if (packets_out <= tp->reordering && |
Nikolay Borisov | 1043e25 | 2016-02-03 09:46:52 +0200 | [diff] [blame] | 2174 | tp->sacked_out >= max_t(__u32, packets_out/2, tcp_reordering) && |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2175 | !tcp_may_send_now(sk)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2176 | /* We have nothing to send. This connection is limited |
| 2177 | * either by receiver window or by application. |
| 2178 | */ |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2179 | return true; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2180 | } |
| 2181 | |
Andreas Petlund | 7e38017 | 2010-02-18 04:48:19 +0000 | [diff] [blame] | 2182 | /* If a thin stream is detected, retransmit after first |
| 2183 | * received dupack. Employ only if SACK is supported in order |
| 2184 | * to avoid possible corner-case series of spurious retransmissions |
| 2185 | * Use only if there are no unsent data. |
| 2186 | */ |
| 2187 | if ((tp->thin_dupack || sysctl_tcp_thin_dupack) && |
| 2188 | tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 && |
| 2189 | tcp_is_sack(tp) && !tcp_send_head(sk)) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2190 | return true; |
Andreas Petlund | 7e38017 | 2010-02-18 04:48:19 +0000 | [diff] [blame] | 2191 | |
Yuchung Cheng | eed530b | 2012-05-02 13:30:03 +0000 | [diff] [blame] | 2192 | /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious |
| 2193 | * retransmissions due to small network reorderings, we implement |
| 2194 | * Mitigation A.3 in the RFC and delay the retransmission for a short |
| 2195 | * interval if appropriate. |
| 2196 | */ |
| 2197 | if (tp->do_early_retrans && !tp->retrans_out && tp->sacked_out && |
Nandita Dukkipati | 6ba8a3b | 2013-03-11 10:00:43 +0000 | [diff] [blame] | 2198 | (tp->packets_out >= (tp->sacked_out + 1) && tp->packets_out < 4) && |
Yuchung Cheng | eed530b | 2012-05-02 13:30:03 +0000 | [diff] [blame] | 2199 | !tcp_may_send_now(sk)) |
Yuchung Cheng | 750ea2b | 2012-05-02 13:30:04 +0000 | [diff] [blame] | 2200 | return !tcp_pause_early_retransmit(sk, flag); |
Yuchung Cheng | eed530b | 2012-05-02 13:30:03 +0000 | [diff] [blame] | 2201 | |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2202 | return false; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2203 | } |
| 2204 | |
Yuchung Cheng | 974c123 | 2012-01-19 14:42:21 +0000 | [diff] [blame] | 2205 | /* Detect loss in event "A" above by marking head of queue up as lost. |
| 2206 | * For FACK or non-SACK(Reno) senders, the first "packets" number of segments |
| 2207 | * are considered lost. For RFC3517 SACK, a segment is considered lost if it |
| 2208 | * has at least tp->reordering SACKed seqments above it; "packets" refers to |
| 2209 | * the maximum SACKed segments to pass before reaching this limit. |
Ilpo Järvinen | 85cc391 | 2007-11-15 19:39:31 -0800 | [diff] [blame] | 2210 | */ |
Ilpo Järvinen | 1fdb936 | 2010-10-14 01:42:30 +0000 | [diff] [blame] | 2211 | static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2212 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2213 | struct tcp_sock *tp = tcp_sk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2214 | struct sk_buff *skb; |
Neal Cardwell | d88270e | 2016-01-25 14:01:53 -0800 | [diff] [blame] | 2215 | int cnt, oldcnt, lost; |
Ilpo Järvinen | c137f3d | 2008-04-07 22:32:38 -0700 | [diff] [blame] | 2216 | unsigned int mss; |
Yuchung Cheng | 974c123 | 2012-01-19 14:42:21 +0000 | [diff] [blame] | 2217 | /* Use SACK to deduce losses of new sequences sent during recovery */ |
| 2218 | const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2219 | |
Ilpo Järvinen | 547b792 | 2008-07-25 21:43:18 -0700 | [diff] [blame] | 2220 | WARN_ON(packets > tp->packets_out); |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 2221 | if (tp->lost_skb_hint) { |
| 2222 | skb = tp->lost_skb_hint; |
| 2223 | cnt = tp->lost_cnt_hint; |
Ilpo Järvinen | 1fdb936 | 2010-10-14 01:42:30 +0000 | [diff] [blame] | 2224 | /* Head already handled? */ |
| 2225 | if (mark_head && skb != tcp_write_queue_head(sk)) |
| 2226 | return; |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 2227 | } else { |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2228 | skb = tcp_write_queue_head(sk); |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 2229 | cnt = 0; |
| 2230 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2231 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2232 | tcp_for_write_queue_from(skb, sk) { |
| 2233 | if (skb == tcp_send_head(sk)) |
| 2234 | break; |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 2235 | /* TODO: do this better */ |
| 2236 | /* this is not the most efficient way to do this... */ |
| 2237 | tp->lost_skb_hint = skb; |
| 2238 | tp->lost_cnt_hint = cnt; |
Ilpo Järvinen | 85cc391 | 2007-11-15 19:39:31 -0800 | [diff] [blame] | 2239 | |
Yuchung Cheng | 974c123 | 2012-01-19 14:42:21 +0000 | [diff] [blame] | 2240 | if (after(TCP_SKB_CB(skb)->end_seq, loss_high)) |
Ilpo Järvinen | c137f3d | 2008-04-07 22:32:38 -0700 | [diff] [blame] | 2241 | break; |
| 2242 | |
| 2243 | oldcnt = cnt; |
Ilpo Järvinen | ad1984e | 2008-01-30 20:06:02 -0800 | [diff] [blame] | 2244 | if (tcp_is_fack(tp) || tcp_is_reno(tp) || |
Ilpo Järvinen | 85cc391 | 2007-11-15 19:39:31 -0800 | [diff] [blame] | 2245 | (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) |
| 2246 | cnt += tcp_skb_pcount(skb); |
| 2247 | |
Ilpo Järvinen | c137f3d | 2008-04-07 22:32:38 -0700 | [diff] [blame] | 2248 | if (cnt > packets) { |
Yuchung Cheng | b3de755 | 2010-09-24 13:22:06 +0000 | [diff] [blame] | 2249 | if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) || |
Neal Cardwell | c0638c2 | 2012-03-02 21:36:51 +0000 | [diff] [blame] | 2250 | (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) || |
Yuchung Cheng | b3de755 | 2010-09-24 13:22:06 +0000 | [diff] [blame] | 2251 | (oldcnt >= packets)) |
Ilpo Järvinen | c137f3d | 2008-04-07 22:32:38 -0700 | [diff] [blame] | 2252 | break; |
| 2253 | |
Eric Dumazet | f69ad29 | 2015-06-11 09:15:18 -0700 | [diff] [blame] | 2254 | mss = tcp_skb_mss(skb); |
Neal Cardwell | d88270e | 2016-01-25 14:01:53 -0800 | [diff] [blame] | 2255 | /* If needed, chop off the prefix to mark as lost. */ |
| 2256 | lost = (packets - oldcnt) * mss; |
| 2257 | if (lost < skb->len && |
| 2258 | tcp_fragment(sk, skb, lost, mss, GFP_ATOMIC) < 0) |
Ilpo Järvinen | c137f3d | 2008-04-07 22:32:38 -0700 | [diff] [blame] | 2259 | break; |
| 2260 | cnt = packets; |
| 2261 | } |
| 2262 | |
Ilpo Järvinen | 41ea36e | 2008-09-20 21:19:22 -0700 | [diff] [blame] | 2263 | tcp_skb_mark_lost(tp, skb); |
Ilpo Järvinen | 1fdb936 | 2010-10-14 01:42:30 +0000 | [diff] [blame] | 2264 | |
| 2265 | if (mark_head) |
| 2266 | break; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2267 | } |
Ilpo Järvinen | 005903b | 2007-08-09 14:44:16 +0300 | [diff] [blame] | 2268 | tcp_verify_left_out(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2269 | } |
| 2270 | |
| 2271 | /* Account newly detected lost packet(s) */ |
| 2272 | |
Ilpo Järvinen | 85cc391 | 2007-11-15 19:39:31 -0800 | [diff] [blame] | 2273 | static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2274 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2275 | struct tcp_sock *tp = tcp_sk(sk); |
| 2276 | |
Ilpo Järvinen | 85cc391 | 2007-11-15 19:39:31 -0800 | [diff] [blame] | 2277 | if (tcp_is_reno(tp)) { |
Ilpo Järvinen | 1fdb936 | 2010-10-14 01:42:30 +0000 | [diff] [blame] | 2278 | tcp_mark_head_lost(sk, 1, 1); |
Ilpo Järvinen | 85cc391 | 2007-11-15 19:39:31 -0800 | [diff] [blame] | 2279 | } else if (tcp_is_fack(tp)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2280 | int lost = tp->fackets_out - tp->reordering; |
| 2281 | if (lost <= 0) |
| 2282 | lost = 1; |
Ilpo Järvinen | 1fdb936 | 2010-10-14 01:42:30 +0000 | [diff] [blame] | 2283 | tcp_mark_head_lost(sk, lost, 0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2284 | } else { |
Ilpo Järvinen | 85cc391 | 2007-11-15 19:39:31 -0800 | [diff] [blame] | 2285 | int sacked_upto = tp->sacked_out - tp->reordering; |
Ilpo Järvinen | 1fdb936 | 2010-10-14 01:42:30 +0000 | [diff] [blame] | 2286 | if (sacked_upto >= 0) |
| 2287 | tcp_mark_head_lost(sk, sacked_upto, 0); |
| 2288 | else if (fast_rexmit) |
| 2289 | tcp_mark_head_lost(sk, 1, 1); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2290 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2291 | } |
| 2292 | |
Yuchung Cheng | 77c6312 | 2015-10-16 21:57:44 -0700 | [diff] [blame] | 2293 | static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when) |
| 2294 | { |
| 2295 | return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && |
| 2296 | before(tp->rx_opt.rcv_tsecr, when); |
| 2297 | } |
| 2298 | |
Yuchung Cheng | 659a8ad | 2015-10-16 21:57:46 -0700 | [diff] [blame] | 2299 | /* skb is spurious retransmitted if the returned timestamp echo |
| 2300 | * reply is prior to the skb transmission time |
| 2301 | */ |
| 2302 | static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp, |
| 2303 | const struct sk_buff *skb) |
| 2304 | { |
| 2305 | return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) && |
| 2306 | tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb)); |
| 2307 | } |
| 2308 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2309 | /* Nothing was retransmitted or returned timestamp is less |
| 2310 | * than timestamp of the first retransmission. |
| 2311 | */ |
Vijay Subramanian | 67b95bd | 2012-07-19 21:32:18 +0000 | [diff] [blame] | 2312 | static inline bool tcp_packet_delayed(const struct tcp_sock *tp) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2313 | { |
| 2314 | return !tp->retrans_stamp || |
Yuchung Cheng | 77c6312 | 2015-10-16 21:57:44 -0700 | [diff] [blame] | 2315 | tcp_tsopt_ecr_before(tp, tp->retrans_stamp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2316 | } |
| 2317 | |
| 2318 | /* Undo procedures. */ |
| 2319 | |
Marcelo Leitner | 1f37bf8 | 2014-11-04 17:15:08 -0200 | [diff] [blame] | 2320 | /* We can clear retrans_stamp when there are no retransmissions in the |
| 2321 | * window. It would seem that it is trivially available for us in |
| 2322 | * tp->retrans_out, however, that kind of assumptions doesn't consider |
| 2323 | * what will happen if errors occur when sending retransmission for the |
| 2324 | * second time. ...It could the that such segment has only |
| 2325 | * TCPCB_EVER_RETRANS set at the present time. It seems that checking |
| 2326 | * the head skb is enough except for some reneging corner cases that |
| 2327 | * are not worth the effort. |
| 2328 | * |
| 2329 | * Main reason for all this complexity is the fact that connection dying |
| 2330 | * time now depends on the validity of the retrans_stamp, in particular, |
| 2331 | * that successive retransmissions of a segment must not advance |
| 2332 | * retrans_stamp under any conditions. |
| 2333 | */ |
| 2334 | static bool tcp_any_retrans_done(const struct sock *sk) |
| 2335 | { |
| 2336 | const struct tcp_sock *tp = tcp_sk(sk); |
| 2337 | struct sk_buff *skb; |
| 2338 | |
| 2339 | if (tp->retrans_out) |
| 2340 | return true; |
| 2341 | |
| 2342 | skb = tcp_write_queue_head(sk); |
| 2343 | if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS)) |
| 2344 | return true; |
| 2345 | |
| 2346 | return false; |
| 2347 | } |
| 2348 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2349 | #if FASTRETRANS_DEBUG > 1 |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2350 | static void DBGUNDO(struct sock *sk, const char *msg) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2351 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2352 | struct tcp_sock *tp = tcp_sk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2353 | struct inet_sock *inet = inet_sk(sk); |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2354 | |
YOSHIFUJI Hideaki | 569508c | 2008-04-14 04:09:36 -0700 | [diff] [blame] | 2355 | if (sk->sk_family == AF_INET) { |
Joe Perches | 91df42b | 2012-05-15 14:11:54 +0000 | [diff] [blame] | 2356 | pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n", |
| 2357 | msg, |
| 2358 | &inet->inet_daddr, ntohs(inet->inet_dport), |
| 2359 | tp->snd_cwnd, tcp_left_out(tp), |
| 2360 | tp->snd_ssthresh, tp->prior_ssthresh, |
| 2361 | tp->packets_out); |
YOSHIFUJI Hideaki | 569508c | 2008-04-14 04:09:36 -0700 | [diff] [blame] | 2362 | } |
Eric Dumazet | dfd56b8 | 2011-12-10 09:48:31 +0000 | [diff] [blame] | 2363 | #if IS_ENABLED(CONFIG_IPV6) |
YOSHIFUJI Hideaki | 569508c | 2008-04-14 04:09:36 -0700 | [diff] [blame] | 2364 | else if (sk->sk_family == AF_INET6) { |
Joe Perches | 91df42b | 2012-05-15 14:11:54 +0000 | [diff] [blame] | 2365 | pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n", |
| 2366 | msg, |
Eric Dumazet | 019b1c9 | 2016-09-22 17:54:00 -0700 | [diff] [blame] | 2367 | &sk->sk_v6_daddr, ntohs(inet->inet_dport), |
Joe Perches | 91df42b | 2012-05-15 14:11:54 +0000 | [diff] [blame] | 2368 | tp->snd_cwnd, tcp_left_out(tp), |
| 2369 | tp->snd_ssthresh, tp->prior_ssthresh, |
| 2370 | tp->packets_out); |
YOSHIFUJI Hideaki | 569508c | 2008-04-14 04:09:36 -0700 | [diff] [blame] | 2371 | } |
| 2372 | #endif |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2373 | } |
| 2374 | #else |
| 2375 | #define DBGUNDO(x...) do { } while (0) |
| 2376 | #endif |
| 2377 | |
Yuchung Cheng | 7026b91 | 2013-05-29 14:20:13 +0000 | [diff] [blame] | 2378 | static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2379 | { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2380 | struct tcp_sock *tp = tcp_sk(sk); |
| 2381 | |
Yuchung Cheng | 6a63df4 | 2013-05-29 14:20:12 +0000 | [diff] [blame] | 2382 | if (unmark_loss) { |
| 2383 | struct sk_buff *skb; |
| 2384 | |
| 2385 | tcp_for_write_queue(skb, sk) { |
| 2386 | if (skb == tcp_send_head(sk)) |
| 2387 | break; |
| 2388 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; |
| 2389 | } |
| 2390 | tp->lost_out = 0; |
| 2391 | tcp_clear_all_retrans_hints(tp); |
| 2392 | } |
| 2393 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2394 | if (tp->prior_ssthresh) { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2395 | const struct inet_connection_sock *icsk = inet_csk(sk); |
| 2396 | |
Florian Westphal | e979918 | 2016-11-21 14:18:38 +0100 | [diff] [blame] | 2397 | tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2398 | |
Yuchung Cheng | 7026b91 | 2013-05-29 14:20:13 +0000 | [diff] [blame] | 2399 | if (tp->prior_ssthresh > tp->snd_ssthresh) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2400 | tp->snd_ssthresh = tp->prior_ssthresh; |
Florian Westphal | 735d383 | 2014-09-29 13:08:30 +0200 | [diff] [blame] | 2401 | tcp_ecn_withdraw_cwr(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2402 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2403 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2404 | tp->snd_cwnd_stamp = tcp_time_stamp; |
Yuchung Cheng | 7026b91 | 2013-05-29 14:20:13 +0000 | [diff] [blame] | 2405 | tp->undo_marker = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2406 | } |
| 2407 | |
Vijay Subramanian | 67b95bd | 2012-07-19 21:32:18 +0000 | [diff] [blame] | 2408 | static inline bool tcp_may_undo(const struct tcp_sock *tp) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2409 | { |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2410 | return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2411 | } |
| 2412 | |
| 2413 | /* People celebrate: "We love our President!" */ |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2414 | static bool tcp_try_undo_recovery(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2415 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2416 | struct tcp_sock *tp = tcp_sk(sk); |
| 2417 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2418 | if (tcp_may_undo(tp)) { |
Pavel Emelyanov | 40b215e | 2008-07-03 01:05:41 -0700 | [diff] [blame] | 2419 | int mib_idx; |
| 2420 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2421 | /* Happy end! We did not retransmit anything |
| 2422 | * or our original transmission succeeded. |
| 2423 | */ |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2424 | DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans"); |
Yuchung Cheng | 7026b91 | 2013-05-29 14:20:13 +0000 | [diff] [blame] | 2425 | tcp_undo_cwnd_reduction(sk, false); |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2426 | if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss) |
Pavel Emelyanov | 40b215e | 2008-07-03 01:05:41 -0700 | [diff] [blame] | 2427 | mib_idx = LINUX_MIB_TCPLOSSUNDO; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2428 | else |
Pavel Emelyanov | 40b215e | 2008-07-03 01:05:41 -0700 | [diff] [blame] | 2429 | mib_idx = LINUX_MIB_TCPFULLUNDO; |
| 2430 | |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 2431 | NET_INC_STATS(sock_net(sk), mib_idx); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2432 | } |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 2433 | if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2434 | /* Hold old state until something *above* high_seq |
| 2435 | * is ACKed. For Reno it is MUST to prevent false |
| 2436 | * fast retransmits (RFC2582). SACK TCP is safe. */ |
Marcelo Leitner | 1f37bf8 | 2014-11-04 17:15:08 -0200 | [diff] [blame] | 2437 | if (!tcp_any_retrans_done(sk)) |
| 2438 | tp->retrans_stamp = 0; |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2439 | return true; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2440 | } |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2441 | tcp_set_ca_state(sk, TCP_CA_Open); |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2442 | return false; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2443 | } |
| 2444 | |
| 2445 | /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */ |
Yuchung Cheng | c7d9d6a | 2013-05-29 14:20:14 +0000 | [diff] [blame] | 2446 | static bool tcp_try_undo_dsack(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2447 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2448 | struct tcp_sock *tp = tcp_sk(sk); |
| 2449 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2450 | if (tp->undo_marker && !tp->undo_retrans) { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2451 | DBGUNDO(sk, "D-SACK"); |
Yuchung Cheng | 7026b91 | 2013-05-29 14:20:13 +0000 | [diff] [blame] | 2452 | tcp_undo_cwnd_reduction(sk, false); |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 2453 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO); |
Yuchung Cheng | c7d9d6a | 2013-05-29 14:20:14 +0000 | [diff] [blame] | 2454 | return true; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2455 | } |
Yuchung Cheng | c7d9d6a | 2013-05-29 14:20:14 +0000 | [diff] [blame] | 2456 | return false; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2457 | } |
| 2458 | |
Yuchung Cheng | e33099f | 2013-03-20 13:33:00 +0000 | [diff] [blame] | 2459 | /* Undo during loss recovery after partial ACK or using F-RTO. */ |
| 2460 | static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2461 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2462 | struct tcp_sock *tp = tcp_sk(sk); |
| 2463 | |
Yuchung Cheng | e33099f | 2013-03-20 13:33:00 +0000 | [diff] [blame] | 2464 | if (frto_undo || tcp_may_undo(tp)) { |
Yuchung Cheng | 7026b91 | 2013-05-29 14:20:13 +0000 | [diff] [blame] | 2465 | tcp_undo_cwnd_reduction(sk, true); |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 2466 | |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2467 | DBGUNDO(sk, "partial loss"); |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 2468 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO); |
Yuchung Cheng | e33099f | 2013-03-20 13:33:00 +0000 | [diff] [blame] | 2469 | if (frto_undo) |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 2470 | NET_INC_STATS(sock_net(sk), |
Eric Dumazet | 02a1d6e | 2016-04-27 16:44:39 -0700 | [diff] [blame] | 2471 | LINUX_MIB_TCPSPURIOUSRTOS); |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2472 | inet_csk(sk)->icsk_retransmits = 0; |
Yuchung Cheng | e33099f | 2013-03-20 13:33:00 +0000 | [diff] [blame] | 2473 | if (frto_undo || tcp_is_sack(tp)) |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2474 | tcp_set_ca_state(sk, TCP_CA_Open); |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2475 | return true; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2476 | } |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2477 | return false; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2478 | } |
| 2479 | |
Yuchung Cheng | 3759824 | 2015-07-01 14:11:15 -0700 | [diff] [blame] | 2480 | /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937. |
Yuchung Cheng | fb4d3d1 | 2012-09-02 17:38:03 +0000 | [diff] [blame] | 2481 | * It computes the number of packets to send (sndcnt) based on packets newly |
| 2482 | * delivered: |
| 2483 | * 1) If the packets in flight is larger than ssthresh, PRR spreads the |
| 2484 | * cwnd reductions across a full RTT. |
Yuchung Cheng | 3759824 | 2015-07-01 14:11:15 -0700 | [diff] [blame] | 2485 | * 2) Otherwise PRR uses packet conservation to send as much as delivered. |
| 2486 | * But when the retransmits are acked without further losses, PRR |
| 2487 | * slow starts cwnd up to ssthresh to speed up the recovery. |
Yuchung Cheng | fb4d3d1 | 2012-09-02 17:38:03 +0000 | [diff] [blame] | 2488 | */ |
Christoph Paasch | 5ee2c94 | 2014-07-14 16:58:32 +0200 | [diff] [blame] | 2489 | static void tcp_init_cwnd_reduction(struct sock *sk) |
Yuchung Cheng | 684bad1 | 2012-09-02 17:38:04 +0000 | [diff] [blame] | 2490 | { |
| 2491 | struct tcp_sock *tp = tcp_sk(sk); |
| 2492 | |
| 2493 | tp->high_seq = tp->snd_nxt; |
Nandita Dukkipati | 9b717a8 | 2013-03-11 10:00:44 +0000 | [diff] [blame] | 2494 | tp->tlp_high_seq = 0; |
Yuchung Cheng | 684bad1 | 2012-09-02 17:38:04 +0000 | [diff] [blame] | 2495 | tp->snd_cwnd_cnt = 0; |
| 2496 | tp->prior_cwnd = tp->snd_cwnd; |
| 2497 | tp->prr_delivered = 0; |
| 2498 | tp->prr_out = 0; |
Christoph Paasch | 5ee2c94 | 2014-07-14 16:58:32 +0200 | [diff] [blame] | 2499 | tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk); |
Florian Westphal | 735d383 | 2014-09-29 13:08:30 +0200 | [diff] [blame] | 2500 | tcp_ecn_queue_cwr(tp); |
Yuchung Cheng | 684bad1 | 2012-09-02 17:38:04 +0000 | [diff] [blame] | 2501 | } |
| 2502 | |
Yuchung Cheng | 31ba0c1 | 2016-02-02 10:33:05 -0800 | [diff] [blame] | 2503 | static void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, |
| 2504 | int flag) |
Yuchung Cheng | fb4d3d1 | 2012-09-02 17:38:03 +0000 | [diff] [blame] | 2505 | { |
| 2506 | struct tcp_sock *tp = tcp_sk(sk); |
| 2507 | int sndcnt = 0; |
| 2508 | int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp); |
| 2509 | |
Yuchung Cheng | 8b8a321 | 2016-01-06 12:42:38 -0800 | [diff] [blame] | 2510 | if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd)) |
| 2511 | return; |
| 2512 | |
Yuchung Cheng | 684bad1 | 2012-09-02 17:38:04 +0000 | [diff] [blame] | 2513 | tp->prr_delivered += newly_acked_sacked; |
Yuchung Cheng | 3759824 | 2015-07-01 14:11:15 -0700 | [diff] [blame] | 2514 | if (delta < 0) { |
Yuchung Cheng | fb4d3d1 | 2012-09-02 17:38:03 +0000 | [diff] [blame] | 2515 | u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered + |
| 2516 | tp->prior_cwnd - 1; |
| 2517 | sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out; |
Yuchung Cheng | 3759824 | 2015-07-01 14:11:15 -0700 | [diff] [blame] | 2518 | } else if ((flag & FLAG_RETRANS_DATA_ACKED) && |
| 2519 | !(flag & FLAG_LOST_RETRANS)) { |
Yuchung Cheng | fb4d3d1 | 2012-09-02 17:38:03 +0000 | [diff] [blame] | 2520 | sndcnt = min_t(int, delta, |
| 2521 | max_t(int, tp->prr_delivered - tp->prr_out, |
| 2522 | newly_acked_sacked) + 1); |
Yuchung Cheng | 3759824 | 2015-07-01 14:11:15 -0700 | [diff] [blame] | 2523 | } else { |
| 2524 | sndcnt = min(delta, newly_acked_sacked); |
Yuchung Cheng | fb4d3d1 | 2012-09-02 17:38:03 +0000 | [diff] [blame] | 2525 | } |
Yuchung Cheng | 31ba0c1 | 2016-02-02 10:33:05 -0800 | [diff] [blame] | 2526 | /* Force a fast retransmit upon entering fast recovery */ |
| 2527 | sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1)); |
Yuchung Cheng | fb4d3d1 | 2012-09-02 17:38:03 +0000 | [diff] [blame] | 2528 | tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt; |
| 2529 | } |
| 2530 | |
Yuchung Cheng | 684bad1 | 2012-09-02 17:38:04 +0000 | [diff] [blame] | 2531 | static inline void tcp_end_cwnd_reduction(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2532 | { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2533 | struct tcp_sock *tp = tcp_sk(sk); |
Nandita Dukkipati | a262f0c | 2011-08-21 20:21:57 +0000 | [diff] [blame] | 2534 | |
Yuchung Cheng | c040276 | 2016-09-19 23:39:21 -0400 | [diff] [blame] | 2535 | if (inet_csk(sk)->icsk_ca_ops->cong_control) |
| 2536 | return; |
| 2537 | |
Yuchung Cheng | 684bad1 | 2012-09-02 17:38:04 +0000 | [diff] [blame] | 2538 | /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */ |
| 2539 | if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || |
| 2540 | (tp->undo_marker && tp->snd_ssthresh < TCP_INFINITE_SSTHRESH)) { |
| 2541 | tp->snd_cwnd = tp->snd_ssthresh; |
| 2542 | tp->snd_cwnd_stamp = tcp_time_stamp; |
Yuchung Cheng | 67d4120 | 2011-03-14 10:57:03 +0000 | [diff] [blame] | 2543 | } |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2544 | tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2545 | } |
| 2546 | |
Yuchung Cheng | 684bad1 | 2012-09-02 17:38:04 +0000 | [diff] [blame] | 2547 | /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */ |
Christoph Paasch | 5ee2c94 | 2014-07-14 16:58:32 +0200 | [diff] [blame] | 2548 | void tcp_enter_cwr(struct sock *sk) |
Yuchung Cheng | 09484d1 | 2012-09-02 17:38:02 +0000 | [diff] [blame] | 2549 | { |
| 2550 | struct tcp_sock *tp = tcp_sk(sk); |
Yuchung Cheng | 09484d1 | 2012-09-02 17:38:02 +0000 | [diff] [blame] | 2551 | |
| 2552 | tp->prior_ssthresh = 0; |
Yuchung Cheng | 684bad1 | 2012-09-02 17:38:04 +0000 | [diff] [blame] | 2553 | if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) { |
Yuchung Cheng | 09484d1 | 2012-09-02 17:38:02 +0000 | [diff] [blame] | 2554 | tp->undo_marker = 0; |
Christoph Paasch | 5ee2c94 | 2014-07-14 16:58:32 +0200 | [diff] [blame] | 2555 | tcp_init_cwnd_reduction(sk); |
Yuchung Cheng | 09484d1 | 2012-09-02 17:38:02 +0000 | [diff] [blame] | 2556 | tcp_set_ca_state(sk, TCP_CA_CWR); |
| 2557 | } |
| 2558 | } |
Kenneth Klette Jonassen | 7782ad8 | 2015-06-10 19:08:16 +0200 | [diff] [blame] | 2559 | EXPORT_SYMBOL(tcp_enter_cwr); |
Yuchung Cheng | 09484d1 | 2012-09-02 17:38:02 +0000 | [diff] [blame] | 2560 | |
Ilpo Järvinen | 8aca6cb | 2008-06-04 11:34:22 -0700 | [diff] [blame] | 2561 | static void tcp_try_keep_open(struct sock *sk) |
| 2562 | { |
| 2563 | struct tcp_sock *tp = tcp_sk(sk); |
| 2564 | int state = TCP_CA_Open; |
| 2565 | |
Neal Cardwell | f698204 | 2011-11-16 08:58:04 +0000 | [diff] [blame] | 2566 | if (tcp_left_out(tp) || tcp_any_retrans_done(sk)) |
Ilpo Järvinen | 8aca6cb | 2008-06-04 11:34:22 -0700 | [diff] [blame] | 2567 | state = TCP_CA_Disorder; |
| 2568 | |
| 2569 | if (inet_csk(sk)->icsk_ca_state != state) { |
| 2570 | tcp_set_ca_state(sk, state); |
| 2571 | tp->high_seq = tp->snd_nxt; |
| 2572 | } |
| 2573 | } |
| 2574 | |
Yuchung Cheng | 31ba0c1 | 2016-02-02 10:33:05 -0800 | [diff] [blame] | 2575 | static void tcp_try_to_open(struct sock *sk, int flag) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2576 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2577 | struct tcp_sock *tp = tcp_sk(sk); |
| 2578 | |
Ilpo Järvinen | 86426c2 | 2007-08-09 14:45:17 +0300 | [diff] [blame] | 2579 | tcp_verify_left_out(tp); |
| 2580 | |
Yuchung Cheng | 9b44190 | 2013-03-20 13:32:58 +0000 | [diff] [blame] | 2581 | if (!tcp_any_retrans_done(sk)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2582 | tp->retrans_stamp = 0; |
| 2583 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2584 | if (flag & FLAG_ECE) |
Christoph Paasch | 5ee2c94 | 2014-07-14 16:58:32 +0200 | [diff] [blame] | 2585 | tcp_enter_cwr(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2586 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2587 | if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) { |
Ilpo Järvinen | 8aca6cb | 2008-06-04 11:34:22 -0700 | [diff] [blame] | 2588 | tcp_try_keep_open(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2589 | } |
| 2590 | } |
| 2591 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 2592 | static void tcp_mtup_probe_failed(struct sock *sk) |
| 2593 | { |
| 2594 | struct inet_connection_sock *icsk = inet_csk(sk); |
| 2595 | |
| 2596 | icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1; |
| 2597 | icsk->icsk_mtup.probe_size = 0; |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 2598 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 2599 | } |
| 2600 | |
Ilpo Järvinen | 72211e9 | 2009-03-14 14:23:04 +0000 | [diff] [blame] | 2601 | static void tcp_mtup_probe_success(struct sock *sk) |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 2602 | { |
| 2603 | struct tcp_sock *tp = tcp_sk(sk); |
| 2604 | struct inet_connection_sock *icsk = inet_csk(sk); |
| 2605 | |
| 2606 | /* FIXME: breaks with very large cwnd */ |
| 2607 | tp->prior_ssthresh = tcp_current_ssthresh(sk); |
| 2608 | tp->snd_cwnd = tp->snd_cwnd * |
| 2609 | tcp_mss_to_mtu(sk, tp->mss_cache) / |
| 2610 | icsk->icsk_mtup.probe_size; |
| 2611 | tp->snd_cwnd_cnt = 0; |
| 2612 | tp->snd_cwnd_stamp = tcp_time_stamp; |
John Heffner | 9c6d5e5 | 2010-10-06 21:18:02 -0700 | [diff] [blame] | 2613 | tp->snd_ssthresh = tcp_current_ssthresh(sk); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 2614 | |
| 2615 | icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size; |
| 2616 | icsk->icsk_mtup.probe_size = 0; |
| 2617 | tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 2618 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 2619 | } |
| 2620 | |
Ilpo Järvinen | e1aa680 | 2008-11-24 21:11:55 -0800 | [diff] [blame] | 2621 | /* Do a simple retransmit without using the backoff mechanisms in |
| 2622 | * tcp_timer. This is used for path mtu discovery. |
| 2623 | * The socket is already locked here. |
| 2624 | */ |
| 2625 | void tcp_simple_retransmit(struct sock *sk) |
| 2626 | { |
| 2627 | const struct inet_connection_sock *icsk = inet_csk(sk); |
| 2628 | struct tcp_sock *tp = tcp_sk(sk); |
| 2629 | struct sk_buff *skb; |
Ilpo Järvinen | 0c54b85 | 2009-03-14 14:23:05 +0000 | [diff] [blame] | 2630 | unsigned int mss = tcp_current_mss(sk); |
Ilpo Järvinen | e1aa680 | 2008-11-24 21:11:55 -0800 | [diff] [blame] | 2631 | u32 prior_lost = tp->lost_out; |
| 2632 | |
| 2633 | tcp_for_write_queue(skb, sk) { |
| 2634 | if (skb == tcp_send_head(sk)) |
| 2635 | break; |
Ilpo Järvinen | 775ffab | 2008-12-05 22:41:26 -0800 | [diff] [blame] | 2636 | if (tcp_skb_seglen(skb) > mss && |
Ilpo Järvinen | e1aa680 | 2008-11-24 21:11:55 -0800 | [diff] [blame] | 2637 | !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) { |
| 2638 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { |
| 2639 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; |
| 2640 | tp->retrans_out -= tcp_skb_pcount(skb); |
| 2641 | } |
| 2642 | tcp_skb_mark_lost_uncond_verify(tp, skb); |
| 2643 | } |
| 2644 | } |
| 2645 | |
| 2646 | tcp_clear_retrans_hints_partial(tp); |
| 2647 | |
| 2648 | if (prior_lost == tp->lost_out) |
| 2649 | return; |
| 2650 | |
| 2651 | if (tcp_is_reno(tp)) |
| 2652 | tcp_limit_reno_sacked(tp); |
| 2653 | |
| 2654 | tcp_verify_left_out(tp); |
| 2655 | |
| 2656 | /* Don't muck with the congestion window here. |
| 2657 | * Reason is that we do not increase amount of _data_ |
| 2658 | * in network, but units changed and effective |
| 2659 | * cwnd/ssthresh really reduced now. |
| 2660 | */ |
| 2661 | if (icsk->icsk_ca_state != TCP_CA_Loss) { |
| 2662 | tp->high_seq = tp->snd_nxt; |
| 2663 | tp->snd_ssthresh = tcp_current_ssthresh(sk); |
| 2664 | tp->prior_ssthresh = 0; |
| 2665 | tp->undo_marker = 0; |
| 2666 | tcp_set_ca_state(sk, TCP_CA_Loss); |
| 2667 | } |
| 2668 | tcp_xmit_retransmit_queue(sk); |
| 2669 | } |
Eric Dumazet | 4bc2f18 | 2010-07-09 21:22:10 +0000 | [diff] [blame] | 2670 | EXPORT_SYMBOL(tcp_simple_retransmit); |
Ilpo Järvinen | e1aa680 | 2008-11-24 21:11:55 -0800 | [diff] [blame] | 2671 | |
Yuchung Cheng | 1fbc340 | 2012-05-02 13:30:02 +0000 | [diff] [blame] | 2672 | static void tcp_enter_recovery(struct sock *sk, bool ece_ack) |
| 2673 | { |
| 2674 | struct tcp_sock *tp = tcp_sk(sk); |
| 2675 | int mib_idx; |
| 2676 | |
| 2677 | if (tcp_is_reno(tp)) |
| 2678 | mib_idx = LINUX_MIB_TCPRENORECOVERY; |
| 2679 | else |
| 2680 | mib_idx = LINUX_MIB_TCPSACKRECOVERY; |
| 2681 | |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 2682 | NET_INC_STATS(sock_net(sk), mib_idx); |
Yuchung Cheng | 1fbc340 | 2012-05-02 13:30:02 +0000 | [diff] [blame] | 2683 | |
Yuchung Cheng | 1fbc340 | 2012-05-02 13:30:02 +0000 | [diff] [blame] | 2684 | tp->prior_ssthresh = 0; |
Yuchung Cheng | 989e04c | 2014-08-22 14:15:22 -0700 | [diff] [blame] | 2685 | tcp_init_undo(tp); |
Yuchung Cheng | 1fbc340 | 2012-05-02 13:30:02 +0000 | [diff] [blame] | 2686 | |
Yuchung Cheng | 291a00d | 2015-07-01 14:11:14 -0700 | [diff] [blame] | 2687 | if (!tcp_in_cwnd_reduction(sk)) { |
Yuchung Cheng | 1fbc340 | 2012-05-02 13:30:02 +0000 | [diff] [blame] | 2688 | if (!ece_ack) |
| 2689 | tp->prior_ssthresh = tcp_current_ssthresh(sk); |
Christoph Paasch | 5ee2c94 | 2014-07-14 16:58:32 +0200 | [diff] [blame] | 2690 | tcp_init_cwnd_reduction(sk); |
Yuchung Cheng | 1fbc340 | 2012-05-02 13:30:02 +0000 | [diff] [blame] | 2691 | } |
Yuchung Cheng | 1fbc340 | 2012-05-02 13:30:02 +0000 | [diff] [blame] | 2692 | tcp_set_ca_state(sk, TCP_CA_Recovery); |
| 2693 | } |
| 2694 | |
Yuchung Cheng | ab42d9e | 2013-03-20 13:32:59 +0000 | [diff] [blame] | 2695 | /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are |
| 2696 | * recovered or spurious. Otherwise retransmits more on partial ACKs. |
| 2697 | */ |
Yuchung Cheng | e662ca4 | 2016-02-02 10:33:04 -0800 | [diff] [blame] | 2698 | static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack, |
| 2699 | int *rexmit) |
Yuchung Cheng | ab42d9e | 2013-03-20 13:32:59 +0000 | [diff] [blame] | 2700 | { |
Yuchung Cheng | ab42d9e | 2013-03-20 13:32:59 +0000 | [diff] [blame] | 2701 | struct tcp_sock *tp = tcp_sk(sk); |
Yuchung Cheng | e33099f | 2013-03-20 13:33:00 +0000 | [diff] [blame] | 2702 | bool recovered = !before(tp->snd_una, tp->high_seq); |
Yuchung Cheng | ab42d9e | 2013-03-20 13:32:59 +0000 | [diff] [blame] | 2703 | |
Yuchung Cheng | da34ac7 | 2015-05-18 12:31:44 -0700 | [diff] [blame] | 2704 | if ((flag & FLAG_SND_UNA_ADVANCED) && |
| 2705 | tcp_try_undo_loss(sk, false)) |
| 2706 | return; |
| 2707 | |
Yuchung Cheng | e33099f | 2013-03-20 13:33:00 +0000 | [diff] [blame] | 2708 | if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */ |
Yuchung Cheng | 0cfa5c0 | 2014-05-30 15:25:59 -0700 | [diff] [blame] | 2709 | /* Step 3.b. A timeout is spurious if not all data are |
| 2710 | * lost, i.e., never-retransmitted data are (s)acked. |
| 2711 | */ |
Yuchung Cheng | da34ac7 | 2015-05-18 12:31:44 -0700 | [diff] [blame] | 2712 | if ((flag & FLAG_ORIG_SACK_ACKED) && |
| 2713 | tcp_try_undo_loss(sk, true)) |
Yuchung Cheng | e33099f | 2013-03-20 13:33:00 +0000 | [diff] [blame] | 2714 | return; |
Yuchung Cheng | 0cfa5c0 | 2014-05-30 15:25:59 -0700 | [diff] [blame] | 2715 | |
Yuchung Cheng | b7b0ed9 | 2015-05-18 12:31:45 -0700 | [diff] [blame] | 2716 | if (after(tp->snd_nxt, tp->high_seq)) { |
| 2717 | if (flag & FLAG_DATA_SACKED || is_dupack) |
| 2718 | tp->frto = 0; /* Step 3.a. loss was real */ |
Yuchung Cheng | e33099f | 2013-03-20 13:33:00 +0000 | [diff] [blame] | 2719 | } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) { |
| 2720 | tp->high_seq = tp->snd_nxt; |
Yuchung Cheng | e662ca4 | 2016-02-02 10:33:04 -0800 | [diff] [blame] | 2721 | /* Step 2.b. Try send new data (but deferred until cwnd |
| 2722 | * is updated in tcp_ack()). Otherwise fall back to |
| 2723 | * the conventional recovery. |
| 2724 | */ |
| 2725 | if (tcp_send_head(sk) && |
| 2726 | after(tcp_wnd_end(tp), tp->snd_nxt)) { |
| 2727 | *rexmit = REXMIT_NEW; |
| 2728 | return; |
| 2729 | } |
Yuchung Cheng | e33099f | 2013-03-20 13:33:00 +0000 | [diff] [blame] | 2730 | tp->frto = 0; |
| 2731 | } |
| 2732 | } |
| 2733 | |
| 2734 | if (recovered) { |
| 2735 | /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */ |
Yuchung Cheng | ab42d9e | 2013-03-20 13:32:59 +0000 | [diff] [blame] | 2736 | tcp_try_undo_recovery(sk); |
| 2737 | return; |
| 2738 | } |
Yuchung Cheng | e33099f | 2013-03-20 13:33:00 +0000 | [diff] [blame] | 2739 | if (tcp_is_reno(tp)) { |
| 2740 | /* A Reno DUPACK means new data in F-RTO step 2.b above are |
| 2741 | * delivered. Lower inflight to clock out (re)tranmissions. |
| 2742 | */ |
| 2743 | if (after(tp->snd_nxt, tp->high_seq) && is_dupack) |
| 2744 | tcp_add_reno_sack(sk); |
| 2745 | else if (flag & FLAG_SND_UNA_ADVANCED) |
| 2746 | tcp_reset_reno_sack(tp); |
| 2747 | } |
Yuchung Cheng | e662ca4 | 2016-02-02 10:33:04 -0800 | [diff] [blame] | 2748 | *rexmit = REXMIT_LOST; |
Yuchung Cheng | ab42d9e | 2013-03-20 13:32:59 +0000 | [diff] [blame] | 2749 | } |
| 2750 | |
Yuchung Cheng | 6a63df4 | 2013-05-29 14:20:12 +0000 | [diff] [blame] | 2751 | /* Undo during fast recovery after partial ACK. */ |
Yuchung Cheng | 31ba0c1 | 2016-02-02 10:33:05 -0800 | [diff] [blame] | 2752 | static bool tcp_try_undo_partial(struct sock *sk, const int acked) |
Yuchung Cheng | 6a63df4 | 2013-05-29 14:20:12 +0000 | [diff] [blame] | 2753 | { |
| 2754 | struct tcp_sock *tp = tcp_sk(sk); |
Yuchung Cheng | 6a63df4 | 2013-05-29 14:20:12 +0000 | [diff] [blame] | 2755 | |
Yuchung Cheng | 7026b91 | 2013-05-29 14:20:13 +0000 | [diff] [blame] | 2756 | if (tp->undo_marker && tcp_packet_delayed(tp)) { |
Yuchung Cheng | 6a63df4 | 2013-05-29 14:20:12 +0000 | [diff] [blame] | 2757 | /* Plain luck! Hole if filled with delayed |
| 2758 | * packet, rather than with a retransmit. |
| 2759 | */ |
Yuchung Cheng | 7026b91 | 2013-05-29 14:20:13 +0000 | [diff] [blame] | 2760 | tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1); |
| 2761 | |
| 2762 | /* We are getting evidence that the reordering degree is higher |
| 2763 | * than we realized. If there are no retransmits out then we |
| 2764 | * can undo. Otherwise we clock out new packets but do not |
| 2765 | * mark more packets lost or retransmit more. |
| 2766 | */ |
Yuchung Cheng | 31ba0c1 | 2016-02-02 10:33:05 -0800 | [diff] [blame] | 2767 | if (tp->retrans_out) |
Yuchung Cheng | 7026b91 | 2013-05-29 14:20:13 +0000 | [diff] [blame] | 2768 | return true; |
Yuchung Cheng | 7026b91 | 2013-05-29 14:20:13 +0000 | [diff] [blame] | 2769 | |
Yuchung Cheng | 6a63df4 | 2013-05-29 14:20:12 +0000 | [diff] [blame] | 2770 | if (!tcp_any_retrans_done(sk)) |
| 2771 | tp->retrans_stamp = 0; |
| 2772 | |
Yuchung Cheng | 7026b91 | 2013-05-29 14:20:13 +0000 | [diff] [blame] | 2773 | DBGUNDO(sk, "partial recovery"); |
| 2774 | tcp_undo_cwnd_reduction(sk, true); |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 2775 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO); |
Yuchung Cheng | 7026b91 | 2013-05-29 14:20:13 +0000 | [diff] [blame] | 2776 | tcp_try_keep_open(sk); |
| 2777 | return true; |
Yuchung Cheng | 6a63df4 | 2013-05-29 14:20:12 +0000 | [diff] [blame] | 2778 | } |
Yuchung Cheng | 7026b91 | 2013-05-29 14:20:13 +0000 | [diff] [blame] | 2779 | return false; |
Yuchung Cheng | 6a63df4 | 2013-05-29 14:20:12 +0000 | [diff] [blame] | 2780 | } |
| 2781 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2782 | /* Process an event, which can update packets-in-flight not trivially. |
| 2783 | * Main goal of this function is to calculate new estimate for left_out, |
| 2784 | * taking into account both packets sitting in receiver's buffer and |
| 2785 | * packets lost by network. |
| 2786 | * |
Yuchung Cheng | 31ba0c1 | 2016-02-02 10:33:05 -0800 | [diff] [blame] | 2787 | * Besides that it updates the congestion state when packet loss or ECN |
| 2788 | * is detected. But it does not reduce the cwnd, it is done by the |
| 2789 | * congestion control later. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2790 | * |
| 2791 | * It does _not_ decide what to send, it is made in function |
| 2792 | * tcp_xmit_retransmit_queue(). |
| 2793 | */ |
Yuchung Cheng | 6804973 | 2013-05-29 14:20:11 +0000 | [diff] [blame] | 2794 | static void tcp_fastretrans_alert(struct sock *sk, const int acked, |
Yuchung Cheng | 31ba0c1 | 2016-02-02 10:33:05 -0800 | [diff] [blame] | 2795 | bool is_dupack, int *ack_flag, int *rexmit) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2796 | { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2797 | struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2798 | struct tcp_sock *tp = tcp_sk(sk); |
Yuchung Cheng | 31ba0c1 | 2016-02-02 10:33:05 -0800 | [diff] [blame] | 2799 | int fast_rexmit = 0, flag = *ack_flag; |
Yuchung Cheng | 6a63df4 | 2013-05-29 14:20:12 +0000 | [diff] [blame] | 2800 | bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) && |
Ilpo Järvinen | 85cc391 | 2007-11-15 19:39:31 -0800 | [diff] [blame] | 2801 | (tcp_fackets_out(tp) > tp->reordering)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2802 | |
Ilpo Järvinen | 3ccd313 | 2007-12-31 04:43:32 -0800 | [diff] [blame] | 2803 | if (WARN_ON(!tp->packets_out && tp->sacked_out)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2804 | tp->sacked_out = 0; |
Ilpo Järvinen | 91fed7a | 2007-10-09 01:24:15 -0700 | [diff] [blame] | 2805 | if (WARN_ON(!tp->sacked_out && tp->fackets_out)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2806 | tp->fackets_out = 0; |
| 2807 | |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2808 | /* Now state machine starts. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2809 | * A. ECE, hence prohibit cwnd undoing, the reduction is required. */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2810 | if (flag & FLAG_ECE) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2811 | tp->prior_ssthresh = 0; |
| 2812 | |
| 2813 | /* B. In all the states check for reneging SACKs. */ |
Ilpo Järvinen | cadbd03 | 2007-12-31 04:49:21 -0800 | [diff] [blame] | 2814 | if (tcp_check_sack_reneging(sk, flag)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2815 | return; |
| 2816 | |
Yuchung Cheng | 974c123 | 2012-01-19 14:42:21 +0000 | [diff] [blame] | 2817 | /* C. Check consistency of the current state. */ |
Ilpo Järvinen | 005903b | 2007-08-09 14:44:16 +0300 | [diff] [blame] | 2818 | tcp_verify_left_out(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2819 | |
Yuchung Cheng | 974c123 | 2012-01-19 14:42:21 +0000 | [diff] [blame] | 2820 | /* D. Check state exit conditions. State can be terminated |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2821 | * when high_seq is ACKed. */ |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2822 | if (icsk->icsk_ca_state == TCP_CA_Open) { |
Ilpo Järvinen | 547b792 | 2008-07-25 21:43:18 -0700 | [diff] [blame] | 2823 | WARN_ON(tp->retrans_out != 0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2824 | tp->retrans_stamp = 0; |
| 2825 | } else if (!before(tp->snd_una, tp->high_seq)) { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2826 | switch (icsk->icsk_ca_state) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2827 | case TCP_CA_CWR: |
| 2828 | /* CWR is to be held something *above* high_seq |
| 2829 | * is ACKed for CWR bit to reach receiver. */ |
| 2830 | if (tp->snd_una != tp->high_seq) { |
Yuchung Cheng | 684bad1 | 2012-09-02 17:38:04 +0000 | [diff] [blame] | 2831 | tcp_end_cwnd_reduction(sk); |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2832 | tcp_set_ca_state(sk, TCP_CA_Open); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2833 | } |
| 2834 | break; |
| 2835 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2836 | case TCP_CA_Recovery: |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 2837 | if (tcp_is_reno(tp)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2838 | tcp_reset_reno_sack(tp); |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2839 | if (tcp_try_undo_recovery(sk)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2840 | return; |
Yuchung Cheng | 684bad1 | 2012-09-02 17:38:04 +0000 | [diff] [blame] | 2841 | tcp_end_cwnd_reduction(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2842 | break; |
| 2843 | } |
| 2844 | } |
| 2845 | |
Yuchung Cheng | 4f41b1c | 2015-10-16 21:57:47 -0700 | [diff] [blame] | 2846 | /* Use RACK to detect loss */ |
| 2847 | if (sysctl_tcp_recovery & TCP_RACK_LOST_RETRANS && |
Yuchung Cheng | 31ba0c1 | 2016-02-02 10:33:05 -0800 | [diff] [blame] | 2848 | tcp_rack_mark_lost(sk)) { |
Yuchung Cheng | 4f41b1c | 2015-10-16 21:57:47 -0700 | [diff] [blame] | 2849 | flag |= FLAG_LOST_RETRANS; |
Yuchung Cheng | 31ba0c1 | 2016-02-02 10:33:05 -0800 | [diff] [blame] | 2850 | *ack_flag |= FLAG_LOST_RETRANS; |
| 2851 | } |
Yuchung Cheng | 4f41b1c | 2015-10-16 21:57:47 -0700 | [diff] [blame] | 2852 | |
Yuchung Cheng | 974c123 | 2012-01-19 14:42:21 +0000 | [diff] [blame] | 2853 | /* E. Process state. */ |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2854 | switch (icsk->icsk_ca_state) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2855 | case TCP_CA_Recovery: |
Ilpo Järvinen | 2e60529 | 2007-08-02 19:46:58 -0700 | [diff] [blame] | 2856 | if (!(flag & FLAG_SND_UNA_ADVANCED)) { |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 2857 | if (tcp_is_reno(tp) && is_dupack) |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2858 | tcp_add_reno_sack(sk); |
Yuchung Cheng | 7026b91 | 2013-05-29 14:20:13 +0000 | [diff] [blame] | 2859 | } else { |
Yuchung Cheng | 31ba0c1 | 2016-02-02 10:33:05 -0800 | [diff] [blame] | 2860 | if (tcp_try_undo_partial(sk, acked)) |
Yuchung Cheng | 7026b91 | 2013-05-29 14:20:13 +0000 | [diff] [blame] | 2861 | return; |
| 2862 | /* Partial ACK arrived. Force fast retransmit. */ |
| 2863 | do_lost = tcp_is_reno(tp) || |
| 2864 | tcp_fackets_out(tp) > tp->reordering; |
| 2865 | } |
Yuchung Cheng | c7d9d6a | 2013-05-29 14:20:14 +0000 | [diff] [blame] | 2866 | if (tcp_try_undo_dsack(sk)) { |
| 2867 | tcp_try_keep_open(sk); |
| 2868 | return; |
| 2869 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2870 | break; |
| 2871 | case TCP_CA_Loss: |
Yuchung Cheng | e662ca4 | 2016-02-02 10:33:04 -0800 | [diff] [blame] | 2872 | tcp_process_loss(sk, flag, is_dupack, rexmit); |
Yuchung Cheng | 291a00d | 2015-07-01 14:11:14 -0700 | [diff] [blame] | 2873 | if (icsk->icsk_ca_state != TCP_CA_Open && |
| 2874 | !(flag & FLAG_LOST_RETRANS)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2875 | return; |
Yuchung Cheng | 291a00d | 2015-07-01 14:11:14 -0700 | [diff] [blame] | 2876 | /* Change state if cwnd is undone or retransmits are lost */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2877 | default: |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 2878 | if (tcp_is_reno(tp)) { |
Ilpo Järvinen | 2e60529 | 2007-08-02 19:46:58 -0700 | [diff] [blame] | 2879 | if (flag & FLAG_SND_UNA_ADVANCED) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2880 | tcp_reset_reno_sack(tp); |
| 2881 | if (is_dupack) |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2882 | tcp_add_reno_sack(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2883 | } |
| 2884 | |
Neal Cardwell | f698204 | 2011-11-16 08:58:04 +0000 | [diff] [blame] | 2885 | if (icsk->icsk_ca_state <= TCP_CA_Disorder) |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2886 | tcp_try_undo_dsack(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2887 | |
Yuchung Cheng | 750ea2b | 2012-05-02 13:30:04 +0000 | [diff] [blame] | 2888 | if (!tcp_time_to_recover(sk, flag)) { |
Yuchung Cheng | 31ba0c1 | 2016-02-02 10:33:05 -0800 | [diff] [blame] | 2889 | tcp_try_to_open(sk, flag); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2890 | return; |
| 2891 | } |
| 2892 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 2893 | /* MTU probe failure: don't reduce cwnd */ |
| 2894 | if (icsk->icsk_ca_state < TCP_CA_CWR && |
| 2895 | icsk->icsk_mtup.probe_size && |
John Heffner | 0e7b136 | 2006-03-20 21:32:58 -0800 | [diff] [blame] | 2896 | tp->snd_una == tp->mtu_probe.probe_seq_start) { |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 2897 | tcp_mtup_probe_failed(sk); |
| 2898 | /* Restores the reduction we did in tcp_mtup_probe() */ |
| 2899 | tp->snd_cwnd++; |
| 2900 | tcp_simple_retransmit(sk); |
| 2901 | return; |
| 2902 | } |
| 2903 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2904 | /* Otherwise enter Recovery state */ |
Yuchung Cheng | 1fbc340 | 2012-05-02 13:30:02 +0000 | [diff] [blame] | 2905 | tcp_enter_recovery(sk, (flag & FLAG_ECE)); |
Ilpo Järvinen | 85cc391 | 2007-11-15 19:39:31 -0800 | [diff] [blame] | 2906 | fast_rexmit = 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2907 | } |
| 2908 | |
Yuchung Cheng | 3e59cb0 | 2013-05-17 13:45:05 +0000 | [diff] [blame] | 2909 | if (do_lost) |
Ilpo Järvinen | 85cc391 | 2007-11-15 19:39:31 -0800 | [diff] [blame] | 2910 | tcp_update_scoreboard(sk, fast_rexmit); |
Yuchung Cheng | e662ca4 | 2016-02-02 10:33:04 -0800 | [diff] [blame] | 2911 | *rexmit = REXMIT_LOST; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2912 | } |
| 2913 | |
Yuchung Cheng | f672258 | 2015-10-16 21:57:42 -0700 | [diff] [blame] | 2914 | static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us) |
| 2915 | { |
Neal Cardwell | 6403389 | 2016-09-19 23:39:10 -0400 | [diff] [blame] | 2916 | struct tcp_sock *tp = tcp_sk(sk); |
| 2917 | u32 wlen = sysctl_tcp_min_rtt_wlen * HZ; |
Yuchung Cheng | f672258 | 2015-10-16 21:57:42 -0700 | [diff] [blame] | 2918 | |
Neal Cardwell | 6403389 | 2016-09-19 23:39:10 -0400 | [diff] [blame] | 2919 | minmax_running_min(&tp->rtt_min, wlen, tcp_time_stamp, |
| 2920 | rtt_us ? : jiffies_to_usecs(1)); |
Yuchung Cheng | f672258 | 2015-10-16 21:57:42 -0700 | [diff] [blame] | 2921 | } |
| 2922 | |
Yuchung Cheng | ed08495 | 2013-07-22 16:20:48 -0700 | [diff] [blame] | 2923 | static inline bool tcp_ack_update_rtt(struct sock *sk, const int flag, |
Yuchung Cheng | f672258 | 2015-10-16 21:57:42 -0700 | [diff] [blame] | 2924 | long seq_rtt_us, long sack_rtt_us, |
| 2925 | long ca_rtt_us) |
Ilpo Järvinen | 41834b7 | 2008-12-05 22:43:26 -0800 | [diff] [blame] | 2926 | { |
Yuchung Cheng | 5b08e47 | 2013-07-22 16:20:46 -0700 | [diff] [blame] | 2927 | const struct tcp_sock *tp = tcp_sk(sk); |
Ilpo Järvinen | 41834b7 | 2008-12-05 22:43:26 -0800 | [diff] [blame] | 2928 | |
Yuchung Cheng | 5b08e47 | 2013-07-22 16:20:46 -0700 | [diff] [blame] | 2929 | /* Prefer RTT measured from ACK's timing to TS-ECR. This is because |
| 2930 | * broken middle-boxes or peers may corrupt TS-ECR fields. But |
| 2931 | * Karn's algorithm forbids taking RTT if some retransmitted data |
| 2932 | * is acked (RFC6298). |
| 2933 | */ |
Eric Dumazet | 740b0f1 | 2014-02-26 14:02:48 -0800 | [diff] [blame] | 2934 | if (seq_rtt_us < 0) |
| 2935 | seq_rtt_us = sack_rtt_us; |
Yuchung Cheng | ed08495 | 2013-07-22 16:20:48 -0700 | [diff] [blame] | 2936 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2937 | /* RTTM Rule: A TSecr value received in a segment is used to |
| 2938 | * update the averaged RTT measurement only if the segment |
| 2939 | * acknowledges some new data, i.e., only if it advances the |
| 2940 | * left edge of the send window. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2941 | * See draft-ietf-tcplw-high-performance-00, section 3.3. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2942 | */ |
Eric Dumazet | 740b0f1 | 2014-02-26 14:02:48 -0800 | [diff] [blame] | 2943 | if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && |
Yuchung Cheng | 2909d87 | 2013-10-24 08:55:25 -0700 | [diff] [blame] | 2944 | flag & FLAG_ACKED) |
Yuchung Cheng | f672258 | 2015-10-16 21:57:42 -0700 | [diff] [blame] | 2945 | seq_rtt_us = ca_rtt_us = jiffies_to_usecs(tcp_time_stamp - |
| 2946 | tp->rx_opt.rcv_tsecr); |
Eric Dumazet | 740b0f1 | 2014-02-26 14:02:48 -0800 | [diff] [blame] | 2947 | if (seq_rtt_us < 0) |
Yuchung Cheng | ed08495 | 2013-07-22 16:20:48 -0700 | [diff] [blame] | 2948 | return false; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2949 | |
Yuchung Cheng | f672258 | 2015-10-16 21:57:42 -0700 | [diff] [blame] | 2950 | /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is |
| 2951 | * always taken together with ACK, SACK, or TS-opts. Any negative |
| 2952 | * values will be skipped with the seq_rtt_us < 0 check above. |
| 2953 | */ |
| 2954 | tcp_update_rtt_min(sk, ca_rtt_us); |
Eric Dumazet | 740b0f1 | 2014-02-26 14:02:48 -0800 | [diff] [blame] | 2955 | tcp_rtt_estimator(sk, seq_rtt_us); |
Yuchung Cheng | 5b08e47 | 2013-07-22 16:20:46 -0700 | [diff] [blame] | 2956 | tcp_set_rto(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2957 | |
Yuchung Cheng | 5b08e47 | 2013-07-22 16:20:46 -0700 | [diff] [blame] | 2958 | /* RFC6298: only reset backoff on valid RTT measurement. */ |
| 2959 | inet_csk(sk)->icsk_backoff = 0; |
Yuchung Cheng | ed08495 | 2013-07-22 16:20:48 -0700 | [diff] [blame] | 2960 | return true; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2961 | } |
| 2962 | |
Yuchung Cheng | 375fe02 | 2013-07-22 16:20:45 -0700 | [diff] [blame] | 2963 | /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */ |
Yuchung Cheng | 0f1c28a | 2015-09-18 11:36:14 -0700 | [diff] [blame] | 2964 | void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req) |
Yuchung Cheng | 375fe02 | 2013-07-22 16:20:45 -0700 | [diff] [blame] | 2965 | { |
Yuchung Cheng | 0f1c28a | 2015-09-18 11:36:14 -0700 | [diff] [blame] | 2966 | long rtt_us = -1L; |
Yuchung Cheng | 375fe02 | 2013-07-22 16:20:45 -0700 | [diff] [blame] | 2967 | |
Yuchung Cheng | 0f1c28a | 2015-09-18 11:36:14 -0700 | [diff] [blame] | 2968 | if (req && !req->num_retrans && tcp_rsk(req)->snt_synack.v64) { |
| 2969 | struct skb_mstamp now; |
Yuchung Cheng | bc15afa | 2013-10-24 08:44:25 -0700 | [diff] [blame] | 2970 | |
Yuchung Cheng | 0f1c28a | 2015-09-18 11:36:14 -0700 | [diff] [blame] | 2971 | skb_mstamp_get(&now); |
| 2972 | rtt_us = skb_mstamp_us_delta(&now, &tcp_rsk(req)->snt_synack); |
| 2973 | } |
| 2974 | |
Yuchung Cheng | f672258 | 2015-10-16 21:57:42 -0700 | [diff] [blame] | 2975 | tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us); |
Yuchung Cheng | 375fe02 | 2013-07-22 16:20:45 -0700 | [diff] [blame] | 2976 | } |
| 2977 | |
Yuchung Cheng | 0f1c28a | 2015-09-18 11:36:14 -0700 | [diff] [blame] | 2978 | |
Eric Dumazet | 2490155 | 2014-05-02 21:18:05 -0700 | [diff] [blame] | 2979 | static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2980 | { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2981 | const struct inet_connection_sock *icsk = inet_csk(sk); |
Eric Dumazet | 2490155 | 2014-05-02 21:18:05 -0700 | [diff] [blame] | 2982 | |
| 2983 | icsk->icsk_ca_ops->cong_avoid(sk, ack, acked); |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2984 | tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2985 | } |
| 2986 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2987 | /* Restart timer after forward progress on connection. |
| 2988 | * RFC2988 recommends to restart timer to now+rto. |
| 2989 | */ |
Yuchung Cheng | 750ea2b | 2012-05-02 13:30:04 +0000 | [diff] [blame] | 2990 | void tcp_rearm_rto(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2991 | { |
Nandita Dukkipati | 6ba8a3b | 2013-03-11 10:00:43 +0000 | [diff] [blame] | 2992 | const struct inet_connection_sock *icsk = inet_csk(sk); |
Yuchung Cheng | 750ea2b | 2012-05-02 13:30:04 +0000 | [diff] [blame] | 2993 | struct tcp_sock *tp = tcp_sk(sk); |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2994 | |
Jerry Chu | 168a8f5 | 2012-08-31 12:29:13 +0000 | [diff] [blame] | 2995 | /* If the retrans timer is currently being used by Fast Open |
| 2996 | * for SYN-ACK retrans purpose, stay put. |
| 2997 | */ |
| 2998 | if (tp->fastopen_rsk) |
| 2999 | return; |
| 3000 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3001 | if (!tp->packets_out) { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3002 | inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3003 | } else { |
Yuchung Cheng | 750ea2b | 2012-05-02 13:30:04 +0000 | [diff] [blame] | 3004 | u32 rto = inet_csk(sk)->icsk_rto; |
| 3005 | /* Offset the time elapsed after installing regular RTO */ |
Nandita Dukkipati | 6ba8a3b | 2013-03-11 10:00:43 +0000 | [diff] [blame] | 3006 | if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS || |
| 3007 | icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) { |
Yuchung Cheng | 750ea2b | 2012-05-02 13:30:04 +0000 | [diff] [blame] | 3008 | struct sk_buff *skb = tcp_write_queue_head(sk); |
Eric Dumazet | 7faee5c | 2014-09-05 15:33:33 -0700 | [diff] [blame] | 3009 | const u32 rto_time_stamp = |
| 3010 | tcp_skb_timestamp(skb) + rto; |
Yuchung Cheng | 750ea2b | 2012-05-02 13:30:04 +0000 | [diff] [blame] | 3011 | s32 delta = (s32)(rto_time_stamp - tcp_time_stamp); |
| 3012 | /* delta may not be positive if the socket is locked |
Nandita Dukkipati | 6ba8a3b | 2013-03-11 10:00:43 +0000 | [diff] [blame] | 3013 | * when the retrans timer fires and is rescheduled. |
Yuchung Cheng | 750ea2b | 2012-05-02 13:30:04 +0000 | [diff] [blame] | 3014 | */ |
| 3015 | if (delta > 0) |
| 3016 | rto = delta; |
| 3017 | } |
| 3018 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto, |
| 3019 | TCP_RTO_MAX); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3020 | } |
Yuchung Cheng | 750ea2b | 2012-05-02 13:30:04 +0000 | [diff] [blame] | 3021 | } |
| 3022 | |
| 3023 | /* This function is called when the delayed ER timer fires. TCP enters |
| 3024 | * fast recovery and performs fast-retransmit. |
| 3025 | */ |
| 3026 | void tcp_resume_early_retransmit(struct sock *sk) |
| 3027 | { |
| 3028 | struct tcp_sock *tp = tcp_sk(sk); |
| 3029 | |
| 3030 | tcp_rearm_rto(sk); |
| 3031 | |
| 3032 | /* Stop if ER is disabled after the delayed ER timer is scheduled */ |
| 3033 | if (!tp->do_early_retrans) |
| 3034 | return; |
| 3035 | |
| 3036 | tcp_enter_recovery(sk, false); |
| 3037 | tcp_update_scoreboard(sk, 1); |
| 3038 | tcp_xmit_retransmit_queue(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3039 | } |
| 3040 | |
Ilpo Järvinen | 7c46a03 | 2007-09-20 11:33:43 -0700 | [diff] [blame] | 3041 | /* If we get here, the whole TSO packet has not been acked. */ |
Ilpo Järvinen | 13fcf85 | 2007-10-09 01:28:45 -0700 | [diff] [blame] | 3042 | static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3043 | { |
| 3044 | struct tcp_sock *tp = tcp_sk(sk); |
Ilpo Järvinen | 7c46a03 | 2007-09-20 11:33:43 -0700 | [diff] [blame] | 3045 | u32 packets_acked; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3046 | |
Ilpo Järvinen | 7c46a03 | 2007-09-20 11:33:43 -0700 | [diff] [blame] | 3047 | BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3048 | |
| 3049 | packets_acked = tcp_skb_pcount(skb); |
Ilpo Järvinen | 7c46a03 | 2007-09-20 11:33:43 -0700 | [diff] [blame] | 3050 | 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] | 3051 | return 0; |
| 3052 | packets_acked -= tcp_skb_pcount(skb); |
| 3053 | |
| 3054 | if (packets_acked) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3055 | BUG_ON(tcp_skb_pcount(skb) == 0); |
Ilpo Järvinen | 7c46a03 | 2007-09-20 11:33:43 -0700 | [diff] [blame] | 3056 | 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] | 3057 | } |
| 3058 | |
Ilpo Järvinen | 13fcf85 | 2007-10-09 01:28:45 -0700 | [diff] [blame] | 3059 | return packets_acked; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3060 | } |
| 3061 | |
Eric Dumazet | ad971f6 | 2014-10-11 15:17:29 -0700 | [diff] [blame] | 3062 | static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb, |
| 3063 | u32 prior_snd_una) |
| 3064 | { |
| 3065 | const struct skb_shared_info *shinfo; |
| 3066 | |
| 3067 | /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */ |
Soheil Hassas Yeganeh | 6b08492 | 2016-04-02 23:08:08 -0400 | [diff] [blame] | 3068 | if (likely(!TCP_SKB_CB(skb)->txstamp_ack)) |
Eric Dumazet | ad971f6 | 2014-10-11 15:17:29 -0700 | [diff] [blame] | 3069 | return; |
| 3070 | |
| 3071 | shinfo = skb_shinfo(skb); |
Soheil Hassas Yeganeh | 0a2cf20 | 2016-04-27 23:39:01 -0400 | [diff] [blame] | 3072 | if (!before(shinfo->tskey, prior_snd_una) && |
Martin KaFai Lau | 479f85c | 2016-04-18 15:39:53 -0700 | [diff] [blame] | 3073 | before(shinfo->tskey, tcp_sk(sk)->snd_una)) |
Eric Dumazet | ad971f6 | 2014-10-11 15:17:29 -0700 | [diff] [blame] | 3074 | __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK); |
| 3075 | } |
| 3076 | |
Ilpo Järvinen | 7c46a03 | 2007-09-20 11:33:43 -0700 | [diff] [blame] | 3077 | /* Remove acknowledged frames from the retransmission queue. If our packet |
| 3078 | * is before the ack sequence we can discard it as it's confirmed to have |
| 3079 | * arrived at the other end. |
| 3080 | */ |
Ilpo Järvinen | 33f5f57 | 2008-10-07 14:43:06 -0700 | [diff] [blame] | 3081 | static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets, |
Yuchung Cheng | 3ebd8871 | 2016-02-02 10:33:07 -0800 | [diff] [blame] | 3082 | u32 prior_snd_una, int *acked, |
Yuchung Cheng | b9f6482 | 2016-09-19 23:39:14 -0400 | [diff] [blame] | 3083 | struct tcp_sacktag_state *sack, |
| 3084 | struct skb_mstamp *now) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3085 | { |
Stephen Hemminger | 2d2abba | 2005-11-10 16:56:12 -0800 | [diff] [blame] | 3086 | const struct inet_connection_sock *icsk = inet_csk(sk); |
Yuchung Cheng | b9f6482 | 2016-09-19 23:39:14 -0400 | [diff] [blame] | 3087 | struct skb_mstamp first_ackt, last_ackt; |
Eric Dumazet | 740b0f1 | 2014-02-26 14:02:48 -0800 | [diff] [blame] | 3088 | struct tcp_sock *tp = tcp_sk(sk); |
Ilpo Järvinen | 90638a0 | 2008-09-20 21:25:52 -0700 | [diff] [blame] | 3089 | u32 prior_sacked = tp->sacked_out; |
Eric Dumazet | 740b0f1 | 2014-02-26 14:02:48 -0800 | [diff] [blame] | 3090 | u32 reord = tp->packets_out; |
| 3091 | bool fully_acked = true; |
Kenneth Klette Jonassen | 31231a8 | 2015-05-01 01:10:58 +0200 | [diff] [blame] | 3092 | long sack_rtt_us = -1L; |
Eric Dumazet | 740b0f1 | 2014-02-26 14:02:48 -0800 | [diff] [blame] | 3093 | long seq_rtt_us = -1L; |
Kenneth Klette Jonassen | 31231a8 | 2015-05-01 01:10:58 +0200 | [diff] [blame] | 3094 | long ca_rtt_us = -1L; |
Eric Dumazet | 740b0f1 | 2014-02-26 14:02:48 -0800 | [diff] [blame] | 3095 | struct sk_buff *skb; |
| 3096 | u32 pkts_acked = 0; |
Lawrence Brakmo | 6f094b9 | 2016-06-08 21:16:44 -0700 | [diff] [blame] | 3097 | u32 last_in_flight = 0; |
Yuchung Cheng | 2f715c1 | 2013-10-24 08:59:27 -0700 | [diff] [blame] | 3098 | bool rtt_update; |
Eric Dumazet | 740b0f1 | 2014-02-26 14:02:48 -0800 | [diff] [blame] | 3099 | int flag = 0; |
| 3100 | |
| 3101 | first_ackt.v64 = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3102 | |
Ilpo Järvinen | 7c46a03 | 2007-09-20 11:33:43 -0700 | [diff] [blame] | 3103 | while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) { |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 3104 | struct tcp_skb_cb *scb = TCP_SKB_CB(skb); |
Ilpo Järvinen | 7c46a03 | 2007-09-20 11:33:43 -0700 | [diff] [blame] | 3105 | u8 sacked = scb->sacked; |
Eric Dumazet | 740b0f1 | 2014-02-26 14:02:48 -0800 | [diff] [blame] | 3106 | u32 acked_pcount; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3107 | |
Eric Dumazet | ad971f6 | 2014-10-11 15:17:29 -0700 | [diff] [blame] | 3108 | tcp_ack_tstamp(sk, skb, prior_snd_una); |
Willem de Bruijn | 712a722 | 2014-08-12 14:53:16 -0400 | [diff] [blame] | 3109 | |
Gavin McCullagh | 2072c22 | 2007-12-29 19:11:21 -0800 | [diff] [blame] | 3110 | /* Determine how many packets and what bytes were acked, tso and else */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3111 | if (after(scb->end_seq, tp->snd_una)) { |
Ilpo Järvinen | 13fcf85 | 2007-10-09 01:28:45 -0700 | [diff] [blame] | 3112 | if (tcp_skb_pcount(skb) == 1 || |
| 3113 | !after(tp->snd_una, scb->seq)) |
| 3114 | break; |
| 3115 | |
Ilpo Järvinen | 7201883 | 2007-12-30 04:37:55 -0800 | [diff] [blame] | 3116 | acked_pcount = tcp_tso_acked(sk, skb); |
| 3117 | if (!acked_pcount) |
Ilpo Järvinen | 13fcf85 | 2007-10-09 01:28:45 -0700 | [diff] [blame] | 3118 | break; |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 3119 | fully_acked = false; |
Ilpo Järvinen | 13fcf85 | 2007-10-09 01:28:45 -0700 | [diff] [blame] | 3120 | } else { |
Eric Dumazet | ad971f6 | 2014-10-11 15:17:29 -0700 | [diff] [blame] | 3121 | /* Speedup tcp_unlink_write_queue() and next loop */ |
| 3122 | prefetchw(skb->next); |
Ilpo Järvinen | 7201883 | 2007-12-30 04:37:55 -0800 | [diff] [blame] | 3123 | acked_pcount = tcp_skb_pcount(skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3124 | } |
| 3125 | |
Eric Dumazet | ad971f6 | 2014-10-11 15:17:29 -0700 | [diff] [blame] | 3126 | if (unlikely(sacked & TCPCB_RETRANS)) { |
Ilpo Järvinen | 89d478f | 2007-12-30 04:35:27 -0800 | [diff] [blame] | 3127 | if (sacked & TCPCB_SACKED_RETRANS) |
Ilpo Järvinen | 7201883 | 2007-12-30 04:37:55 -0800 | [diff] [blame] | 3128 | tp->retrans_out -= acked_pcount; |
Ilpo Järvinen | 89d478f | 2007-12-30 04:35:27 -0800 | [diff] [blame] | 3129 | flag |= FLAG_RETRANS_DATA_ACKED; |
Kenneth Klette Jonassen | 3d0d26c | 2015-04-11 02:17:49 +0200 | [diff] [blame] | 3130 | } else if (!(sacked & TCPCB_SACKED_ACKED)) { |
Eric Dumazet | 740b0f1 | 2014-02-26 14:02:48 -0800 | [diff] [blame] | 3131 | last_ackt = skb->skb_mstamp; |
Eric Dumazet | 431a912 | 2014-03-09 17:36:02 -0700 | [diff] [blame] | 3132 | WARN_ON_ONCE(last_ackt.v64 == 0); |
Eric Dumazet | 740b0f1 | 2014-02-26 14:02:48 -0800 | [diff] [blame] | 3133 | if (!first_ackt.v64) |
| 3134 | first_ackt = last_ackt; |
| 3135 | |
Lawrence Brakmo | 6f094b9 | 2016-06-08 21:16:44 -0700 | [diff] [blame] | 3136 | last_in_flight = TCP_SKB_CB(skb)->tx.in_flight; |
Kenneth Klette Jonassen | 3d0d26c | 2015-04-11 02:17:49 +0200 | [diff] [blame] | 3137 | reord = min(pkts_acked, reord); |
| 3138 | if (!after(scb->end_seq, tp->high_seq)) |
| 3139 | flag |= FLAG_ORIG_SACK_ACKED; |
Stephen Hemminger | 2d2abba | 2005-11-10 16:56:12 -0800 | [diff] [blame] | 3140 | } |
Ilpo Järvinen | 89d478f | 2007-12-30 04:35:27 -0800 | [diff] [blame] | 3141 | |
Yuchung Cheng | ddf1af6 | 2016-02-02 10:33:06 -0800 | [diff] [blame] | 3142 | if (sacked & TCPCB_SACKED_ACKED) { |
Ilpo Järvinen | 7201883 | 2007-12-30 04:37:55 -0800 | [diff] [blame] | 3143 | tp->sacked_out -= acked_pcount; |
Yuchung Cheng | ddf1af6 | 2016-02-02 10:33:06 -0800 | [diff] [blame] | 3144 | } else if (tcp_is_sack(tp)) { |
| 3145 | tp->delivered += acked_pcount; |
| 3146 | if (!tcp_skb_spurious_retrans(tp, skb)) |
| 3147 | tcp_rack_advance(tp, &skb->skb_mstamp, sacked); |
| 3148 | } |
Ilpo Järvinen | 89d478f | 2007-12-30 04:35:27 -0800 | [diff] [blame] | 3149 | if (sacked & TCPCB_LOST) |
Ilpo Järvinen | 7201883 | 2007-12-30 04:37:55 -0800 | [diff] [blame] | 3150 | tp->lost_out -= acked_pcount; |
Ilpo Järvinen | 89d478f | 2007-12-30 04:35:27 -0800 | [diff] [blame] | 3151 | |
Ilpo Järvinen | 7201883 | 2007-12-30 04:37:55 -0800 | [diff] [blame] | 3152 | tp->packets_out -= acked_pcount; |
| 3153 | pkts_acked += acked_pcount; |
Yuchung Cheng | b9f6482 | 2016-09-19 23:39:14 -0400 | [diff] [blame] | 3154 | tcp_rate_skb_delivered(sk, skb, sack->rate); |
Ilpo Järvinen | 13fcf85 | 2007-10-09 01:28:45 -0700 | [diff] [blame] | 3155 | |
Ilpo Järvinen | 009a2e3 | 2007-09-20 11:34:38 -0700 | [diff] [blame] | 3156 | /* Initial outgoing SYN's get put onto the write_queue |
| 3157 | * just like anything else we transmit. It is not |
| 3158 | * true data, and if we misinform our callers that |
| 3159 | * this ACK acks real data, we will erroneously exit |
| 3160 | * connection startup slow start one packet too |
| 3161 | * quickly. This is severely frowned upon behavior. |
| 3162 | */ |
Eric Dumazet | ad971f6 | 2014-10-11 15:17:29 -0700 | [diff] [blame] | 3163 | if (likely(!(scb->tcp_flags & TCPHDR_SYN))) { |
Ilpo Järvinen | 009a2e3 | 2007-09-20 11:34:38 -0700 | [diff] [blame] | 3164 | flag |= FLAG_DATA_ACKED; |
| 3165 | } else { |
| 3166 | flag |= FLAG_SYN_ACKED; |
| 3167 | tp->retrans_stamp = 0; |
| 3168 | } |
| 3169 | |
Ilpo Järvinen | 13fcf85 | 2007-10-09 01:28:45 -0700 | [diff] [blame] | 3170 | if (!fully_acked) |
| 3171 | break; |
| 3172 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 3173 | tcp_unlink_write_queue(skb, sk); |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 3174 | sk_wmem_free_skb(sk, skb); |
Eric Dumazet | ad971f6 | 2014-10-11 15:17:29 -0700 | [diff] [blame] | 3175 | if (unlikely(skb == tp->retransmit_skb_hint)) |
Ilpo Järvinen | ef9da47 | 2008-09-20 21:25:15 -0700 | [diff] [blame] | 3176 | tp->retransmit_skb_hint = NULL; |
Eric Dumazet | ad971f6 | 2014-10-11 15:17:29 -0700 | [diff] [blame] | 3177 | if (unlikely(skb == tp->lost_skb_hint)) |
Ilpo Järvinen | 90638a0 | 2008-09-20 21:25:52 -0700 | [diff] [blame] | 3178 | tp->lost_skb_hint = NULL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3179 | } |
| 3180 | |
Francis Yan | 0f87230 | 2016-11-27 23:07:14 -0800 | [diff] [blame] | 3181 | if (!skb) |
| 3182 | tcp_chrono_stop(sk, TCP_CHRONO_BUSY); |
| 3183 | |
Ilpo Järvinen | 33f5f57 | 2008-10-07 14:43:06 -0700 | [diff] [blame] | 3184 | if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una))) |
| 3185 | tp->snd_up = tp->snd_una; |
| 3186 | |
Ilpo Järvinen | cadbd03 | 2007-12-31 04:49:21 -0800 | [diff] [blame] | 3187 | if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) |
| 3188 | flag |= FLAG_SACK_RENEGING; |
| 3189 | |
Yuchung Cheng | 9e45a3e | 2015-10-16 21:57:41 -0700 | [diff] [blame] | 3190 | if (likely(first_ackt.v64) && !(flag & FLAG_RETRANS_DATA_ACKED)) { |
Yuchung Cheng | b9f6482 | 2016-09-19 23:39:14 -0400 | [diff] [blame] | 3191 | seq_rtt_us = skb_mstamp_us_delta(now, &first_ackt); |
| 3192 | ca_rtt_us = skb_mstamp_us_delta(now, &last_ackt); |
Kenneth Klette Jonassen | 31231a8 | 2015-05-01 01:10:58 +0200 | [diff] [blame] | 3193 | } |
| 3194 | if (sack->first_sackt.v64) { |
Yuchung Cheng | b9f6482 | 2016-09-19 23:39:14 -0400 | [diff] [blame] | 3195 | sack_rtt_us = skb_mstamp_us_delta(now, &sack->first_sackt); |
| 3196 | ca_rtt_us = skb_mstamp_us_delta(now, &sack->last_sackt); |
Eric Dumazet | 740b0f1 | 2014-02-26 14:02:48 -0800 | [diff] [blame] | 3197 | } |
Yuchung Cheng | b9f6482 | 2016-09-19 23:39:14 -0400 | [diff] [blame] | 3198 | sack->rate->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet, or -1 */ |
Yuchung Cheng | f672258 | 2015-10-16 21:57:42 -0700 | [diff] [blame] | 3199 | rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us, |
| 3200 | ca_rtt_us); |
Yuchung Cheng | ed08495 | 2013-07-22 16:20:48 -0700 | [diff] [blame] | 3201 | |
Ilpo Järvinen | 7c46a03 | 2007-09-20 11:33:43 -0700 | [diff] [blame] | 3202 | if (flag & FLAG_ACKED) { |
Yuchung Cheng | 2f715c1 | 2013-10-24 08:59:27 -0700 | [diff] [blame] | 3203 | tcp_rearm_rto(sk); |
Ilpo Järvinen | 72211e9 | 2009-03-14 14:23:04 +0000 | [diff] [blame] | 3204 | if (unlikely(icsk->icsk_mtup.probe_size && |
| 3205 | !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) { |
| 3206 | tcp_mtup_probe_success(sk); |
| 3207 | } |
| 3208 | |
Ilpo Järvinen | c7caf8d | 2007-11-10 21:22:18 -0800 | [diff] [blame] | 3209 | if (tcp_is_reno(tp)) { |
| 3210 | tcp_remove_reno_sacks(sk, pkts_acked); |
| 3211 | } else { |
Ilpo Järvinen | 59a08cb | 2009-02-28 04:44:28 +0000 | [diff] [blame] | 3212 | int delta; |
| 3213 | |
Ilpo Järvinen | c7caf8d | 2007-11-10 21:22:18 -0800 | [diff] [blame] | 3214 | /* Non-retransmitted hole got filled? That's reordering */ |
| 3215 | if (reord < prior_fackets) |
| 3216 | tcp_update_reordering(sk, tp->fackets_out - reord, 0); |
Ilpo Järvinen | 90638a0 | 2008-09-20 21:25:52 -0700 | [diff] [blame] | 3217 | |
Ilpo Järvinen | 59a08cb | 2009-02-28 04:44:28 +0000 | [diff] [blame] | 3218 | delta = tcp_is_fack(tp) ? pkts_acked : |
| 3219 | prior_sacked - tp->sacked_out; |
| 3220 | tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta); |
Ilpo Järvinen | c7caf8d | 2007-11-10 21:22:18 -0800 | [diff] [blame] | 3221 | } |
| 3222 | |
Ilpo Järvinen | 91fed7a | 2007-10-09 01:24:15 -0700 | [diff] [blame] | 3223 | tp->fackets_out -= min(pkts_acked, tp->fackets_out); |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 3224 | |
Eric Dumazet | 740b0f1 | 2014-02-26 14:02:48 -0800 | [diff] [blame] | 3225 | } else if (skb && rtt_update && sack_rtt_us >= 0 && |
Yuchung Cheng | b9f6482 | 2016-09-19 23:39:14 -0400 | [diff] [blame] | 3226 | sack_rtt_us > skb_mstamp_us_delta(now, &skb->skb_mstamp)) { |
Yuchung Cheng | 2f715c1 | 2013-10-24 08:59:27 -0700 | [diff] [blame] | 3227 | /* Do not re-arm RTO if the sack RTT is measured from data sent |
| 3228 | * after when the head was last (re)transmitted. Otherwise the |
| 3229 | * timeout may continue to extend in loss recovery. |
| 3230 | */ |
| 3231 | tcp_rearm_rto(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3232 | } |
| 3233 | |
Lawrence Brakmo | 756ee17 | 2016-05-11 10:02:13 -0700 | [diff] [blame] | 3234 | if (icsk->icsk_ca_ops->pkts_acked) { |
| 3235 | struct ack_sample sample = { .pkts_acked = pkts_acked, |
Lawrence Brakmo | 6f094b9 | 2016-06-08 21:16:44 -0700 | [diff] [blame] | 3236 | .rtt_us = ca_rtt_us, |
| 3237 | .in_flight = last_in_flight }; |
Lawrence Brakmo | 756ee17 | 2016-05-11 10:02:13 -0700 | [diff] [blame] | 3238 | |
| 3239 | icsk->icsk_ca_ops->pkts_acked(sk, &sample); |
| 3240 | } |
Kenneth Klette Jonassen | 138998f | 2015-05-01 01:10:59 +0200 | [diff] [blame] | 3241 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3242 | #if FASTRETRANS_DEBUG > 0 |
Ilpo Järvinen | 547b792 | 2008-07-25 21:43:18 -0700 | [diff] [blame] | 3243 | WARN_ON((int)tp->sacked_out < 0); |
| 3244 | WARN_ON((int)tp->lost_out < 0); |
| 3245 | WARN_ON((int)tp->retrans_out < 0); |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 3246 | if (!tp->packets_out && tcp_is_sack(tp)) { |
Stephen Hemminger | cfcabdc | 2007-10-09 01:59:42 -0700 | [diff] [blame] | 3247 | icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3248 | if (tp->lost_out) { |
Joe Perches | 91df42b | 2012-05-15 14:11:54 +0000 | [diff] [blame] | 3249 | pr_debug("Leak l=%u %d\n", |
| 3250 | tp->lost_out, icsk->icsk_ca_state); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3251 | tp->lost_out = 0; |
| 3252 | } |
| 3253 | if (tp->sacked_out) { |
Joe Perches | 91df42b | 2012-05-15 14:11:54 +0000 | [diff] [blame] | 3254 | pr_debug("Leak s=%u %d\n", |
| 3255 | tp->sacked_out, icsk->icsk_ca_state); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3256 | tp->sacked_out = 0; |
| 3257 | } |
| 3258 | if (tp->retrans_out) { |
Joe Perches | 91df42b | 2012-05-15 14:11:54 +0000 | [diff] [blame] | 3259 | pr_debug("Leak r=%u %d\n", |
| 3260 | tp->retrans_out, icsk->icsk_ca_state); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3261 | tp->retrans_out = 0; |
| 3262 | } |
| 3263 | } |
| 3264 | #endif |
Yuchung Cheng | 3ebd8871 | 2016-02-02 10:33:07 -0800 | [diff] [blame] | 3265 | *acked = pkts_acked; |
Ilpo Järvinen | 7c46a03 | 2007-09-20 11:33:43 -0700 | [diff] [blame] | 3266 | return flag; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3267 | } |
| 3268 | |
| 3269 | static void tcp_ack_probe(struct sock *sk) |
| 3270 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3271 | const struct tcp_sock *tp = tcp_sk(sk); |
| 3272 | struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3273 | |
| 3274 | /* Was it a usable window open? */ |
| 3275 | |
Ilpo Järvinen | 90840de | 2007-12-31 04:48:41 -0800 | [diff] [blame] | 3276 | if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3277 | icsk->icsk_backoff = 0; |
| 3278 | inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3279 | /* Socket must be waked up by subsequent tcp_data_snd_check(). |
| 3280 | * This function is not for random using! |
| 3281 | */ |
| 3282 | } else { |
Eric Dumazet | 21c8fe9 | 2015-05-06 14:26:24 -0700 | [diff] [blame] | 3283 | unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX); |
Eric Dumazet | fcdd1cf | 2014-09-22 13:19:44 -0700 | [diff] [blame] | 3284 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3285 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, |
Eric Dumazet | fcdd1cf | 2014-09-22 13:19:44 -0700 | [diff] [blame] | 3286 | when, TCP_RTO_MAX); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3287 | } |
| 3288 | } |
| 3289 | |
Vijay Subramanian | 67b95bd | 2012-07-19 21:32:18 +0000 | [diff] [blame] | 3290 | static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3291 | { |
Eric Dumazet | a02cec2 | 2010-09-22 20:43:57 +0000 | [diff] [blame] | 3292 | return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) || |
| 3293 | inet_csk(sk)->icsk_ca_state != TCP_CA_Open; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3294 | } |
| 3295 | |
Yuchung Cheng | 0f7cc9a | 2013-08-21 17:29:23 -0700 | [diff] [blame] | 3296 | /* Decide wheather to run the increase function of congestion control. */ |
Vijay Subramanian | 67b95bd | 2012-07-19 21:32:18 +0000 | [diff] [blame] | 3297 | static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3298 | { |
Yuchung Cheng | 0f7cc9a | 2013-08-21 17:29:23 -0700 | [diff] [blame] | 3299 | /* If reordering is high then always grow cwnd whenever data is |
| 3300 | * delivered regardless of its ordering. Otherwise stay conservative |
Yuchung Cheng | 16edfe7 | 2013-09-05 15:36:09 -0700 | [diff] [blame] | 3301 | * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/ |
Yuchung Cheng | 0f7cc9a | 2013-08-21 17:29:23 -0700 | [diff] [blame] | 3302 | * new SACK or ECE mark may first advance cwnd here and later reduce |
| 3303 | * cwnd in tcp_fastretrans_alert() based on more states. |
| 3304 | */ |
Nikolay Borisov | 1043e25 | 2016-02-03 09:46:52 +0200 | [diff] [blame] | 3305 | if (tcp_sk(sk)->reordering > sock_net(sk)->ipv4.sysctl_tcp_reordering) |
Yuchung Cheng | 0f7cc9a | 2013-08-21 17:29:23 -0700 | [diff] [blame] | 3306 | return flag & FLAG_FORWARD_PROGRESS; |
| 3307 | |
Yuchung Cheng | 16edfe7 | 2013-09-05 15:36:09 -0700 | [diff] [blame] | 3308 | return flag & FLAG_DATA_ACKED; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3309 | } |
| 3310 | |
Yuchung Cheng | d452e6c | 2016-02-02 10:33:09 -0800 | [diff] [blame] | 3311 | /* The "ultimate" congestion control function that aims to replace the rigid |
| 3312 | * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction). |
| 3313 | * It's called toward the end of processing an ACK with precise rate |
| 3314 | * information. All transmission or retransmission are delayed afterwards. |
| 3315 | */ |
| 3316 | static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked, |
Yuchung Cheng | c040276 | 2016-09-19 23:39:21 -0400 | [diff] [blame] | 3317 | int flag, const struct rate_sample *rs) |
Yuchung Cheng | d452e6c | 2016-02-02 10:33:09 -0800 | [diff] [blame] | 3318 | { |
Yuchung Cheng | c040276 | 2016-09-19 23:39:21 -0400 | [diff] [blame] | 3319 | const struct inet_connection_sock *icsk = inet_csk(sk); |
| 3320 | |
| 3321 | if (icsk->icsk_ca_ops->cong_control) { |
| 3322 | icsk->icsk_ca_ops->cong_control(sk, rs); |
| 3323 | return; |
| 3324 | } |
| 3325 | |
Yuchung Cheng | d452e6c | 2016-02-02 10:33:09 -0800 | [diff] [blame] | 3326 | if (tcp_in_cwnd_reduction(sk)) { |
| 3327 | /* Reduce cwnd if state mandates */ |
| 3328 | tcp_cwnd_reduction(sk, acked_sacked, flag); |
| 3329 | } else if (tcp_may_raise_cwnd(sk, flag)) { |
| 3330 | /* Advance cwnd if state allows */ |
| 3331 | tcp_cong_avoid(sk, ack, acked_sacked); |
| 3332 | } |
| 3333 | tcp_update_pacing_rate(sk); |
| 3334 | } |
| 3335 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3336 | /* Check that window update is acceptable. |
| 3337 | * The function assumes that snd_una<=ack<=snd_next. |
| 3338 | */ |
Vijay Subramanian | 67b95bd | 2012-07-19 21:32:18 +0000 | [diff] [blame] | 3339 | static inline bool tcp_may_update_window(const struct tcp_sock *tp, |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 3340 | const u32 ack, const u32 ack_seq, |
| 3341 | const u32 nwin) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3342 | { |
Eric Dumazet | a02cec2 | 2010-09-22 20:43:57 +0000 | [diff] [blame] | 3343 | return after(ack, tp->snd_una) || |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3344 | after(ack_seq, tp->snd_wl1) || |
Eric Dumazet | a02cec2 | 2010-09-22 20:43:57 +0000 | [diff] [blame] | 3345 | (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3346 | } |
| 3347 | |
Eric Dumazet | 0df48c2 | 2015-04-28 15:28:17 -0700 | [diff] [blame] | 3348 | /* If we update tp->snd_una, also update tp->bytes_acked */ |
| 3349 | static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack) |
| 3350 | { |
| 3351 | u32 delta = ack - tp->snd_una; |
| 3352 | |
Eric Dumazet | 46cc6e4 | 2016-05-03 16:56:03 -0700 | [diff] [blame] | 3353 | sock_owned_by_me((struct sock *)tp); |
Eric Dumazet | 0df48c2 | 2015-04-28 15:28:17 -0700 | [diff] [blame] | 3354 | tp->bytes_acked += delta; |
| 3355 | tp->snd_una = ack; |
| 3356 | } |
| 3357 | |
Eric Dumazet | bdd1f9e | 2015-04-28 15:28:18 -0700 | [diff] [blame] | 3358 | /* If we update tp->rcv_nxt, also update tp->bytes_received */ |
| 3359 | static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq) |
| 3360 | { |
| 3361 | u32 delta = seq - tp->rcv_nxt; |
| 3362 | |
Eric Dumazet | 46cc6e4 | 2016-05-03 16:56:03 -0700 | [diff] [blame] | 3363 | sock_owned_by_me((struct sock *)tp); |
Eric Dumazet | bdd1f9e | 2015-04-28 15:28:18 -0700 | [diff] [blame] | 3364 | tp->bytes_received += delta; |
| 3365 | tp->rcv_nxt = seq; |
| 3366 | } |
| 3367 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3368 | /* Update our send window. |
| 3369 | * |
| 3370 | * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2 |
| 3371 | * and in FreeBSD. NetBSD's one is even worse.) is wrong. |
| 3372 | */ |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 3373 | static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack, |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 3374 | u32 ack_seq) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3375 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 3376 | struct tcp_sock *tp = tcp_sk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3377 | int flag = 0; |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 3378 | u32 nwin = ntohs(tcp_hdr(skb)->window); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3379 | |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 3380 | if (likely(!tcp_hdr(skb)->syn)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3381 | nwin <<= tp->rx_opt.snd_wscale; |
| 3382 | |
| 3383 | if (tcp_may_update_window(tp, ack, ack_seq, nwin)) { |
| 3384 | flag |= FLAG_WIN_UPDATE; |
Hantzis Fotis | ee7537b | 2009-03-02 22:42:02 -0800 | [diff] [blame] | 3385 | tcp_update_wl(tp, ack_seq); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3386 | |
| 3387 | if (tp->snd_wnd != nwin) { |
| 3388 | tp->snd_wnd = nwin; |
| 3389 | |
| 3390 | /* Note, it is the only place, where |
| 3391 | * fast path is recovered for sending TCP. |
| 3392 | */ |
Herbert Xu | 2ad4106 | 2005-10-27 18:47:46 +1000 | [diff] [blame] | 3393 | tp->pred_flags = 0; |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 3394 | tcp_fast_path_check(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3395 | |
Eric Dumazet | 6f021c6 | 2015-08-21 12:30:00 -0700 | [diff] [blame] | 3396 | if (tcp_send_head(sk)) |
| 3397 | tcp_slow_start_after_idle_check(sk); |
| 3398 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3399 | if (nwin > tp->max_window) { |
| 3400 | tp->max_window = nwin; |
Arnaldo Carvalho de Melo | d83d846 | 2005-12-13 23:26:10 -0800 | [diff] [blame] | 3401 | tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3402 | } |
| 3403 | } |
| 3404 | } |
| 3405 | |
Eric Dumazet | 0df48c2 | 2015-04-28 15:28:17 -0700 | [diff] [blame] | 3406 | tcp_snd_una_update(tp, ack); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3407 | |
| 3408 | return flag; |
| 3409 | } |
| 3410 | |
Jason Baron | 083ae30 | 2016-07-14 11:38:40 -0400 | [diff] [blame] | 3411 | static bool __tcp_oow_rate_limited(struct net *net, int mib_idx, |
| 3412 | u32 *last_oow_ack_time) |
| 3413 | { |
| 3414 | if (*last_oow_ack_time) { |
| 3415 | s32 elapsed = (s32)(tcp_time_stamp - *last_oow_ack_time); |
| 3416 | |
| 3417 | if (0 <= elapsed && elapsed < sysctl_tcp_invalid_ratelimit) { |
| 3418 | NET_INC_STATS(net, mib_idx); |
| 3419 | return true; /* rate-limited: don't send yet! */ |
| 3420 | } |
| 3421 | } |
| 3422 | |
| 3423 | *last_oow_ack_time = tcp_time_stamp; |
| 3424 | |
| 3425 | return false; /* not rate-limited: go ahead, send dupack now! */ |
| 3426 | } |
| 3427 | |
Eric Dumazet | 7970ddc | 2015-03-16 21:06:20 -0700 | [diff] [blame] | 3428 | /* Return true if we're currently rate-limiting out-of-window ACKs and |
| 3429 | * thus shouldn't send a dupack right now. We rate-limit dupacks in |
| 3430 | * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS |
| 3431 | * attacks that send repeated SYNs or ACKs for the same connection. To |
| 3432 | * do this, we do not send a duplicate SYNACK or ACK if the remote |
| 3433 | * endpoint is sending out-of-window SYNs or pure ACKs at a high rate. |
| 3434 | */ |
| 3435 | bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb, |
| 3436 | int mib_idx, u32 *last_oow_ack_time) |
| 3437 | { |
| 3438 | /* Data packets without SYNs are not likely part of an ACK loop. */ |
| 3439 | if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) && |
| 3440 | !tcp_hdr(skb)->syn) |
Jason Baron | 083ae30 | 2016-07-14 11:38:40 -0400 | [diff] [blame] | 3441 | return false; |
Eric Dumazet | 7970ddc | 2015-03-16 21:06:20 -0700 | [diff] [blame] | 3442 | |
Jason Baron | 083ae30 | 2016-07-14 11:38:40 -0400 | [diff] [blame] | 3443 | return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time); |
Eric Dumazet | 7970ddc | 2015-03-16 21:06:20 -0700 | [diff] [blame] | 3444 | } |
| 3445 | |
Eric Dumazet | 354e4aa | 2012-10-21 19:57:11 +0000 | [diff] [blame] | 3446 | /* RFC 5961 7 [ACK Throttling] */ |
Neal Cardwell | f2b2c58 | 2015-02-06 16:04:40 -0500 | [diff] [blame] | 3447 | static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb) |
Eric Dumazet | 354e4aa | 2012-10-21 19:57:11 +0000 | [diff] [blame] | 3448 | { |
| 3449 | /* unprotected vars, we dont care of overwrites */ |
| 3450 | static u32 challenge_timestamp; |
| 3451 | static unsigned int challenge_count; |
Neal Cardwell | f2b2c58 | 2015-02-06 16:04:40 -0500 | [diff] [blame] | 3452 | struct tcp_sock *tp = tcp_sk(sk); |
Eric Dumazet | 75ff39c | 2016-07-10 10:04:02 +0200 | [diff] [blame] | 3453 | u32 count, now; |
Eric Dumazet | 354e4aa | 2012-10-21 19:57:11 +0000 | [diff] [blame] | 3454 | |
Neal Cardwell | f2b2c58 | 2015-02-06 16:04:40 -0500 | [diff] [blame] | 3455 | /* First check our per-socket dupack rate limit. */ |
Jason Baron | 083ae30 | 2016-07-14 11:38:40 -0400 | [diff] [blame] | 3456 | if (__tcp_oow_rate_limited(sock_net(sk), |
| 3457 | LINUX_MIB_TCPACKSKIPPEDCHALLENGE, |
| 3458 | &tp->last_oow_ack_time)) |
Neal Cardwell | f2b2c58 | 2015-02-06 16:04:40 -0500 | [diff] [blame] | 3459 | return; |
| 3460 | |
Eric Dumazet | 75ff39c | 2016-07-10 10:04:02 +0200 | [diff] [blame] | 3461 | /* Then check host-wide RFC 5961 rate limit. */ |
Neal Cardwell | f2b2c58 | 2015-02-06 16:04:40 -0500 | [diff] [blame] | 3462 | now = jiffies / HZ; |
Eric Dumazet | 354e4aa | 2012-10-21 19:57:11 +0000 | [diff] [blame] | 3463 | if (now != challenge_timestamp) { |
Eric Dumazet | 75ff39c | 2016-07-10 10:04:02 +0200 | [diff] [blame] | 3464 | u32 half = (sysctl_tcp_challenge_ack_limit + 1) >> 1; |
| 3465 | |
Eric Dumazet | 354e4aa | 2012-10-21 19:57:11 +0000 | [diff] [blame] | 3466 | challenge_timestamp = now; |
Eric Dumazet | 75ff39c | 2016-07-10 10:04:02 +0200 | [diff] [blame] | 3467 | WRITE_ONCE(challenge_count, half + |
| 3468 | prandom_u32_max(sysctl_tcp_challenge_ack_limit)); |
Eric Dumazet | 354e4aa | 2012-10-21 19:57:11 +0000 | [diff] [blame] | 3469 | } |
Eric Dumazet | 75ff39c | 2016-07-10 10:04:02 +0200 | [diff] [blame] | 3470 | count = READ_ONCE(challenge_count); |
| 3471 | if (count > 0) { |
| 3472 | WRITE_ONCE(challenge_count, count - 1); |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 3473 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK); |
Eric Dumazet | 354e4aa | 2012-10-21 19:57:11 +0000 | [diff] [blame] | 3474 | tcp_send_ack(sk); |
| 3475 | } |
| 3476 | } |
| 3477 | |
Eric Dumazet | 12fb3dd | 2013-04-19 07:19:48 +0000 | [diff] [blame] | 3478 | static void tcp_store_ts_recent(struct tcp_sock *tp) |
| 3479 | { |
| 3480 | tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval; |
| 3481 | tp->rx_opt.ts_recent_stamp = get_seconds(); |
| 3482 | } |
| 3483 | |
| 3484 | static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq) |
| 3485 | { |
| 3486 | if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) { |
| 3487 | /* PAWS bug workaround wrt. ACK frames, the PAWS discard |
| 3488 | * extra check below makes sure this can only happen |
| 3489 | * for pure ACK frames. -DaveM |
| 3490 | * |
| 3491 | * Not only, also it occurs for expired timestamps. |
| 3492 | */ |
| 3493 | |
| 3494 | if (tcp_paws_check(&tp->rx_opt, 0)) |
| 3495 | tcp_store_ts_recent(tp); |
| 3496 | } |
| 3497 | } |
| 3498 | |
Nandita Dukkipati | 9b717a8 | 2013-03-11 10:00:44 +0000 | [diff] [blame] | 3499 | /* This routine deals with acks during a TLP episode. |
Sébastien Barré | 08abdff | 2015-01-12 10:30:40 +0100 | [diff] [blame] | 3500 | * We mark the end of a TLP episode on receiving TLP dupack or when |
| 3501 | * ack is after tlp_high_seq. |
Nandita Dukkipati | 9b717a8 | 2013-03-11 10:00:44 +0000 | [diff] [blame] | 3502 | * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe. |
| 3503 | */ |
| 3504 | static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag) |
| 3505 | { |
| 3506 | struct tcp_sock *tp = tcp_sk(sk); |
Nandita Dukkipati | 9b717a8 | 2013-03-11 10:00:44 +0000 | [diff] [blame] | 3507 | |
Sébastien Barré | 08abdff | 2015-01-12 10:30:40 +0100 | [diff] [blame] | 3508 | if (before(ack, tp->tlp_high_seq)) |
Nandita Dukkipati | 9b717a8 | 2013-03-11 10:00:44 +0000 | [diff] [blame] | 3509 | return; |
Nandita Dukkipati | 9b717a8 | 2013-03-11 10:00:44 +0000 | [diff] [blame] | 3510 | |
Sébastien Barré | 08abdff | 2015-01-12 10:30:40 +0100 | [diff] [blame] | 3511 | if (flag & FLAG_DSACKING_ACK) { |
| 3512 | /* This DSACK means original and TLP probe arrived; no loss */ |
Nandita Dukkipati | 9b717a8 | 2013-03-11 10:00:44 +0000 | [diff] [blame] | 3513 | tp->tlp_high_seq = 0; |
Sébastien Barré | 08abdff | 2015-01-12 10:30:40 +0100 | [diff] [blame] | 3514 | } else if (after(ack, tp->tlp_high_seq)) { |
| 3515 | /* ACK advances: there was a loss, so reduce cwnd. Reset |
| 3516 | * tlp_high_seq in tcp_init_cwnd_reduction() |
| 3517 | */ |
| 3518 | tcp_init_cwnd_reduction(sk); |
| 3519 | tcp_set_ca_state(sk, TCP_CA_CWR); |
| 3520 | tcp_end_cwnd_reduction(sk); |
| 3521 | tcp_try_keep_open(sk); |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 3522 | NET_INC_STATS(sock_net(sk), |
Eric Dumazet | 02a1d6e | 2016-04-27 16:44:39 -0700 | [diff] [blame] | 3523 | LINUX_MIB_TCPLOSSPROBERECOVERY); |
Sébastien Barré | 08abdff | 2015-01-12 10:30:40 +0100 | [diff] [blame] | 3524 | } else if (!(flag & (FLAG_SND_UNA_ADVANCED | |
| 3525 | FLAG_NOT_DUP | FLAG_DATA_SACKED))) { |
| 3526 | /* Pure dupack: original and TLP probe arrived; no loss */ |
| 3527 | tp->tlp_high_seq = 0; |
Nandita Dukkipati | 9b717a8 | 2013-03-11 10:00:44 +0000 | [diff] [blame] | 3528 | } |
| 3529 | } |
| 3530 | |
Florian Westphal | 7354c8c | 2014-09-26 22:37:34 +0200 | [diff] [blame] | 3531 | static inline void tcp_in_ack_event(struct sock *sk, u32 flags) |
| 3532 | { |
| 3533 | const struct inet_connection_sock *icsk = inet_csk(sk); |
| 3534 | |
| 3535 | if (icsk->icsk_ca_ops->in_ack_event) |
| 3536 | icsk->icsk_ca_ops->in_ack_event(sk, flags); |
| 3537 | } |
| 3538 | |
Yuchung Cheng | e662ca4 | 2016-02-02 10:33:04 -0800 | [diff] [blame] | 3539 | /* Congestion control has updated the cwnd already. So if we're in |
| 3540 | * loss recovery then now we do any new sends (for FRTO) or |
| 3541 | * retransmits (for CA_Loss or CA_recovery) that make sense. |
| 3542 | */ |
| 3543 | static void tcp_xmit_recovery(struct sock *sk, int rexmit) |
| 3544 | { |
| 3545 | struct tcp_sock *tp = tcp_sk(sk); |
| 3546 | |
| 3547 | if (rexmit == REXMIT_NONE) |
| 3548 | return; |
| 3549 | |
| 3550 | if (unlikely(rexmit == 2)) { |
| 3551 | __tcp_push_pending_frames(sk, tcp_current_mss(sk), |
| 3552 | TCP_NAGLE_OFF); |
| 3553 | if (after(tp->snd_nxt, tp->high_seq)) |
| 3554 | return; |
| 3555 | tp->frto = 0; |
| 3556 | } |
| 3557 | tcp_xmit_retransmit_queue(sk); |
| 3558 | } |
| 3559 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3560 | /* This routine deals with incoming acks, but not outgoing ones. */ |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 3561 | static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3562 | { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 3563 | struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3564 | struct tcp_sock *tp = tcp_sk(sk); |
Kenneth Klette Jonassen | 196da97 | 2015-05-01 01:10:57 +0200 | [diff] [blame] | 3565 | struct tcp_sacktag_state sack_state; |
Yuchung Cheng | b9f6482 | 2016-09-19 23:39:14 -0400 | [diff] [blame] | 3566 | struct rate_sample rs = { .prior_delivered = 0 }; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3567 | u32 prior_snd_una = tp->snd_una; |
| 3568 | u32 ack_seq = TCP_SKB_CB(skb)->seq; |
| 3569 | u32 ack = TCP_SKB_CB(skb)->ack_seq; |
Neal Cardwell | 7d2b55f | 2011-11-16 08:58:01 +0000 | [diff] [blame] | 3570 | bool is_dupack = false; |
Ilpo Järvinen | c7caf8d | 2007-11-10 21:22:18 -0800 | [diff] [blame] | 3571 | u32 prior_fackets; |
Nandita Dukkipati | 35f079e | 2013-05-21 15:12:07 +0000 | [diff] [blame] | 3572 | int prior_packets = tp->packets_out; |
Yuchung Cheng | b9f6482 | 2016-09-19 23:39:14 -0400 | [diff] [blame] | 3573 | u32 delivered = tp->delivered; |
| 3574 | u32 lost = tp->lost; |
Yuchung Cheng | 6804973 | 2013-05-29 14:20:11 +0000 | [diff] [blame] | 3575 | int acked = 0; /* Number of packets newly acked */ |
Yuchung Cheng | e662ca4 | 2016-02-02 10:33:04 -0800 | [diff] [blame] | 3576 | int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */ |
Yuchung Cheng | b9f6482 | 2016-09-19 23:39:14 -0400 | [diff] [blame] | 3577 | struct skb_mstamp now; |
Kenneth Klette Jonassen | 196da97 | 2015-05-01 01:10:57 +0200 | [diff] [blame] | 3578 | |
Kenneth Klette Jonassen | 31231a8 | 2015-05-01 01:10:58 +0200 | [diff] [blame] | 3579 | sack_state.first_sackt.v64 = 0; |
Yuchung Cheng | b9f6482 | 2016-09-19 23:39:14 -0400 | [diff] [blame] | 3580 | sack_state.rate = &rs; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3581 | |
Eric Dumazet | ad971f6 | 2014-10-11 15:17:29 -0700 | [diff] [blame] | 3582 | /* We very likely will need to access write queue head. */ |
| 3583 | prefetchw(sk->sk_write_queue.next); |
| 3584 | |
John Dykstra | 96e0bf4 | 2009-03-22 21:49:57 -0700 | [diff] [blame] | 3585 | /* If the ack is older than previous acks |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3586 | * then we can probably ignore it. |
| 3587 | */ |
Eric Dumazet | 354e4aa | 2012-10-21 19:57:11 +0000 | [diff] [blame] | 3588 | if (before(ack, prior_snd_una)) { |
| 3589 | /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */ |
| 3590 | if (before(ack, prior_snd_una - tp->max_window)) { |
Neal Cardwell | f2b2c58 | 2015-02-06 16:04:40 -0500 | [diff] [blame] | 3591 | tcp_send_challenge_ack(sk, skb); |
Eric Dumazet | 354e4aa | 2012-10-21 19:57:11 +0000 | [diff] [blame] | 3592 | return -1; |
| 3593 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3594 | goto old_ack; |
Eric Dumazet | 354e4aa | 2012-10-21 19:57:11 +0000 | [diff] [blame] | 3595 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3596 | |
John Dykstra | 96e0bf4 | 2009-03-22 21:49:57 -0700 | [diff] [blame] | 3597 | /* If the ack includes data we haven't sent yet, discard |
| 3598 | * this segment (RFC793 Section 3.9). |
| 3599 | */ |
| 3600 | if (after(ack, tp->snd_nxt)) |
| 3601 | goto invalid_ack; |
| 3602 | |
Yuchung Cheng | b9f6482 | 2016-09-19 23:39:14 -0400 | [diff] [blame] | 3603 | skb_mstamp_get(&now); |
| 3604 | |
Nandita Dukkipati | 6ba8a3b | 2013-03-11 10:00:43 +0000 | [diff] [blame] | 3605 | if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS || |
| 3606 | icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) |
Yuchung Cheng | 750ea2b | 2012-05-02 13:30:04 +0000 | [diff] [blame] | 3607 | tcp_rearm_rto(sk); |
| 3608 | |
Neal Cardwell | 0c9ab09 | 2014-08-14 16:13:07 -0400 | [diff] [blame] | 3609 | if (after(ack, prior_snd_una)) { |
Ilpo Järvinen | 2e60529 | 2007-08-02 19:46:58 -0700 | [diff] [blame] | 3610 | flag |= FLAG_SND_UNA_ADVANCED; |
Neal Cardwell | 0c9ab09 | 2014-08-14 16:13:07 -0400 | [diff] [blame] | 3611 | icsk->icsk_retransmits = 0; |
| 3612 | } |
Ilpo Järvinen | 2e60529 | 2007-08-02 19:46:58 -0700 | [diff] [blame] | 3613 | |
Ilpo Järvinen | c7caf8d | 2007-11-10 21:22:18 -0800 | [diff] [blame] | 3614 | prior_fackets = tp->fackets_out; |
Yuchung Cheng | b9f6482 | 2016-09-19 23:39:14 -0400 | [diff] [blame] | 3615 | rs.prior_in_flight = tcp_packets_in_flight(tp); |
Ilpo Järvinen | c7caf8d | 2007-11-10 21:22:18 -0800 | [diff] [blame] | 3616 | |
Eric Dumazet | 12fb3dd | 2013-04-19 07:19:48 +0000 | [diff] [blame] | 3617 | /* ts_recent update must be made after we are sure that the packet |
| 3618 | * is in window. |
| 3619 | */ |
| 3620 | if (flag & FLAG_UPDATE_TS_RECENT) |
| 3621 | tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); |
| 3622 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 3623 | if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3624 | /* Window is constant, pure forward advance. |
| 3625 | * No more checks are required. |
| 3626 | * Note, we use the fact that SND.UNA>=SND.WL2. |
| 3627 | */ |
Hantzis Fotis | ee7537b | 2009-03-02 22:42:02 -0800 | [diff] [blame] | 3628 | tcp_update_wl(tp, ack_seq); |
Eric Dumazet | 0df48c2 | 2015-04-28 15:28:17 -0700 | [diff] [blame] | 3629 | tcp_snd_una_update(tp, ack); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3630 | flag |= FLAG_WIN_UPDATE; |
| 3631 | |
Florian Westphal | 9890092 | 2014-09-26 22:37:35 +0200 | [diff] [blame] | 3632 | tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE); |
Stephen Hemminger | 317a76f | 2005-06-23 12:19:55 -0700 | [diff] [blame] | 3633 | |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 3634 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3635 | } else { |
Florian Westphal | 9890092 | 2014-09-26 22:37:35 +0200 | [diff] [blame] | 3636 | u32 ack_ev_flags = CA_ACK_SLOWPATH; |
| 3637 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3638 | if (ack_seq != TCP_SKB_CB(skb)->end_seq) |
| 3639 | flag |= FLAG_DATA; |
| 3640 | else |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 3641 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3642 | |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 3643 | flag |= tcp_ack_update_window(sk, skb, ack, ack_seq); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3644 | |
| 3645 | if (TCP_SKB_CB(skb)->sacked) |
Yuchung Cheng | 59c9af4 | 2013-07-22 16:20:47 -0700 | [diff] [blame] | 3646 | flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una, |
Kenneth Klette Jonassen | 196da97 | 2015-05-01 01:10:57 +0200 | [diff] [blame] | 3647 | &sack_state); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3648 | |
Florian Westphal | 735d383 | 2014-09-29 13:08:30 +0200 | [diff] [blame] | 3649 | if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3650 | flag |= FLAG_ECE; |
Florian Westphal | 9890092 | 2014-09-26 22:37:35 +0200 | [diff] [blame] | 3651 | ack_ev_flags |= CA_ACK_ECE; |
| 3652 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3653 | |
Florian Westphal | 9890092 | 2014-09-26 22:37:35 +0200 | [diff] [blame] | 3654 | if (flag & FLAG_WIN_UPDATE) |
| 3655 | ack_ev_flags |= CA_ACK_WIN_UPDATE; |
| 3656 | |
| 3657 | tcp_in_ack_event(sk, ack_ev_flags); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3658 | } |
| 3659 | |
| 3660 | /* We passed data and got it acked, remove any soft error |
| 3661 | * log. Something worked... |
| 3662 | */ |
| 3663 | sk->sk_err_soft = 0; |
David S. Miller | 4b53fb6 | 2008-07-23 16:38:45 -0700 | [diff] [blame] | 3664 | icsk->icsk_probes_out = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3665 | tp->rcv_tstamp = tcp_time_stamp; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3666 | if (!prior_packets) |
| 3667 | goto no_queue; |
| 3668 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3669 | /* See if we can take anything off of the retransmit queue. */ |
Yuchung Cheng | 3ebd8871 | 2016-02-02 10:33:07 -0800 | [diff] [blame] | 3670 | flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una, &acked, |
Yuchung Cheng | b9f6482 | 2016-09-19 23:39:14 -0400 | [diff] [blame] | 3671 | &sack_state, &now); |
Nandita Dukkipati | a262f0c | 2011-08-21 20:21:57 +0000 | [diff] [blame] | 3672 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 3673 | if (tcp_ack_is_dubious(sk, flag)) { |
Neal Cardwell | 7d2b55f | 2011-11-16 08:58:01 +0000 | [diff] [blame] | 3674 | is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP)); |
Yuchung Cheng | 31ba0c1 | 2016-02-02 10:33:05 -0800 | [diff] [blame] | 3675 | tcp_fastretrans_alert(sk, acked, is_dupack, &flag, &rexmit); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3676 | } |
Nandita Dukkipati | 9b717a8 | 2013-03-11 10:00:44 +0000 | [diff] [blame] | 3677 | if (tp->tlp_high_seq) |
| 3678 | tcp_process_tlp_ack(sk, ack, flag); |
| 3679 | |
David S. Miller | 5110effe | 2012-07-02 02:21:03 -0700 | [diff] [blame] | 3680 | if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) { |
| 3681 | struct dst_entry *dst = __sk_dst_get(sk); |
| 3682 | if (dst) |
| 3683 | dst_confirm(dst); |
| 3684 | } |
Nandita Dukkipati | 6ba8a3b | 2013-03-11 10:00:43 +0000 | [diff] [blame] | 3685 | |
| 3686 | if (icsk->icsk_pending == ICSK_TIME_RETRANS) |
| 3687 | tcp_schedule_loss_probe(sk); |
Yuchung Cheng | b9f6482 | 2016-09-19 23:39:14 -0400 | [diff] [blame] | 3688 | delivered = tp->delivered - delivered; /* freshly ACKed or SACKed */ |
| 3689 | lost = tp->lost - lost; /* freshly marked lost */ |
| 3690 | tcp_rate_gen(sk, delivered, lost, &now, &rs); |
Yuchung Cheng | c040276 | 2016-09-19 23:39:21 -0400 | [diff] [blame] | 3691 | tcp_cong_control(sk, ack, delivered, flag, &rs); |
Yuchung Cheng | e662ca4 | 2016-02-02 10:33:04 -0800 | [diff] [blame] | 3692 | tcp_xmit_recovery(sk, rexmit); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3693 | return 1; |
| 3694 | |
| 3695 | no_queue: |
Neal Cardwell | 5628adf | 2011-11-16 08:58:02 +0000 | [diff] [blame] | 3696 | /* If data was DSACKed, see if we can undo a cwnd reduction. */ |
| 3697 | if (flag & FLAG_DSACKING_ACK) |
Yuchung Cheng | 31ba0c1 | 2016-02-02 10:33:05 -0800 | [diff] [blame] | 3698 | tcp_fastretrans_alert(sk, acked, is_dupack, &flag, &rexmit); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3699 | /* If this ack opens up a zero window, clear backoff. It was |
| 3700 | * being used to time the probes, and is probably far higher than |
| 3701 | * it needs to be for normal retransmission. |
| 3702 | */ |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 3703 | if (tcp_send_head(sk)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3704 | tcp_ack_probe(sk); |
Nandita Dukkipati | 9b717a8 | 2013-03-11 10:00:44 +0000 | [diff] [blame] | 3705 | |
| 3706 | if (tp->tlp_high_seq) |
| 3707 | tcp_process_tlp_ack(sk, ack, flag); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3708 | return 1; |
| 3709 | |
John Dykstra | 96e0bf4 | 2009-03-22 21:49:57 -0700 | [diff] [blame] | 3710 | invalid_ack: |
| 3711 | SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt); |
| 3712 | return -1; |
| 3713 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3714 | old_ack: |
Neal Cardwell | e95ae2f | 2011-11-16 08:58:03 +0000 | [diff] [blame] | 3715 | /* If data was SACKed, tag it and see if we should send more data. |
| 3716 | * If data was DSACKed, see if we can undo a cwnd reduction. |
| 3717 | */ |
Ilpo Järvinen | 8aca6cb | 2008-06-04 11:34:22 -0700 | [diff] [blame] | 3718 | if (TCP_SKB_CB(skb)->sacked) { |
Yuchung Cheng | 59c9af4 | 2013-07-22 16:20:47 -0700 | [diff] [blame] | 3719 | flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una, |
Kenneth Klette Jonassen | 196da97 | 2015-05-01 01:10:57 +0200 | [diff] [blame] | 3720 | &sack_state); |
Yuchung Cheng | 31ba0c1 | 2016-02-02 10:33:05 -0800 | [diff] [blame] | 3721 | tcp_fastretrans_alert(sk, acked, is_dupack, &flag, &rexmit); |
Yuchung Cheng | e662ca4 | 2016-02-02 10:33:04 -0800 | [diff] [blame] | 3722 | tcp_xmit_recovery(sk, rexmit); |
Ilpo Järvinen | 8aca6cb | 2008-06-04 11:34:22 -0700 | [diff] [blame] | 3723 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3724 | |
John Dykstra | 96e0bf4 | 2009-03-22 21:49:57 -0700 | [diff] [blame] | 3725 | SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3726 | return 0; |
| 3727 | } |
| 3728 | |
Daniel Lee | 7f9b838 | 2015-04-06 14:37:26 -0700 | [diff] [blame] | 3729 | static void tcp_parse_fastopen_option(int len, const unsigned char *cookie, |
| 3730 | bool syn, struct tcp_fastopen_cookie *foc, |
| 3731 | bool exp_opt) |
| 3732 | { |
| 3733 | /* Valid only in SYN or SYN-ACK with an even length. */ |
| 3734 | if (!foc || !syn || len < 0 || (len & 1)) |
| 3735 | return; |
| 3736 | |
| 3737 | if (len >= TCP_FASTOPEN_COOKIE_MIN && |
| 3738 | len <= TCP_FASTOPEN_COOKIE_MAX) |
| 3739 | memcpy(foc->val, cookie, len); |
| 3740 | else if (len != 0) |
| 3741 | len = -1; |
| 3742 | foc->len = len; |
| 3743 | foc->exp = exp_opt; |
| 3744 | } |
| 3745 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3746 | /* Look for tcp options. Normally only called on SYN and SYNACK packets. |
| 3747 | * But, this can also be called on packets in the established flow when |
| 3748 | * the fast version below fails. |
| 3749 | */ |
Christoph Paasch | 1a2c618 | 2013-03-17 08:23:34 +0000 | [diff] [blame] | 3750 | void tcp_parse_options(const struct sk_buff *skb, |
| 3751 | struct tcp_options_received *opt_rx, int estab, |
Yuchung Cheng | 2100c8d | 2012-07-19 06:43:05 +0000 | [diff] [blame] | 3752 | struct tcp_fastopen_cookie *foc) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3753 | { |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 3754 | const unsigned char *ptr; |
| 3755 | const struct tcphdr *th = tcp_hdr(skb); |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 3756 | int length = (th->doff * 4) - sizeof(struct tcphdr); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3757 | |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 3758 | ptr = (const unsigned char *)(th + 1); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3759 | opt_rx->saw_tstamp = 0; |
| 3760 | |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 3761 | while (length > 0) { |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 3762 | int opcode = *ptr++; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3763 | int opsize; |
| 3764 | |
| 3765 | switch (opcode) { |
Ilpo Järvinen | f038ac8 | 2008-01-03 20:36:55 -0800 | [diff] [blame] | 3766 | case TCPOPT_EOL: |
| 3767 | return; |
| 3768 | case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ |
| 3769 | length--; |
| 3770 | continue; |
| 3771 | default: |
| 3772 | opsize = *ptr++; |
| 3773 | if (opsize < 2) /* "silly options" */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3774 | return; |
Ilpo Järvinen | f038ac8 | 2008-01-03 20:36:55 -0800 | [diff] [blame] | 3775 | if (opsize > length) |
| 3776 | return; /* don't parse partial options */ |
| 3777 | switch (opcode) { |
| 3778 | case TCPOPT_MSS: |
| 3779 | if (opsize == TCPOLEN_MSS && th->syn && !estab) { |
Harvey Harrison | d3e2ce3 | 2008-05-02 16:26:16 -0700 | [diff] [blame] | 3780 | u16 in_mss = get_unaligned_be16(ptr); |
Ilpo Järvinen | f038ac8 | 2008-01-03 20:36:55 -0800 | [diff] [blame] | 3781 | if (in_mss) { |
| 3782 | if (opt_rx->user_mss && |
| 3783 | opt_rx->user_mss < in_mss) |
| 3784 | in_mss = opt_rx->user_mss; |
| 3785 | opt_rx->mss_clamp = in_mss; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3786 | } |
Stephen Hemminger | 3ff50b7 | 2007-04-20 17:09:22 -0700 | [diff] [blame] | 3787 | } |
Ilpo Järvinen | f038ac8 | 2008-01-03 20:36:55 -0800 | [diff] [blame] | 3788 | break; |
| 3789 | case TCPOPT_WINDOW: |
| 3790 | if (opsize == TCPOLEN_WINDOW && th->syn && |
David S. Miller | bb5b7c1 | 2009-12-15 20:56:42 -0800 | [diff] [blame] | 3791 | !estab && sysctl_tcp_window_scaling) { |
Ilpo Järvinen | f038ac8 | 2008-01-03 20:36:55 -0800 | [diff] [blame] | 3792 | __u8 snd_wscale = *(__u8 *)ptr; |
| 3793 | opt_rx->wscale_ok = 1; |
| 3794 | if (snd_wscale > 14) { |
Joe Perches | e87cc47 | 2012-05-13 21:56:26 +0000 | [diff] [blame] | 3795 | net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n", |
| 3796 | __func__, |
| 3797 | snd_wscale); |
Ilpo Järvinen | f038ac8 | 2008-01-03 20:36:55 -0800 | [diff] [blame] | 3798 | snd_wscale = 14; |
| 3799 | } |
| 3800 | opt_rx->snd_wscale = snd_wscale; |
| 3801 | } |
| 3802 | break; |
| 3803 | case TCPOPT_TIMESTAMP: |
| 3804 | if ((opsize == TCPOLEN_TIMESTAMP) && |
| 3805 | ((estab && opt_rx->tstamp_ok) || |
David S. Miller | bb5b7c1 | 2009-12-15 20:56:42 -0800 | [diff] [blame] | 3806 | (!estab && sysctl_tcp_timestamps))) { |
Ilpo Järvinen | f038ac8 | 2008-01-03 20:36:55 -0800 | [diff] [blame] | 3807 | opt_rx->saw_tstamp = 1; |
Harvey Harrison | d3e2ce3 | 2008-05-02 16:26:16 -0700 | [diff] [blame] | 3808 | opt_rx->rcv_tsval = get_unaligned_be32(ptr); |
| 3809 | opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4); |
Ilpo Järvinen | f038ac8 | 2008-01-03 20:36:55 -0800 | [diff] [blame] | 3810 | } |
| 3811 | break; |
| 3812 | case TCPOPT_SACK_PERM: |
| 3813 | if (opsize == TCPOLEN_SACK_PERM && th->syn && |
David S. Miller | bb5b7c1 | 2009-12-15 20:56:42 -0800 | [diff] [blame] | 3814 | !estab && sysctl_tcp_sack) { |
Vijay Subramanian | ab56222 | 2011-12-20 13:23:24 +0000 | [diff] [blame] | 3815 | opt_rx->sack_ok = TCP_SACK_SEEN; |
Ilpo Järvinen | f038ac8 | 2008-01-03 20:36:55 -0800 | [diff] [blame] | 3816 | tcp_sack_reset(opt_rx); |
| 3817 | } |
| 3818 | break; |
Stephen Hemminger | 3ff50b7 | 2007-04-20 17:09:22 -0700 | [diff] [blame] | 3819 | |
Ilpo Järvinen | f038ac8 | 2008-01-03 20:36:55 -0800 | [diff] [blame] | 3820 | case TCPOPT_SACK: |
| 3821 | if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) && |
| 3822 | !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) && |
| 3823 | opt_rx->sack_ok) { |
| 3824 | TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th; |
| 3825 | } |
| 3826 | break; |
| 3827 | #ifdef CONFIG_TCP_MD5SIG |
| 3828 | case TCPOPT_MD5SIG: |
| 3829 | /* |
| 3830 | * The MD5 Hash has already been |
| 3831 | * checked (see tcp_v{4,6}_do_rcv()). |
| 3832 | */ |
| 3833 | break; |
| 3834 | #endif |
Daniel Lee | 7f9b838 | 2015-04-06 14:37:26 -0700 | [diff] [blame] | 3835 | case TCPOPT_FASTOPEN: |
| 3836 | tcp_parse_fastopen_option( |
| 3837 | opsize - TCPOLEN_FASTOPEN_BASE, |
| 3838 | ptr, th->syn, foc, false); |
| 3839 | break; |
| 3840 | |
Yuchung Cheng | 2100c8d | 2012-07-19 06:43:05 +0000 | [diff] [blame] | 3841 | case TCPOPT_EXP: |
| 3842 | /* Fast Open option shares code 254 using a |
Daniel Lee | 7f9b838 | 2015-04-06 14:37:26 -0700 | [diff] [blame] | 3843 | * 16 bits magic number. |
Yuchung Cheng | 2100c8d | 2012-07-19 06:43:05 +0000 | [diff] [blame] | 3844 | */ |
Daniel Lee | 7f9b838 | 2015-04-06 14:37:26 -0700 | [diff] [blame] | 3845 | if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE && |
| 3846 | get_unaligned_be16(ptr) == |
| 3847 | TCPOPT_FASTOPEN_MAGIC) |
| 3848 | tcp_parse_fastopen_option(opsize - |
| 3849 | TCPOLEN_EXP_FASTOPEN_BASE, |
| 3850 | ptr + 2, th->syn, foc, true); |
Yuchung Cheng | 2100c8d | 2012-07-19 06:43:05 +0000 | [diff] [blame] | 3851 | break; |
| 3852 | |
| 3853 | } |
Ilpo Järvinen | f038ac8 | 2008-01-03 20:36:55 -0800 | [diff] [blame] | 3854 | ptr += opsize-2; |
| 3855 | length -= opsize; |
Stephen Hemminger | 3ff50b7 | 2007-04-20 17:09:22 -0700 | [diff] [blame] | 3856 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3857 | } |
| 3858 | } |
Eric Dumazet | 4bc2f18 | 2010-07-09 21:22:10 +0000 | [diff] [blame] | 3859 | EXPORT_SYMBOL(tcp_parse_options); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3860 | |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 3861 | static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th) |
Ilpo Järvinen | a4356b2 | 2008-08-23 05:12:29 -0700 | [diff] [blame] | 3862 | { |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 3863 | const __be32 *ptr = (const __be32 *)(th + 1); |
Ilpo Järvinen | a4356b2 | 2008-08-23 05:12:29 -0700 | [diff] [blame] | 3864 | |
| 3865 | if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
| 3866 | | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) { |
| 3867 | tp->rx_opt.saw_tstamp = 1; |
| 3868 | ++ptr; |
| 3869 | tp->rx_opt.rcv_tsval = ntohl(*ptr); |
| 3870 | ++ptr; |
Andrew Vagin | e3e1202 | 2013-08-27 12:21:55 +0400 | [diff] [blame] | 3871 | if (*ptr) |
| 3872 | tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset; |
| 3873 | else |
| 3874 | tp->rx_opt.rcv_tsecr = 0; |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 3875 | return true; |
Ilpo Järvinen | a4356b2 | 2008-08-23 05:12:29 -0700 | [diff] [blame] | 3876 | } |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 3877 | return false; |
Ilpo Järvinen | a4356b2 | 2008-08-23 05:12:29 -0700 | [diff] [blame] | 3878 | } |
| 3879 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3880 | /* Fast parse options. This hopes to only see timestamps. |
| 3881 | * If it is wrong it falls back on tcp_parse_options(). |
| 3882 | */ |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 3883 | static bool tcp_fast_parse_options(const struct sk_buff *skb, |
Christoph Paasch | 1a2c618 | 2013-03-17 08:23:34 +0000 | [diff] [blame] | 3884 | const struct tcphdr *th, struct tcp_sock *tp) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3885 | { |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 3886 | /* In the spirit of fast parsing, compare doff directly to constant |
| 3887 | * values. Because equality is used, short doff can be ignored here. |
| 3888 | */ |
| 3889 | if (th->doff == (sizeof(*th) / 4)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3890 | tp->rx_opt.saw_tstamp = 0; |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 3891 | return false; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3892 | } else if (tp->rx_opt.tstamp_ok && |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 3893 | th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) { |
Ilpo Järvinen | a4356b2 | 2008-08-23 05:12:29 -0700 | [diff] [blame] | 3894 | if (tcp_parse_aligned_timestamp(tp, th)) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 3895 | return true; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3896 | } |
Andrey Vagin | ee684b6 | 2013-02-11 05:50:19 +0000 | [diff] [blame] | 3897 | |
Christoph Paasch | 1a2c618 | 2013-03-17 08:23:34 +0000 | [diff] [blame] | 3898 | tcp_parse_options(skb, &tp->rx_opt, 1, NULL); |
Andrew Vagin | e3e1202 | 2013-08-27 12:21:55 +0400 | [diff] [blame] | 3899 | if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) |
Andrey Vagin | ee684b6 | 2013-02-11 05:50:19 +0000 | [diff] [blame] | 3900 | tp->rx_opt.rcv_tsecr -= tp->tsoffset; |
| 3901 | |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 3902 | return true; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3903 | } |
| 3904 | |
YOSHIFUJI Hideaki | 7d5d552 | 2008-04-17 12:29:53 +0900 | [diff] [blame] | 3905 | #ifdef CONFIG_TCP_MD5SIG |
| 3906 | /* |
| 3907 | * Parse MD5 Signature option |
| 3908 | */ |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 3909 | const u8 *tcp_parse_md5sig_option(const struct tcphdr *th) |
YOSHIFUJI Hideaki | 7d5d552 | 2008-04-17 12:29:53 +0900 | [diff] [blame] | 3910 | { |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 3911 | int length = (th->doff << 2) - sizeof(*th); |
| 3912 | const u8 *ptr = (const u8 *)(th + 1); |
YOSHIFUJI Hideaki | 7d5d552 | 2008-04-17 12:29:53 +0900 | [diff] [blame] | 3913 | |
| 3914 | /* If the TCP option is too short, we can short cut */ |
| 3915 | if (length < TCPOLEN_MD5SIG) |
| 3916 | return NULL; |
| 3917 | |
| 3918 | while (length > 0) { |
| 3919 | int opcode = *ptr++; |
| 3920 | int opsize; |
| 3921 | |
Weilong Chen | a22318e | 2013-12-23 14:37:26 +0800 | [diff] [blame] | 3922 | switch (opcode) { |
YOSHIFUJI Hideaki | 7d5d552 | 2008-04-17 12:29:53 +0900 | [diff] [blame] | 3923 | case TCPOPT_EOL: |
| 3924 | return NULL; |
| 3925 | case TCPOPT_NOP: |
| 3926 | length--; |
| 3927 | continue; |
| 3928 | default: |
| 3929 | opsize = *ptr++; |
| 3930 | if (opsize < 2 || opsize > length) |
| 3931 | return NULL; |
| 3932 | if (opcode == TCPOPT_MD5SIG) |
Dmitry Popov | ba78e2d | 2010-08-07 20:24:28 -0700 | [diff] [blame] | 3933 | return opsize == TCPOLEN_MD5SIG ? ptr : NULL; |
YOSHIFUJI Hideaki | 7d5d552 | 2008-04-17 12:29:53 +0900 | [diff] [blame] | 3934 | } |
| 3935 | ptr += opsize - 2; |
| 3936 | length -= opsize; |
| 3937 | } |
| 3938 | return NULL; |
| 3939 | } |
Eric Dumazet | 4bc2f18 | 2010-07-09 21:22:10 +0000 | [diff] [blame] | 3940 | EXPORT_SYMBOL(tcp_parse_md5sig_option); |
YOSHIFUJI Hideaki | 7d5d552 | 2008-04-17 12:29:53 +0900 | [diff] [blame] | 3941 | #endif |
| 3942 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3943 | /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM |
| 3944 | * |
| 3945 | * It is not fatal. If this ACK does _not_ change critical state (seqs, window) |
| 3946 | * it can pass through stack. So, the following predicate verifies that |
| 3947 | * this segment is not used for anything but congestion avoidance or |
| 3948 | * fast retransmit. Moreover, we even are able to eliminate most of such |
| 3949 | * second order effects, if we apply some small "replay" window (~RTO) |
| 3950 | * to timestamp space. |
| 3951 | * |
| 3952 | * All these measures still do not guarantee that we reject wrapped ACKs |
| 3953 | * on networks with high bandwidth, when sequence space is recycled fastly, |
| 3954 | * but it guarantees that such events will be very rare and do not affect |
| 3955 | * connection seriously. This doesn't look nice, but alas, PAWS is really |
| 3956 | * buggy extension. |
| 3957 | * |
| 3958 | * [ Later note. Even worse! It is buggy for segments _with_ data. RFC |
| 3959 | * states that events when retransmit arrives after original data are rare. |
| 3960 | * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is |
| 3961 | * the biggest problem on large power networks even with minor reordering. |
| 3962 | * OK, let's give it small replay window. If peer clock is even 1hz, it is safe |
| 3963 | * up to bandwidth of 18Gigabit/sec. 8) ] |
| 3964 | */ |
| 3965 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3966 | 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] | 3967 | { |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 3968 | const struct tcp_sock *tp = tcp_sk(sk); |
| 3969 | const struct tcphdr *th = tcp_hdr(skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3970 | u32 seq = TCP_SKB_CB(skb)->seq; |
| 3971 | u32 ack = TCP_SKB_CB(skb)->ack_seq; |
| 3972 | |
| 3973 | return (/* 1. Pure ACK with correct sequence number. */ |
| 3974 | (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) && |
| 3975 | |
| 3976 | /* 2. ... and duplicate ACK. */ |
| 3977 | ack == tp->snd_una && |
| 3978 | |
| 3979 | /* 3. ... and does not update window. */ |
| 3980 | !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) && |
| 3981 | |
| 3982 | /* 4. ... and sits in replay window. */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3983 | (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] | 3984 | } |
| 3985 | |
Vijay Subramanian | 67b95bd | 2012-07-19 21:32:18 +0000 | [diff] [blame] | 3986 | static inline bool tcp_paws_discard(const struct sock *sk, |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 3987 | const struct sk_buff *skb) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3988 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3989 | const struct tcp_sock *tp = tcp_sk(sk); |
Ilpo Järvinen | c887e6d | 2009-03-14 14:23:03 +0000 | [diff] [blame] | 3990 | |
| 3991 | return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) && |
| 3992 | !tcp_disordered_ack(sk, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3993 | } |
| 3994 | |
| 3995 | /* Check segment sequence number for validity. |
| 3996 | * |
| 3997 | * Segment controls are considered valid, if the segment |
| 3998 | * fits to the window after truncation to the window. Acceptability |
| 3999 | * of data (and SYN, FIN, of course) is checked separately. |
| 4000 | * See tcp_data_queue(), for example. |
| 4001 | * |
| 4002 | * Also, controls (RST is main one) are accepted using RCV.WUP instead |
| 4003 | * of RCV.NXT. Peer still did not advance his SND.UNA when we |
| 4004 | * delayed ACK, so that hisSND.UNA<=ourRCV.WUP. |
| 4005 | * (borrowed from freebsd) |
| 4006 | */ |
| 4007 | |
Vijay Subramanian | 67b95bd | 2012-07-19 21:32:18 +0000 | [diff] [blame] | 4008 | static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4009 | { |
| 4010 | return !before(end_seq, tp->rcv_wup) && |
| 4011 | !after(seq, tp->rcv_nxt + tcp_receive_window(tp)); |
| 4012 | } |
| 4013 | |
| 4014 | /* When we get a reset we do this. */ |
Jerry Chu | 1046716 | 2012-08-31 12:29:11 +0000 | [diff] [blame] | 4015 | void tcp_reset(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4016 | { |
| 4017 | /* We want the right error as BSD sees it (and indeed as we do). */ |
| 4018 | switch (sk->sk_state) { |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 4019 | case TCP_SYN_SENT: |
| 4020 | sk->sk_err = ECONNREFUSED; |
| 4021 | break; |
| 4022 | case TCP_CLOSE_WAIT: |
| 4023 | sk->sk_err = EPIPE; |
| 4024 | break; |
| 4025 | case TCP_CLOSE: |
| 4026 | return; |
| 4027 | default: |
| 4028 | sk->sk_err = ECONNRESET; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4029 | } |
Tom Marshall | a4d2580 | 2010-09-20 15:42:05 -0700 | [diff] [blame] | 4030 | /* This barrier is coupled with smp_rmb() in tcp_poll() */ |
| 4031 | smp_wmb(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4032 | |
| 4033 | if (!sock_flag(sk, SOCK_DEAD)) |
| 4034 | sk->sk_error_report(sk); |
| 4035 | |
| 4036 | tcp_done(sk); |
| 4037 | } |
| 4038 | |
| 4039 | /* |
| 4040 | * Process the FIN bit. This now behaves as it is supposed to work |
| 4041 | * and the FIN takes effect when it is validly part of sequence |
| 4042 | * space. Not before when we get holes. |
| 4043 | * |
| 4044 | * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT |
| 4045 | * (and thence onto LAST-ACK and finally, CLOSE, we never enter |
| 4046 | * TIME-WAIT) |
| 4047 | * |
| 4048 | * If we are in FINWAIT-1, a received FIN indicates simultaneous |
| 4049 | * close and we go into CLOSING (and later onto TIME-WAIT) |
| 4050 | * |
| 4051 | * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT. |
| 4052 | */ |
Eric Dumazet | e3e17b7 | 2016-02-06 11:16:28 -0800 | [diff] [blame] | 4053 | void tcp_fin(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4054 | { |
| 4055 | struct tcp_sock *tp = tcp_sk(sk); |
| 4056 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 4057 | inet_csk_schedule_ack(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4058 | |
| 4059 | sk->sk_shutdown |= RCV_SHUTDOWN; |
| 4060 | sock_set_flag(sk, SOCK_DONE); |
| 4061 | |
| 4062 | switch (sk->sk_state) { |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 4063 | case TCP_SYN_RECV: |
| 4064 | case TCP_ESTABLISHED: |
| 4065 | /* Move to CLOSE_WAIT */ |
| 4066 | tcp_set_state(sk, TCP_CLOSE_WAIT); |
Jon Maxwell | 2251ae4 | 2015-07-08 10:12:28 +1000 | [diff] [blame] | 4067 | inet_csk(sk)->icsk_ack.pingpong = 1; |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 4068 | break; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4069 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 4070 | case TCP_CLOSE_WAIT: |
| 4071 | case TCP_CLOSING: |
| 4072 | /* Received a retransmission of the FIN, do |
| 4073 | * nothing. |
| 4074 | */ |
| 4075 | break; |
| 4076 | case TCP_LAST_ACK: |
| 4077 | /* RFC793: Remain in the LAST-ACK state. */ |
| 4078 | break; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4079 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 4080 | case TCP_FIN_WAIT1: |
| 4081 | /* This case occurs when a simultaneous close |
| 4082 | * happens, we must ack the received FIN and |
| 4083 | * enter the CLOSING state. |
| 4084 | */ |
| 4085 | tcp_send_ack(sk); |
| 4086 | tcp_set_state(sk, TCP_CLOSING); |
| 4087 | break; |
| 4088 | case TCP_FIN_WAIT2: |
| 4089 | /* Received a FIN -- send ACK and enter TIME_WAIT. */ |
| 4090 | tcp_send_ack(sk); |
| 4091 | tcp_time_wait(sk, TCP_TIME_WAIT, 0); |
| 4092 | break; |
| 4093 | default: |
| 4094 | /* Only TCP_LISTEN and TCP_CLOSE are left, in these |
| 4095 | * cases we should never reach this piece of code. |
| 4096 | */ |
Joe Perches | 058bd4d | 2012-03-11 18:36:11 +0000 | [diff] [blame] | 4097 | pr_err("%s: Impossible, sk->sk_state=%d\n", |
Harvey Harrison | 0dc4787 | 2008-03-05 20:47:47 -0800 | [diff] [blame] | 4098 | __func__, sk->sk_state); |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 4099 | break; |
Stephen Hemminger | 3ff50b7 | 2007-04-20 17:09:22 -0700 | [diff] [blame] | 4100 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4101 | |
| 4102 | /* It _is_ possible, that we have something out-of-order _after_ FIN. |
| 4103 | * Probably, we should reset in this case. For now drop them. |
| 4104 | */ |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4105 | skb_rbtree_purge(&tp->out_of_order_queue); |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 4106 | if (tcp_is_sack(tp)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4107 | tcp_sack_reset(&tp->rx_opt); |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 4108 | sk_mem_reclaim(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4109 | |
| 4110 | if (!sock_flag(sk, SOCK_DEAD)) { |
| 4111 | sk->sk_state_change(sk); |
| 4112 | |
| 4113 | /* Do not send POLL_HUP for half duplex close. */ |
| 4114 | if (sk->sk_shutdown == SHUTDOWN_MASK || |
| 4115 | sk->sk_state == TCP_CLOSE) |
Pavel Emelyanov | 8d8ad9d | 2007-11-26 20:10:50 +0800 | [diff] [blame] | 4116 | sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4117 | else |
Pavel Emelyanov | 8d8ad9d | 2007-11-26 20:10:50 +0800 | [diff] [blame] | 4118 | sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4119 | } |
| 4120 | } |
| 4121 | |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 4122 | static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 4123 | u32 end_seq) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4124 | { |
| 4125 | if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) { |
| 4126 | if (before(seq, sp->start_seq)) |
| 4127 | sp->start_seq = seq; |
| 4128 | if (after(end_seq, sp->end_seq)) |
| 4129 | sp->end_seq = end_seq; |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 4130 | return true; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4131 | } |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 4132 | return false; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4133 | } |
| 4134 | |
Pavel Emelyanov | 1ed8346 | 2008-07-16 20:29:51 -0700 | [diff] [blame] | 4135 | static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4136 | { |
Pavel Emelyanov | 1ed8346 | 2008-07-16 20:29:51 -0700 | [diff] [blame] | 4137 | struct tcp_sock *tp = tcp_sk(sk); |
| 4138 | |
David S. Miller | bb5b7c1 | 2009-12-15 20:56:42 -0800 | [diff] [blame] | 4139 | if (tcp_is_sack(tp) && sysctl_tcp_dsack) { |
Pavel Emelyanov | 40b215e | 2008-07-03 01:05:41 -0700 | [diff] [blame] | 4140 | int mib_idx; |
| 4141 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4142 | if (before(seq, tp->rcv_nxt)) |
Pavel Emelyanov | 40b215e | 2008-07-03 01:05:41 -0700 | [diff] [blame] | 4143 | mib_idx = LINUX_MIB_TCPDSACKOLDSENT; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4144 | else |
Pavel Emelyanov | 40b215e | 2008-07-03 01:05:41 -0700 | [diff] [blame] | 4145 | mib_idx = LINUX_MIB_TCPDSACKOFOSENT; |
| 4146 | |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 4147 | NET_INC_STATS(sock_net(sk), mib_idx); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4148 | |
| 4149 | tp->rx_opt.dsack = 1; |
| 4150 | tp->duplicate_sack[0].start_seq = seq; |
| 4151 | tp->duplicate_sack[0].end_seq = end_seq; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4152 | } |
| 4153 | } |
| 4154 | |
Pavel Emelyanov | 1ed8346 | 2008-07-16 20:29:51 -0700 | [diff] [blame] | 4155 | static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4156 | { |
Pavel Emelyanov | 1ed8346 | 2008-07-16 20:29:51 -0700 | [diff] [blame] | 4157 | struct tcp_sock *tp = tcp_sk(sk); |
| 4158 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4159 | if (!tp->rx_opt.dsack) |
Pavel Emelyanov | 1ed8346 | 2008-07-16 20:29:51 -0700 | [diff] [blame] | 4160 | tcp_dsack_set(sk, seq, end_seq); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4161 | else |
| 4162 | tcp_sack_extend(tp->duplicate_sack, seq, end_seq); |
| 4163 | } |
| 4164 | |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 4165 | static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4166 | { |
| 4167 | struct tcp_sock *tp = tcp_sk(sk); |
| 4168 | |
| 4169 | if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && |
| 4170 | before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 4171 | NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 4172 | tcp_enter_quickack_mode(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4173 | |
David S. Miller | bb5b7c1 | 2009-12-15 20:56:42 -0800 | [diff] [blame] | 4174 | if (tcp_is_sack(tp) && sysctl_tcp_dsack) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4175 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; |
| 4176 | |
| 4177 | if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) |
| 4178 | end_seq = tp->rcv_nxt; |
Pavel Emelyanov | 1ed8346 | 2008-07-16 20:29:51 -0700 | [diff] [blame] | 4179 | tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4180 | } |
| 4181 | } |
| 4182 | |
| 4183 | tcp_send_ack(sk); |
| 4184 | } |
| 4185 | |
| 4186 | /* These routines update the SACK block as out-of-order packets arrive or |
| 4187 | * in-order packets close up the sequence space. |
| 4188 | */ |
| 4189 | static void tcp_sack_maybe_coalesce(struct tcp_sock *tp) |
| 4190 | { |
| 4191 | int this_sack; |
| 4192 | struct tcp_sack_block *sp = &tp->selective_acks[0]; |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 4193 | struct tcp_sack_block *swalk = sp + 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4194 | |
| 4195 | /* See if the recent change to the first SACK eats into |
| 4196 | * or hits the sequence space of other SACK blocks, if so coalesce. |
| 4197 | */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 4198 | for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4199 | if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) { |
| 4200 | int i; |
| 4201 | |
| 4202 | /* Zap SWALK, by moving every further SACK up by one slot. |
| 4203 | * Decrease num_sacks. |
| 4204 | */ |
| 4205 | tp->rx_opt.num_sacks--; |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 4206 | for (i = this_sack; i < tp->rx_opt.num_sacks; i++) |
| 4207 | sp[i] = sp[i + 1]; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4208 | continue; |
| 4209 | } |
| 4210 | this_sack++, swalk++; |
| 4211 | } |
| 4212 | } |
| 4213 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4214 | static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq) |
| 4215 | { |
| 4216 | struct tcp_sock *tp = tcp_sk(sk); |
| 4217 | struct tcp_sack_block *sp = &tp->selective_acks[0]; |
| 4218 | int cur_sacks = tp->rx_opt.num_sacks; |
| 4219 | int this_sack; |
| 4220 | |
| 4221 | if (!cur_sacks) |
| 4222 | goto new_sack; |
| 4223 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 4224 | for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4225 | if (tcp_sack_extend(sp, seq, end_seq)) { |
| 4226 | /* Rotate this_sack to the first one. */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 4227 | for (; this_sack > 0; this_sack--, sp--) |
Ilpo Järvinen | a0bffff | 2009-03-21 13:36:17 -0700 | [diff] [blame] | 4228 | swap(*sp, *(sp - 1)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4229 | if (cur_sacks > 1) |
| 4230 | tcp_sack_maybe_coalesce(tp); |
| 4231 | return; |
| 4232 | } |
| 4233 | } |
| 4234 | |
| 4235 | /* Could not find an adjacent existing SACK, build a new one, |
| 4236 | * put it at the front, and shift everyone else down. We |
| 4237 | * always know there is at least one SACK present already here. |
| 4238 | * |
| 4239 | * If the sack array is full, forget about the last one. |
| 4240 | */ |
Adam Langley | 4389dde | 2008-07-19 00:07:02 -0700 | [diff] [blame] | 4241 | if (this_sack >= TCP_NUM_SACKS) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4242 | this_sack--; |
| 4243 | tp->rx_opt.num_sacks--; |
| 4244 | sp--; |
| 4245 | } |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 4246 | for (; this_sack > 0; this_sack--, sp--) |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 4247 | *sp = *(sp - 1); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4248 | |
| 4249 | new_sack: |
| 4250 | /* Build the new head SACK, and we're done. */ |
| 4251 | sp->start_seq = seq; |
| 4252 | sp->end_seq = end_seq; |
| 4253 | tp->rx_opt.num_sacks++; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4254 | } |
| 4255 | |
| 4256 | /* RCV.NXT advances, some SACKs should be eaten. */ |
| 4257 | |
| 4258 | static void tcp_sack_remove(struct tcp_sock *tp) |
| 4259 | { |
| 4260 | struct tcp_sack_block *sp = &tp->selective_acks[0]; |
| 4261 | int num_sacks = tp->rx_opt.num_sacks; |
| 4262 | int this_sack; |
| 4263 | |
| 4264 | /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */ |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4265 | if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4266 | tp->rx_opt.num_sacks = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4267 | return; |
| 4268 | } |
| 4269 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 4270 | for (this_sack = 0; this_sack < num_sacks;) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4271 | /* Check if the start of the sack is covered by RCV.NXT. */ |
| 4272 | if (!before(tp->rcv_nxt, sp->start_seq)) { |
| 4273 | int i; |
| 4274 | |
| 4275 | /* RCV.NXT must cover all the block! */ |
Ilpo Järvinen | 547b792 | 2008-07-25 21:43:18 -0700 | [diff] [blame] | 4276 | WARN_ON(before(tp->rcv_nxt, sp->end_seq)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4277 | |
| 4278 | /* Zap this SACK, by moving forward any other SACKS. */ |
Weilong Chen | a22318e | 2013-12-23 14:37:26 +0800 | [diff] [blame] | 4279 | for (i = this_sack+1; i < num_sacks; i++) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4280 | tp->selective_acks[i-1] = tp->selective_acks[i]; |
| 4281 | num_sacks--; |
| 4282 | continue; |
| 4283 | } |
| 4284 | this_sack++; |
| 4285 | sp++; |
| 4286 | } |
Ilpo Järvinen | 5861f8e | 2009-03-14 14:23:01 +0000 | [diff] [blame] | 4287 | tp->rx_opt.num_sacks = num_sacks; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4288 | } |
| 4289 | |
Eric Dumazet | 1402d36 | 2012-04-23 07:11:42 +0000 | [diff] [blame] | 4290 | /** |
| 4291 | * tcp_try_coalesce - try to merge skb to prior one |
| 4292 | * @sk: socket |
| 4293 | * @to: prior buffer |
| 4294 | * @from: buffer to add in queue |
Eric Dumazet | 923dd34 | 2012-05-02 07:55:58 +0000 | [diff] [blame] | 4295 | * @fragstolen: pointer to boolean |
Eric Dumazet | 1402d36 | 2012-04-23 07:11:42 +0000 | [diff] [blame] | 4296 | * |
| 4297 | * Before queueing skb @from after @to, try to merge them |
| 4298 | * to reduce overall memory use and queue lengths, if cost is small. |
| 4299 | * Packets in ofo or receive queues can stay a long time. |
| 4300 | * Better try to coalesce them right now to avoid future collapses. |
Eric Dumazet | 783c175 | 2012-04-23 17:34:36 +0000 | [diff] [blame] | 4301 | * Returns true if caller should free @from instead of queueing it |
Eric Dumazet | 1402d36 | 2012-04-23 07:11:42 +0000 | [diff] [blame] | 4302 | */ |
Eric Dumazet | 783c175 | 2012-04-23 17:34:36 +0000 | [diff] [blame] | 4303 | static bool tcp_try_coalesce(struct sock *sk, |
| 4304 | struct sk_buff *to, |
Eric Dumazet | 329033f | 2012-04-27 00:38:33 +0000 | [diff] [blame] | 4305 | struct sk_buff *from, |
| 4306 | bool *fragstolen) |
Eric Dumazet | 1402d36 | 2012-04-23 07:11:42 +0000 | [diff] [blame] | 4307 | { |
Eric Dumazet | bad43ca | 2012-05-19 03:02:02 +0000 | [diff] [blame] | 4308 | int delta; |
Eric Dumazet | 1402d36 | 2012-04-23 07:11:42 +0000 | [diff] [blame] | 4309 | |
Eric Dumazet | 329033f | 2012-04-27 00:38:33 +0000 | [diff] [blame] | 4310 | *fragstolen = false; |
Alexander Duyck | 34a802a | 2012-05-02 21:19:09 +0000 | [diff] [blame] | 4311 | |
Eric Dumazet | 1ca7ee3 | 2012-05-23 17:51:37 +0000 | [diff] [blame] | 4312 | /* Its possible this segment overlaps with prior segment in queue */ |
| 4313 | if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq) |
| 4314 | return false; |
| 4315 | |
Eric Dumazet | bad43ca | 2012-05-19 03:02:02 +0000 | [diff] [blame] | 4316 | if (!skb_try_coalesce(to, from, fragstolen, &delta)) |
Eric Dumazet | 783c175 | 2012-04-23 17:34:36 +0000 | [diff] [blame] | 4317 | return false; |
Alexander Duyck | 34a802a | 2012-05-02 21:19:09 +0000 | [diff] [blame] | 4318 | |
Alexander Duyck | 34a802a | 2012-05-02 21:19:09 +0000 | [diff] [blame] | 4319 | atomic_add(delta, &sk->sk_rmem_alloc); |
| 4320 | sk_mem_charge(sk, delta); |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 4321 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE); |
Alexander Duyck | 34a802a | 2012-05-02 21:19:09 +0000 | [diff] [blame] | 4322 | TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq; |
| 4323 | TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq; |
Eric Dumazet | e93a043 | 2014-09-15 04:19:52 -0700 | [diff] [blame] | 4324 | TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags; |
Alexander Duyck | 34a802a | 2012-05-02 21:19:09 +0000 | [diff] [blame] | 4325 | return true; |
Eric Dumazet | 1402d36 | 2012-04-23 07:11:42 +0000 | [diff] [blame] | 4326 | } |
| 4327 | |
Eric Dumazet | 532182c | 2016-04-01 08:52:19 -0700 | [diff] [blame] | 4328 | static void tcp_drop(struct sock *sk, struct sk_buff *skb) |
| 4329 | { |
| 4330 | sk_drops_add(sk, skb); |
| 4331 | __kfree_skb(skb); |
| 4332 | } |
| 4333 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4334 | /* This one checks to see if we can put data from the |
| 4335 | * out_of_order queue into the receive_queue. |
| 4336 | */ |
| 4337 | static void tcp_ofo_queue(struct sock *sk) |
| 4338 | { |
| 4339 | struct tcp_sock *tp = tcp_sk(sk); |
| 4340 | __u32 dsack_high = tp->rcv_nxt; |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4341 | bool fin, fragstolen, eaten; |
Eric Dumazet | bd1e75a | 2014-09-19 08:26:20 -0700 | [diff] [blame] | 4342 | struct sk_buff *skb, *tail; |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4343 | struct rb_node *p; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4344 | |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4345 | p = rb_first(&tp->out_of_order_queue); |
| 4346 | while (p) { |
| 4347 | skb = rb_entry(p, struct sk_buff, rbnode); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4348 | if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) |
| 4349 | break; |
| 4350 | |
| 4351 | if (before(TCP_SKB_CB(skb)->seq, dsack_high)) { |
| 4352 | __u32 dsack = dsack_high; |
| 4353 | if (before(TCP_SKB_CB(skb)->end_seq, dsack_high)) |
| 4354 | dsack_high = TCP_SKB_CB(skb)->end_seq; |
| 4355 | tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack); |
| 4356 | } |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4357 | p = rb_next(p); |
| 4358 | rb_erase(&skb->rbnode, &tp->out_of_order_queue); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4359 | |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4360 | if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4361 | SOCK_DEBUG(sk, "ofo packet was already received\n"); |
Eric Dumazet | 532182c | 2016-04-01 08:52:19 -0700 | [diff] [blame] | 4362 | tcp_drop(sk, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4363 | continue; |
| 4364 | } |
| 4365 | SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n", |
| 4366 | tp->rcv_nxt, TCP_SKB_CB(skb)->seq, |
| 4367 | TCP_SKB_CB(skb)->end_seq); |
| 4368 | |
Eric Dumazet | bd1e75a | 2014-09-19 08:26:20 -0700 | [diff] [blame] | 4369 | tail = skb_peek_tail(&sk->sk_receive_queue); |
| 4370 | eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen); |
Eric Dumazet | bdd1f9e | 2015-04-28 15:28:18 -0700 | [diff] [blame] | 4371 | tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq); |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4372 | fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN; |
Eric Dumazet | bd1e75a | 2014-09-19 08:26:20 -0700 | [diff] [blame] | 4373 | if (!eaten) |
| 4374 | __skb_queue_tail(&sk->sk_receive_queue, skb); |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4375 | else |
Eric Dumazet | bd1e75a | 2014-09-19 08:26:20 -0700 | [diff] [blame] | 4376 | kfree_skb_partial(skb, fragstolen); |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4377 | |
| 4378 | if (unlikely(fin)) { |
| 4379 | tcp_fin(sk); |
| 4380 | /* tcp_fin() purges tp->out_of_order_queue, |
| 4381 | * so we must end this loop right now. |
| 4382 | */ |
| 4383 | break; |
| 4384 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4385 | } |
| 4386 | } |
| 4387 | |
| 4388 | static bool tcp_prune_ofo_queue(struct sock *sk); |
| 4389 | static int tcp_prune_queue(struct sock *sk); |
| 4390 | |
| 4391 | static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb, |
| 4392 | unsigned int size) |
| 4393 | { |
| 4394 | if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || |
| 4395 | !sk_rmem_schedule(sk, skb, size)) { |
| 4396 | |
| 4397 | if (tcp_prune_queue(sk) < 0) |
| 4398 | return -1; |
Pavel Emelyanov | 1ed8346 | 2008-07-16 20:29:51 -0700 | [diff] [blame] | 4399 | |
Eric Dumazet | 36a6503 | 2016-08-17 14:17:09 -0700 | [diff] [blame] | 4400 | while (!sk_rmem_schedule(sk, skb, size)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4401 | if (!tcp_prune_ofo_queue(sk)) |
| 4402 | return -1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4403 | } |
| 4404 | } |
| 4405 | return 0; |
| 4406 | } |
| 4407 | |
Eric Dumazet | e86b2919 | 2012-03-18 11:06:44 +0000 | [diff] [blame] | 4408 | static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb) |
| 4409 | { |
| 4410 | struct tcp_sock *tp = tcp_sk(sk); |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4411 | struct rb_node **p, *q, *parent; |
Eric Dumazet | e86b2919 | 2012-03-18 11:06:44 +0000 | [diff] [blame] | 4412 | struct sk_buff *skb1; |
| 4413 | u32 seq, end_seq; |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4414 | bool fragstolen; |
Eric Dumazet | e86b2919 | 2012-03-18 11:06:44 +0000 | [diff] [blame] | 4415 | |
Florian Westphal | 735d383 | 2014-09-29 13:08:30 +0200 | [diff] [blame] | 4416 | tcp_ecn_check_ce(tp, skb); |
Eric Dumazet | e86b2919 | 2012-03-18 11:06:44 +0000 | [diff] [blame] | 4417 | |
Mel Gorman | c76562b | 2012-07-31 16:44:41 -0700 | [diff] [blame] | 4418 | if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) { |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 4419 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP); |
Eric Dumazet | 532182c | 2016-04-01 08:52:19 -0700 | [diff] [blame] | 4420 | tcp_drop(sk, skb); |
Eric Dumazet | e86b2919 | 2012-03-18 11:06:44 +0000 | [diff] [blame] | 4421 | return; |
| 4422 | } |
| 4423 | |
| 4424 | /* Disable header prediction. */ |
| 4425 | tp->pred_flags = 0; |
| 4426 | inet_csk_schedule_ack(sk); |
| 4427 | |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 4428 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE); |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4429 | seq = TCP_SKB_CB(skb)->seq; |
| 4430 | end_seq = TCP_SKB_CB(skb)->end_seq; |
Eric Dumazet | e86b2919 | 2012-03-18 11:06:44 +0000 | [diff] [blame] | 4431 | SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n", |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4432 | tp->rcv_nxt, seq, end_seq); |
Eric Dumazet | e86b2919 | 2012-03-18 11:06:44 +0000 | [diff] [blame] | 4433 | |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4434 | p = &tp->out_of_order_queue.rb_node; |
| 4435 | if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) { |
Eric Dumazet | e86b2919 | 2012-03-18 11:06:44 +0000 | [diff] [blame] | 4436 | /* Initial out of order segment, build 1 SACK. */ |
| 4437 | if (tcp_is_sack(tp)) { |
| 4438 | tp->rx_opt.num_sacks = 1; |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4439 | tp->selective_acks[0].start_seq = seq; |
| 4440 | tp->selective_acks[0].end_seq = end_seq; |
Eric Dumazet | e86b2919 | 2012-03-18 11:06:44 +0000 | [diff] [blame] | 4441 | } |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4442 | rb_link_node(&skb->rbnode, NULL, p); |
| 4443 | rb_insert_color(&skb->rbnode, &tp->out_of_order_queue); |
| 4444 | tp->ooo_last_skb = skb; |
Eric Dumazet | e86b2919 | 2012-03-18 11:06:44 +0000 | [diff] [blame] | 4445 | goto end; |
| 4446 | } |
| 4447 | |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4448 | /* In the typical case, we are adding an skb to the end of the list. |
| 4449 | * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup. |
| 4450 | */ |
| 4451 | if (tcp_try_coalesce(sk, tp->ooo_last_skb, skb, &fragstolen)) { |
| 4452 | coalesce_done: |
| 4453 | tcp_grow_window(sk, skb); |
| 4454 | kfree_skb_partial(skb, fragstolen); |
| 4455 | skb = NULL; |
| 4456 | goto add_sack; |
Eric Dumazet | e86b2919 | 2012-03-18 11:06:44 +0000 | [diff] [blame] | 4457 | } |
Eric Dumazet | 2594a2a | 2016-09-09 14:22:45 -0700 | [diff] [blame] | 4458 | /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */ |
| 4459 | if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) { |
| 4460 | parent = &tp->ooo_last_skb->rbnode; |
| 4461 | p = &parent->rb_right; |
| 4462 | goto insert; |
Eric Dumazet | e86b2919 | 2012-03-18 11:06:44 +0000 | [diff] [blame] | 4463 | } |
| 4464 | |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4465 | /* Find place to insert this segment. Handle overlaps on the way. */ |
| 4466 | parent = NULL; |
| 4467 | while (*p) { |
| 4468 | parent = *p; |
| 4469 | skb1 = rb_entry(parent, struct sk_buff, rbnode); |
| 4470 | if (before(seq, TCP_SKB_CB(skb1)->seq)) { |
| 4471 | p = &parent->rb_left; |
| 4472 | continue; |
Eric Dumazet | e86b2919 | 2012-03-18 11:06:44 +0000 | [diff] [blame] | 4473 | } |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4474 | if (before(seq, TCP_SKB_CB(skb1)->end_seq)) { |
| 4475 | if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) { |
| 4476 | /* All the bits are present. Drop. */ |
| 4477 | NET_INC_STATS(sock_net(sk), |
| 4478 | LINUX_MIB_TCPOFOMERGE); |
| 4479 | __kfree_skb(skb); |
| 4480 | skb = NULL; |
| 4481 | tcp_dsack_set(sk, seq, end_seq); |
| 4482 | goto add_sack; |
| 4483 | } |
| 4484 | if (after(seq, TCP_SKB_CB(skb1)->seq)) { |
| 4485 | /* Partial overlap. */ |
| 4486 | tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq); |
| 4487 | } else { |
| 4488 | /* skb's seq == skb1's seq and skb covers skb1. |
| 4489 | * Replace skb1 with skb. |
| 4490 | */ |
| 4491 | rb_replace_node(&skb1->rbnode, &skb->rbnode, |
| 4492 | &tp->out_of_order_queue); |
| 4493 | tcp_dsack_extend(sk, |
| 4494 | TCP_SKB_CB(skb1)->seq, |
| 4495 | TCP_SKB_CB(skb1)->end_seq); |
| 4496 | NET_INC_STATS(sock_net(sk), |
| 4497 | LINUX_MIB_TCPOFOMERGE); |
| 4498 | __kfree_skb(skb1); |
Eric Dumazet | 76f0dcb | 2016-09-13 22:55:05 -0700 | [diff] [blame] | 4499 | goto merge_right; |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4500 | } |
| 4501 | } else if (tcp_try_coalesce(sk, skb1, skb, &fragstolen)) { |
| 4502 | goto coalesce_done; |
| 4503 | } |
| 4504 | p = &parent->rb_right; |
Eric Dumazet | e86b2919 | 2012-03-18 11:06:44 +0000 | [diff] [blame] | 4505 | } |
Eric Dumazet | 2594a2a | 2016-09-09 14:22:45 -0700 | [diff] [blame] | 4506 | insert: |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4507 | /* Insert segment into RB tree. */ |
| 4508 | rb_link_node(&skb->rbnode, parent, p); |
| 4509 | rb_insert_color(&skb->rbnode, &tp->out_of_order_queue); |
Eric Dumazet | e86b2919 | 2012-03-18 11:06:44 +0000 | [diff] [blame] | 4510 | |
Eric Dumazet | 76f0dcb | 2016-09-13 22:55:05 -0700 | [diff] [blame] | 4511 | merge_right: |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4512 | /* Remove other segments covered by skb. */ |
| 4513 | while ((q = rb_next(&skb->rbnode)) != NULL) { |
| 4514 | skb1 = rb_entry(q, struct sk_buff, rbnode); |
Eric Dumazet | e86b2919 | 2012-03-18 11:06:44 +0000 | [diff] [blame] | 4515 | |
| 4516 | if (!after(end_seq, TCP_SKB_CB(skb1)->seq)) |
| 4517 | break; |
| 4518 | if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) { |
| 4519 | tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, |
| 4520 | end_seq); |
| 4521 | break; |
| 4522 | } |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4523 | rb_erase(&skb1->rbnode, &tp->out_of_order_queue); |
Eric Dumazet | e86b2919 | 2012-03-18 11:06:44 +0000 | [diff] [blame] | 4524 | tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, |
| 4525 | TCP_SKB_CB(skb1)->end_seq); |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 4526 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE); |
Eric Dumazet | 532182c | 2016-04-01 08:52:19 -0700 | [diff] [blame] | 4527 | tcp_drop(sk, skb1); |
Eric Dumazet | e86b2919 | 2012-03-18 11:06:44 +0000 | [diff] [blame] | 4528 | } |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4529 | /* If there is no skb after us, we are the last_skb ! */ |
| 4530 | if (!q) |
| 4531 | tp->ooo_last_skb = skb; |
Eric Dumazet | e86b2919 | 2012-03-18 11:06:44 +0000 | [diff] [blame] | 4532 | |
| 4533 | add_sack: |
| 4534 | if (tcp_is_sack(tp)) |
| 4535 | tcp_sack_new_ofo_skb(sk, seq, end_seq); |
| 4536 | end: |
Eric Dumazet | 4e4f1fc | 2013-09-06 10:35:58 -0700 | [diff] [blame] | 4537 | if (skb) { |
| 4538 | tcp_grow_window(sk, skb); |
Eric Dumazet | e86b2919 | 2012-03-18 11:06:44 +0000 | [diff] [blame] | 4539 | skb_set_owner_r(skb, sk); |
Eric Dumazet | 4e4f1fc | 2013-09-06 10:35:58 -0700 | [diff] [blame] | 4540 | } |
Eric Dumazet | e86b2919 | 2012-03-18 11:06:44 +0000 | [diff] [blame] | 4541 | } |
| 4542 | |
Pavel Emelyanov | 292e8d8 | 2012-05-10 01:49:41 +0000 | [diff] [blame] | 4543 | static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen, |
Eric Dumazet | b081f85 | 2012-05-02 09:58:29 +0000 | [diff] [blame] | 4544 | bool *fragstolen) |
| 4545 | { |
| 4546 | int eaten; |
| 4547 | struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue); |
| 4548 | |
| 4549 | __skb_pull(skb, hdrlen); |
| 4550 | eaten = (tail && |
| 4551 | tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0; |
Eric Dumazet | bdd1f9e | 2015-04-28 15:28:18 -0700 | [diff] [blame] | 4552 | tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq); |
Eric Dumazet | b081f85 | 2012-05-02 09:58:29 +0000 | [diff] [blame] | 4553 | if (!eaten) { |
| 4554 | __skb_queue_tail(&sk->sk_receive_queue, skb); |
| 4555 | skb_set_owner_r(skb, sk); |
| 4556 | } |
| 4557 | return eaten; |
| 4558 | } |
Eric Dumazet | e86b2919 | 2012-03-18 11:06:44 +0000 | [diff] [blame] | 4559 | |
Pavel Emelyanov | 292e8d8 | 2012-05-10 01:49:41 +0000 | [diff] [blame] | 4560 | int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size) |
| 4561 | { |
Eric Dumazet | cb93471 | 2014-09-17 03:14:42 -0700 | [diff] [blame] | 4562 | struct sk_buff *skb; |
Eric Dumazet | 5d4c9bf | 2015-11-18 21:03:33 -0800 | [diff] [blame] | 4563 | int err = -ENOMEM; |
| 4564 | int data_len = 0; |
Pavel Emelyanov | 292e8d8 | 2012-05-10 01:49:41 +0000 | [diff] [blame] | 4565 | bool fragstolen; |
| 4566 | |
Pavel Emelyanov | c454e61 | 2012-10-29 05:05:33 +0000 | [diff] [blame] | 4567 | if (size == 0) |
| 4568 | return 0; |
| 4569 | |
Eric Dumazet | 5d4c9bf | 2015-11-18 21:03:33 -0800 | [diff] [blame] | 4570 | if (size > PAGE_SIZE) { |
| 4571 | int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS); |
| 4572 | |
| 4573 | data_len = npages << PAGE_SHIFT; |
| 4574 | size = data_len + (size & ~PAGE_MASK); |
| 4575 | } |
| 4576 | skb = alloc_skb_with_frags(size - data_len, data_len, |
| 4577 | PAGE_ALLOC_COSTLY_ORDER, |
| 4578 | &err, sk->sk_allocation); |
Pavel Emelyanov | 292e8d8 | 2012-05-10 01:49:41 +0000 | [diff] [blame] | 4579 | if (!skb) |
| 4580 | goto err; |
| 4581 | |
Eric Dumazet | 5d4c9bf | 2015-11-18 21:03:33 -0800 | [diff] [blame] | 4582 | skb_put(skb, size - data_len); |
| 4583 | skb->data_len = data_len; |
| 4584 | skb->len = size; |
| 4585 | |
Eric Dumazet | cb93471 | 2014-09-17 03:14:42 -0700 | [diff] [blame] | 4586 | if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) |
Mel Gorman | c76562b | 2012-07-31 16:44:41 -0700 | [diff] [blame] | 4587 | goto err_free; |
| 4588 | |
Eric Dumazet | 5d4c9bf | 2015-11-18 21:03:33 -0800 | [diff] [blame] | 4589 | err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size); |
| 4590 | if (err) |
Pavel Emelyanov | 292e8d8 | 2012-05-10 01:49:41 +0000 | [diff] [blame] | 4591 | goto err_free; |
| 4592 | |
| 4593 | TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt; |
| 4594 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size; |
| 4595 | TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1; |
| 4596 | |
Eric Dumazet | cb93471 | 2014-09-17 03:14:42 -0700 | [diff] [blame] | 4597 | if (tcp_queue_rcv(sk, skb, 0, &fragstolen)) { |
Pavel Emelyanov | 292e8d8 | 2012-05-10 01:49:41 +0000 | [diff] [blame] | 4598 | WARN_ON_ONCE(fragstolen); /* should not happen */ |
| 4599 | __kfree_skb(skb); |
| 4600 | } |
| 4601 | return size; |
| 4602 | |
| 4603 | err_free: |
| 4604 | kfree_skb(skb); |
| 4605 | err: |
Eric Dumazet | 5d4c9bf | 2015-11-18 21:03:33 -0800 | [diff] [blame] | 4606 | return err; |
| 4607 | |
Pavel Emelyanov | 292e8d8 | 2012-05-10 01:49:41 +0000 | [diff] [blame] | 4608 | } |
| 4609 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4610 | static void tcp_data_queue(struct sock *sk, struct sk_buff *skb) |
| 4611 | { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4612 | struct tcp_sock *tp = tcp_sk(sk); |
Eric Dumazet | 329033f | 2012-04-27 00:38:33 +0000 | [diff] [blame] | 4613 | bool fragstolen = false; |
Eric Dumazet | 532182c | 2016-04-01 08:52:19 -0700 | [diff] [blame] | 4614 | int eaten = -1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4615 | |
Eric Dumazet | 532182c | 2016-04-01 08:52:19 -0700 | [diff] [blame] | 4616 | if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) { |
| 4617 | __kfree_skb(skb); |
| 4618 | return; |
| 4619 | } |
Eric Dumazet | f84af32 | 2010-04-28 15:31:51 -0700 | [diff] [blame] | 4620 | skb_dst_drop(skb); |
Peter Pan(潘卫平) | 155c6e1 | 2014-09-24 22:17:02 +0800 | [diff] [blame] | 4621 | __skb_pull(skb, tcp_hdr(skb)->doff * 4); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4622 | |
Florian Westphal | 735d383 | 2014-09-29 13:08:30 +0200 | [diff] [blame] | 4623 | tcp_ecn_accept_cwr(tp, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4624 | |
Ilpo Järvinen | 5861f8e | 2009-03-14 14:23:01 +0000 | [diff] [blame] | 4625 | tp->rx_opt.dsack = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4626 | |
| 4627 | /* Queue data for delivery to the user. |
| 4628 | * Packets in sequence go to the receive queue. |
| 4629 | * Out of sequence packets to the out_of_order_queue. |
| 4630 | */ |
| 4631 | if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { |
| 4632 | if (tcp_receive_window(tp) == 0) |
| 4633 | goto out_of_window; |
| 4634 | |
| 4635 | /* Ok. In sequence. In window. */ |
| 4636 | if (tp->ucopy.task == current && |
| 4637 | tp->copied_seq == tp->rcv_nxt && tp->ucopy.len && |
| 4638 | sock_owned_by_user(sk) && !tp->urg_data) { |
| 4639 | int chunk = min_t(unsigned int, skb->len, |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 4640 | tp->ucopy.len); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4641 | |
| 4642 | __set_current_state(TASK_RUNNING); |
| 4643 | |
Al Viro | f4362a2 | 2014-11-24 13:26:06 -0500 | [diff] [blame] | 4644 | if (!skb_copy_datagram_msg(skb, 0, tp->ucopy.msg, chunk)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4645 | tp->ucopy.len -= chunk; |
| 4646 | tp->copied_seq += chunk; |
Jerry Chu | 44f5324 | 2011-01-25 13:46:30 -0800 | [diff] [blame] | 4647 | eaten = (chunk == skb->len); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4648 | tcp_rcv_space_adjust(sk); |
| 4649 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4650 | } |
| 4651 | |
| 4652 | if (eaten <= 0) { |
| 4653 | queue_and_out: |
Eric Dumazet | 76dfa60 | 2015-05-15 12:39:29 -0700 | [diff] [blame] | 4654 | if (eaten < 0) { |
| 4655 | if (skb_queue_len(&sk->sk_receive_queue) == 0) |
| 4656 | sk_forced_mem_schedule(sk, skb->truesize); |
| 4657 | else if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) |
| 4658 | goto drop; |
| 4659 | } |
Eric Dumazet | b081f85 | 2012-05-02 09:58:29 +0000 | [diff] [blame] | 4660 | eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4661 | } |
Eric Dumazet | bdd1f9e | 2015-04-28 15:28:18 -0700 | [diff] [blame] | 4662 | tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq); |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 4663 | if (skb->len) |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 4664 | tcp_event_data_recv(sk, skb); |
Peter Pan(潘卫平) | 155c6e1 | 2014-09-24 22:17:02 +0800 | [diff] [blame] | 4665 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) |
Eric Dumazet | 20c4cb7 | 2011-10-20 17:44:03 -0400 | [diff] [blame] | 4666 | tcp_fin(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4667 | |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4668 | if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4669 | tcp_ofo_queue(sk); |
| 4670 | |
| 4671 | /* RFC2581. 4.2. SHOULD send immediate ACK, when |
| 4672 | * gap in queue is filled. |
| 4673 | */ |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4674 | if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 4675 | inet_csk(sk)->icsk_ack.pingpong = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4676 | } |
| 4677 | |
| 4678 | if (tp->rx_opt.num_sacks) |
| 4679 | tcp_sack_remove(tp); |
| 4680 | |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 4681 | tcp_fast_path_check(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4682 | |
Eric Dumazet | 923dd34 | 2012-05-02 07:55:58 +0000 | [diff] [blame] | 4683 | if (eaten > 0) |
| 4684 | kfree_skb_partial(skb, fragstolen); |
Eric Dumazet | 1d57f19 | 2012-09-17 12:51:39 +0000 | [diff] [blame] | 4685 | if (!sock_flag(sk, SOCK_DEAD)) |
David S. Miller | 676d236 | 2014-04-11 16:15:36 -0400 | [diff] [blame] | 4686 | sk->sk_data_ready(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4687 | return; |
| 4688 | } |
| 4689 | |
| 4690 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { |
| 4691 | /* A retransmit, 2nd most common case. Force an immediate ack. */ |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 4692 | NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); |
Pavel Emelyanov | 1ed8346 | 2008-07-16 20:29:51 -0700 | [diff] [blame] | 4693 | tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4694 | |
| 4695 | out_of_window: |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 4696 | tcp_enter_quickack_mode(sk); |
| 4697 | inet_csk_schedule_ack(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4698 | drop: |
Eric Dumazet | 532182c | 2016-04-01 08:52:19 -0700 | [diff] [blame] | 4699 | tcp_drop(sk, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4700 | return; |
| 4701 | } |
| 4702 | |
| 4703 | /* Out of window. F.e. zero window probe. */ |
| 4704 | if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp))) |
| 4705 | goto out_of_window; |
| 4706 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 4707 | tcp_enter_quickack_mode(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4708 | |
| 4709 | if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { |
| 4710 | /* Partial packet, seq < rcv_next < end_seq */ |
| 4711 | SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n", |
| 4712 | tp->rcv_nxt, TCP_SKB_CB(skb)->seq, |
| 4713 | TCP_SKB_CB(skb)->end_seq); |
| 4714 | |
Pavel Emelyanov | 1ed8346 | 2008-07-16 20:29:51 -0700 | [diff] [blame] | 4715 | tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt); |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 4716 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4717 | /* If window is closed, drop tail of packet. But after |
| 4718 | * remembering D-SACK for its head made in previous line. |
| 4719 | */ |
| 4720 | if (!tcp_receive_window(tp)) |
| 4721 | goto out_of_window; |
| 4722 | goto queue_and_out; |
| 4723 | } |
| 4724 | |
Eric Dumazet | e86b2919 | 2012-03-18 11:06:44 +0000 | [diff] [blame] | 4725 | tcp_data_queue_ofo(sk, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4726 | } |
| 4727 | |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4728 | static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list) |
Ilpo Järvinen | 2cf4663 | 2008-08-23 05:11:41 -0700 | [diff] [blame] | 4729 | { |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4730 | if (list) |
| 4731 | return !skb_queue_is_last(list, skb) ? skb->next : NULL; |
David S. Miller | 9152194 | 2009-05-28 21:35:47 -0700 | [diff] [blame] | 4732 | |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4733 | return rb_entry_safe(rb_next(&skb->rbnode), struct sk_buff, rbnode); |
| 4734 | } |
Ilpo Järvinen | 2cf4663 | 2008-08-23 05:11:41 -0700 | [diff] [blame] | 4735 | |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4736 | static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb, |
| 4737 | struct sk_buff_head *list, |
| 4738 | struct rb_root *root) |
| 4739 | { |
| 4740 | struct sk_buff *next = tcp_skb_next(skb, list); |
| 4741 | |
| 4742 | if (list) |
| 4743 | __skb_unlink(skb, list); |
| 4744 | else |
| 4745 | rb_erase(&skb->rbnode, root); |
| 4746 | |
Ilpo Järvinen | 2cf4663 | 2008-08-23 05:11:41 -0700 | [diff] [blame] | 4747 | __kfree_skb(skb); |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 4748 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED); |
Ilpo Järvinen | 2cf4663 | 2008-08-23 05:11:41 -0700 | [diff] [blame] | 4749 | |
| 4750 | return next; |
| 4751 | } |
| 4752 | |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4753 | /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */ |
| 4754 | static void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb) |
| 4755 | { |
| 4756 | struct rb_node **p = &root->rb_node; |
| 4757 | struct rb_node *parent = NULL; |
| 4758 | struct sk_buff *skb1; |
| 4759 | |
| 4760 | while (*p) { |
| 4761 | parent = *p; |
| 4762 | skb1 = rb_entry(parent, struct sk_buff, rbnode); |
| 4763 | if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq)) |
| 4764 | p = &parent->rb_left; |
| 4765 | else |
| 4766 | p = &parent->rb_right; |
| 4767 | } |
| 4768 | rb_link_node(&skb->rbnode, parent, p); |
| 4769 | rb_insert_color(&skb->rbnode, root); |
| 4770 | } |
| 4771 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4772 | /* Collapse contiguous sequence of skbs head..tail with |
| 4773 | * sequence numbers start..end. |
David S. Miller | 9152194 | 2009-05-28 21:35:47 -0700 | [diff] [blame] | 4774 | * |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4775 | * If tail is NULL, this means until the end of the queue. |
David S. Miller | 9152194 | 2009-05-28 21:35:47 -0700 | [diff] [blame] | 4776 | * |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4777 | * Segments with FIN/SYN are not collapsed (only because this |
| 4778 | * simplifies code) |
| 4779 | */ |
| 4780 | static void |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4781 | tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root, |
| 4782 | struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4783 | { |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4784 | struct sk_buff *skb = head, *n; |
| 4785 | struct sk_buff_head tmp; |
David S. Miller | 9152194 | 2009-05-28 21:35:47 -0700 | [diff] [blame] | 4786 | bool end_of_skbs; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4787 | |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 4788 | /* First, check that queue is collapsible and find |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4789 | * the point where collapsing can be useful. |
| 4790 | */ |
David S. Miller | 9152194 | 2009-05-28 21:35:47 -0700 | [diff] [blame] | 4791 | restart: |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4792 | for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) { |
| 4793 | n = tcp_skb_next(skb, list); |
| 4794 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4795 | /* No new bits? It is possible on ofo queue. */ |
| 4796 | if (!before(start, TCP_SKB_CB(skb)->end_seq)) { |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4797 | skb = tcp_collapse_one(sk, skb, list, root); |
David S. Miller | 9152194 | 2009-05-28 21:35:47 -0700 | [diff] [blame] | 4798 | if (!skb) |
| 4799 | break; |
| 4800 | goto restart; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4801 | } |
| 4802 | |
| 4803 | /* The first skb to collapse is: |
| 4804 | * - not SYN/FIN and |
| 4805 | * - bloated or contains data before "start" or |
| 4806 | * overlaps to the next one. |
| 4807 | */ |
Eric Dumazet | e11ecdd | 2014-09-15 04:19:51 -0700 | [diff] [blame] | 4808 | if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) && |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4809 | (tcp_win_from_space(skb->truesize) > skb->len || |
David S. Miller | 9152194 | 2009-05-28 21:35:47 -0700 | [diff] [blame] | 4810 | before(TCP_SKB_CB(skb)->seq, start))) { |
| 4811 | end_of_skbs = false; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4812 | break; |
David S. Miller | 9152194 | 2009-05-28 21:35:47 -0700 | [diff] [blame] | 4813 | } |
| 4814 | |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4815 | if (n && n != tail && |
| 4816 | TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) { |
| 4817 | end_of_skbs = false; |
| 4818 | break; |
David S. Miller | 9152194 | 2009-05-28 21:35:47 -0700 | [diff] [blame] | 4819 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4820 | |
| 4821 | /* Decided to skip this, advance start seq. */ |
| 4822 | start = TCP_SKB_CB(skb)->end_seq; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4823 | } |
Eric Dumazet | e11ecdd | 2014-09-15 04:19:51 -0700 | [diff] [blame] | 4824 | if (end_of_skbs || |
| 4825 | (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN))) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4826 | return; |
| 4827 | |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4828 | __skb_queue_head_init(&tmp); |
| 4829 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4830 | while (before(start, end)) { |
Eric Dumazet | b3d6cb9 | 2014-09-15 04:19:53 -0700 | [diff] [blame] | 4831 | int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4832 | struct sk_buff *nskb; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4833 | |
Eric Dumazet | b3d6cb9 | 2014-09-15 04:19:53 -0700 | [diff] [blame] | 4834 | nskb = alloc_skb(copy, GFP_ATOMIC); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4835 | if (!nskb) |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4836 | break; |
Arnaldo Carvalho de Melo | c51957d | 2007-03-10 12:47:22 -0300 | [diff] [blame] | 4837 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4838 | memcpy(nskb->cb, skb->cb, sizeof(skb->cb)); |
| 4839 | TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start; |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4840 | if (list) |
| 4841 | __skb_queue_before(list, skb, nskb); |
| 4842 | else |
| 4843 | __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */ |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 4844 | skb_set_owner_r(nskb, sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4845 | |
| 4846 | /* Copy data, releasing collapsed skbs. */ |
| 4847 | while (copy > 0) { |
| 4848 | int offset = start - TCP_SKB_CB(skb)->seq; |
| 4849 | int size = TCP_SKB_CB(skb)->end_seq - start; |
| 4850 | |
Kris Katterjohn | 09a6266 | 2006-01-08 22:24:28 -0800 | [diff] [blame] | 4851 | BUG_ON(offset < 0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4852 | if (size > 0) { |
| 4853 | size = min(copy, size); |
| 4854 | if (skb_copy_bits(skb, offset, skb_put(nskb, size), size)) |
| 4855 | BUG(); |
| 4856 | TCP_SKB_CB(nskb)->end_seq += size; |
| 4857 | copy -= size; |
| 4858 | start += size; |
| 4859 | } |
| 4860 | if (!before(start, TCP_SKB_CB(skb)->end_seq)) { |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4861 | skb = tcp_collapse_one(sk, skb, list, root); |
David S. Miller | 9152194 | 2009-05-28 21:35:47 -0700 | [diff] [blame] | 4862 | if (!skb || |
| 4863 | skb == tail || |
Eric Dumazet | e11ecdd | 2014-09-15 04:19:51 -0700 | [diff] [blame] | 4864 | (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN))) |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4865 | goto end; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4866 | } |
| 4867 | } |
| 4868 | } |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4869 | end: |
| 4870 | skb_queue_walk_safe(&tmp, skb, n) |
| 4871 | tcp_rbtree_insert(root, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4872 | } |
| 4873 | |
| 4874 | /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs |
| 4875 | * and tcp_collapse() them until all the queue is collapsed. |
| 4876 | */ |
| 4877 | static void tcp_collapse_ofo_queue(struct sock *sk) |
| 4878 | { |
| 4879 | struct tcp_sock *tp = tcp_sk(sk); |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4880 | struct sk_buff *skb, *head; |
| 4881 | struct rb_node *p; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4882 | u32 start, end; |
| 4883 | |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4884 | p = rb_first(&tp->out_of_order_queue); |
| 4885 | skb = rb_entry_safe(p, struct sk_buff, rbnode); |
| 4886 | new_range: |
| 4887 | if (!skb) { |
| 4888 | p = rb_last(&tp->out_of_order_queue); |
| 4889 | /* Note: This is possible p is NULL here. We do not |
| 4890 | * use rb_entry_safe(), as ooo_last_skb is valid only |
| 4891 | * if rbtree is not empty. |
| 4892 | */ |
| 4893 | tp->ooo_last_skb = rb_entry(p, struct sk_buff, rbnode); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4894 | return; |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4895 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4896 | start = TCP_SKB_CB(skb)->seq; |
| 4897 | end = TCP_SKB_CB(skb)->end_seq; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4898 | |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4899 | for (head = skb;;) { |
| 4900 | skb = tcp_skb_next(skb, NULL); |
David S. Miller | 9152194 | 2009-05-28 21:35:47 -0700 | [diff] [blame] | 4901 | |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4902 | /* Range is terminated when we see a gap or when |
| 4903 | * we are at the queue end. |
| 4904 | */ |
David S. Miller | 9152194 | 2009-05-28 21:35:47 -0700 | [diff] [blame] | 4905 | if (!skb || |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4906 | after(TCP_SKB_CB(skb)->seq, end) || |
| 4907 | before(TCP_SKB_CB(skb)->end_seq, start)) { |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4908 | tcp_collapse(sk, NULL, &tp->out_of_order_queue, |
David S. Miller | 8728b83 | 2005-08-09 19:25:21 -0700 | [diff] [blame] | 4909 | head, skb, start, end); |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4910 | goto new_range; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4911 | } |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4912 | |
| 4913 | if (unlikely(before(TCP_SKB_CB(skb)->seq, start))) |
| 4914 | start = TCP_SKB_CB(skb)->seq; |
| 4915 | if (after(TCP_SKB_CB(skb)->end_seq, end)) |
| 4916 | end = TCP_SKB_CB(skb)->end_seq; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4917 | } |
| 4918 | } |
| 4919 | |
Vitaliy Gusev | b000cd3 | 2008-04-15 00:33:38 -0700 | [diff] [blame] | 4920 | /* |
Eric Dumazet | 36a6503 | 2016-08-17 14:17:09 -0700 | [diff] [blame] | 4921 | * Clean the out-of-order queue to make room. |
| 4922 | * We drop high sequences packets to : |
| 4923 | * 1) Let a chance for holes to be filled. |
| 4924 | * 2) not add too big latencies if thousands of packets sit there. |
| 4925 | * (But if application shrinks SO_RCVBUF, we could still end up |
| 4926 | * freeing whole queue here) |
| 4927 | * |
| 4928 | * Return true if queue has shrunk. |
Vitaliy Gusev | b000cd3 | 2008-04-15 00:33:38 -0700 | [diff] [blame] | 4929 | */ |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 4930 | static bool tcp_prune_ofo_queue(struct sock *sk) |
Vitaliy Gusev | b000cd3 | 2008-04-15 00:33:38 -0700 | [diff] [blame] | 4931 | { |
| 4932 | struct tcp_sock *tp = tcp_sk(sk); |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4933 | struct rb_node *node, *prev; |
Vitaliy Gusev | b000cd3 | 2008-04-15 00:33:38 -0700 | [diff] [blame] | 4934 | |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4935 | if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) |
Eric Dumazet | 36a6503 | 2016-08-17 14:17:09 -0700 | [diff] [blame] | 4936 | return false; |
Vitaliy Gusev | b000cd3 | 2008-04-15 00:33:38 -0700 | [diff] [blame] | 4937 | |
Eric Dumazet | 36a6503 | 2016-08-17 14:17:09 -0700 | [diff] [blame] | 4938 | NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED); |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4939 | node = &tp->ooo_last_skb->rbnode; |
| 4940 | do { |
| 4941 | prev = rb_prev(node); |
| 4942 | rb_erase(node, &tp->out_of_order_queue); |
| 4943 | tcp_drop(sk, rb_entry(node, struct sk_buff, rbnode)); |
Vitaliy Gusev | b000cd3 | 2008-04-15 00:33:38 -0700 | [diff] [blame] | 4944 | sk_mem_reclaim(sk); |
Eric Dumazet | 36a6503 | 2016-08-17 14:17:09 -0700 | [diff] [blame] | 4945 | if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf && |
| 4946 | !tcp_under_memory_pressure(sk)) |
| 4947 | break; |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4948 | node = prev; |
| 4949 | } while (node); |
| 4950 | tp->ooo_last_skb = rb_entry(prev, struct sk_buff, rbnode); |
Eric Dumazet | 36a6503 | 2016-08-17 14:17:09 -0700 | [diff] [blame] | 4951 | |
| 4952 | /* Reset SACK state. A conforming SACK implementation will |
| 4953 | * do the same at a timeout based retransmit. When a connection |
| 4954 | * is in a sad state like this, we care only about integrity |
| 4955 | * of the connection not performance. |
| 4956 | */ |
| 4957 | if (tp->rx_opt.sack_ok) |
| 4958 | tcp_sack_reset(&tp->rx_opt); |
| 4959 | return true; |
Vitaliy Gusev | b000cd3 | 2008-04-15 00:33:38 -0700 | [diff] [blame] | 4960 | } |
| 4961 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4962 | /* Reduce allocated memory if we can, trying to get |
| 4963 | * the socket within its memory limits again. |
| 4964 | * |
| 4965 | * Return less than zero if we should start dropping frames |
| 4966 | * until the socket owning process reads some of the data |
| 4967 | * to stabilize the situation. |
| 4968 | */ |
| 4969 | static int tcp_prune_queue(struct sock *sk) |
| 4970 | { |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 4971 | struct tcp_sock *tp = tcp_sk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4972 | |
| 4973 | SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq); |
| 4974 | |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 4975 | NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4976 | |
| 4977 | if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf) |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 4978 | tcp_clamp_window(sk); |
Eric Dumazet | b8da51e | 2015-05-15 12:39:27 -0700 | [diff] [blame] | 4979 | else if (tcp_under_memory_pressure(sk)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4980 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss); |
| 4981 | |
| 4982 | tcp_collapse_ofo_queue(sk); |
David S. Miller | 9152194 | 2009-05-28 21:35:47 -0700 | [diff] [blame] | 4983 | if (!skb_queue_empty(&sk->sk_receive_queue)) |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 4984 | tcp_collapse(sk, &sk->sk_receive_queue, NULL, |
David S. Miller | 9152194 | 2009-05-28 21:35:47 -0700 | [diff] [blame] | 4985 | skb_peek(&sk->sk_receive_queue), |
| 4986 | NULL, |
| 4987 | tp->copied_seq, tp->rcv_nxt); |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 4988 | sk_mem_reclaim(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4989 | |
| 4990 | if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) |
| 4991 | return 0; |
| 4992 | |
| 4993 | /* Collapsing did not help, destructive actions follow. |
| 4994 | * This must not ever occur. */ |
| 4995 | |
Vitaliy Gusev | b000cd3 | 2008-04-15 00:33:38 -0700 | [diff] [blame] | 4996 | tcp_prune_ofo_queue(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4997 | |
| 4998 | if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) |
| 4999 | return 0; |
| 5000 | |
| 5001 | /* If we are really being abused, tell the caller to silently |
| 5002 | * drop receive data on the floor. It will get retransmitted |
| 5003 | * and hopefully then we'll have sufficient space. |
| 5004 | */ |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 5005 | NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5006 | |
| 5007 | /* Massive buffer overcommit. */ |
| 5008 | tp->pred_flags = 0; |
| 5009 | return -1; |
| 5010 | } |
| 5011 | |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 5012 | static bool tcp_should_expand_sndbuf(const struct sock *sk) |
David S. Miller | 0d9901d | 2005-07-05 15:21:10 -0700 | [diff] [blame] | 5013 | { |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 5014 | const struct tcp_sock *tp = tcp_sk(sk); |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 5015 | |
David S. Miller | 0d9901d | 2005-07-05 15:21:10 -0700 | [diff] [blame] | 5016 | /* If the user specified a specific send buffer setting, do |
| 5017 | * not modify it. |
| 5018 | */ |
| 5019 | if (sk->sk_userlocks & SOCK_SNDBUF_LOCK) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 5020 | return false; |
David S. Miller | 0d9901d | 2005-07-05 15:21:10 -0700 | [diff] [blame] | 5021 | |
| 5022 | /* If we are under global TCP memory pressure, do not expand. */ |
Eric Dumazet | b8da51e | 2015-05-15 12:39:27 -0700 | [diff] [blame] | 5023 | if (tcp_under_memory_pressure(sk)) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 5024 | return false; |
David S. Miller | 0d9901d | 2005-07-05 15:21:10 -0700 | [diff] [blame] | 5025 | |
| 5026 | /* If we are under soft global TCP memory pressure, do not expand. */ |
Glauber Costa | 180d8cd | 2011-12-11 21:47:02 +0000 | [diff] [blame] | 5027 | if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0)) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 5028 | return false; |
David S. Miller | 0d9901d | 2005-07-05 15:21:10 -0700 | [diff] [blame] | 5029 | |
| 5030 | /* If we filled the congestion window, do not expand. */ |
Neal Cardwell | 6514890 | 2015-02-20 13:33:16 -0500 | [diff] [blame] | 5031 | if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 5032 | return false; |
David S. Miller | 0d9901d | 2005-07-05 15:21:10 -0700 | [diff] [blame] | 5033 | |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 5034 | return true; |
David S. Miller | 0d9901d | 2005-07-05 15:21:10 -0700 | [diff] [blame] | 5035 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5036 | |
| 5037 | /* When incoming ACK allowed to free some skb from write_queue, |
| 5038 | * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket |
| 5039 | * on the exit from tcp input handler. |
| 5040 | * |
| 5041 | * PROBLEM: sndbuf expansion does not work well with largesend. |
| 5042 | */ |
| 5043 | static void tcp_new_space(struct sock *sk) |
| 5044 | { |
| 5045 | struct tcp_sock *tp = tcp_sk(sk); |
| 5046 | |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 5047 | if (tcp_should_expand_sndbuf(sk)) { |
Eric Dumazet | 6ae7053 | 2013-10-01 10:23:44 -0700 | [diff] [blame] | 5048 | tcp_sndbuf_expand(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5049 | tp->snd_cwnd_stamp = tcp_time_stamp; |
| 5050 | } |
| 5051 | |
| 5052 | sk->sk_write_space(sk); |
| 5053 | } |
| 5054 | |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 5055 | static void tcp_check_space(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5056 | { |
| 5057 | if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) { |
| 5058 | sock_reset_flag(sk, SOCK_QUEUE_SHRUNK); |
jbaron@akamai.com | 3c71512 | 2015-04-20 20:05:07 +0000 | [diff] [blame] | 5059 | /* pairs with tcp_poll() */ |
| 5060 | smp_mb__after_atomic(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5061 | if (sk->sk_socket && |
| 5062 | test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) |
| 5063 | tcp_new_space(sk); |
| 5064 | } |
| 5065 | } |
| 5066 | |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 5067 | static inline void tcp_data_snd_check(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5068 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 5069 | tcp_push_pending_frames(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5070 | tcp_check_space(sk); |
| 5071 | } |
| 5072 | |
| 5073 | /* |
| 5074 | * Check if sending an ack is needed. |
| 5075 | */ |
| 5076 | static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible) |
| 5077 | { |
| 5078 | struct tcp_sock *tp = tcp_sk(sk); |
| 5079 | |
| 5080 | /* More than one full frame received... */ |
Joe Perches | 9d4fb27 | 2009-11-23 10:41:23 -0800 | [diff] [blame] | 5081 | 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] | 5082 | /* ... and right edge of window advances far enough. |
| 5083 | * (tcp_recvmsg() will send ACK otherwise). Or... |
| 5084 | */ |
Joe Perches | 9d4fb27 | 2009-11-23 10:41:23 -0800 | [diff] [blame] | 5085 | __tcp_select_window(sk) >= tp->rcv_wnd) || |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5086 | /* We ACK each frame or... */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 5087 | tcp_in_quickack_mode(sk) || |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5088 | /* We have out of order data. */ |
Yaogong Wang | 9f5afea | 2016-09-07 14:49:28 -0700 | [diff] [blame] | 5089 | (ofo_possible && !RB_EMPTY_ROOT(&tp->out_of_order_queue))) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5090 | /* Then ack it now */ |
| 5091 | tcp_send_ack(sk); |
| 5092 | } else { |
| 5093 | /* Else, send delayed ack. */ |
| 5094 | tcp_send_delayed_ack(sk); |
| 5095 | } |
| 5096 | } |
| 5097 | |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 5098 | static inline void tcp_ack_snd_check(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5099 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 5100 | if (!inet_csk_ack_scheduled(sk)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5101 | /* We sent a data segment already. */ |
| 5102 | return; |
| 5103 | } |
| 5104 | __tcp_ack_snd_check(sk, 1); |
| 5105 | } |
| 5106 | |
| 5107 | /* |
| 5108 | * This routine is only called when we have urgent data |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 5109 | * signaled. Its the 'slow' part of tcp_urg. It could be |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5110 | * moved inline now as tcp_urg is only called from one |
| 5111 | * place. We handle URGent data wrong. We have to - as |
| 5112 | * BSD still doesn't use the correction from RFC961. |
| 5113 | * For 1003.1g we should support a new option TCP_STDURG to permit |
| 5114 | * either form (or just set the sysctl tcp_stdurg). |
| 5115 | */ |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 5116 | |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 5117 | static void tcp_check_urg(struct sock *sk, const struct tcphdr *th) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5118 | { |
| 5119 | struct tcp_sock *tp = tcp_sk(sk); |
| 5120 | u32 ptr = ntohs(th->urg_ptr); |
| 5121 | |
| 5122 | if (ptr && !sysctl_tcp_stdurg) |
| 5123 | ptr--; |
| 5124 | ptr += ntohl(th->seq); |
| 5125 | |
| 5126 | /* Ignore urgent data that we've already seen and read. */ |
| 5127 | if (after(tp->copied_seq, ptr)) |
| 5128 | return; |
| 5129 | |
| 5130 | /* Do not replay urg ptr. |
| 5131 | * |
| 5132 | * NOTE: interesting situation not covered by specs. |
| 5133 | * Misbehaving sender may send urg ptr, pointing to segment, |
| 5134 | * which we already have in ofo queue. We are not able to fetch |
| 5135 | * such data and will stay in TCP_URG_NOTYET until will be eaten |
| 5136 | * by recvmsg(). Seems, we are not obliged to handle such wicked |
| 5137 | * situations. But it is worth to think about possibility of some |
| 5138 | * DoSes using some hypothetical application level deadlock. |
| 5139 | */ |
| 5140 | if (before(ptr, tp->rcv_nxt)) |
| 5141 | return; |
| 5142 | |
| 5143 | /* Do we already have a newer (or duplicate) urgent pointer? */ |
| 5144 | if (tp->urg_data && !after(ptr, tp->urg_seq)) |
| 5145 | return; |
| 5146 | |
| 5147 | /* Tell the world about our new urgent pointer. */ |
| 5148 | sk_send_sigurg(sk); |
| 5149 | |
| 5150 | /* We may be adding urgent data when the last byte read was |
| 5151 | * urgent. To do this requires some care. We cannot just ignore |
| 5152 | * tp->copied_seq since we would read the last urgent byte again |
| 5153 | * as data, nor can we alter copied_seq until this data arrives |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 5154 | * or we break the semantics of SIOCATMARK (and thus sockatmark()) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5155 | * |
| 5156 | * NOTE. Double Dutch. Rendering to plain English: author of comment |
| 5157 | * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB); |
| 5158 | * and expect that both A and B disappear from stream. This is _wrong_. |
| 5159 | * Though this happens in BSD with high probability, this is occasional. |
| 5160 | * Any application relying on this is buggy. Note also, that fix "works" |
| 5161 | * only in this artificial test. Insert some normal data between A and B and we will |
| 5162 | * decline of BSD again. Verdict: it is better to remove to trap |
| 5163 | * buggy users. |
| 5164 | */ |
| 5165 | if (tp->urg_seq == tp->copied_seq && tp->urg_data && |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 5166 | !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5167 | struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); |
| 5168 | tp->copied_seq++; |
| 5169 | 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] | 5170 | __skb_unlink(skb, &sk->sk_receive_queue); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5171 | __kfree_skb(skb); |
| 5172 | } |
| 5173 | } |
| 5174 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 5175 | tp->urg_data = TCP_URG_NOTYET; |
| 5176 | tp->urg_seq = ptr; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5177 | |
| 5178 | /* Disable header prediction. */ |
| 5179 | tp->pred_flags = 0; |
| 5180 | } |
| 5181 | |
| 5182 | /* This is the 'fast' part of urgent handling. */ |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 5183 | static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5184 | { |
| 5185 | struct tcp_sock *tp = tcp_sk(sk); |
| 5186 | |
| 5187 | /* Check if we get a new urgent pointer - normally not. */ |
| 5188 | if (th->urg) |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 5189 | tcp_check_urg(sk, th); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5190 | |
| 5191 | /* Do we wait for any urgent data? - normally not... */ |
| 5192 | if (tp->urg_data == TCP_URG_NOTYET) { |
| 5193 | u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) - |
| 5194 | th->syn; |
| 5195 | |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 5196 | /* Is the urgent pointer pointing into this packet? */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5197 | if (ptr < skb->len) { |
| 5198 | u8 tmp; |
| 5199 | if (skb_copy_bits(skb, ptr, &tmp, 1)) |
| 5200 | BUG(); |
| 5201 | tp->urg_data = TCP_URG_VALID | tmp; |
| 5202 | if (!sock_flag(sk, SOCK_DEAD)) |
David S. Miller | 676d236 | 2014-04-11 16:15:36 -0400 | [diff] [blame] | 5203 | sk->sk_data_ready(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5204 | } |
| 5205 | } |
| 5206 | } |
| 5207 | |
| 5208 | static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen) |
| 5209 | { |
| 5210 | struct tcp_sock *tp = tcp_sk(sk); |
| 5211 | int chunk = skb->len - hlen; |
| 5212 | int err; |
| 5213 | |
Herbert Xu | 6047637 | 2007-04-09 11:59:39 -0700 | [diff] [blame] | 5214 | if (skb_csum_unnecessary(skb)) |
Al Viro | f4362a2 | 2014-11-24 13:26:06 -0500 | [diff] [blame] | 5215 | err = skb_copy_datagram_msg(skb, hlen, tp->ucopy.msg, chunk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5216 | else |
Al Viro | f4362a2 | 2014-11-24 13:26:06 -0500 | [diff] [blame] | 5217 | err = skb_copy_and_csum_datagram_msg(skb, hlen, tp->ucopy.msg); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5218 | |
| 5219 | if (!err) { |
| 5220 | tp->ucopy.len -= chunk; |
| 5221 | tp->copied_seq += chunk; |
| 5222 | tcp_rcv_space_adjust(sk); |
| 5223 | } |
| 5224 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5225 | return err; |
| 5226 | } |
| 5227 | |
Ilpo Järvinen | cbe2d12 | 2008-08-23 05:10:12 -0700 | [diff] [blame] | 5228 | /* Does PAWS and seqno based validation of an incoming segment, flags will |
| 5229 | * play significant role here. |
| 5230 | */ |
Eric Dumazet | 0c24604 | 2012-07-17 01:41:30 +0000 | [diff] [blame] | 5231 | static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb, |
| 5232 | const struct tcphdr *th, int syn_inerr) |
Ilpo Järvinen | cbe2d12 | 2008-08-23 05:10:12 -0700 | [diff] [blame] | 5233 | { |
| 5234 | struct tcp_sock *tp = tcp_sk(sk); |
Pau Espin Pedrol | e00431b | 2016-06-07 16:30:34 +0200 | [diff] [blame] | 5235 | bool rst_seq_match = false; |
Ilpo Järvinen | cbe2d12 | 2008-08-23 05:10:12 -0700 | [diff] [blame] | 5236 | |
| 5237 | /* RFC1323: H1. Apply PAWS check first. */ |
Christoph Paasch | 1a2c618 | 2013-03-17 08:23:34 +0000 | [diff] [blame] | 5238 | if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp && |
Ilpo Järvinen | cbe2d12 | 2008-08-23 05:10:12 -0700 | [diff] [blame] | 5239 | tcp_paws_discard(sk, skb)) { |
| 5240 | if (!th->rst) { |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 5241 | NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED); |
Neal Cardwell | f2b2c58 | 2015-02-06 16:04:40 -0500 | [diff] [blame] | 5242 | if (!tcp_oow_rate_limited(sock_net(sk), skb, |
| 5243 | LINUX_MIB_TCPACKSKIPPEDPAWS, |
| 5244 | &tp->last_oow_ack_time)) |
| 5245 | tcp_send_dupack(sk, skb); |
Ilpo Järvinen | cbe2d12 | 2008-08-23 05:10:12 -0700 | [diff] [blame] | 5246 | goto discard; |
| 5247 | } |
| 5248 | /* Reset is accepted even if it did not pass PAWS. */ |
| 5249 | } |
| 5250 | |
| 5251 | /* Step 1: check sequence number */ |
| 5252 | if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) { |
| 5253 | /* RFC793, page 37: "In all states except SYN-SENT, all reset |
| 5254 | * (RST) segments are validated by checking their SEQ-fields." |
| 5255 | * And page 69: "If an incoming segment is not acceptable, |
| 5256 | * an acknowledgment should be sent in reply (unless the RST |
| 5257 | * bit is set, if so drop the segment and return)". |
| 5258 | */ |
Eric Dumazet | e371589 | 2012-07-17 12:29:30 +0000 | [diff] [blame] | 5259 | if (!th->rst) { |
| 5260 | if (th->syn) |
| 5261 | goto syn_challenge; |
Neal Cardwell | f2b2c58 | 2015-02-06 16:04:40 -0500 | [diff] [blame] | 5262 | if (!tcp_oow_rate_limited(sock_net(sk), skb, |
| 5263 | LINUX_MIB_TCPACKSKIPPEDSEQ, |
| 5264 | &tp->last_oow_ack_time)) |
| 5265 | tcp_send_dupack(sk, skb); |
Eric Dumazet | e371589 | 2012-07-17 12:29:30 +0000 | [diff] [blame] | 5266 | } |
Ilpo Järvinen | cbe2d12 | 2008-08-23 05:10:12 -0700 | [diff] [blame] | 5267 | goto discard; |
| 5268 | } |
| 5269 | |
| 5270 | /* Step 2: check RST bit */ |
| 5271 | if (th->rst) { |
Pau Espin Pedrol | e00431b | 2016-06-07 16:30:34 +0200 | [diff] [blame] | 5272 | /* RFC 5961 3.2 (extend to match against SACK too if available): |
| 5273 | * If seq num matches RCV.NXT or the right-most SACK block, |
| 5274 | * then |
Eric Dumazet | 282f23c | 2012-07-17 10:13:05 +0200 | [diff] [blame] | 5275 | * RESET the connection |
| 5276 | * else |
| 5277 | * Send a challenge ACK |
| 5278 | */ |
Pau Espin Pedrol | e00431b | 2016-06-07 16:30:34 +0200 | [diff] [blame] | 5279 | if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { |
| 5280 | rst_seq_match = true; |
| 5281 | } else if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) { |
| 5282 | struct tcp_sack_block *sp = &tp->selective_acks[0]; |
| 5283 | int max_sack = sp[0].end_seq; |
| 5284 | int this_sack; |
| 5285 | |
| 5286 | for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; |
| 5287 | ++this_sack) { |
| 5288 | max_sack = after(sp[this_sack].end_seq, |
| 5289 | max_sack) ? |
| 5290 | sp[this_sack].end_seq : max_sack; |
| 5291 | } |
| 5292 | |
| 5293 | if (TCP_SKB_CB(skb)->seq == max_sack) |
| 5294 | rst_seq_match = true; |
| 5295 | } |
| 5296 | |
| 5297 | if (rst_seq_match) |
Eric Dumazet | 282f23c | 2012-07-17 10:13:05 +0200 | [diff] [blame] | 5298 | tcp_reset(sk); |
| 5299 | else |
Neal Cardwell | f2b2c58 | 2015-02-06 16:04:40 -0500 | [diff] [blame] | 5300 | tcp_send_challenge_ack(sk, skb); |
Ilpo Järvinen | cbe2d12 | 2008-08-23 05:10:12 -0700 | [diff] [blame] | 5301 | goto discard; |
| 5302 | } |
| 5303 | |
Ilpo Järvinen | cbe2d12 | 2008-08-23 05:10:12 -0700 | [diff] [blame] | 5304 | /* step 3: check security and precedence [ignored] */ |
| 5305 | |
Eric Dumazet | 0c24604 | 2012-07-17 01:41:30 +0000 | [diff] [blame] | 5306 | /* step 4: Check for a SYN |
Sowmini Varadhan | cd21453 | 2014-10-30 12:48:08 -0400 | [diff] [blame] | 5307 | * RFC 5961 4.2 : Send a challenge ack |
Eric Dumazet | 0c24604 | 2012-07-17 01:41:30 +0000 | [diff] [blame] | 5308 | */ |
| 5309 | if (th->syn) { |
Eric Dumazet | e371589 | 2012-07-17 12:29:30 +0000 | [diff] [blame] | 5310 | syn_challenge: |
Ilpo Järvinen | cbe2d12 | 2008-08-23 05:10:12 -0700 | [diff] [blame] | 5311 | if (syn_inerr) |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 5312 | TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS); |
| 5313 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE); |
Neal Cardwell | f2b2c58 | 2015-02-06 16:04:40 -0500 | [diff] [blame] | 5314 | tcp_send_challenge_ack(sk, skb); |
Eric Dumazet | 0c24604 | 2012-07-17 01:41:30 +0000 | [diff] [blame] | 5315 | goto discard; |
Ilpo Järvinen | cbe2d12 | 2008-08-23 05:10:12 -0700 | [diff] [blame] | 5316 | } |
| 5317 | |
Eric Dumazet | 0c24604 | 2012-07-17 01:41:30 +0000 | [diff] [blame] | 5318 | return true; |
Ilpo Järvinen | cbe2d12 | 2008-08-23 05:10:12 -0700 | [diff] [blame] | 5319 | |
| 5320 | discard: |
Eric Dumazet | 532182c | 2016-04-01 08:52:19 -0700 | [diff] [blame] | 5321 | tcp_drop(sk, skb); |
Eric Dumazet | 0c24604 | 2012-07-17 01:41:30 +0000 | [diff] [blame] | 5322 | return false; |
Ilpo Järvinen | cbe2d12 | 2008-08-23 05:10:12 -0700 | [diff] [blame] | 5323 | } |
| 5324 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5325 | /* |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 5326 | * TCP receive function for the ESTABLISHED state. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5327 | * |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 5328 | * 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] | 5329 | * disabled when: |
| 5330 | * - A zero window was announced from us - zero window probing |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 5331 | * is only handled properly in the slow path. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5332 | * - Out of order segments arrived. |
| 5333 | * - Urgent data is expected. |
| 5334 | * - There is no buffer space left |
| 5335 | * - Unexpected TCP flags/window values/header lengths are received |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 5336 | * (detected by checking the TCP header against pred_flags) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5337 | * - Data is sent in both directions. Fast path only supports pure senders |
| 5338 | * or pure receivers (this means either the sequence number or the ack |
| 5339 | * value must stay constant) |
| 5340 | * - Unexpected TCP option. |
| 5341 | * |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 5342 | * When these conditions are not satisfied it drops into a standard |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5343 | * receive procedure patterned after RFC793 to handle all cases. |
| 5344 | * The first three cases are guaranteed by proper pred_flags setting, |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 5345 | * the rest is checked inline. Fast processing is turned on in |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5346 | * tcp_data_queue when everything is OK. |
| 5347 | */ |
Vijay Subramanian | c995ae2 | 2013-09-03 12:23:22 -0700 | [diff] [blame] | 5348 | void tcp_rcv_established(struct sock *sk, struct sk_buff *skb, |
| 5349 | const struct tcphdr *th, unsigned int len) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5350 | { |
| 5351 | struct tcp_sock *tp = tcp_sk(sk); |
| 5352 | |
Ian Morris | 51456b2 | 2015-04-03 09:17:26 +0100 | [diff] [blame] | 5353 | if (unlikely(!sk->sk_rx_dst)) |
Eric Dumazet | 5d299f3 | 2012-08-06 05:09:33 +0000 | [diff] [blame] | 5354 | inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5355 | /* |
| 5356 | * Header prediction. |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 5357 | * The code loosely follows the one in the famous |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5358 | * "30 instruction TCP receive" Van Jacobson mail. |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 5359 | * |
| 5360 | * Van's trick is to deposit buffers into socket queue |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5361 | * on a device interrupt, to call tcp_recv function |
| 5362 | * on the receive process context and checksum and copy |
| 5363 | * the buffer to user space. smart... |
| 5364 | * |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 5365 | * Our current scheme is not silly either but we take the |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5366 | * extra cost of the net_bh soft interrupt processing... |
| 5367 | * We do checksum and copy also but from device to kernel. |
| 5368 | */ |
| 5369 | |
| 5370 | tp->rx_opt.saw_tstamp = 0; |
| 5371 | |
| 5372 | /* pred_flags is 0xS?10 << 16 + snd_wnd |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 5373 | * if header_prediction is to be made |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5374 | * 'S' will always be tp->tcp_header_len >> 2 |
| 5375 | * '?' 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] | 5376 | * turn it off (when there are holes in the receive |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5377 | * space for instance) |
| 5378 | * PSH flag is ignored. |
| 5379 | */ |
| 5380 | |
| 5381 | if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags && |
John Dykstra | 96e0bf4 | 2009-03-22 21:49:57 -0700 | [diff] [blame] | 5382 | TCP_SKB_CB(skb)->seq == tp->rcv_nxt && |
| 5383 | !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5384 | int tcp_header_len = tp->tcp_header_len; |
| 5385 | |
| 5386 | /* Timestamp header prediction: tcp_header_len |
| 5387 | * is automatically equal to th->doff*4 due to pred_flags |
| 5388 | * match. |
| 5389 | */ |
| 5390 | |
| 5391 | /* Check timestamp */ |
| 5392 | if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5393 | /* No? Slow path! */ |
Ilpo Järvinen | a4356b2 | 2008-08-23 05:12:29 -0700 | [diff] [blame] | 5394 | if (!tcp_parse_aligned_timestamp(tp, th)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5395 | goto slow_path; |
| 5396 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5397 | /* If PAWS failed, check it more carefully in slow path */ |
| 5398 | if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0) |
| 5399 | goto slow_path; |
| 5400 | |
| 5401 | /* DO NOT update ts_recent here, if checksum fails |
| 5402 | * and timestamp was corrupted part, it will result |
| 5403 | * in a hung connection since we will drop all |
| 5404 | * future packets due to the PAWS test. |
| 5405 | */ |
| 5406 | } |
| 5407 | |
| 5408 | if (len <= tcp_header_len) { |
| 5409 | /* Bulk data transfer: sender */ |
| 5410 | if (len == tcp_header_len) { |
| 5411 | /* Predicted packet is in window by definition. |
| 5412 | * seq == rcv_nxt and rcv_wup <= rcv_nxt. |
| 5413 | * Hence, check seq<=rcv_wup reduces to: |
| 5414 | */ |
| 5415 | if (tcp_header_len == |
| 5416 | (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && |
| 5417 | tp->rcv_nxt == tp->rcv_wup) |
| 5418 | tcp_store_ts_recent(tp); |
| 5419 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5420 | /* We know that such packets are checksummed |
| 5421 | * on entry. |
| 5422 | */ |
| 5423 | tcp_ack(sk, skb, 0); |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 5424 | __kfree_skb(skb); |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 5425 | tcp_data_snd_check(sk); |
Vijay Subramanian | c995ae2 | 2013-09-03 12:23:22 -0700 | [diff] [blame] | 5426 | return; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5427 | } else { /* Header too small */ |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 5428 | TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5429 | goto discard; |
| 5430 | } |
| 5431 | } else { |
| 5432 | int eaten = 0; |
Eric Dumazet | b081f85 | 2012-05-02 09:58:29 +0000 | [diff] [blame] | 5433 | bool fragstolen = false; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5434 | |
Dan Williams | d27f9bc | 2013-12-30 11:37:15 -0800 | [diff] [blame] | 5435 | if (tp->ucopy.task == current && |
| 5436 | tp->copied_seq == tp->rcv_nxt && |
| 5437 | len - tcp_header_len <= tp->ucopy.len && |
| 5438 | sock_owned_by_user(sk)) { |
| 5439 | __set_current_state(TASK_RUNNING); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5440 | |
Dan Williams | d27f9bc | 2013-12-30 11:37:15 -0800 | [diff] [blame] | 5441 | if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5442 | /* Predicted packet is in window by definition. |
| 5443 | * seq == rcv_nxt and rcv_wup <= rcv_nxt. |
| 5444 | * Hence, check seq<=rcv_wup reduces to: |
| 5445 | */ |
| 5446 | if (tcp_header_len == |
| 5447 | (sizeof(struct tcphdr) + |
| 5448 | TCPOLEN_TSTAMP_ALIGNED) && |
| 5449 | tp->rcv_nxt == tp->rcv_wup) |
| 5450 | tcp_store_ts_recent(tp); |
| 5451 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 5452 | tcp_rcv_rtt_measure_ts(sk, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5453 | |
| 5454 | __skb_pull(skb, tcp_header_len); |
Eric Dumazet | bdd1f9e | 2015-04-28 15:28:18 -0700 | [diff] [blame] | 5455 | tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq); |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 5456 | NET_INC_STATS(sock_net(sk), |
Eric Dumazet | 02a1d6e | 2016-04-27 16:44:39 -0700 | [diff] [blame] | 5457 | LINUX_MIB_TCPHPHITSTOUSER); |
Dan Williams | d27f9bc | 2013-12-30 11:37:15 -0800 | [diff] [blame] | 5458 | eaten = 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5459 | } |
| 5460 | } |
| 5461 | if (!eaten) { |
Eric Dumazet | fb3477c | 2016-04-29 14:16:48 -0700 | [diff] [blame] | 5462 | if (tcp_checksum_complete(skb)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5463 | goto csum_error; |
| 5464 | |
Neal Cardwell | aab2b4b | 2013-03-04 06:23:05 +0000 | [diff] [blame] | 5465 | if ((int)skb->truesize > sk->sk_forward_alloc) |
| 5466 | goto step5; |
| 5467 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5468 | /* Predicted packet is in window by definition. |
| 5469 | * seq == rcv_nxt and rcv_wup <= rcv_nxt. |
| 5470 | * Hence, check seq<=rcv_wup reduces to: |
| 5471 | */ |
| 5472 | if (tcp_header_len == |
| 5473 | (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && |
| 5474 | tp->rcv_nxt == tp->rcv_wup) |
| 5475 | tcp_store_ts_recent(tp); |
| 5476 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 5477 | tcp_rcv_rtt_measure_ts(sk, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5478 | |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 5479 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5480 | |
| 5481 | /* Bulk data transfer: receiver */ |
Eric Dumazet | b081f85 | 2012-05-02 09:58:29 +0000 | [diff] [blame] | 5482 | eaten = tcp_queue_rcv(sk, skb, tcp_header_len, |
| 5483 | &fragstolen); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5484 | } |
| 5485 | |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 5486 | tcp_event_data_recv(sk, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5487 | |
| 5488 | if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) { |
| 5489 | /* Well, only one small jumplet in fast path... */ |
| 5490 | tcp_ack(sk, skb, FLAG_DATA); |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 5491 | tcp_data_snd_check(sk); |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 5492 | if (!inet_csk_ack_scheduled(sk)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5493 | goto no_ack; |
| 5494 | } |
| 5495 | |
Dan Williams | d27f9bc | 2013-12-30 11:37:15 -0800 | [diff] [blame] | 5496 | __tcp_ack_snd_check(sk, 0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5497 | no_ack: |
| 5498 | if (eaten) |
Eric Dumazet | b081f85 | 2012-05-02 09:58:29 +0000 | [diff] [blame] | 5499 | kfree_skb_partial(skb, fragstolen); |
David S. Miller | 676d236 | 2014-04-11 16:15:36 -0400 | [diff] [blame] | 5500 | sk->sk_data_ready(sk); |
Vijay Subramanian | c995ae2 | 2013-09-03 12:23:22 -0700 | [diff] [blame] | 5501 | return; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5502 | } |
| 5503 | } |
| 5504 | |
| 5505 | slow_path: |
Eric Dumazet | fb3477c | 2016-04-29 14:16:48 -0700 | [diff] [blame] | 5506 | if (len < (th->doff << 2) || tcp_checksum_complete(skb)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5507 | goto csum_error; |
| 5508 | |
Calvin Owens | 0c228e8 | 2014-11-20 15:09:53 -0800 | [diff] [blame] | 5509 | if (!th->ack && !th->rst && !th->syn) |
Eric Dumazet | c3ae62af | 2012-12-26 12:44:34 +0000 | [diff] [blame] | 5510 | goto discard; |
| 5511 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5512 | /* |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5513 | * Standard slow path. |
| 5514 | */ |
| 5515 | |
Eric Dumazet | 0c24604 | 2012-07-17 01:41:30 +0000 | [diff] [blame] | 5516 | if (!tcp_validate_incoming(sk, skb, th, 1)) |
Vijay Subramanian | c995ae2 | 2013-09-03 12:23:22 -0700 | [diff] [blame] | 5517 | return; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5518 | |
| 5519 | step5: |
Eric Dumazet | 12fb3dd | 2013-04-19 07:19:48 +0000 | [diff] [blame] | 5520 | if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0) |
John Dykstra | 96e0bf4 | 2009-03-22 21:49:57 -0700 | [diff] [blame] | 5521 | goto discard; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5522 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 5523 | tcp_rcv_rtt_measure_ts(sk, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5524 | |
| 5525 | /* Process urgent data. */ |
| 5526 | tcp_urg(sk, skb, th); |
| 5527 | |
| 5528 | /* step 7: process the segment text */ |
| 5529 | tcp_data_queue(sk, skb); |
| 5530 | |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 5531 | tcp_data_snd_check(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5532 | tcp_ack_snd_check(sk); |
Vijay Subramanian | c995ae2 | 2013-09-03 12:23:22 -0700 | [diff] [blame] | 5533 | return; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5534 | |
| 5535 | csum_error: |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 5536 | TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS); |
| 5537 | TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5538 | |
| 5539 | discard: |
Eric Dumazet | 532182c | 2016-04-01 08:52:19 -0700 | [diff] [blame] | 5540 | tcp_drop(sk, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5541 | } |
Eric Dumazet | 4bc2f18 | 2010-07-09 21:22:10 +0000 | [diff] [blame] | 5542 | EXPORT_SYMBOL(tcp_rcv_established); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5543 | |
Pavel Emelyanov | 370816a | 2012-04-19 03:40:01 +0000 | [diff] [blame] | 5544 | void tcp_finish_connect(struct sock *sk, struct sk_buff *skb) |
| 5545 | { |
| 5546 | struct tcp_sock *tp = tcp_sk(sk); |
| 5547 | struct inet_connection_sock *icsk = inet_csk(sk); |
| 5548 | |
| 5549 | tcp_set_state(sk, TCP_ESTABLISHED); |
| 5550 | |
Ian Morris | 00db412 | 2015-04-03 09:17:27 +0100 | [diff] [blame] | 5551 | if (skb) { |
Eric Dumazet | 5d299f3 | 2012-08-06 05:09:33 +0000 | [diff] [blame] | 5552 | icsk->icsk_af_ops->sk_rx_dst_set(sk, skb); |
Pavel Emelyanov | 370816a | 2012-04-19 03:40:01 +0000 | [diff] [blame] | 5553 | security_inet_conn_established(sk, skb); |
David S. Miller | 41063e9 | 2012-06-19 21:22:05 -0700 | [diff] [blame] | 5554 | } |
Pavel Emelyanov | 370816a | 2012-04-19 03:40:01 +0000 | [diff] [blame] | 5555 | |
| 5556 | /* Make sure socket is routed, for correct metrics. */ |
| 5557 | icsk->icsk_af_ops->rebuild_header(sk); |
| 5558 | |
| 5559 | tcp_init_metrics(sk); |
| 5560 | |
| 5561 | tcp_init_congestion_control(sk); |
| 5562 | |
| 5563 | /* Prevent spurious tcp_cwnd_restart() on first data |
| 5564 | * packet. |
| 5565 | */ |
| 5566 | tp->lsndtime = tcp_time_stamp; |
| 5567 | |
| 5568 | tcp_init_buffer_space(sk); |
| 5569 | |
| 5570 | if (sock_flag(sk, SOCK_KEEPOPEN)) |
| 5571 | inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp)); |
| 5572 | |
| 5573 | if (!tp->rx_opt.snd_wscale) |
| 5574 | __tcp_fast_path_on(tp, tp->snd_wnd); |
| 5575 | else |
| 5576 | tp->pred_flags = 0; |
| 5577 | |
| 5578 | if (!sock_flag(sk, SOCK_DEAD)) { |
| 5579 | sk->sk_state_change(sk); |
| 5580 | sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT); |
| 5581 | } |
| 5582 | } |
| 5583 | |
Yuchung Cheng | 8e4178c | 2012-07-19 06:43:08 +0000 | [diff] [blame] | 5584 | static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack, |
| 5585 | struct tcp_fastopen_cookie *cookie) |
| 5586 | { |
| 5587 | struct tcp_sock *tp = tcp_sk(sk); |
Yuchung Cheng | 67da22d | 2012-07-19 06:43:11 +0000 | [diff] [blame] | 5588 | struct sk_buff *data = tp->syn_data ? tcp_write_queue_head(sk) : NULL; |
Daniel Lee | 2646c83 | 2015-04-06 14:37:27 -0700 | [diff] [blame] | 5589 | u16 mss = tp->rx_opt.mss_clamp, try_exp = 0; |
| 5590 | bool syn_drop = false; |
Yuchung Cheng | 8e4178c | 2012-07-19 06:43:08 +0000 | [diff] [blame] | 5591 | |
| 5592 | if (mss == tp->rx_opt.user_mss) { |
| 5593 | struct tcp_options_received opt; |
Yuchung Cheng | 8e4178c | 2012-07-19 06:43:08 +0000 | [diff] [blame] | 5594 | |
| 5595 | /* Get original SYNACK MSS value if user MSS sets mss_clamp */ |
| 5596 | tcp_clear_options(&opt); |
| 5597 | opt.user_mss = opt.mss_clamp = 0; |
Christoph Paasch | 1a2c618 | 2013-03-17 08:23:34 +0000 | [diff] [blame] | 5598 | tcp_parse_options(synack, &opt, 0, NULL); |
Yuchung Cheng | 8e4178c | 2012-07-19 06:43:08 +0000 | [diff] [blame] | 5599 | mss = opt.mss_clamp; |
| 5600 | } |
| 5601 | |
Daniel Lee | 2646c83 | 2015-04-06 14:37:27 -0700 | [diff] [blame] | 5602 | if (!tp->syn_fastopen) { |
| 5603 | /* Ignore an unsolicited cookie */ |
Yuchung Cheng | 67da22d | 2012-07-19 06:43:11 +0000 | [diff] [blame] | 5604 | cookie->len = -1; |
Daniel Lee | 2646c83 | 2015-04-06 14:37:27 -0700 | [diff] [blame] | 5605 | } else if (tp->total_retrans) { |
| 5606 | /* SYN timed out and the SYN-ACK neither has a cookie nor |
| 5607 | * acknowledges data. Presumably the remote received only |
| 5608 | * the retransmitted (regular) SYNs: either the original |
| 5609 | * SYN-data or the corresponding SYN-ACK was dropped. |
| 5610 | */ |
| 5611 | syn_drop = (cookie->len < 0 && data); |
| 5612 | } else if (cookie->len < 0 && !tp->syn_data) { |
| 5613 | /* We requested a cookie but didn't get it. If we did not use |
| 5614 | * the (old) exp opt format then try so next time (try_exp=1). |
| 5615 | * Otherwise we go back to use the RFC7413 opt (try_exp=2). |
| 5616 | */ |
| 5617 | try_exp = tp->syn_fastopen_exp ? 2 : 1; |
| 5618 | } |
Yuchung Cheng | 67da22d | 2012-07-19 06:43:11 +0000 | [diff] [blame] | 5619 | |
Daniel Lee | 2646c83 | 2015-04-06 14:37:27 -0700 | [diff] [blame] | 5620 | tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp); |
Yuchung Cheng | 8e4178c | 2012-07-19 06:43:08 +0000 | [diff] [blame] | 5621 | |
| 5622 | if (data) { /* Retransmit unacked data in SYN */ |
Yuchung Cheng | 93b174a | 2012-12-06 08:45:32 +0000 | [diff] [blame] | 5623 | tcp_for_write_queue_from(data, sk) { |
| 5624 | if (data == tcp_send_head(sk) || |
Eric Dumazet | 10d3be5 | 2016-04-21 10:55:23 -0700 | [diff] [blame] | 5625 | __tcp_retransmit_skb(sk, data, 1)) |
Yuchung Cheng | 93b174a | 2012-12-06 08:45:32 +0000 | [diff] [blame] | 5626 | break; |
| 5627 | } |
Yuchung Cheng | 8e4178c | 2012-07-19 06:43:08 +0000 | [diff] [blame] | 5628 | tcp_rearm_rto(sk); |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 5629 | NET_INC_STATS(sock_net(sk), |
Eric Dumazet | 02a1d6e | 2016-04-27 16:44:39 -0700 | [diff] [blame] | 5630 | LINUX_MIB_TCPFASTOPENACTIVEFAIL); |
Yuchung Cheng | 8e4178c | 2012-07-19 06:43:08 +0000 | [diff] [blame] | 5631 | return true; |
| 5632 | } |
Yuchung Cheng | 6f73601 | 2012-10-19 15:14:44 +0000 | [diff] [blame] | 5633 | tp->syn_data_acked = tp->syn_data; |
Yuchung Cheng | f19c29e | 2014-03-03 12:31:36 -0800 | [diff] [blame] | 5634 | if (tp->syn_data_acked) |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 5635 | NET_INC_STATS(sock_net(sk), |
Eric Dumazet | 02a1d6e | 2016-04-27 16:44:39 -0700 | [diff] [blame] | 5636 | LINUX_MIB_TCPFASTOPENACTIVE); |
Eric Dumazet | 61d2bca | 2016-02-01 21:03:07 -0800 | [diff] [blame] | 5637 | |
| 5638 | tcp_fastopen_add_skb(sk, synack); |
| 5639 | |
Yuchung Cheng | 8e4178c | 2012-07-19 06:43:08 +0000 | [diff] [blame] | 5640 | return false; |
| 5641 | } |
| 5642 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5643 | static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb, |
Eric Dumazet | bda07a6 | 2015-09-29 07:42:40 -0700 | [diff] [blame] | 5644 | const struct tcphdr *th) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5645 | { |
Arnaldo Carvalho de Melo | d83d846 | 2005-12-13 23:26:10 -0800 | [diff] [blame] | 5646 | struct inet_connection_sock *icsk = inet_csk(sk); |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 5647 | struct tcp_sock *tp = tcp_sk(sk); |
Yuchung Cheng | 8e4178c | 2012-07-19 06:43:08 +0000 | [diff] [blame] | 5648 | struct tcp_fastopen_cookie foc = { .len = -1 }; |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 5649 | int saved_clamp = tp->rx_opt.mss_clamp; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5650 | |
Christoph Paasch | 1a2c618 | 2013-03-17 08:23:34 +0000 | [diff] [blame] | 5651 | tcp_parse_options(skb, &tp->rx_opt, 0, &foc); |
Andrew Vagin | e3e1202 | 2013-08-27 12:21:55 +0400 | [diff] [blame] | 5652 | if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) |
Andrey Vagin | ee684b6 | 2013-02-11 05:50:19 +0000 | [diff] [blame] | 5653 | tp->rx_opt.rcv_tsecr -= tp->tsoffset; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5654 | |
| 5655 | if (th->ack) { |
| 5656 | /* rfc793: |
| 5657 | * "If the state is SYN-SENT then |
| 5658 | * first check the ACK bit |
| 5659 | * If the ACK bit is set |
| 5660 | * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send |
| 5661 | * a reset (unless the RST bit is set, if so drop |
| 5662 | * the segment and return)" |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5663 | */ |
Yuchung Cheng | 8e4178c | 2012-07-19 06:43:08 +0000 | [diff] [blame] | 5664 | if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) || |
| 5665 | after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5666 | goto reset_and_undo; |
| 5667 | |
| 5668 | if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && |
| 5669 | !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp, |
| 5670 | tcp_time_stamp)) { |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 5671 | NET_INC_STATS(sock_net(sk), |
Eric Dumazet | 02a1d6e | 2016-04-27 16:44:39 -0700 | [diff] [blame] | 5672 | LINUX_MIB_PAWSACTIVEREJECTED); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5673 | goto reset_and_undo; |
| 5674 | } |
| 5675 | |
| 5676 | /* Now ACK is acceptable. |
| 5677 | * |
| 5678 | * "If the RST bit is set |
| 5679 | * If the ACK was acceptable then signal the user "error: |
| 5680 | * connection reset", drop the segment, enter CLOSED state, |
| 5681 | * delete TCB, and return." |
| 5682 | */ |
| 5683 | |
| 5684 | if (th->rst) { |
| 5685 | tcp_reset(sk); |
| 5686 | goto discard; |
| 5687 | } |
| 5688 | |
| 5689 | /* rfc793: |
| 5690 | * "fifth, if neither of the SYN or RST bits is set then |
| 5691 | * drop the segment and return." |
| 5692 | * |
| 5693 | * See note below! |
| 5694 | * --ANK(990513) |
| 5695 | */ |
| 5696 | if (!th->syn) |
| 5697 | goto discard_and_undo; |
| 5698 | |
| 5699 | /* rfc793: |
| 5700 | * "If the SYN bit is on ... |
| 5701 | * are acceptable then ... |
| 5702 | * (our SYN has been ACKed), change the connection |
| 5703 | * state to ESTABLISHED..." |
| 5704 | */ |
| 5705 | |
Florian Westphal | 735d383 | 2014-09-29 13:08:30 +0200 | [diff] [blame] | 5706 | tcp_ecn_rcv_synack(tp, th); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5707 | |
Razvan Ghitulete | 1b3a692 | 2012-08-14 16:30:20 +0300 | [diff] [blame] | 5708 | tcp_init_wl(tp, TCP_SKB_CB(skb)->seq); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5709 | tcp_ack(sk, skb, FLAG_SLOWPATH); |
| 5710 | |
| 5711 | /* Ok.. it's good. Set up sequence numbers and |
| 5712 | * move to established. |
| 5713 | */ |
| 5714 | tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; |
| 5715 | tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; |
| 5716 | |
| 5717 | /* RFC1323: The window in SYN & SYN/ACK segments is |
| 5718 | * never scaled. |
| 5719 | */ |
| 5720 | tp->snd_wnd = ntohs(th->window); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5721 | |
| 5722 | if (!tp->rx_opt.wscale_ok) { |
| 5723 | tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0; |
| 5724 | tp->window_clamp = min(tp->window_clamp, 65535U); |
| 5725 | } |
| 5726 | |
| 5727 | if (tp->rx_opt.saw_tstamp) { |
| 5728 | tp->rx_opt.tstamp_ok = 1; |
| 5729 | tp->tcp_header_len = |
| 5730 | sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; |
| 5731 | tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; |
| 5732 | tcp_store_ts_recent(tp); |
| 5733 | } else { |
| 5734 | tp->tcp_header_len = sizeof(struct tcphdr); |
| 5735 | } |
| 5736 | |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 5737 | if (tcp_is_sack(tp) && sysctl_tcp_fack) |
| 5738 | tcp_enable_fack(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5739 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 5740 | tcp_mtup_init(sk); |
Arnaldo Carvalho de Melo | d83d846 | 2005-12-13 23:26:10 -0800 | [diff] [blame] | 5741 | tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5742 | tcp_initialize_rcv_mss(sk); |
| 5743 | |
| 5744 | /* Remember, tcp_poll() does not lock socket! |
| 5745 | * Change state from SYN-SENT only after copied_seq |
| 5746 | * is initialized. */ |
| 5747 | tp->copied_seq = tp->rcv_nxt; |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 5748 | |
Ralf Baechle | e16aa20 | 2006-12-07 00:11:33 -0800 | [diff] [blame] | 5749 | smp_mb(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5750 | |
Pavel Emelyanov | 370816a | 2012-04-19 03:40:01 +0000 | [diff] [blame] | 5751 | tcp_finish_connect(sk, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5752 | |
Yuchung Cheng | 67da22d | 2012-07-19 06:43:11 +0000 | [diff] [blame] | 5753 | if ((tp->syn_fastopen || tp->syn_data) && |
| 5754 | tcp_rcv_fastopen_synack(sk, skb, &foc)) |
Yuchung Cheng | 8e4178c | 2012-07-19 06:43:08 +0000 | [diff] [blame] | 5755 | return -1; |
| 5756 | |
Arnaldo Carvalho de Melo | 295f732 | 2005-08-09 20:11:56 -0700 | [diff] [blame] | 5757 | if (sk->sk_write_pending || |
| 5758 | icsk->icsk_accept_queue.rskq_defer_accept || |
| 5759 | icsk->icsk_ack.pingpong) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5760 | /* Save one ACK. Data will be ready after |
| 5761 | * several ticks, if write_pending is set. |
| 5762 | * |
| 5763 | * It may be deleted, but with this feature tcpdumps |
| 5764 | * look so _wonderfully_ clever, that I was not able |
| 5765 | * to stand against the temptation 8) --ANK |
| 5766 | */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 5767 | inet_csk_schedule_ack(sk); |
Arnaldo Carvalho de Melo | 295f732 | 2005-08-09 20:11:56 -0700 | [diff] [blame] | 5768 | icsk->icsk_ack.lrcvtime = tcp_time_stamp; |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 5769 | tcp_enter_quickack_mode(sk); |
Arnaldo Carvalho de Melo | 3f421ba | 2005-08-09 20:11:08 -0700 | [diff] [blame] | 5770 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, |
| 5771 | TCP_DELACK_MAX, TCP_RTO_MAX); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5772 | |
| 5773 | discard: |
Eric Dumazet | 532182c | 2016-04-01 08:52:19 -0700 | [diff] [blame] | 5774 | tcp_drop(sk, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5775 | return 0; |
| 5776 | } else { |
| 5777 | tcp_send_ack(sk); |
| 5778 | } |
| 5779 | return -1; |
| 5780 | } |
| 5781 | |
| 5782 | /* No ACK in the segment */ |
| 5783 | |
| 5784 | if (th->rst) { |
| 5785 | /* rfc793: |
| 5786 | * "If the RST bit is set |
| 5787 | * |
| 5788 | * Otherwise (no ACK) drop the segment and return." |
| 5789 | */ |
| 5790 | |
| 5791 | goto discard_and_undo; |
| 5792 | } |
| 5793 | |
| 5794 | /* PAWS check. */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 5795 | if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && |
Ilpo Järvinen | c887e6d | 2009-03-14 14:23:03 +0000 | [diff] [blame] | 5796 | tcp_paws_reject(&tp->rx_opt, 0)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5797 | goto discard_and_undo; |
| 5798 | |
| 5799 | if (th->syn) { |
| 5800 | /* We see SYN without ACK. It is attempt of |
| 5801 | * simultaneous connect with crossed SYNs. |
| 5802 | * Particularly, it can be connect to self. |
| 5803 | */ |
| 5804 | tcp_set_state(sk, TCP_SYN_RECV); |
| 5805 | |
| 5806 | if (tp->rx_opt.saw_tstamp) { |
| 5807 | tp->rx_opt.tstamp_ok = 1; |
| 5808 | tcp_store_ts_recent(tp); |
| 5809 | tp->tcp_header_len = |
| 5810 | sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; |
| 5811 | } else { |
| 5812 | tp->tcp_header_len = sizeof(struct tcphdr); |
| 5813 | } |
| 5814 | |
| 5815 | tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; |
Eric Dumazet | 142a2e7 | 2015-11-26 08:18:14 -0800 | [diff] [blame] | 5816 | tp->copied_seq = tp->rcv_nxt; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5817 | tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; |
| 5818 | |
| 5819 | /* RFC1323: The window in SYN & SYN/ACK segments is |
| 5820 | * never scaled. |
| 5821 | */ |
| 5822 | tp->snd_wnd = ntohs(th->window); |
| 5823 | tp->snd_wl1 = TCP_SKB_CB(skb)->seq; |
| 5824 | tp->max_window = tp->snd_wnd; |
| 5825 | |
Florian Westphal | 735d383 | 2014-09-29 13:08:30 +0200 | [diff] [blame] | 5826 | tcp_ecn_rcv_syn(tp, th); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5827 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 5828 | tcp_mtup_init(sk); |
Arnaldo Carvalho de Melo | d83d846 | 2005-12-13 23:26:10 -0800 | [diff] [blame] | 5829 | tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5830 | tcp_initialize_rcv_mss(sk); |
| 5831 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5832 | tcp_send_synack(sk); |
| 5833 | #if 0 |
| 5834 | /* Note, we could accept data and URG from this segment. |
Jerry Chu | 168a8f5 | 2012-08-31 12:29:13 +0000 | [diff] [blame] | 5835 | * There are no obstacles to make this (except that we must |
| 5836 | * either change tcp_recvmsg() to prevent it from returning data |
| 5837 | * before 3WHS completes per RFC793, or employ TCP Fast Open). |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5838 | * |
| 5839 | * However, if we ignore data in ACKless segments sometimes, |
| 5840 | * we have no reasons to accept it sometimes. |
| 5841 | * Also, seems the code doing it in step6 of tcp_rcv_state_process |
| 5842 | * is not flawless. So, discard packet for sanity. |
| 5843 | * Uncomment this return to process the data. |
| 5844 | */ |
| 5845 | return -1; |
| 5846 | #else |
| 5847 | goto discard; |
| 5848 | #endif |
| 5849 | } |
| 5850 | /* "fifth, if neither of the SYN or RST bits is set then |
| 5851 | * drop the segment and return." |
| 5852 | */ |
| 5853 | |
| 5854 | discard_and_undo: |
| 5855 | tcp_clear_options(&tp->rx_opt); |
| 5856 | tp->rx_opt.mss_clamp = saved_clamp; |
| 5857 | goto discard; |
| 5858 | |
| 5859 | reset_and_undo: |
| 5860 | tcp_clear_options(&tp->rx_opt); |
| 5861 | tp->rx_opt.mss_clamp = saved_clamp; |
| 5862 | return 1; |
| 5863 | } |
| 5864 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5865 | /* |
| 5866 | * This function implements the receiving procedure of RFC 793 for |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 5867 | * all states except ESTABLISHED and TIME_WAIT. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5868 | * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be |
| 5869 | * address independent. |
| 5870 | */ |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 5871 | |
Eric Dumazet | 72ab4a8 | 2015-09-29 07:42:41 -0700 | [diff] [blame] | 5872 | int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5873 | { |
| 5874 | struct tcp_sock *tp = tcp_sk(sk); |
Arnaldo Carvalho de Melo | 8292a17 | 2005-12-13 23:15:52 -0800 | [diff] [blame] | 5875 | struct inet_connection_sock *icsk = inet_csk(sk); |
Eric Dumazet | 72ab4a8 | 2015-09-29 07:42:41 -0700 | [diff] [blame] | 5876 | const struct tcphdr *th = tcp_hdr(skb); |
Jerry Chu | 168a8f5 | 2012-08-31 12:29:13 +0000 | [diff] [blame] | 5877 | struct request_sock *req; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5878 | int queued = 0; |
Eric Dumazet | 1f6afc8 | 2013-05-24 15:03:54 +0000 | [diff] [blame] | 5879 | bool acceptable; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5880 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5881 | switch (sk->sk_state) { |
| 5882 | case TCP_CLOSE: |
| 5883 | goto discard; |
| 5884 | |
| 5885 | case TCP_LISTEN: |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 5886 | if (th->ack) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5887 | return 1; |
| 5888 | |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 5889 | if (th->rst) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5890 | goto discard; |
| 5891 | |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 5892 | if (th->syn) { |
Eric Dumazet | fdf5af0 | 2011-12-02 23:41:42 +0000 | [diff] [blame] | 5893 | if (th->fin) |
| 5894 | goto discard; |
Arnaldo Carvalho de Melo | 8292a17 | 2005-12-13 23:15:52 -0800 | [diff] [blame] | 5895 | if (icsk->icsk_af_ops->conn_request(sk, skb) < 0) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5896 | return 1; |
| 5897 | |
Eric Dumazet | 0aea76d | 2016-04-21 22:13:01 -0700 | [diff] [blame] | 5898 | consume_skb(skb); |
Masayuki Nakagawa | fb7e239 | 2007-01-23 20:15:06 -0800 | [diff] [blame] | 5899 | return 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5900 | } |
| 5901 | goto discard; |
| 5902 | |
| 5903 | case TCP_SYN_SENT: |
Eric Dumazet | 8804b27 | 2016-04-13 22:05:40 -0700 | [diff] [blame] | 5904 | tp->rx_opt.saw_tstamp = 0; |
Eric Dumazet | bda07a6 | 2015-09-29 07:42:40 -0700 | [diff] [blame] | 5905 | queued = tcp_rcv_synsent_state_process(sk, skb, th); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5906 | if (queued >= 0) |
| 5907 | return queued; |
| 5908 | |
| 5909 | /* Do step6 onward by hand. */ |
| 5910 | tcp_urg(sk, skb, th); |
| 5911 | __kfree_skb(skb); |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 5912 | tcp_data_snd_check(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5913 | return 0; |
| 5914 | } |
| 5915 | |
Eric Dumazet | 8804b27 | 2016-04-13 22:05:40 -0700 | [diff] [blame] | 5916 | tp->rx_opt.saw_tstamp = 0; |
Jerry Chu | 168a8f5 | 2012-08-31 12:29:13 +0000 | [diff] [blame] | 5917 | req = tp->fastopen_rsk; |
Ian Morris | 00db412 | 2015-04-03 09:17:27 +0100 | [diff] [blame] | 5918 | if (req) { |
Jerry Chu | 37561f6 | 2012-10-22 11:26:36 +0000 | [diff] [blame] | 5919 | WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV && |
Jerry Chu | 168a8f5 | 2012-08-31 12:29:13 +0000 | [diff] [blame] | 5920 | sk->sk_state != TCP_FIN_WAIT1); |
| 5921 | |
Ian Morris | 51456b2 | 2015-04-03 09:17:26 +0100 | [diff] [blame] | 5922 | if (!tcp_check_req(sk, skb, req, true)) |
Jerry Chu | 168a8f5 | 2012-08-31 12:29:13 +0000 | [diff] [blame] | 5923 | goto discard; |
Neal Cardwell | e69bebd | 2012-09-22 04:18:57 +0000 | [diff] [blame] | 5924 | } |
Eric Dumazet | c3ae62af | 2012-12-26 12:44:34 +0000 | [diff] [blame] | 5925 | |
Calvin Owens | 0c228e8 | 2014-11-20 15:09:53 -0800 | [diff] [blame] | 5926 | if (!th->ack && !th->rst && !th->syn) |
Eric Dumazet | c3ae62af | 2012-12-26 12:44:34 +0000 | [diff] [blame] | 5927 | goto discard; |
| 5928 | |
Neal Cardwell | e69bebd | 2012-09-22 04:18:57 +0000 | [diff] [blame] | 5929 | if (!tcp_validate_incoming(sk, skb, th, 0)) |
Eric Dumazet | 0c24604 | 2012-07-17 01:41:30 +0000 | [diff] [blame] | 5930 | return 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5931 | |
| 5932 | /* step 5: check the ACK field */ |
Eric Dumazet | 1f6afc8 | 2013-05-24 15:03:54 +0000 | [diff] [blame] | 5933 | acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH | |
| 5934 | FLAG_UPDATE_TS_RECENT) > 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 5935 | |
Eric Dumazet | 1f6afc8 | 2013-05-24 15:03:54 +0000 | [diff] [blame] | 5936 | switch (sk->sk_state) { |
| 5937 | case TCP_SYN_RECV: |
Joe Perches | 61eb9003 | 2013-05-24 18:36:13 +0000 | [diff] [blame] | 5938 | if (!acceptable) |
Eric Dumazet | 1f6afc8 | 2013-05-24 15:03:54 +0000 | [diff] [blame] | 5939 | return 1; |
Joe Perches | 61eb9003 | 2013-05-24 18:36:13 +0000 | [diff] [blame] | 5940 | |
Yuchung Cheng | 0f1c28a | 2015-09-18 11:36:14 -0700 | [diff] [blame] | 5941 | if (!tp->srtt_us) |
| 5942 | tcp_synack_rtt_meas(sk, req); |
| 5943 | |
Joe Perches | 61eb9003 | 2013-05-24 18:36:13 +0000 | [diff] [blame] | 5944 | /* Once we leave TCP_SYN_RECV, we no longer need req |
| 5945 | * so release it. |
| 5946 | */ |
| 5947 | if (req) { |
Yuchung Cheng | 7e32b44 | 2016-09-21 16:16:15 -0700 | [diff] [blame] | 5948 | inet_csk(sk)->icsk_retransmits = 0; |
Joe Perches | 61eb9003 | 2013-05-24 18:36:13 +0000 | [diff] [blame] | 5949 | reqsk_fastopen_remove(sk, req, false); |
| 5950 | } else { |
| 5951 | /* Make sure socket is routed, for correct metrics. */ |
| 5952 | icsk->icsk_af_ops->rebuild_header(sk); |
| 5953 | tcp_init_congestion_control(sk); |
| 5954 | |
| 5955 | tcp_mtup_init(sk); |
Joe Perches | 61eb9003 | 2013-05-24 18:36:13 +0000 | [diff] [blame] | 5956 | tp->copied_seq = tp->rcv_nxt; |
Eric Dumazet | b0983d3 | 2013-09-20 13:56:58 -0700 | [diff] [blame] | 5957 | tcp_init_buffer_space(sk); |
Eric Dumazet | 1f6afc8 | 2013-05-24 15:03:54 +0000 | [diff] [blame] | 5958 | } |
Joe Perches | 61eb9003 | 2013-05-24 18:36:13 +0000 | [diff] [blame] | 5959 | smp_mb(); |
| 5960 | tcp_set_state(sk, TCP_ESTABLISHED); |
| 5961 | sk->sk_state_change(sk); |
| 5962 | |
| 5963 | /* Note, that this wakeup is only for marginal crossed SYN case. |
| 5964 | * Passively open sockets are not waked up, because |
| 5965 | * sk->sk_sleep == NULL and sk->sk_socket == NULL. |
| 5966 | */ |
| 5967 | if (sk->sk_socket) |
| 5968 | sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT); |
| 5969 | |
| 5970 | tp->snd_una = TCP_SKB_CB(skb)->ack_seq; |
| 5971 | tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale; |
| 5972 | tcp_init_wl(tp, TCP_SKB_CB(skb)->seq); |
| 5973 | |
| 5974 | if (tp->rx_opt.tstamp_ok) |
| 5975 | tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; |
| 5976 | |
| 5977 | if (req) { |
| 5978 | /* Re-arm the timer because data may have been sent out. |
| 5979 | * This is similar to the regular data transmission case |
| 5980 | * when new data has just been ack'ed. |
| 5981 | * |
| 5982 | * (TFO) - we could try to be more aggressive and |
| 5983 | * retransmitting any data sooner based on when they |
| 5984 | * are sent out. |
| 5985 | */ |
| 5986 | tcp_rearm_rto(sk); |
| 5987 | } else |
| 5988 | tcp_init_metrics(sk); |
| 5989 | |
Yuchung Cheng | c040276 | 2016-09-19 23:39:21 -0400 | [diff] [blame] | 5990 | if (!inet_csk(sk)->icsk_ca_ops->cong_control) |
| 5991 | tcp_update_pacing_rate(sk); |
Neal Cardwell | 02cf4eb | 2013-10-21 15:40:19 -0400 | [diff] [blame] | 5992 | |
Joe Perches | 61eb9003 | 2013-05-24 18:36:13 +0000 | [diff] [blame] | 5993 | /* Prevent spurious tcp_cwnd_restart() on first data packet */ |
| 5994 | tp->lsndtime = tcp_time_stamp; |
| 5995 | |
| 5996 | tcp_initialize_rcv_mss(sk); |
| 5997 | tcp_fast_path_on(tp); |
Eric Dumazet | 1f6afc8 | 2013-05-24 15:03:54 +0000 | [diff] [blame] | 5998 | break; |
| 5999 | |
Joe Perches | c48b22d | 2013-05-24 18:06:58 +0000 | [diff] [blame] | 6000 | case TCP_FIN_WAIT1: { |
| 6001 | struct dst_entry *dst; |
| 6002 | int tmo; |
| 6003 | |
Eric Dumazet | 1f6afc8 | 2013-05-24 15:03:54 +0000 | [diff] [blame] | 6004 | /* If we enter the TCP_FIN_WAIT1 state and we are a |
| 6005 | * Fast Open socket and this is the first acceptable |
| 6006 | * ACK we have received, this would have acknowledged |
| 6007 | * our SYNACK so stop the SYNACK timer. |
| 6008 | */ |
Ian Morris | 00db412 | 2015-04-03 09:17:27 +0100 | [diff] [blame] | 6009 | if (req) { |
Eric Dumazet | 1f6afc8 | 2013-05-24 15:03:54 +0000 | [diff] [blame] | 6010 | /* Return RST if ack_seq is invalid. |
| 6011 | * Note that RFC793 only says to generate a |
| 6012 | * DUPACK for it but for TCP Fast Open it seems |
| 6013 | * better to treat this case like TCP_SYN_RECV |
| 6014 | * above. |
| 6015 | */ |
| 6016 | if (!acceptable) |
| 6017 | return 1; |
| 6018 | /* We no longer need the request sock. */ |
| 6019 | reqsk_fastopen_remove(sk, req, false); |
| 6020 | tcp_rearm_rto(sk); |
| 6021 | } |
Joe Perches | c48b22d | 2013-05-24 18:06:58 +0000 | [diff] [blame] | 6022 | if (tp->snd_una != tp->write_seq) |
| 6023 | break; |
Eric Dumazet | 1f6afc8 | 2013-05-24 15:03:54 +0000 | [diff] [blame] | 6024 | |
Joe Perches | c48b22d | 2013-05-24 18:06:58 +0000 | [diff] [blame] | 6025 | tcp_set_state(sk, TCP_FIN_WAIT2); |
| 6026 | sk->sk_shutdown |= SEND_SHUTDOWN; |
Eric Dumazet | 1f6afc8 | 2013-05-24 15:03:54 +0000 | [diff] [blame] | 6027 | |
Joe Perches | c48b22d | 2013-05-24 18:06:58 +0000 | [diff] [blame] | 6028 | dst = __sk_dst_get(sk); |
| 6029 | if (dst) |
| 6030 | dst_confirm(dst); |
Eric Dumazet | 1f6afc8 | 2013-05-24 15:03:54 +0000 | [diff] [blame] | 6031 | |
Joe Perches | c48b22d | 2013-05-24 18:06:58 +0000 | [diff] [blame] | 6032 | if (!sock_flag(sk, SOCK_DEAD)) { |
| 6033 | /* Wake up lingering close() */ |
| 6034 | sk->sk_state_change(sk); |
| 6035 | break; |
| 6036 | } |
Eric Dumazet | 1f6afc8 | 2013-05-24 15:03:54 +0000 | [diff] [blame] | 6037 | |
Joe Perches | c48b22d | 2013-05-24 18:06:58 +0000 | [diff] [blame] | 6038 | if (tp->linger2 < 0 || |
| 6039 | (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && |
| 6040 | after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) { |
| 6041 | tcp_done(sk); |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 6042 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA); |
Joe Perches | c48b22d | 2013-05-24 18:06:58 +0000 | [diff] [blame] | 6043 | return 1; |
| 6044 | } |
Eric Dumazet | 1f6afc8 | 2013-05-24 15:03:54 +0000 | [diff] [blame] | 6045 | |
Joe Perches | c48b22d | 2013-05-24 18:06:58 +0000 | [diff] [blame] | 6046 | tmo = tcp_fin_time(sk); |
| 6047 | if (tmo > TCP_TIMEWAIT_LEN) { |
| 6048 | inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN); |
| 6049 | } else if (th->fin || sock_owned_by_user(sk)) { |
| 6050 | /* Bad case. We could lose such FIN otherwise. |
| 6051 | * It is not a big problem, but it looks confusing |
| 6052 | * and not so rare event. We still can lose it now, |
| 6053 | * if it spins in bh_lock_sock(), but it is really |
| 6054 | * marginal case. |
| 6055 | */ |
| 6056 | inet_csk_reset_keepalive_timer(sk, tmo); |
| 6057 | } else { |
| 6058 | tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); |
| 6059 | goto discard; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 6060 | } |
Eric Dumazet | 1f6afc8 | 2013-05-24 15:03:54 +0000 | [diff] [blame] | 6061 | break; |
Joe Perches | c48b22d | 2013-05-24 18:06:58 +0000 | [diff] [blame] | 6062 | } |
Eric Dumazet | 1f6afc8 | 2013-05-24 15:03:54 +0000 | [diff] [blame] | 6063 | |
| 6064 | case TCP_CLOSING: |
| 6065 | if (tp->snd_una == tp->write_seq) { |
| 6066 | tcp_time_wait(sk, TCP_TIME_WAIT, 0); |
| 6067 | goto discard; |
| 6068 | } |
| 6069 | break; |
| 6070 | |
| 6071 | case TCP_LAST_ACK: |
| 6072 | if (tp->snd_una == tp->write_seq) { |
| 6073 | tcp_update_metrics(sk); |
| 6074 | tcp_done(sk); |
| 6075 | goto discard; |
| 6076 | } |
| 6077 | break; |
Eric Dumazet | c3ae62af | 2012-12-26 12:44:34 +0000 | [diff] [blame] | 6078 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 6079 | |
| 6080 | /* step 6: check the URG bit */ |
| 6081 | tcp_urg(sk, skb, th); |
| 6082 | |
| 6083 | /* step 7: process the segment text */ |
| 6084 | switch (sk->sk_state) { |
| 6085 | case TCP_CLOSE_WAIT: |
| 6086 | case TCP_CLOSING: |
| 6087 | case TCP_LAST_ACK: |
| 6088 | if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) |
| 6089 | break; |
| 6090 | case TCP_FIN_WAIT1: |
| 6091 | case TCP_FIN_WAIT2: |
| 6092 | /* RFC 793 says to queue data in these states, |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 6093 | * RFC 1122 says we MUST send a reset. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 6094 | * BSD 4.4 also does reset. |
| 6095 | */ |
| 6096 | if (sk->sk_shutdown & RCV_SHUTDOWN) { |
| 6097 | if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && |
| 6098 | after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 6099 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 6100 | tcp_reset(sk); |
| 6101 | return 1; |
| 6102 | } |
| 6103 | } |
| 6104 | /* Fall through */ |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 6105 | case TCP_ESTABLISHED: |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 6106 | tcp_data_queue(sk, skb); |
| 6107 | queued = 1; |
| 6108 | break; |
| 6109 | } |
| 6110 | |
| 6111 | /* tcp_data could move socket to TIME-WAIT */ |
| 6112 | if (sk->sk_state != TCP_CLOSE) { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 6113 | tcp_data_snd_check(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 6114 | tcp_ack_snd_check(sk); |
| 6115 | } |
| 6116 | |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 6117 | if (!queued) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 6118 | discard: |
Eric Dumazet | 532182c | 2016-04-01 08:52:19 -0700 | [diff] [blame] | 6119 | tcp_drop(sk, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 6120 | } |
| 6121 | return 0; |
| 6122 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 6123 | EXPORT_SYMBOL(tcp_rcv_state_process); |
Octavian Purdila | 1fb6f15 | 2014-06-25 17:10:02 +0300 | [diff] [blame] | 6124 | |
| 6125 | static inline void pr_drop_req(struct request_sock *req, __u16 port, int family) |
| 6126 | { |
| 6127 | struct inet_request_sock *ireq = inet_rsk(req); |
| 6128 | |
| 6129 | if (family == AF_INET) |
Joe Perches | ba7a46f | 2014-11-11 10:59:17 -0800 | [diff] [blame] | 6130 | net_dbg_ratelimited("drop open request from %pI4/%u\n", |
| 6131 | &ireq->ir_rmt_addr, port); |
Octavian Purdila | 4135ab8 | 2014-06-28 21:20:54 +0300 | [diff] [blame] | 6132 | #if IS_ENABLED(CONFIG_IPV6) |
| 6133 | else if (family == AF_INET6) |
Joe Perches | ba7a46f | 2014-11-11 10:59:17 -0800 | [diff] [blame] | 6134 | net_dbg_ratelimited("drop open request from %pI6/%u\n", |
| 6135 | &ireq->ir_v6_rmt_addr, port); |
Octavian Purdila | 4135ab8 | 2014-06-28 21:20:54 +0300 | [diff] [blame] | 6136 | #endif |
Octavian Purdila | 1fb6f15 | 2014-06-25 17:10:02 +0300 | [diff] [blame] | 6137 | } |
| 6138 | |
Florian Westphal | d82bd12 | 2014-09-29 13:08:29 +0200 | [diff] [blame] | 6139 | /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set |
| 6140 | * |
| 6141 | * If we receive a SYN packet with these bits set, it means a |
| 6142 | * network is playing bad games with TOS bits. In order to |
| 6143 | * avoid possible false congestion notifications, we disable |
stephen hemminger | f4e715c | 2014-10-29 16:05:06 -0700 | [diff] [blame] | 6144 | * TCP ECN negotiation. |
Florian Westphal | d82bd12 | 2014-09-29 13:08:29 +0200 | [diff] [blame] | 6145 | * |
| 6146 | * Exception: tcp_ca wants ECN. This is required for DCTCP |
Florian Westphal | 843c2fd | 2015-01-30 20:45:20 +0100 | [diff] [blame] | 6147 | * congestion control: Linux DCTCP asserts ECT on all packets, |
| 6148 | * including SYN, which is most optimal solution; however, |
| 6149 | * others, such as FreeBSD do not. |
Florian Westphal | d82bd12 | 2014-09-29 13:08:29 +0200 | [diff] [blame] | 6150 | */ |
| 6151 | static void tcp_ecn_create_request(struct request_sock *req, |
| 6152 | const struct sk_buff *skb, |
Florian Westphal | f7b3bec | 2014-11-03 17:35:03 +0100 | [diff] [blame] | 6153 | const struct sock *listen_sk, |
| 6154 | const struct dst_entry *dst) |
Florian Westphal | d82bd12 | 2014-09-29 13:08:29 +0200 | [diff] [blame] | 6155 | { |
| 6156 | const struct tcphdr *th = tcp_hdr(skb); |
| 6157 | const struct net *net = sock_net(listen_sk); |
| 6158 | bool th_ecn = th->ece && th->cwr; |
Florian Westphal | 843c2fd | 2015-01-30 20:45:20 +0100 | [diff] [blame] | 6159 | bool ect, ecn_ok; |
Daniel Borkmann | c3a8d94 | 2015-08-31 15:58:47 +0200 | [diff] [blame] | 6160 | u32 ecn_ok_dst; |
Florian Westphal | d82bd12 | 2014-09-29 13:08:29 +0200 | [diff] [blame] | 6161 | |
| 6162 | if (!th_ecn) |
| 6163 | return; |
| 6164 | |
| 6165 | ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield); |
Daniel Borkmann | c3a8d94 | 2015-08-31 15:58:47 +0200 | [diff] [blame] | 6166 | ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK); |
| 6167 | ecn_ok = net->ipv4.sysctl_tcp_ecn || ecn_ok_dst; |
Florian Westphal | d82bd12 | 2014-09-29 13:08:29 +0200 | [diff] [blame] | 6168 | |
Daniel Borkmann | c3a8d94 | 2015-08-31 15:58:47 +0200 | [diff] [blame] | 6169 | if ((!ect && ecn_ok) || tcp_ca_needs_ecn(listen_sk) || |
| 6170 | (ecn_ok_dst & DST_FEATURE_ECN_CA)) |
Florian Westphal | d82bd12 | 2014-09-29 13:08:29 +0200 | [diff] [blame] | 6171 | inet_rsk(req)->ecn_ok = 1; |
| 6172 | } |
| 6173 | |
Eric Dumazet | 1bfc443 | 2015-03-16 21:06:19 -0700 | [diff] [blame] | 6174 | static void tcp_openreq_init(struct request_sock *req, |
| 6175 | const struct tcp_options_received *rx_opt, |
| 6176 | struct sk_buff *skb, const struct sock *sk) |
| 6177 | { |
| 6178 | struct inet_request_sock *ireq = inet_rsk(req); |
| 6179 | |
Eric Dumazet | ed53d0a | 2015-10-08 19:33:23 -0700 | [diff] [blame] | 6180 | req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */ |
Eric Dumazet | 1bfc443 | 2015-03-16 21:06:19 -0700 | [diff] [blame] | 6181 | req->cookie_ts = 0; |
| 6182 | tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq; |
| 6183 | tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; |
Yuchung Cheng | 0f1c28a | 2015-09-18 11:36:14 -0700 | [diff] [blame] | 6184 | skb_mstamp_get(&tcp_rsk(req)->snt_synack); |
Eric Dumazet | 1bfc443 | 2015-03-16 21:06:19 -0700 | [diff] [blame] | 6185 | tcp_rsk(req)->last_oow_ack_time = 0; |
| 6186 | req->mss = rx_opt->mss_clamp; |
| 6187 | req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0; |
| 6188 | ireq->tstamp_ok = rx_opt->tstamp_ok; |
| 6189 | ireq->sack_ok = rx_opt->sack_ok; |
| 6190 | ireq->snd_wscale = rx_opt->snd_wscale; |
| 6191 | ireq->wscale_ok = rx_opt->wscale_ok; |
| 6192 | ireq->acked = 0; |
| 6193 | ireq->ecn_ok = 0; |
| 6194 | ireq->ir_rmt_port = tcp_hdr(skb)->source; |
| 6195 | ireq->ir_num = ntohs(tcp_hdr(skb)->dest); |
| 6196 | ireq->ir_mark = inet_request_mark(sk, skb); |
| 6197 | } |
| 6198 | |
Eric Dumazet | e49bb33 | 2015-03-17 18:32:27 -0700 | [diff] [blame] | 6199 | struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops, |
Eric Dumazet | a1a5344 | 2015-10-04 21:08:11 -0700 | [diff] [blame] | 6200 | struct sock *sk_listener, |
| 6201 | bool attach_listener) |
Eric Dumazet | e49bb33 | 2015-03-17 18:32:27 -0700 | [diff] [blame] | 6202 | { |
Eric Dumazet | a1a5344 | 2015-10-04 21:08:11 -0700 | [diff] [blame] | 6203 | struct request_sock *req = reqsk_alloc(ops, sk_listener, |
| 6204 | attach_listener); |
Eric Dumazet | e49bb33 | 2015-03-17 18:32:27 -0700 | [diff] [blame] | 6205 | |
| 6206 | if (req) { |
| 6207 | struct inet_request_sock *ireq = inet_rsk(req); |
| 6208 | |
| 6209 | kmemcheck_annotate_bitfield(ireq, flags); |
| 6210 | ireq->opt = NULL; |
Huw Davies | 56ac42b | 2016-06-27 15:05:28 -0400 | [diff] [blame] | 6211 | #if IS_ENABLED(CONFIG_IPV6) |
| 6212 | ireq->pktopts = NULL; |
| 6213 | #endif |
Eric Dumazet | e49bb33 | 2015-03-17 18:32:27 -0700 | [diff] [blame] | 6214 | atomic64_set(&ireq->ir_cookie, 0); |
| 6215 | ireq->ireq_state = TCP_NEW_SYN_RECV; |
| 6216 | write_pnet(&ireq->ireq_net, sock_net(sk_listener)); |
Eric Dumazet | 0144a81 | 2015-03-24 21:45:56 -0700 | [diff] [blame] | 6217 | ireq->ireq_family = sk_listener->sk_family; |
Eric Dumazet | e49bb33 | 2015-03-17 18:32:27 -0700 | [diff] [blame] | 6218 | } |
| 6219 | |
| 6220 | return req; |
| 6221 | } |
| 6222 | EXPORT_SYMBOL(inet_reqsk_alloc); |
| 6223 | |
Eric Dumazet | 41d25fe | 2015-03-25 15:08:47 -0700 | [diff] [blame] | 6224 | /* |
| 6225 | * Return true if a syncookie should be sent |
| 6226 | */ |
Eric Dumazet | 2985aaa | 2015-09-29 07:42:51 -0700 | [diff] [blame] | 6227 | static bool tcp_syn_flood_action(const struct sock *sk, |
Eric Dumazet | 41d25fe | 2015-03-25 15:08:47 -0700 | [diff] [blame] | 6228 | const struct sk_buff *skb, |
| 6229 | const char *proto) |
| 6230 | { |
Eric Dumazet | 8d2675f | 2015-10-02 11:43:25 -0700 | [diff] [blame] | 6231 | struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue; |
Eric Dumazet | 41d25fe | 2015-03-25 15:08:47 -0700 | [diff] [blame] | 6232 | const char *msg = "Dropping request"; |
| 6233 | bool want_cookie = false; |
Nikolay Borisov | 12ed824 | 2016-02-03 09:46:51 +0200 | [diff] [blame] | 6234 | struct net *net = sock_net(sk); |
Eric Dumazet | 41d25fe | 2015-03-25 15:08:47 -0700 | [diff] [blame] | 6235 | |
| 6236 | #ifdef CONFIG_SYN_COOKIES |
Nikolay Borisov | 12ed824 | 2016-02-03 09:46:51 +0200 | [diff] [blame] | 6237 | if (net->ipv4.sysctl_tcp_syncookies) { |
Eric Dumazet | 41d25fe | 2015-03-25 15:08:47 -0700 | [diff] [blame] | 6238 | msg = "Sending cookies"; |
| 6239 | want_cookie = true; |
Eric Dumazet | 02a1d6e | 2016-04-27 16:44:39 -0700 | [diff] [blame] | 6240 | __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES); |
Eric Dumazet | 41d25fe | 2015-03-25 15:08:47 -0700 | [diff] [blame] | 6241 | } else |
| 6242 | #endif |
Eric Dumazet | 02a1d6e | 2016-04-27 16:44:39 -0700 | [diff] [blame] | 6243 | __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP); |
Eric Dumazet | 41d25fe | 2015-03-25 15:08:47 -0700 | [diff] [blame] | 6244 | |
Eric Dumazet | 8d2675f | 2015-10-02 11:43:25 -0700 | [diff] [blame] | 6245 | if (!queue->synflood_warned && |
Nikolay Borisov | 12ed824 | 2016-02-03 09:46:51 +0200 | [diff] [blame] | 6246 | net->ipv4.sysctl_tcp_syncookies != 2 && |
Eric Dumazet | 8d2675f | 2015-10-02 11:43:25 -0700 | [diff] [blame] | 6247 | xchg(&queue->synflood_warned, 1) == 0) |
Eric Dumazet | 41d25fe | 2015-03-25 15:08:47 -0700 | [diff] [blame] | 6248 | pr_info("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n", |
| 6249 | proto, ntohs(tcp_hdr(skb)->dest), msg); |
Eric Dumazet | 2985aaa | 2015-09-29 07:42:51 -0700 | [diff] [blame] | 6250 | |
Eric Dumazet | 41d25fe | 2015-03-25 15:08:47 -0700 | [diff] [blame] | 6251 | return want_cookie; |
| 6252 | } |
| 6253 | |
Eric Dumazet | cd8ae85 | 2015-05-03 21:34:46 -0700 | [diff] [blame] | 6254 | static void tcp_reqsk_record_syn(const struct sock *sk, |
| 6255 | struct request_sock *req, |
| 6256 | const struct sk_buff *skb) |
| 6257 | { |
| 6258 | if (tcp_sk(sk)->save_syn) { |
| 6259 | u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb); |
| 6260 | u32 *copy; |
| 6261 | |
| 6262 | copy = kmalloc(len + sizeof(u32), GFP_ATOMIC); |
| 6263 | if (copy) { |
| 6264 | copy[0] = len; |
| 6265 | memcpy(©[1], skb_network_header(skb), len); |
| 6266 | req->saved_syn = copy; |
| 6267 | } |
| 6268 | } |
| 6269 | } |
| 6270 | |
Octavian Purdila | 1fb6f15 | 2014-06-25 17:10:02 +0300 | [diff] [blame] | 6271 | int tcp_conn_request(struct request_sock_ops *rsk_ops, |
| 6272 | const struct tcp_request_sock_ops *af_ops, |
| 6273 | struct sock *sk, struct sk_buff *skb) |
| 6274 | { |
Octavian Purdila | 1fb6f15 | 2014-06-25 17:10:02 +0300 | [diff] [blame] | 6275 | struct tcp_fastopen_cookie foc = { .len = -1 }; |
Eric Dumazet | 7c85af8 | 2015-09-24 17:16:05 -0700 | [diff] [blame] | 6276 | __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn; |
| 6277 | struct tcp_options_received tmp_opt; |
| 6278 | struct tcp_sock *tp = tcp_sk(sk); |
Nikolay Borisov | 12ed824 | 2016-02-03 09:46:51 +0200 | [diff] [blame] | 6279 | struct net *net = sock_net(sk); |
Eric Dumazet | 7c85af8 | 2015-09-24 17:16:05 -0700 | [diff] [blame] | 6280 | struct sock *fastopen_sk = NULL; |
| 6281 | struct dst_entry *dst = NULL; |
| 6282 | struct request_sock *req; |
| 6283 | bool want_cookie = false; |
| 6284 | struct flowi fl; |
Octavian Purdila | 1fb6f15 | 2014-06-25 17:10:02 +0300 | [diff] [blame] | 6285 | |
| 6286 | /* TW buckets are converted to open requests without |
| 6287 | * limitations, they conserve resources and peer is |
| 6288 | * evidently real one. |
| 6289 | */ |
Nikolay Borisov | 12ed824 | 2016-02-03 09:46:51 +0200 | [diff] [blame] | 6290 | if ((net->ipv4.sysctl_tcp_syncookies == 2 || |
Octavian Purdila | 1fb6f15 | 2014-06-25 17:10:02 +0300 | [diff] [blame] | 6291 | inet_csk_reqsk_queue_is_full(sk)) && !isn) { |
| 6292 | want_cookie = tcp_syn_flood_action(sk, skb, rsk_ops->slab_name); |
| 6293 | if (!want_cookie) |
| 6294 | goto drop; |
| 6295 | } |
| 6296 | |
Eric Dumazet | 5ea8ea2 | 2016-10-26 09:27:57 -0700 | [diff] [blame] | 6297 | if (sk_acceptq_is_full(sk)) { |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 6298 | NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS); |
Octavian Purdila | 1fb6f15 | 2014-06-25 17:10:02 +0300 | [diff] [blame] | 6299 | goto drop; |
| 6300 | } |
| 6301 | |
Eric Dumazet | a1a5344 | 2015-10-04 21:08:11 -0700 | [diff] [blame] | 6302 | req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie); |
Octavian Purdila | 1fb6f15 | 2014-06-25 17:10:02 +0300 | [diff] [blame] | 6303 | if (!req) |
| 6304 | goto drop; |
| 6305 | |
| 6306 | tcp_rsk(req)->af_specific = af_ops; |
| 6307 | |
| 6308 | tcp_clear_options(&tmp_opt); |
| 6309 | tmp_opt.mss_clamp = af_ops->mss_clamp; |
| 6310 | tmp_opt.user_mss = tp->rx_opt.user_mss; |
| 6311 | tcp_parse_options(skb, &tmp_opt, 0, want_cookie ? NULL : &foc); |
| 6312 | |
| 6313 | if (want_cookie && !tmp_opt.saw_tstamp) |
| 6314 | tcp_clear_options(&tmp_opt); |
| 6315 | |
| 6316 | tmp_opt.tstamp_ok = tmp_opt.saw_tstamp; |
| 6317 | tcp_openreq_init(req, &tmp_opt, skb, sk); |
KOVACS Krisztian | 7a68257 | 2016-09-23 11:27:42 +0200 | [diff] [blame] | 6318 | inet_rsk(req)->no_srccheck = inet_sk(sk)->transparent; |
Octavian Purdila | 1fb6f15 | 2014-06-25 17:10:02 +0300 | [diff] [blame] | 6319 | |
Eric Dumazet | 16f8616 | 2015-03-13 15:51:10 -0700 | [diff] [blame] | 6320 | /* Note: tcp_v6_init_req() might override ir_iif for link locals */ |
David Ahern | 6dd9a14 | 2015-12-16 13:20:44 -0800 | [diff] [blame] | 6321 | inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb); |
Eric Dumazet | 16f8616 | 2015-03-13 15:51:10 -0700 | [diff] [blame] | 6322 | |
Octavian Purdila | 1fb6f15 | 2014-06-25 17:10:02 +0300 | [diff] [blame] | 6323 | af_ops->init_req(req, sk, skb); |
| 6324 | |
| 6325 | if (security_inet_conn_request(sk, skb, req)) |
| 6326 | goto drop_and_free; |
| 6327 | |
Florian Westphal | f7b3bec | 2014-11-03 17:35:03 +0100 | [diff] [blame] | 6328 | if (!want_cookie && !isn) { |
Octavian Purdila | 1fb6f15 | 2014-06-25 17:10:02 +0300 | [diff] [blame] | 6329 | /* VJ's idea. We save last timestamp seen |
| 6330 | * from the destination in peer table, when entering |
| 6331 | * state TIME-WAIT, and check against it before |
| 6332 | * accepting new connection request. |
| 6333 | * |
| 6334 | * If "isn" is not zero, this request hit alive |
| 6335 | * timewait bucket, so that all the necessary checks |
| 6336 | * are made in the function processing timewait state. |
| 6337 | */ |
Hannes Frederic Sowa | a26552a | 2014-08-14 22:06:12 +0200 | [diff] [blame] | 6338 | if (tcp_death_row.sysctl_tw_recycle) { |
Octavian Purdila | 1fb6f15 | 2014-06-25 17:10:02 +0300 | [diff] [blame] | 6339 | bool strict; |
| 6340 | |
| 6341 | dst = af_ops->route_req(sk, &fl, req, &strict); |
Hannes Frederic Sowa | a26552a | 2014-08-14 22:06:12 +0200 | [diff] [blame] | 6342 | |
Octavian Purdila | 1fb6f15 | 2014-06-25 17:10:02 +0300 | [diff] [blame] | 6343 | if (dst && strict && |
Hannes Frederic Sowa | a26552a | 2014-08-14 22:06:12 +0200 | [diff] [blame] | 6344 | !tcp_peer_is_proven(req, dst, true, |
| 6345 | tmp_opt.saw_tstamp)) { |
Eric Dumazet | c10d931 | 2016-04-29 14:16:47 -0700 | [diff] [blame] | 6346 | NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSPASSIVEREJECTED); |
Octavian Purdila | 1fb6f15 | 2014-06-25 17:10:02 +0300 | [diff] [blame] | 6347 | goto drop_and_release; |
| 6348 | } |
| 6349 | } |
| 6350 | /* Kill the following clause, if you dislike this way. */ |
Nikolay Borisov | 12ed824 | 2016-02-03 09:46:51 +0200 | [diff] [blame] | 6351 | else if (!net->ipv4.sysctl_tcp_syncookies && |
Octavian Purdila | 1fb6f15 | 2014-06-25 17:10:02 +0300 | [diff] [blame] | 6352 | (sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) < |
| 6353 | (sysctl_max_syn_backlog >> 2)) && |
Hannes Frederic Sowa | a26552a | 2014-08-14 22:06:12 +0200 | [diff] [blame] | 6354 | !tcp_peer_is_proven(req, dst, false, |
| 6355 | tmp_opt.saw_tstamp)) { |
Octavian Purdila | 1fb6f15 | 2014-06-25 17:10:02 +0300 | [diff] [blame] | 6356 | /* Without syncookies last quarter of |
| 6357 | * backlog is filled with destinations, |
| 6358 | * proven to be alive. |
| 6359 | * It means that we continue to communicate |
| 6360 | * to destinations, already remembered |
| 6361 | * to the moment of synflood. |
| 6362 | */ |
| 6363 | pr_drop_req(req, ntohs(tcp_hdr(skb)->source), |
| 6364 | rsk_ops->family); |
| 6365 | goto drop_and_release; |
| 6366 | } |
| 6367 | |
| 6368 | isn = af_ops->init_seq(skb); |
| 6369 | } |
| 6370 | if (!dst) { |
| 6371 | dst = af_ops->route_req(sk, &fl, req, NULL); |
| 6372 | if (!dst) |
| 6373 | goto drop_and_free; |
| 6374 | } |
| 6375 | |
Florian Westphal | f7b3bec | 2014-11-03 17:35:03 +0100 | [diff] [blame] | 6376 | tcp_ecn_create_request(req, skb, sk, dst); |
| 6377 | |
| 6378 | if (want_cookie) { |
| 6379 | isn = cookie_init_sequence(af_ops, sk, skb, &req->mss); |
| 6380 | req->cookie_ts = tmp_opt.tstamp_ok; |
| 6381 | if (!tmp_opt.tstamp_ok) |
| 6382 | inet_rsk(req)->ecn_ok = 0; |
| 6383 | } |
| 6384 | |
Octavian Purdila | 1fb6f15 | 2014-06-25 17:10:02 +0300 | [diff] [blame] | 6385 | tcp_rsk(req)->snt_isn = isn; |
Eric Dumazet | 58d607d | 2015-09-15 15:24:20 -0700 | [diff] [blame] | 6386 | tcp_rsk(req)->txhash = net_tx_rndhash(); |
Octavian Purdila | 1fb6f15 | 2014-06-25 17:10:02 +0300 | [diff] [blame] | 6387 | tcp_openreq_init_rwin(req, sk, dst); |
Eric Dumazet | ca6fb06 | 2015-10-02 11:43:35 -0700 | [diff] [blame] | 6388 | if (!want_cookie) { |
Eric Dumazet | ca6fb06 | 2015-10-02 11:43:35 -0700 | [diff] [blame] | 6389 | tcp_reqsk_record_syn(sk, req, skb); |
Eric Dumazet | 7656d84 | 2015-10-04 21:08:07 -0700 | [diff] [blame] | 6390 | fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst); |
Eric Dumazet | ca6fb06 | 2015-10-02 11:43:35 -0700 | [diff] [blame] | 6391 | } |
Eric Dumazet | 7c85af8 | 2015-09-24 17:16:05 -0700 | [diff] [blame] | 6392 | if (fastopen_sk) { |
Eric Dumazet | ca6fb06 | 2015-10-02 11:43:35 -0700 | [diff] [blame] | 6393 | af_ops->send_synack(fastopen_sk, dst, &fl, req, |
Eric Dumazet | b3d0514 | 2016-04-13 22:05:39 -0700 | [diff] [blame] | 6394 | &foc, TCP_SYNACK_FASTOPEN); |
Eric Dumazet | 7656d84 | 2015-10-04 21:08:07 -0700 | [diff] [blame] | 6395 | /* Add the child socket directly into the accept queue */ |
| 6396 | inet_csk_reqsk_queue_add(sk, req, fastopen_sk); |
| 6397 | sk->sk_data_ready(sk); |
| 6398 | bh_unlock_sock(fastopen_sk); |
Eric Dumazet | 7c85af8 | 2015-09-24 17:16:05 -0700 | [diff] [blame] | 6399 | sock_put(fastopen_sk); |
| 6400 | } else { |
Eric Dumazet | 9439ce0 | 2015-03-17 18:32:29 -0700 | [diff] [blame] | 6401 | tcp_rsk(req)->tfo_listener = false; |
Eric Dumazet | ca6fb06 | 2015-10-02 11:43:35 -0700 | [diff] [blame] | 6402 | if (!want_cookie) |
| 6403 | inet_csk_reqsk_queue_hash_add(sk, req, TCP_TIMEOUT_INIT); |
Eric Dumazet | b3d0514 | 2016-04-13 22:05:39 -0700 | [diff] [blame] | 6404 | af_ops->send_synack(sk, dst, &fl, req, &foc, |
| 6405 | !want_cookie ? TCP_SYNACK_NORMAL : |
| 6406 | TCP_SYNACK_COOKIE); |
Eric Dumazet | 9caad86 | 2016-04-01 08:52:20 -0700 | [diff] [blame] | 6407 | if (want_cookie) { |
| 6408 | reqsk_free(req); |
| 6409 | return 0; |
| 6410 | } |
Octavian Purdila | 1fb6f15 | 2014-06-25 17:10:02 +0300 | [diff] [blame] | 6411 | } |
Eric Dumazet | ca6fb06 | 2015-10-02 11:43:35 -0700 | [diff] [blame] | 6412 | reqsk_put(req); |
Octavian Purdila | 1fb6f15 | 2014-06-25 17:10:02 +0300 | [diff] [blame] | 6413 | return 0; |
| 6414 | |
| 6415 | drop_and_release: |
| 6416 | dst_release(dst); |
| 6417 | drop_and_free: |
| 6418 | reqsk_free(req); |
| 6419 | drop: |
Eric Dumazet | 9caad86 | 2016-04-01 08:52:20 -0700 | [diff] [blame] | 6420 | tcp_listendrop(sk); |
Octavian Purdila | 1fb6f15 | 2014-06-25 17:10:02 +0300 | [diff] [blame] | 6421 | return 0; |
| 6422 | } |
| 6423 | EXPORT_SYMBOL(tcp_conn_request); |