blob: 3f7bd7ae7d7af8628b06c124fd8795cbf30e4ce0 [file] [log] [blame]
Greg Kroah-Hartmanb2441312017-11-01 15:07:57 +01001// SPDX-License-Identifier: GPL-2.0
Linus Torvalds1da177e2005-04-16 15:20:36 -07002/*
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
6 *
7 * Implementation of the Transmission Control Protocol(TCP).
8 *
Jesper Juhl02c30a82005-05-05 16:16:16 -07009 * Authors: Ross Biro
Linus Torvalds1da177e2005-04-16 15:20:36 -070010 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
20 */
21
22/*
23 * Changes:
24 * Pedro Roque : Fast Retransmit/Recovery.
25 * Two receive queues.
26 * Retransmit queue handled by TCP.
27 * Better retransmit timer handling.
28 * New congestion avoidance.
29 * Header prediction.
30 * Variable renaming.
31 *
32 * Eric : Fast Retransmit.
33 * Randy Scott : MSS option defines.
34 * Eric Schenk : Fixes to slow start algorithm.
35 * Eric Schenk : Yet another double ACK bug.
36 * Eric Schenk : Delayed ACK bug fixes.
37 * Eric Schenk : Floyd style fast retrans war avoidance.
38 * David S. Miller : Don't allow zero congestion window.
39 * Eric Schenk : Fix retransmitter so that it sends
40 * next packet on ack of previous packet.
41 * Andi Kleen : Moved open_request checking here
42 * and process RSTs for open_requests.
43 * Andi Kleen : Better prune_queue, and other fixes.
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -080044 * Andrey Savochkin: Fix RTT measurements in the presence of
Linus Torvalds1da177e2005-04-16 15:20:36 -070045 * timestamps.
46 * Andrey Savochkin: Check sequence numbers correctly when
47 * removing SACKs due to in sequence incoming
48 * data segments.
49 * Andi Kleen: Make sure we never ack data there is not
50 * enough room for. Also make this condition
51 * a fatal error if it might still happen.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +090052 * Andi Kleen: Add tcp_measure_rcv_mss to make
Linus Torvalds1da177e2005-04-16 15:20:36 -070053 * connections with MSS<min(MTU,ann. MSS)
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +090054 * work without delayed acks.
Linus Torvalds1da177e2005-04-16 15:20:36 -070055 * Andi Kleen: Process packets with PSH set in the
56 * fast path.
57 * J Hadi Salim: ECN support
58 * Andrei Gurtov,
59 * Pasi Sarolahti,
60 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
61 * engine. Lots of bugs are found.
62 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
Linus Torvalds1da177e2005-04-16 15:20:36 -070063 */
64
Joe Perchesafd465032012-03-12 07:03:32 +000065#define pr_fmt(fmt) "TCP: " fmt
66
Linus Torvalds1da177e2005-04-16 15:20:36 -070067#include <linux/mm.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090068#include <linux/slab.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070069#include <linux/module.h>
70#include <linux/sysctl.h>
Ilpo Järvinena0bffff2009-03-21 13:36:17 -070071#include <linux/kernel.h>
Eric Dumazetad971f62014-10-11 15:17:29 -070072#include <linux/prefetch.h>
Satoru SATOH5ffc02a2008-05-04 22:14:42 -070073#include <net/dst.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070074#include <net/tcp.h>
75#include <net/inet_common.h>
76#include <linux/ipsec.h>
77#include <asm/unaligned.h>
Willem de Bruijne1c8a602014-08-04 22:11:50 -040078#include <linux/errqueue.h>
Song Liu59415212017-10-23 09:20:25 -070079#include <trace/events/tcp.h>
Jakub Kicinski494bc1d2019-05-08 16:46:14 -070080#include <linux/jump_label_ratelimit.h>
Amritha Nambiarc6345ce2018-06-29 21:26:57 -070081#include <net/busy_poll.h>
Peter Krystadeda7acd2020-01-21 16:56:16 -080082#include <net/mptcp.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070083
Brian Haleyab32ea52006-09-22 14:15:41 -070084int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
Linus Torvalds1da177e2005-04-16 15:20:36 -070085
Linus Torvalds1da177e2005-04-16 15:20:36 -070086#define FLAG_DATA 0x01 /* Incoming frame contained data. */
87#define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
88#define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
89#define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
90#define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
91#define FLAG_DATA_SACKED 0x20 /* New SACK. */
92#define FLAG_ECE 0x40 /* ECE in this ACK */
Yuchung Cheng291a00d2015-07-01 14:11:14 -070093#define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
Florian Westphal31770e32017-08-30 19:24:58 +020094#define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
Yuchung Chenge33099f2013-03-20 13:33:00 +000095#define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
Ilpo Järvinen2e605292007-08-02 19:46:58 -070096#define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
Ryousei Takano564262c2007-10-25 23:03:52 -070097#define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
Neal Cardwelldf92c832017-08-03 09:19:54 -040098#define FLAG_SET_XMIT_TIMER 0x1000 /* Set TLP or RTO timer */
Ilpo Järvinencadbd032007-12-31 04:49:21 -080099#define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
Eric Dumazet12fb3dd2013-04-19 07:19:48 +0000100#define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
Eric Dumazetd0e1a1b2017-05-23 15:24:46 -0700101#define FLAG_NO_CHALLENGE_ACK 0x8000 /* do not call tcp_send_challenge_ack() */
Yuchung Chengeb36be02018-01-17 12:11:00 -0800102#define FLAG_ACK_MAYBE_DELAYED 0x10000 /* Likely a delayed ACK */
Yuchung Cheng63f367d2021-07-27 10:42:57 -0400103#define FLAG_DSACK_TLP 0x20000 /* DSACK for tail loss probe */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700104
105#define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
106#define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
Priyaranjan Jhad09b9e62017-11-03 17:46:55 -0700107#define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE|FLAG_DSACKING_ACK)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700108#define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
109
Linus Torvalds1da177e2005-04-16 15:20:36 -0700110#define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700111#define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700112
Yuchung Chenge662ca42016-02-02 10:33:04 -0800113#define REXMIT_NONE 0 /* no loss recovery to do */
114#define REXMIT_LOST 1 /* retransmit packets marked lost */
115#define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */
116
Ilya Lesokhin6dac1522018-04-30 10:16:10 +0300117#if IS_ENABLED(CONFIG_TLS_DEVICE)
Jakub Kicinski494bc1d2019-05-08 16:46:14 -0700118static DEFINE_STATIC_KEY_DEFERRED_FALSE(clean_acked_data_enabled, HZ);
Ilya Lesokhin6dac1522018-04-30 10:16:10 +0300119
120void clean_acked_data_enable(struct inet_connection_sock *icsk,
121 void (*cad)(struct sock *sk, u32 ack_seq))
122{
123 icsk->icsk_clean_acked = cad;
Willem de Bruijn7b581392019-06-13 11:08:16 -0400124 static_branch_deferred_inc(&clean_acked_data_enabled);
Ilya Lesokhin6dac1522018-04-30 10:16:10 +0300125}
126EXPORT_SYMBOL_GPL(clean_acked_data_enable);
127
128void clean_acked_data_disable(struct inet_connection_sock *icsk)
129{
Jakub Kicinski494bc1d2019-05-08 16:46:14 -0700130 static_branch_slow_dec_deferred(&clean_acked_data_enabled);
Ilya Lesokhin6dac1522018-04-30 10:16:10 +0300131 icsk->icsk_clean_acked = NULL;
132}
133EXPORT_SYMBOL_GPL(clean_acked_data_disable);
Jakub Kicinski494bc1d2019-05-08 16:46:14 -0700134
135void clean_acked_data_flush(void)
136{
137 static_key_deferred_flush(&clean_acked_data_enabled);
138}
139EXPORT_SYMBOL_GPL(clean_acked_data_flush);
Ilya Lesokhin6dac1522018-04-30 10:16:10 +0300140#endif
141
Martin KaFai Lau72be0fe2020-08-20 12:00:39 -0700142#ifdef CONFIG_CGROUP_BPF
Martin KaFai Lau00d211a2020-08-20 12:00:46 -0700143static void bpf_skops_parse_hdr(struct sock *sk, struct sk_buff *skb)
144{
145 bool unknown_opt = tcp_sk(sk)->rx_opt.saw_unknown &&
146 BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk),
147 BPF_SOCK_OPS_PARSE_UNKNOWN_HDR_OPT_CB_FLAG);
148 bool parse_all_opt = BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk),
149 BPF_SOCK_OPS_PARSE_ALL_HDR_OPT_CB_FLAG);
Martin KaFai Lau0813a842020-08-20 12:01:04 -0700150 struct bpf_sock_ops_kern sock_ops;
Martin KaFai Lau00d211a2020-08-20 12:00:46 -0700151
152 if (likely(!unknown_opt && !parse_all_opt))
153 return;
154
155 /* The skb will be handled in the
156 * bpf_skops_established() or
157 * bpf_skops_write_hdr_opt().
158 */
159 switch (sk->sk_state) {
160 case TCP_SYN_RECV:
161 case TCP_SYN_SENT:
162 case TCP_LISTEN:
163 return;
164 }
165
Martin KaFai Lau0813a842020-08-20 12:01:04 -0700166 sock_owned_by_me(sk);
167
168 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
169 sock_ops.op = BPF_SOCK_OPS_PARSE_HDR_OPT_CB;
170 sock_ops.is_fullsock = 1;
171 sock_ops.sk = sk;
172 bpf_skops_init_skb(&sock_ops, skb, tcp_hdrlen(skb));
173
174 BPF_CGROUP_RUN_PROG_SOCK_OPS(&sock_ops);
Martin KaFai Lau00d211a2020-08-20 12:00:46 -0700175}
176
Martin KaFai Lau72be0fe2020-08-20 12:00:39 -0700177static void bpf_skops_established(struct sock *sk, int bpf_op,
178 struct sk_buff *skb)
179{
180 struct bpf_sock_ops_kern sock_ops;
181
182 sock_owned_by_me(sk);
183
184 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
185 sock_ops.op = bpf_op;
186 sock_ops.is_fullsock = 1;
187 sock_ops.sk = sk;
Martin KaFai Lau0813a842020-08-20 12:01:04 -0700188 /* sk with TCP_REPAIR_ON does not have skb in tcp_finish_connect */
189 if (skb)
190 bpf_skops_init_skb(&sock_ops, skb, tcp_hdrlen(skb));
Martin KaFai Lau72be0fe2020-08-20 12:00:39 -0700191
192 BPF_CGROUP_RUN_PROG_SOCK_OPS(&sock_ops);
193}
194#else
Martin KaFai Lau00d211a2020-08-20 12:00:46 -0700195static void bpf_skops_parse_hdr(struct sock *sk, struct sk_buff *skb)
196{
197}
198
Martin KaFai Lau72be0fe2020-08-20 12:00:39 -0700199static void bpf_skops_established(struct sock *sk, int bpf_op,
200 struct sk_buff *skb)
201{
202}
203#endif
204
Marcelo Ricardo Leitner0b9aefe2017-04-01 11:00:21 -0300205static void tcp_gro_dev_warn(struct sock *sk, const struct sk_buff *skb,
206 unsigned int len)
Marcelo Ricardo Leitnerdcb17d22016-12-05 18:37:13 -0200207{
208 static bool __once __read_mostly;
209
210 if (!__once) {
211 struct net_device *dev;
212
213 __once = true;
214
215 rcu_read_lock();
216 dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif);
Marcelo Ricardo Leitner0b9aefe2017-04-01 11:00:21 -0300217 if (!dev || len >= dev->mtu)
218 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n",
219 dev ? dev->name : "Unknown driver");
Marcelo Ricardo Leitnerdcb17d22016-12-05 18:37:13 -0200220 rcu_read_unlock();
221 }
222}
223
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900224/* Adapt the MSS value used to make delayed ack decision to the
Linus Torvalds1da177e2005-04-16 15:20:36 -0700225 * real world.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900226 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800227static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700228{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700229 struct inet_connection_sock *icsk = inet_csk(sk);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900230 const unsigned int lss = icsk->icsk_ack.last_seg_size;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700231 unsigned int len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700232
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900233 icsk->icsk_ack.last_seg_size = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700234
235 /* skb->len may jitter because of SACKs, even if peer
236 * sends good full-sized frames.
237 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800238 len = skb_shinfo(skb)->gso_size ? : skb->len;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700239 if (len >= icsk->icsk_ack.rcv_mss) {
Marcelo Ricardo Leitnerdcb17d22016-12-05 18:37:13 -0200240 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len,
241 tcp_sk(sk)->advmss);
Marcelo Ricardo Leitner0b9aefe2017-04-01 11:00:21 -0300242 /* Account for possibly-removed options */
243 if (unlikely(len > icsk->icsk_ack.rcv_mss +
244 MAX_TCP_OPTION_SPACE))
245 tcp_gro_dev_warn(sk, skb, len);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700246 } else {
247 /* Otherwise, we make more careful check taking into account,
248 * that SACKs block is variable.
249 *
250 * "len" is invariant segment length, including TCP header.
251 */
Arnaldo Carvalho de Melo9c702202007-04-25 18:04:18 -0700252 len += skb->data - skb_transport_header(skb);
William Allen Simpsonbee7ca92009-11-10 09:51:18 +0000253 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -0700254 /* If PSH is not set, packet should be
255 * full sized, provided peer TCP is not badly broken.
256 * This observation (if it is correct 8)) allows
257 * to handle super-low mtu links fairly.
258 */
259 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -0700260 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700261 /* Subtract also invariant (if peer is RFC compliant),
262 * tcp header plus fixed timestamp option length.
263 * Resulting "len" is MSS free of SACK jitter.
264 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700265 len -= tcp_sk(sk)->tcp_header_len;
266 icsk->icsk_ack.last_seg_size = len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700267 if (len == lss) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700268 icsk->icsk_ack.rcv_mss = len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700269 return;
270 }
271 }
Alexey Kuznetsov1ef96962006-09-19 12:52:50 -0700272 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
273 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700274 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700275 }
276}
277
Eric Dumazet9a9c9b52018-05-21 15:08:56 -0700278static void tcp_incr_quickack(struct sock *sk, unsigned int max_quickacks)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700279{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700280 struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazet95c96172012-04-15 05:58:06 +0000281 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700282
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800283 if (quickacks == 0)
284 quickacks = 2;
Eric Dumazet9a9c9b52018-05-21 15:08:56 -0700285 quickacks = min(quickacks, max_quickacks);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700286 if (quickacks > icsk->icsk_ack.quick)
Eric Dumazet9a9c9b52018-05-21 15:08:56 -0700287 icsk->icsk_ack.quick = quickacks;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700288}
289
Yuchung Chenga0496ef22018-07-18 13:56:36 -0700290void tcp_enter_quickack_mode(struct sock *sk, unsigned int max_quickacks)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700291{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700292 struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazet9a9c9b52018-05-21 15:08:56 -0700293
294 tcp_incr_quickack(sk, max_quickacks);
Wei Wang31954cd2019-01-25 10:53:19 -0800295 inet_csk_exit_pingpong_mode(sk);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700296 icsk->icsk_ack.ato = TCP_ATO_MIN;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700297}
Yuchung Chenga0496ef22018-07-18 13:56:36 -0700298EXPORT_SYMBOL(tcp_enter_quickack_mode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700299
300/* Send ACKs quickly, if "quick" count is not exhausted
301 * and the session is not interactive.
302 */
303
Jon Maxwell2251ae42015-07-08 10:12:28 +1000304static bool tcp_in_quickack_mode(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700305{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700306 const struct inet_connection_sock *icsk = inet_csk(sk);
Jon Maxwell2251ae42015-07-08 10:12:28 +1000307 const struct dst_entry *dst = __sk_dst_get(sk);
Eric Dumazeta2a385d2012-05-16 23:15:34 +0000308
Jon Maxwell2251ae42015-07-08 10:12:28 +1000309 return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
Wei Wang31954cd2019-01-25 10:53:19 -0800310 (icsk->icsk_ack.quick && !inet_csk_in_pingpong_mode(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700311}
312
Florian Westphal735d3832014-09-29 13:08:30 +0200313static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700314{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800315 if (tp->ecn_flags & TCP_ECN_OK)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700316 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
317}
318
Yuchung Chengfd2123a2018-08-09 09:38:12 -0700319static void tcp_ecn_accept_cwr(struct sock *sk, const struct sk_buff *skb)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700320{
Lawrence Brakmo9aee4002018-07-23 17:49:39 -0700321 if (tcp_hdr(skb)->cwr) {
Yuchung Chengfd2123a2018-08-09 09:38:12 -0700322 tcp_sk(sk)->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
Lawrence Brakmo9aee4002018-07-23 17:49:39 -0700323
324 /* If the sender is telling us it has entered CWR, then its
325 * cwnd may be very low (even just 1 packet), so we should ACK
326 * immediately.
327 */
Denis Kirjanov257028402020-06-25 14:51:06 +0300328 if (TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq)
329 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
Lawrence Brakmo9aee4002018-07-23 17:49:39 -0700330 }
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700331}
332
Florian Westphal735d3832014-09-29 13:08:30 +0200333static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700334{
Neal Cardwellaf38d072019-09-09 16:56:02 -0400335 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700336}
337
Yousuk Seungf4c9f852018-06-04 15:29:51 -0700338static void __tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700339{
Yousuk Seungf4c9f852018-06-04 15:29:51 -0700340 struct tcp_sock *tp = tcp_sk(sk);
341
Eric Dumazetb82d1bb2011-09-27 02:20:08 -0400342 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
Eric Dumazet7a269ff2011-09-22 20:02:19 +0000343 case INET_ECN_NOT_ECT:
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700344 /* Funny extension: if ECT is not set on a segment,
Eric Dumazet7a269ff2011-09-22 20:02:19 +0000345 * and we already seen ECT on a previous segment,
346 * it is probably a retransmit.
347 */
348 if (tp->ecn_flags & TCP_ECN_SEEN)
Eric Dumazet15ecbe92018-06-27 08:47:21 -0700349 tcp_enter_quickack_mode(sk, 2);
Eric Dumazet7a269ff2011-09-22 20:02:19 +0000350 break;
351 case INET_ECN_CE:
Yousuk Seungf4c9f852018-06-04 15:29:51 -0700352 if (tcp_ca_needs_ecn(sk))
353 tcp_ca_event(sk, CA_EVENT_ECN_IS_CE);
Florian Westphal98900922014-09-26 22:37:35 +0200354
Eric Dumazetaae06bf2012-08-06 11:04:43 +0000355 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
356 /* Better not delay acks, sender can have a very low cwnd */
Eric Dumazet15ecbe92018-06-27 08:47:21 -0700357 tcp_enter_quickack_mode(sk, 2);
Eric Dumazetaae06bf2012-08-06 11:04:43 +0000358 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
359 }
Eric Dumazet7a269ff2011-09-22 20:02:19 +0000360 tp->ecn_flags |= TCP_ECN_SEEN;
Florian Westphal98900922014-09-26 22:37:35 +0200361 break;
362 default:
Yousuk Seungf4c9f852018-06-04 15:29:51 -0700363 if (tcp_ca_needs_ecn(sk))
364 tcp_ca_event(sk, CA_EVENT_ECN_NO_CE);
Florian Westphal98900922014-09-26 22:37:35 +0200365 tp->ecn_flags |= TCP_ECN_SEEN;
366 break;
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700367 }
368}
369
Yousuk Seungf4c9f852018-06-04 15:29:51 -0700370static void tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
Florian Westphal735d3832014-09-29 13:08:30 +0200371{
Yousuk Seungf4c9f852018-06-04 15:29:51 -0700372 if (tcp_sk(sk)->ecn_flags & TCP_ECN_OK)
373 __tcp_ecn_check_ce(sk, skb);
Florian Westphal735d3832014-09-29 13:08:30 +0200374}
375
376static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700377{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800378 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700379 tp->ecn_flags &= ~TCP_ECN_OK;
380}
381
Florian Westphal735d3832014-09-29 13:08:30 +0200382static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700383{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800384 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700385 tp->ecn_flags &= ~TCP_ECN_OK;
386}
387
Florian Westphal735d3832014-09-29 13:08:30 +0200388static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700389{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800390 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
Eric Dumazeta2a385d2012-05-16 23:15:34 +0000391 return true;
392 return false;
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700393}
394
Linus Torvalds1da177e2005-04-16 15:20:36 -0700395/* Buffer size and advertised window tuning.
396 *
397 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
398 */
399
Eric Dumazet6ae70532013-10-01 10:23:44 -0700400static void tcp_sndbuf_expand(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700401{
Eric Dumazet6ae70532013-10-01 10:23:44 -0700402 const struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng77bfc172016-09-19 23:39:20 -0400403 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
Eric Dumazet6ae70532013-10-01 10:23:44 -0700404 int sndmem, per_mss;
405 u32 nr_segs;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700406
Eric Dumazet6ae70532013-10-01 10:23:44 -0700407 /* Worst case is non GSO/TSO : each frame consumes one skb
408 * and skb->head is kmalloced using power of two area of memory
409 */
410 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
411 MAX_TCP_HEADER +
412 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
413
414 per_mss = roundup_pow_of_two(per_mss) +
415 SKB_DATA_ALIGN(sizeof(struct sk_buff));
416
417 nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
418 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
419
420 /* Fast Recovery (RFC 5681 3.2) :
421 * Cubic needs 1.7 factor, rounded to 2 to include
Linus Torvaldsa9a08842018-02-11 14:34:03 -0800422 * extra cushion (application might react slowly to EPOLLOUT)
Eric Dumazet6ae70532013-10-01 10:23:44 -0700423 */
Yuchung Cheng77bfc172016-09-19 23:39:20 -0400424 sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2;
425 sndmem *= nr_segs * per_mss;
Eric Dumazet6ae70532013-10-01 10:23:44 -0700426
Eric Dumazet06a59ec2011-10-13 18:24:42 +0000427 if (sk->sk_sndbuf < sndmem)
Eric Dumazete292f052019-10-10 20:17:45 -0700428 WRITE_ONCE(sk->sk_sndbuf,
429 min(sndmem, sock_net(sk)->ipv4.sysctl_tcp_wmem[2]));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700430}
431
432/* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
433 *
434 * All tcp_full_space() is split to two parts: "network" buffer, allocated
435 * forward and advertised in receiver window (tp->rcv_wnd) and
436 * "application buffer", required to isolate scheduling/application
437 * latencies from network.
438 * window_clamp is maximal advertised window. It can be less than
439 * tcp_full_space(), in this case tcp_full_space() - window_clamp
440 * is reserved for "application" buffer. The less window_clamp is
441 * the smoother our behaviour from viewpoint of network, but the lower
442 * throughput and the higher sensitivity of the connection to losses. 8)
443 *
444 * rcv_ssthresh is more strict window_clamp used at "slow start"
445 * phase to predict further behaviour of this connection.
446 * It is used for two goals:
447 * - to enforce header prediction at sender, even when application
448 * requires some significant "application buffer". It is check #1.
449 * - to prevent pruning of receive queue because of misprediction
450 * of receiver window. Check #2.
451 *
452 * The scheme does not work when sender sends good segments opening
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800453 * window and then starts to feed us spaghetti. But it should work
Linus Torvalds1da177e2005-04-16 15:20:36 -0700454 * in common situations. Otherwise, we have to rely on queue collapsing.
455 */
456
457/* Slow part of check#2. */
Eric Dumazet240bfd12021-07-21 03:15:28 -0700458static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb,
459 unsigned int skbtruesize)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700460{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700461 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462 /* Optimize this! */
Eric Dumazet240bfd12021-07-21 03:15:28 -0700463 int truesize = tcp_win_from_space(sk, skbtruesize) >> 1;
Eric Dumazet356d1832017-11-07 00:29:28 -0800464 int window = tcp_win_from_space(sk, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]) >> 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700465
466 while (tp->rcv_ssthresh <= window) {
467 if (truesize <= skb->len)
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700468 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700469
470 truesize >>= 1;
471 window >>= 1;
472 }
473 return 0;
474}
475
Eric Dumazet240bfd12021-07-21 03:15:28 -0700476/* Even if skb appears to have a bad len/truesize ratio, TCP coalescing
477 * can play nice with us, as sk_buff and skb->head might be either
478 * freed or shared with up to MAX_SKB_FRAGS segments.
479 * Only give a boost to drivers using page frag(s) to hold the frame(s),
480 * and if no payload was pulled in skb->head before reaching us.
481 */
482static u32 truesize_adjust(bool adjust, const struct sk_buff *skb)
483{
484 u32 truesize = skb->truesize;
485
486 if (adjust && !skb_headlen(skb)) {
487 truesize -= SKB_TRUESIZE(skb_end_offset(skb));
488 /* paranoid check, some drivers might be buggy */
489 if (unlikely((int)truesize < (int)skb->len))
490 truesize = skb->truesize;
491 }
492 return truesize;
493}
494
495static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb,
496 bool adjust)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700497{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700498 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet50ce1632019-04-16 10:55:20 -0700499 int room;
500
501 room = min_t(int, tp->window_clamp, tcp_space(sk)) - tp->rcv_ssthresh;
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700502
Linus Torvalds1da177e2005-04-16 15:20:36 -0700503 /* Check #1 */
Eric Dumazet50ce1632019-04-16 10:55:20 -0700504 if (room > 0 && !tcp_under_memory_pressure(sk)) {
Eric Dumazet240bfd12021-07-21 03:15:28 -0700505 unsigned int truesize = truesize_adjust(adjust, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700506 int incr;
507
508 /* Check #2. Increase window, if skb with such overhead
509 * will fit to rcvbuf in future.
510 */
Eric Dumazet240bfd12021-07-21 03:15:28 -0700511 if (tcp_win_from_space(sk, truesize) <= skb->len)
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800512 incr = 2 * tp->advmss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700513 else
Eric Dumazet240bfd12021-07-21 03:15:28 -0700514 incr = __tcp_grow_window(sk, skb, truesize);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700515
516 if (incr) {
Eric Dumazet4d846f02012-04-16 23:28:07 +0000517 incr = max_t(int, incr, 2 * skb->len);
Eric Dumazet50ce1632019-04-16 10:55:20 -0700518 tp->rcv_ssthresh += min(room, incr);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700519 inet_csk(sk)->icsk_ack.quick |= 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700520 }
521 }
522}
523
Yuchung Chenga3375312018-09-27 11:21:19 -0700524/* 3. Try to fixup all. It is made immediately after connection enters
Linus Torvalds1da177e2005-04-16 15:20:36 -0700525 * established state.
526 */
Florian Westphalbc183de2020-05-29 00:01:52 +0200527static void tcp_init_buffer_space(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700528{
Eric Dumazet0c126542017-10-26 21:55:08 -0700529 int tcp_app_win = sock_net(sk)->ipv4.sysctl_tcp_app_win;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700530 struct tcp_sock *tp = tcp_sk(sk);
531 int maxwin;
532
Linus Torvalds1da177e2005-04-16 15:20:36 -0700533 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
Eric Dumazet6ae70532013-10-01 10:23:44 -0700534 tcp_sndbuf_expand(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700535
Eric Dumazet9a568de2017-05-16 14:00:14 -0700536 tcp_mstamp_refresh(tp);
Eric Dumazet645f4c62017-04-25 10:15:41 -0700537 tp->rcvq_space.time = tp->tcp_mstamp;
Eric Dumazetb0983d32013-09-20 13:56:58 -0700538 tp->rcvq_space.seq = tp->copied_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700539
540 maxwin = tcp_full_space(sk);
541
542 if (tp->window_clamp >= maxwin) {
543 tp->window_clamp = maxwin;
544
Eric Dumazet0c126542017-10-26 21:55:08 -0700545 if (tcp_app_win && maxwin > 4 * tp->advmss)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700546 tp->window_clamp = max(maxwin -
Eric Dumazet0c126542017-10-26 21:55:08 -0700547 (maxwin >> tcp_app_win),
Linus Torvalds1da177e2005-04-16 15:20:36 -0700548 4 * tp->advmss);
549 }
550
551 /* Force reservation of one segment. */
Eric Dumazet0c126542017-10-26 21:55:08 -0700552 if (tcp_app_win &&
Linus Torvalds1da177e2005-04-16 15:20:36 -0700553 tp->window_clamp > 2 * tp->advmss &&
554 tp->window_clamp + tp->advmss > maxwin)
555 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
556
557 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
Eric Dumazetc2203cf2017-05-16 14:00:04 -0700558 tp->snd_cwnd_stamp = tcp_jiffies32;
Eric Dumazet72d05c02020-12-08 08:21:31 -0800559 tp->rcvq_space.space = min3(tp->rcv_ssthresh, tp->rcv_wnd,
560 (u32)TCP_INIT_CWND * tp->advmss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700561}
562
Yuchung Chenga3375312018-09-27 11:21:19 -0700563/* 4. Recalculate window clamp after socket hit its memory bounds. */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700564static void tcp_clamp_window(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700565{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700566 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300567 struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazet356d1832017-11-07 00:29:28 -0800568 struct net *net = sock_net(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700569
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300570 icsk->icsk_ack.quick = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700571
Eric Dumazet356d1832017-11-07 00:29:28 -0800572 if (sk->sk_rcvbuf < net->ipv4.sysctl_tcp_rmem[2] &&
John Heffner326f36e2005-11-10 17:11:48 -0800573 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
Eric Dumazetb8da51e2015-05-15 12:39:27 -0700574 !tcp_under_memory_pressure(sk) &&
Glauber Costa180d8cd2011-12-11 21:47:02 +0000575 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
Eric Dumazetebb3b782019-10-10 20:17:44 -0700576 WRITE_ONCE(sk->sk_rcvbuf,
577 min(atomic_read(&sk->sk_rmem_alloc),
578 net->ipv4.sysctl_tcp_rmem[2]));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700579 }
John Heffner326f36e2005-11-10 17:11:48 -0800580 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800581 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700582}
583
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800584/* Initialize RCV_MSS value.
585 * RCV_MSS is an our guess about MSS used by the peer.
586 * We haven't any direct information about the MSS.
587 * It's better to underestimate the RCV_MSS rather than overestimate.
588 * Overestimations make us ACKing less frequently than needed.
589 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
590 */
591void tcp_initialize_rcv_mss(struct sock *sk)
592{
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400593 const struct tcp_sock *tp = tcp_sk(sk);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800594 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
595
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800596 hint = min(hint, tp->rcv_wnd / 2);
William Allen Simpsonbee7ca92009-11-10 09:51:18 +0000597 hint = min(hint, TCP_MSS_DEFAULT);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800598 hint = max(hint, TCP_MIN_MSS);
599
600 inet_csk(sk)->icsk_ack.rcv_mss = hint;
601}
Eric Dumazet4bc2f182010-07-09 21:22:10 +0000602EXPORT_SYMBOL(tcp_initialize_rcv_mss);
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800603
Linus Torvalds1da177e2005-04-16 15:20:36 -0700604/* Receiver "autotuning" code.
605 *
606 * The algorithm for RTT estimation w/o timestamps is based on
607 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
Alexander A. Klimov7a6498e2020-07-06 19:38:50 +0200608 * <https://public.lanl.gov/radiant/pubs.html#DRS>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700609 *
610 * More detail on this code can be found at
Justin P. Mattock631dd1a2010-10-18 11:03:14 +0200611 * <http://staff.psc.edu/jheffner/>,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700612 * though this reference is out of date. A new paper
613 * is pending.
614 */
615static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
616{
Eric Dumazet645f4c62017-04-25 10:15:41 -0700617 u32 new_sample = tp->rcv_rtt_est.rtt_us;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700618 long m = sample;
619
Linus Torvalds1da177e2005-04-16 15:20:36 -0700620 if (new_sample != 0) {
621 /* If we sample in larger samples in the non-timestamp
622 * case, we could grossly overestimate the RTT especially
623 * with chatty applications or bulk transfer apps which
624 * are stalled on filesystem I/O.
625 *
626 * Also, since we are only going for a minimum in the
Stephen Hemminger31f34262005-11-15 15:17:10 -0800627 * non-timestamp case, we do not smooth things out
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800628 * else with timestamps disabled convergence takes too
Linus Torvalds1da177e2005-04-16 15:20:36 -0700629 * long.
630 */
631 if (!win_dep) {
632 m -= (new_sample >> 3);
633 new_sample += m;
Neal Cardwell18a223e2012-04-10 07:59:20 +0000634 } else {
635 m <<= 3;
636 if (m < new_sample)
637 new_sample = m;
638 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700639 } else {
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800640 /* No previous measure. */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700641 new_sample = m << 3;
642 }
643
Eric Dumazet645f4c62017-04-25 10:15:41 -0700644 tp->rcv_rtt_est.rtt_us = new_sample;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700645}
646
647static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
648{
Eric Dumazet645f4c62017-04-25 10:15:41 -0700649 u32 delta_us;
650
Eric Dumazet9a568de2017-05-16 14:00:14 -0700651 if (tp->rcv_rtt_est.time == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700652 goto new_measure;
653 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
654 return;
Eric Dumazet9a568de2017-05-16 14:00:14 -0700655 delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time);
Wei Wang9ee11bd2017-12-12 16:28:58 -0800656 if (!delta_us)
657 delta_us = 1;
Eric Dumazet645f4c62017-04-25 10:15:41 -0700658 tcp_rcv_rtt_update(tp, delta_us, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700659
660new_measure:
661 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
Eric Dumazet645f4c62017-04-25 10:15:41 -0700662 tp->rcv_rtt_est.time = tp->tcp_mstamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700663}
664
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800665static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
666 const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700667{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700668 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet9a568de2017-05-16 14:00:14 -0700669
Wei Wang3f6c65d2018-06-19 21:42:50 -0700670 if (tp->rx_opt.rcv_tsecr == tp->rcv_rtt_last_tsecr)
671 return;
672 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
673
674 if (TCP_SKB_CB(skb)->end_seq -
675 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss) {
Eric Dumazet9a568de2017-05-16 14:00:14 -0700676 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
Wei Wang9ee11bd2017-12-12 16:28:58 -0800677 u32 delta_us;
Eric Dumazet9a568de2017-05-16 14:00:14 -0700678
Eric Dumazet9efdda42018-11-24 09:12:24 -0800679 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
680 if (!delta)
681 delta = 1;
682 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
683 tcp_rcv_rtt_update(tp, delta_us, 0);
684 }
Eric Dumazet9a568de2017-05-16 14:00:14 -0700685 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700686}
687
688/*
689 * This function should be called every time data is copied to user space.
690 * It calculates the appropriate TCP receive buffer space.
691 */
692void tcp_rcv_space_adjust(struct sock *sk)
693{
694 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet607065b2017-12-10 17:55:03 -0800695 u32 copied;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700696 int time;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900697
Yafang Shao61638492018-04-20 23:18:26 +0800698 trace_tcp_rcv_space_adjust(sk);
699
Eric Dumazet86323852017-12-06 11:08:19 -0800700 tcp_mstamp_refresh(tp);
Eric Dumazet9a568de2017-05-16 14:00:14 -0700701 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time);
Eric Dumazet645f4c62017-04-25 10:15:41 -0700702 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700703 return;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900704
Eric Dumazetb0983d32013-09-20 13:56:58 -0700705 /* Number of bytes copied to user in last RTT */
706 copied = tp->copied_seq - tp->rcvq_space.seq;
707 if (copied <= tp->rcvq_space.space)
708 goto new_measure;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700709
Eric Dumazetb0983d32013-09-20 13:56:58 -0700710 /* A bit of theory :
711 * copied = bytes received in previous RTT, our base window
712 * To cope with packet losses, we need a 2x factor
713 * To cope with slow start, and sender growing its cwin by 100 %
714 * every RTT, we need a 4x factor, because the ACK we are sending
715 * now is for the next RTT, not the current one :
716 * <prev RTT . ><current RTT .. ><next RTT .... >
717 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700718
Eric Dumazet4540c0c2017-10-27 07:47:22 -0700719 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf &&
Eric Dumazetb0983d32013-09-20 13:56:58 -0700720 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
Eric Dumazet607065b2017-12-10 17:55:03 -0800721 int rcvmem, rcvbuf;
Eric Dumazetc3916ad2017-12-10 17:55:04 -0800722 u64 rcvwin, grow;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700723
Eric Dumazetb0983d32013-09-20 13:56:58 -0700724 /* minimal window to cope with packet losses, assuming
725 * steady state. Add some cushion because of small variations.
726 */
Eric Dumazet607065b2017-12-10 17:55:03 -0800727 rcvwin = ((u64)copied << 1) + 16 * tp->advmss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700728
Eric Dumazetc3916ad2017-12-10 17:55:04 -0800729 /* Accommodate for sender rate increase (eg. slow start) */
730 grow = rcvwin * (copied - tp->rcvq_space.space);
731 do_div(grow, tp->rcvq_space.space);
732 rcvwin += (grow << 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700733
Eric Dumazetb0983d32013-09-20 13:56:58 -0700734 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
Eric Dumazet94f08932017-10-26 21:55:09 -0700735 while (tcp_win_from_space(sk, rcvmem) < tp->advmss)
Eric Dumazetb0983d32013-09-20 13:56:58 -0700736 rcvmem += 128;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700737
Eric Dumazet607065b2017-12-10 17:55:03 -0800738 do_div(rcvwin, tp->advmss);
739 rcvbuf = min_t(u64, rcvwin * rcvmem,
740 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
Eric Dumazetb0983d32013-09-20 13:56:58 -0700741 if (rcvbuf > sk->sk_rcvbuf) {
Eric Dumazetebb3b782019-10-10 20:17:44 -0700742 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
Eric Dumazetb0983d32013-09-20 13:56:58 -0700743
744 /* Make the window clamp follow along. */
Eric Dumazet02db5572017-12-10 17:55:02 -0800745 tp->window_clamp = tcp_win_from_space(sk, rcvbuf);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700746 }
747 }
Eric Dumazetb0983d32013-09-20 13:56:58 -0700748 tp->rcvq_space.space = copied;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900749
Linus Torvalds1da177e2005-04-16 15:20:36 -0700750new_measure:
751 tp->rcvq_space.seq = tp->copied_seq;
Eric Dumazet645f4c62017-04-25 10:15:41 -0700752 tp->rcvq_space.time = tp->tcp_mstamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700753}
754
755/* There is something which you must keep in mind when you analyze the
756 * behavior of the tp->ato delayed ack timeout interval. When a
757 * connection starts up, we want to ack as quickly as possible. The
758 * problem is that "good" TCP's do slow start at the beginning of data
759 * transmission. The means that until we send the first few ACK's the
760 * sender will sit on his end and only queue most of his data, because
761 * he can only send snd_cwnd unacked packets at any given time. For
762 * each ACK we send, he increments snd_cwnd and transmits more of his
763 * queue. -DaveM
764 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700765static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700766{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700767 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700768 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700769 u32 now;
770
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700771 inet_csk_schedule_ack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700772
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700773 tcp_measure_rcv_mss(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700774
775 tcp_rcv_rtt_measure(tp);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900776
Eric Dumazet70eabf02017-05-16 14:00:07 -0700777 now = tcp_jiffies32;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700778
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700779 if (!icsk->icsk_ack.ato) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700780 /* The _first_ data packet received, initialize
781 * delayed ACK engine.
782 */
Eric Dumazet9a9c9b52018-05-21 15:08:56 -0700783 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700784 icsk->icsk_ack.ato = TCP_ATO_MIN;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700785 } else {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700786 int m = now - icsk->icsk_ack.lrcvtime;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700787
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800788 if (m <= TCP_ATO_MIN / 2) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700789 /* The fastest case is the first. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700790 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
791 } else if (m < icsk->icsk_ack.ato) {
792 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
793 if (icsk->icsk_ack.ato > icsk->icsk_rto)
794 icsk->icsk_ack.ato = icsk->icsk_rto;
795 } else if (m > icsk->icsk_rto) {
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800796 /* Too long gap. Apparently sender failed to
Linus Torvalds1da177e2005-04-16 15:20:36 -0700797 * restart window, so that we send ACKs quickly.
798 */
Eric Dumazet9a9c9b52018-05-21 15:08:56 -0700799 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
Hideo Aoki3ab224b2007-12-31 00:11:19 -0800800 sk_mem_reclaim(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700801 }
802 }
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700803 icsk->icsk_ack.lrcvtime = now;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700804
Yousuk Seungf4c9f852018-06-04 15:29:51 -0700805 tcp_ecn_check_ce(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700806
807 if (skb->len >= 128)
Eric Dumazet240bfd12021-07-21 03:15:28 -0700808 tcp_grow_window(sk, skb, true);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700809}
810
Linus Torvalds1da177e2005-04-16 15:20:36 -0700811/* Called to compute a smoothed rtt estimate. The data fed to this
812 * routine either comes from timestamps, or from segments that were
813 * known _not_ to have been retransmitted [see Karn/Partridge
814 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
815 * piece by Van Jacobson.
816 * NOTE: the next three routines used to be one big routine.
817 * To save cycles in the RFC 1323 implementation it was better to break
818 * it up into three procedures. -- erics
819 */
Eric Dumazet740b0f12014-02-26 14:02:48 -0800820static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700821{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300822 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet740b0f12014-02-26 14:02:48 -0800823 long m = mrtt_us; /* RTT */
824 u32 srtt = tp->srtt_us;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700825
Linus Torvalds1da177e2005-04-16 15:20:36 -0700826 /* The following amusing code comes from Jacobson's
827 * article in SIGCOMM '88. Note that rtt and mdev
828 * are scaled versions of rtt and mean deviation.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900829 * This is designed to be as fast as possible
Linus Torvalds1da177e2005-04-16 15:20:36 -0700830 * m stands for "measurement".
831 *
832 * On a 1990 paper the rto value is changed to:
833 * RTO = rtt + 4 * mdev
834 *
835 * Funny. This algorithm seems to be very broken.
836 * These formulae increase RTO, when it should be decreased, increase
Stephen Hemminger31f34262005-11-15 15:17:10 -0800837 * too slowly, when it should be increased quickly, decrease too quickly
Linus Torvalds1da177e2005-04-16 15:20:36 -0700838 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
839 * does not matter how to _calculate_ it. Seems, it was trap
840 * that VJ failed to avoid. 8)
841 */
Eric Dumazet4a5ab4e2014-02-06 15:57:10 -0800842 if (srtt != 0) {
843 m -= (srtt >> 3); /* m is now error in rtt est */
844 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700845 if (m < 0) {
846 m = -m; /* m is now abs(error) */
Eric Dumazet740b0f12014-02-26 14:02:48 -0800847 m -= (tp->mdev_us >> 2); /* similar update on mdev */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700848 /* This is similar to one of Eifel findings.
849 * Eifel blocks mdev updates when rtt decreases.
850 * This solution is a bit different: we use finer gain
851 * for mdev in this case (alpha*beta).
852 * Like Eifel it also prevents growth of rto,
853 * but also it limits too fast rto decreases,
854 * happening in pure Eifel.
855 */
856 if (m > 0)
857 m >>= 3;
858 } else {
Eric Dumazet740b0f12014-02-26 14:02:48 -0800859 m -= (tp->mdev_us >> 2); /* similar update on mdev */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700860 }
Eric Dumazet740b0f12014-02-26 14:02:48 -0800861 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
862 if (tp->mdev_us > tp->mdev_max_us) {
863 tp->mdev_max_us = tp->mdev_us;
864 if (tp->mdev_max_us > tp->rttvar_us)
865 tp->rttvar_us = tp->mdev_max_us;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700866 }
867 if (after(tp->snd_una, tp->rtt_seq)) {
Eric Dumazet740b0f12014-02-26 14:02:48 -0800868 if (tp->mdev_max_us < tp->rttvar_us)
869 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700870 tp->rtt_seq = tp->snd_nxt;
Eric Dumazet740b0f12014-02-26 14:02:48 -0800871 tp->mdev_max_us = tcp_rto_min_us(sk);
Stanislav Fomichev23729ff2019-07-02 09:13:56 -0700872
873 tcp_bpf_rtt(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700874 }
875 } else {
876 /* no previous measure. */
Eric Dumazet4a5ab4e2014-02-06 15:57:10 -0800877 srtt = m << 3; /* take the measured time to be rtt */
Eric Dumazet740b0f12014-02-26 14:02:48 -0800878 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
879 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
880 tp->mdev_max_us = tp->rttvar_us;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700881 tp->rtt_seq = tp->snd_nxt;
Stanislav Fomichev23729ff2019-07-02 09:13:56 -0700882
883 tcp_bpf_rtt(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700884 }
Eric Dumazet740b0f12014-02-26 14:02:48 -0800885 tp->srtt_us = max(1U, srtt);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700886}
887
Eric Dumazet95bd09e2013-08-27 05:46:32 -0700888static void tcp_update_pacing_rate(struct sock *sk)
889{
890 const struct tcp_sock *tp = tcp_sk(sk);
891 u64 rate;
892
893 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
Eric Dumazet43e122b2015-08-21 17:38:02 -0700894 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
895
896 /* current rate is (cwnd * mss) / srtt
897 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
898 * In Congestion Avoidance phase, set it to 120 % the current rate.
899 *
900 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
901 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
902 * end of slow start and should slow down.
903 */
904 if (tp->snd_cwnd < tp->snd_ssthresh / 2)
Eric Dumazet23a7102a2017-10-27 07:47:31 -0700905 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio;
Eric Dumazet43e122b2015-08-21 17:38:02 -0700906 else
Eric Dumazetc26e91f2017-10-27 07:47:32 -0700907 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio;
Eric Dumazet95bd09e2013-08-27 05:46:32 -0700908
909 rate *= max(tp->snd_cwnd, tp->packets_out);
910
Eric Dumazet740b0f12014-02-26 14:02:48 -0800911 if (likely(tp->srtt_us))
912 do_div(rate, tp->srtt_us);
Eric Dumazet95bd09e2013-08-27 05:46:32 -0700913
Mark Rutlanda9da6f22017-10-23 14:07:18 -0700914 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate
Eric Dumazetba537422013-10-09 17:14:52 -0700915 * without any lock. We want to make sure compiler wont store
916 * intermediate values in this location.
917 */
Mark Rutlanda9da6f22017-10-23 14:07:18 -0700918 WRITE_ONCE(sk->sk_pacing_rate, min_t(u64, rate,
919 sk->sk_max_pacing_rate));
Eric Dumazet95bd09e2013-08-27 05:46:32 -0700920}
921
Linus Torvalds1da177e2005-04-16 15:20:36 -0700922/* Calculate rto without backoff. This is the second half of Van Jacobson's
923 * routine referred to above.
924 */
stephen hemmingerf7e56a72013-12-29 11:39:51 -0800925static void tcp_set_rto(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700926{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700927 const struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700928 /* Old crap is replaced with new one. 8)
929 *
930 * More seriously:
931 * 1. If rtt variance happened to be less 50msec, it is hallucination.
932 * It cannot be less due to utterly erratic ACK generation made
933 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
934 * to do with delayed acks, because at cwnd>2 true delack timeout
935 * is invisible. Actually, Linux-2.4 also generates erratic
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800936 * ACKs in some circumstances.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700937 */
Damian Lukowskif1ecd5d2009-08-26 00:16:31 +0000938 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700939
940 /* 2. Fixups made earlier cannot be right.
941 * If we do not estimate RTO correctly without them,
942 * all the algo is pure shit and should be replaced
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800943 * with correct one. It is exactly, which we pretend to do.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700944 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700945
Ilpo Järvinenee6aac52008-12-05 22:43:08 -0800946 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
947 * guarantees that rto is higher.
948 */
Damian Lukowskif1ecd5d2009-08-26 00:16:31 +0000949 tcp_bound_rto(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700950}
951
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400952__u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700953{
954 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
955
Gerrit Renker22b71c82010-08-29 19:23:12 +0000956 if (!cwnd)
David S. Miller442b9632011-02-02 17:05:11 -0800957 cwnd = TCP_INIT_CWND;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700958 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
959}
960
Priyaranjan Jhaa71d77e2020-07-16 12:12:34 -0700961struct tcp_sacktag_state {
962 /* Timestamps for earliest and latest never-retransmitted segment
963 * that was SACKed. RTO needs the earliest RTT to stay conservative,
964 * but congestion control should still get an accurate delay signal.
965 */
966 u64 first_sackt;
967 u64 last_sackt;
968 u32 reord;
969 u32 sack_delivered;
970 int flag;
971 unsigned int mss_now;
972 struct rate_sample *rate;
973};
974
Priyaranjan Jhaad2b9b02020-09-24 15:23:14 -0700975/* Take a notice that peer is sending D-SACKs. Skip update of data delivery
976 * and spurious retransmission information if this DSACK is unlikely caused by
977 * sender's action:
978 * - DSACKed sequence range is larger than maximum receiver's window.
979 * - Total no. of DSACKed segments exceed the total no. of retransmitted segs.
980 */
Priyaranjan Jhaa71d77e2020-07-16 12:12:34 -0700981static u32 tcp_dsack_seen(struct tcp_sock *tp, u32 start_seq,
982 u32 end_seq, struct tcp_sacktag_state *state)
Ilpo Järvinene60402d2007-08-09 15:14:46 +0300983{
Priyaranjan Jhaa71d77e2020-07-16 12:12:34 -0700984 u32 seq_len, dup_segs = 1;
985
Priyaranjan Jhaad2b9b02020-09-24 15:23:14 -0700986 if (!before(start_seq, end_seq))
987 return 0;
988
989 seq_len = end_seq - start_seq;
990 /* Dubious DSACK: DSACKed range greater than maximum advertised rwnd */
991 if (seq_len > tp->max_window)
992 return 0;
993 if (seq_len > tp->mss_cache)
994 dup_segs = DIV_ROUND_UP(seq_len, tp->mss_cache);
Yuchung Cheng63f367d2021-07-27 10:42:57 -0400995 else if (tp->tlp_high_seq && tp->tlp_high_seq == end_seq)
996 state->flag |= FLAG_DSACK_TLP;
Priyaranjan Jhaad2b9b02020-09-24 15:23:14 -0700997
998 tp->dsack_dups += dup_segs;
999 /* Skip the DSACK if dup segs weren't retransmitted by sender */
1000 if (tp->dsack_dups > tp->total_retrans)
1001 return 0;
Priyaranjan Jhaa71d77e2020-07-16 12:12:34 -07001002
Vijay Subramanianab562222011-12-20 13:23:24 +00001003 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
Neal Cardwella657db02021-07-27 10:42:58 -04001004 /* We increase the RACK ordering window in rounds where we receive
1005 * DSACKs that may have been due to reordering causing RACK to trigger
1006 * a spurious fast recovery. Thus RACK ignores DSACKs that happen
1007 * without having seen reordering, or that match TLP probes (TLP
1008 * is timer-driven, not triggered by RACK).
1009 */
1010 if (tp->reord_seen && !(state->flag & FLAG_DSACK_TLP))
1011 tp->rack.dsack_seen = 1;
Priyaranjan Jhaa71d77e2020-07-16 12:12:34 -07001012
1013 state->flag |= FLAG_DSACKING_ACK;
1014 /* A spurious retransmission is delivered */
1015 state->sack_delivered += dup_segs;
1016
1017 return dup_segs;
Ilpo Järvinene60402d2007-08-09 15:14:46 +03001018}
1019
Yuchung Cheng737ff312017-11-08 13:01:27 -08001020/* It's reordering when higher sequence was delivered (i.e. sacked) before
1021 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering
1022 * distance is approximated in full-mss packet distance ("reordering").
1023 */
1024static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq,
1025 const int ts)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001026{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001027 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng737ff312017-11-08 13:01:27 -08001028 const u32 mss = tp->mss_cache;
1029 u32 fack, metric;
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001030
Yuchung Cheng737ff312017-11-08 13:01:27 -08001031 fack = tcp_highest_sack_seq(tp);
1032 if (!before(low_seq, fack))
Soheil Hassas Yeganeh6f5b24e2017-05-16 17:39:02 -04001033 return;
1034
Yuchung Cheng737ff312017-11-08 13:01:27 -08001035 metric = fack - low_seq;
1036 if ((metric > tp->reordering * mss) && mss) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001037#if FASTRETRANS_DEBUG > 1
Joe Perches91df42b2012-05-15 14:11:54 +00001038 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
1039 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
1040 tp->reordering,
Yuchung Cheng737ff312017-11-08 13:01:27 -08001041 0,
Joe Perches91df42b2012-05-15 14:11:54 +00001042 tp->sacked_out,
1043 tp->undo_marker ? tp->undo_retrans : 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001044#endif
Yuchung Cheng737ff312017-11-08 13:01:27 -08001045 tp->reordering = min_t(u32, (metric + mss - 1) / mss,
1046 sock_net(sk)->ipv4.sysctl_tcp_max_reordering);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001047 }
Yuchung Chengeed530b2012-05-02 13:30:03 +00001048
Yuchung Cheng2d2517e2017-04-04 14:15:40 -07001049 /* This exciting event is worth to be remembered. 8) */
Wei Wang7ec65372018-07-31 17:46:24 -07001050 tp->reord_seen++;
Yuchung Cheng737ff312017-11-08 13:01:27 -08001051 NET_INC_STATS(sock_net(sk),
1052 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001053}
1054
Yuchung Cheng68698972020-09-25 10:04:30 -07001055 /* This must be called before lost_out or retrans_out are updated
1056 * on a new loss, because we want to know if all skbs previously
1057 * known to be lost have already been retransmitted, indicating
1058 * that this newly lost skb is our next skb to retransmit.
1059 */
Ilpo Järvinenc8c213f2008-09-20 21:18:55 -07001060static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
1061{
Pengcheng Yange176b1b2020-01-14 17:23:40 +08001062 if ((!tp->retransmit_skb_hint && tp->retrans_out >= tp->lost_out) ||
1063 (tp->retransmit_skb_hint &&
1064 before(TCP_SKB_CB(skb)->seq,
1065 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)))
Ilpo Järvinen006f5822008-09-20 21:20:20 -07001066 tp->retransmit_skb_hint = skb;
Ilpo Järvinenc8c213f2008-09-20 21:18:55 -07001067}
1068
Yuchung Cheng9cd8b6c2020-10-01 14:05:18 -07001069/* Sum the number of packets on the wire we have marked as lost, and
1070 * notify the congestion control module that the given skb was marked lost.
1071 */
1072static void tcp_notify_skb_loss_event(struct tcp_sock *tp, const struct sk_buff *skb)
1073{
1074 tp->lost += tcp_skb_pcount(skb);
1075}
1076
Yuchung Chengfd2146742020-09-25 10:04:29 -07001077void tcp_mark_skb_lost(struct sock *sk, struct sk_buff *skb)
1078{
Yuchung Cheng68698972020-09-25 10:04:30 -07001079 __u8 sacked = TCP_SKB_CB(skb)->sacked;
Yuchung Chengfd2146742020-09-25 10:04:29 -07001080 struct tcp_sock *tp = tcp_sk(sk);
1081
Yuchung Cheng68698972020-09-25 10:04:30 -07001082 if (sacked & TCPCB_SACKED_ACKED)
1083 return;
1084
1085 tcp_verify_retransmit_hint(tp, skb);
1086 if (sacked & TCPCB_LOST) {
1087 if (sacked & TCPCB_SACKED_RETRANS) {
1088 /* Account for retransmits that are lost again */
1089 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1090 tp->retrans_out -= tcp_skb_pcount(skb);
1091 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT,
1092 tcp_skb_pcount(skb));
Yuchung Cheng9cd8b6c2020-10-01 14:05:18 -07001093 tcp_notify_skb_loss_event(tp, skb);
Yuchung Cheng68698972020-09-25 10:04:30 -07001094 }
1095 } else {
1096 tp->lost_out += tcp_skb_pcount(skb);
1097 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
Yuchung Cheng9cd8b6c2020-10-01 14:05:18 -07001098 tcp_notify_skb_loss_event(tp, skb);
Yuchung Chengfd2146742020-09-25 10:04:29 -07001099 }
1100}
1101
Yousuk Seung082d4fa2020-06-26 21:05:35 -07001102/* Updates the delivered and delivered_ce counts */
1103static void tcp_count_delivered(struct tcp_sock *tp, u32 delivered,
1104 bool ece_ack)
1105{
1106 tp->delivered += delivered;
1107 if (ece_ack)
1108 tp->delivered_ce += delivered;
1109}
1110
Linus Torvalds1da177e2005-04-16 15:20:36 -07001111/* This procedure tags the retransmission queue when SACKs arrive.
1112 *
1113 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1114 * Packets in queue with these bits set are counted in variables
1115 * sacked_out, retrans_out and lost_out, correspondingly.
1116 *
1117 * Valid combinations are:
1118 * Tag InFlight Description
1119 * 0 1 - orig segment is in flight.
1120 * S 0 - nothing flies, orig reached receiver.
1121 * L 0 - nothing flies, orig lost by net.
1122 * R 2 - both orig and retransmit are in flight.
1123 * L|R 1 - orig is lost, retransmit is in flight.
1124 * S|R 1 - orig reached receiver, retrans is still in flight.
1125 * (L|S|R is logically valid, it could occur when L|R is sacked,
1126 * but it is equivalent to plain S and code short-curcuits it to S.
1127 * L|S is logically invalid, it would mean -1 packet in flight 8))
1128 *
1129 * These 6 states form finite state machine, controlled by the following events:
1130 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1131 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
Yuchung Cheng974c1232012-01-19 14:42:21 +00001132 * 3. Loss detection event of two flavors:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001133 * A. Scoreboard estimator decided the packet is lost.
1134 * A'. Reno "three dupacks" marks head of queue lost.
Yuchung Cheng974c1232012-01-19 14:42:21 +00001135 * B. SACK arrives sacking SND.NXT at the moment, when the
Linus Torvalds1da177e2005-04-16 15:20:36 -07001136 * segment was retransmitted.
1137 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1138 *
1139 * It is pleasant to note, that state diagram turns out to be commutative,
1140 * so that we are allowed not to be bothered by order of our actions,
1141 * when multiple events arrive simultaneously. (see the function below).
1142 *
1143 * Reordering detection.
1144 * --------------------
1145 * Reordering metric is maximal distance, which a packet can be displaced
1146 * in packet stream. With SACKs we can estimate it:
1147 *
1148 * 1. SACK fills old hole and the corresponding segment was not
1149 * ever retransmitted -> reordering. Alas, we cannot use it
1150 * when segment was retransmitted.
1151 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1152 * for retransmitted and already SACKed segment -> reordering..
1153 * Both of these heuristics are not used in Loss state, when we cannot
1154 * account for retransmits accurately.
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001155 *
1156 * SACK block validation.
1157 * ----------------------
1158 *
1159 * SACK block range validation checks that the received SACK block fits to
1160 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1161 * Note that SND.UNA is not included to the range though being valid because
Ilpo Järvinen0e835332007-10-01 15:28:17 -07001162 * it means that the receiver is rather inconsistent with itself reporting
1163 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1164 * perfectly valid, however, in light of RFC2018 which explicitly states
1165 * that "SACK block MUST reflect the newest segment. Even if the newest
1166 * segment is going to be discarded ...", not that it looks very clever
1167 * in case of head skb. Due to potentional receiver driven attacks, we
1168 * choose to avoid immediate execution of a walk in write queue due to
1169 * reneging and defer head skb's loss recovery to standard loss recovery
1170 * procedure that will eventually trigger (nothing forbids us doing this).
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001171 *
1172 * Implements also blockage to start_seq wrap-around. Problem lies in the
1173 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1174 * there's no guarantee that it will be before snd_nxt (n). The problem
1175 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1176 * wrap (s_w):
1177 *
1178 * <- outs wnd -> <- wrapzone ->
1179 * u e n u_w e_w s n_w
1180 * | | | | | | |
1181 * |<------------+------+----- TCP seqno space --------------+---------->|
1182 * ...-- <2^31 ->| |<--------...
1183 * ...---- >2^31 ------>| |<--------...
1184 *
1185 * Current code wouldn't be vulnerable but it's better still to discard such
1186 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1187 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1188 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1189 * equal to the ideal case (infinite seqno space without wrap caused issues).
1190 *
1191 * With D-SACK the lower bound is extended to cover sequence space below
1192 * SND.UNA down to undo_marker, which is the last point of interest. Yet
Ryousei Takano564262c2007-10-25 23:03:52 -07001193 * again, D-SACK block must not to go across snd_una (for the same reason as
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001194 * for the normal SACK blocks, explained above). But there all simplicity
1195 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1196 * fully below undo_marker they do not affect behavior in anyway and can
1197 * therefore be safely ignored. In rare cases (which are more or less
1198 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1199 * fragmentation and packet reordering past skb's retransmission. To consider
1200 * them correctly, the acceptable range must be extended even more though
1201 * the exact amount is rather hard to quantify. However, tp->max_window can
1202 * be used as an exaggerated estimate.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001203 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001204static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1205 u32 start_seq, u32 end_seq)
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001206{
1207 /* Too far in future, or reversed (interpretation is ambiguous) */
1208 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001209 return false;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001210
1211 /* Nasty start_seq wrap-around check (see comments above) */
1212 if (!before(start_seq, tp->snd_nxt))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001213 return false;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001214
Ryousei Takano564262c2007-10-25 23:03:52 -07001215 /* In outstanding window? ...This is valid exit for D-SACKs too.
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001216 * start_seq == snd_una is non-sensical (see comments above)
1217 */
1218 if (after(start_seq, tp->snd_una))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001219 return true;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001220
1221 if (!is_dsack || !tp->undo_marker)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001222 return false;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001223
1224 /* ...Then it's D-SACK, and must reside below snd_una completely */
Zheng Yanf779b2d2011-09-18 22:37:34 -04001225 if (after(end_seq, tp->snd_una))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001226 return false;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001227
1228 if (!before(start_seq, tp->undo_marker))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001229 return true;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001230
1231 /* Too old */
1232 if (!after(end_seq, tp->undo_marker))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001233 return false;
Ilpo Järvinen5b3c9882007-08-24 22:54:44 -07001234
1235 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1236 * start_seq < undo_marker and end_seq >= undo_marker.
1237 */
1238 return !before(start_seq, end_seq - tp->max_window);
1239}
1240
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001241static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1242 struct tcp_sack_block_wire *sp, int num_sacks,
Priyaranjan Jhaa71d77e2020-07-16 12:12:34 -07001243 u32 prior_snd_una, struct tcp_sacktag_state *state)
David S. Millerd06e0212007-06-18 22:43:06 -07001244{
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07001245 struct tcp_sock *tp = tcp_sk(sk);
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07001246 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1247 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
Priyaranjan Jhaa71d77e2020-07-16 12:12:34 -07001248 u32 dup_segs;
David S. Millerd06e0212007-06-18 22:43:06 -07001249
1250 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
Eric Dumazetc10d9312016-04-29 14:16:47 -07001251 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
David S. Millerd06e0212007-06-18 22:43:06 -07001252 } else if (num_sacks > 1) {
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07001253 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1254 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
David S. Millerd06e0212007-06-18 22:43:06 -07001255
Priyaranjan Jhaa71d77e2020-07-16 12:12:34 -07001256 if (after(end_seq_0, end_seq_1) || before(start_seq_0, start_seq_1))
1257 return false;
1258 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKOFORECV);
1259 } else {
1260 return false;
David S. Millerd06e0212007-06-18 22:43:06 -07001261 }
1262
Priyaranjan Jhaa71d77e2020-07-16 12:12:34 -07001263 dup_segs = tcp_dsack_seen(tp, start_seq_0, end_seq_0, state);
Priyaranjan Jhaad2b9b02020-09-24 15:23:14 -07001264 if (!dup_segs) { /* Skip dubious DSACK */
1265 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKIGNOREDDUBIOUS);
1266 return false;
1267 }
1268
Priyaranjan Jhae3a5a1e2020-07-16 12:12:35 -07001269 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECVSEGS, dup_segs);
Priyaranjan Jhaa71d77e2020-07-16 12:12:34 -07001270
David S. Millerd06e0212007-06-18 22:43:06 -07001271 /* D-SACK for already forgotten data... Do dumb counting. */
Priyaranjan Jhaa71d77e2020-07-16 12:12:34 -07001272 if (tp->undo_marker && tp->undo_retrans > 0 &&
David S. Millerd06e0212007-06-18 22:43:06 -07001273 !after(end_seq_0, prior_snd_una) &&
1274 after(end_seq_0, tp->undo_marker))
Priyaranjan Jhaa71d77e2020-07-16 12:12:34 -07001275 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - dup_segs);
David S. Millerd06e0212007-06-18 22:43:06 -07001276
Priyaranjan Jhaa71d77e2020-07-16 12:12:34 -07001277 return true;
David S. Millerd06e0212007-06-18 22:43:06 -07001278}
1279
Ilpo Järvinend1935942007-10-11 17:34:25 -07001280/* Check if skb is fully within the SACK block. In presence of GSO skbs,
1281 * the incoming SACK may not exactly match but we can find smaller MSS
1282 * aligned portion of it that matches. Therefore we might need to fragment
1283 * which may fail and creates some hassle (caller must handle error case
1284 * returns).
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001285 *
1286 * FIXME: this could be merged to shift decision code
Ilpo Järvinend1935942007-10-11 17:34:25 -07001287 */
Adrian Bunk0f79efd2007-10-26 03:57:36 -07001288static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001289 u32 start_seq, u32 end_seq)
Ilpo Järvinend1935942007-10-11 17:34:25 -07001290{
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001291 int err;
1292 bool in_sack;
Ilpo Järvinend1935942007-10-11 17:34:25 -07001293 unsigned int pkt_len;
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001294 unsigned int mss;
Ilpo Järvinend1935942007-10-11 17:34:25 -07001295
1296 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1297 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1298
1299 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1300 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001301 mss = tcp_skb_mss(skb);
Ilpo Järvinend1935942007-10-11 17:34:25 -07001302 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1303
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001304 if (!in_sack) {
Ilpo Järvinend1935942007-10-11 17:34:25 -07001305 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001306 if (pkt_len < mss)
1307 pkt_len = mss;
1308 } else {
Ilpo Järvinend1935942007-10-11 17:34:25 -07001309 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001310 if (pkt_len < mss)
1311 return -EINVAL;
1312 }
1313
1314 /* Round if necessary so that SACKs cover only full MSSes
1315 * and/or the remaining small portion (if present)
1316 */
1317 if (pkt_len > mss) {
1318 unsigned int new_len = (pkt_len / mss) * mss;
Yuchung Chengb451e5d2017-05-10 17:01:27 -07001319 if (!in_sack && new_len < pkt_len)
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001320 new_len += mss;
Ilpo Järvinenadb92db2008-11-24 21:13:50 -08001321 pkt_len = new_len;
1322 }
Yuchung Chengb451e5d2017-05-10 17:01:27 -07001323
1324 if (pkt_len >= skb->len && !in_sack)
1325 return 0;
1326
Eric Dumazet75c119a2017-10-05 22:21:27 -07001327 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
1328 pkt_len, mss, GFP_ATOMIC);
Ilpo Järvinend1935942007-10-11 17:34:25 -07001329 if (err < 0)
1330 return err;
1331 }
1332
1333 return in_sack;
1334}
1335
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001336/* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1337static u8 tcp_sacktag_one(struct sock *sk,
1338 struct tcp_sacktag_state *state, u8 sacked,
1339 u32 start_seq, u32 end_seq,
Eric Dumazet740b0f12014-02-26 14:02:48 -08001340 int dup_sack, int pcount,
Eric Dumazet9a568de2017-05-16 14:00:14 -07001341 u64 xmit_time)
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001342{
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001343 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001344
1345 /* Account D-SACK for retransmitted packet. */
1346 if (dup_sack && (sacked & TCPCB_RETRANS)) {
Yuchung Cheng6e08d5e2014-07-02 12:07:16 -07001347 if (tp->undo_marker && tp->undo_retrans > 0 &&
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001348 after(end_seq, tp->undo_marker))
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001349 tp->undo_retrans--;
Yuchung Cheng737ff312017-11-08 13:01:27 -08001350 if ((sacked & TCPCB_SACKED_ACKED) &&
1351 before(start_seq, state->reord))
1352 state->reord = start_seq;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001353 }
1354
1355 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001356 if (!after(end_seq, tp->snd_una))
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001357 return sacked;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001358
1359 if (!(sacked & TCPCB_SACKED_ACKED)) {
Eric Dumazetd2329f12017-04-25 10:15:38 -07001360 tcp_rack_advance(tp, sacked, end_seq, xmit_time);
Yuchung Cheng659a8ad2015-10-16 21:57:46 -07001361
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001362 if (sacked & TCPCB_SACKED_RETRANS) {
1363 /* If the segment is not tagged as lost,
1364 * we do not clear RETRANS, believing
1365 * that retransmission is still in flight.
1366 */
1367 if (sacked & TCPCB_LOST) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001368 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001369 tp->lost_out -= pcount;
1370 tp->retrans_out -= pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001371 }
1372 } else {
1373 if (!(sacked & TCPCB_RETRANS)) {
1374 /* New sack for not retransmitted frame,
1375 * which was in hole. It is reordering.
1376 */
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001377 if (before(start_seq,
Yuchung Cheng737ff312017-11-08 13:01:27 -08001378 tcp_highest_sack_seq(tp)) &&
1379 before(start_seq, state->reord))
1380 state->reord = start_seq;
1381
Yuchung Chenge33099f2013-03-20 13:33:00 +00001382 if (!after(end_seq, tp->high_seq))
1383 state->flag |= FLAG_ORIG_SACK_ACKED;
Eric Dumazet9a568de2017-05-16 14:00:14 -07001384 if (state->first_sackt == 0)
1385 state->first_sackt = xmit_time;
1386 state->last_sackt = xmit_time;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001387 }
1388
1389 if (sacked & TCPCB_LOST) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001390 sacked &= ~TCPCB_LOST;
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001391 tp->lost_out -= pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001392 }
1393 }
1394
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001395 sacked |= TCPCB_SACKED_ACKED;
1396 state->flag |= FLAG_DATA_SACKED;
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001397 tp->sacked_out += pcount;
Yousuk Seung082d4fa2020-06-26 21:05:35 -07001398 /* Out-of-order packets delivered */
Yousuk Seungf00394c2020-06-26 21:05:34 -07001399 state->sack_delivered += pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001400
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001401 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
Yuchung Cheng713bafe2017-11-08 13:01:26 -08001402 if (tp->lost_skb_hint &&
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001403 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001404 tp->lost_cnt_hint += pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001405 }
1406
1407 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1408 * frames and clear it. undo_retrans is decreased above, L|R frames
1409 * are accounted above as well.
1410 */
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001411 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1412 sacked &= ~TCPCB_SACKED_RETRANS;
Ilpo Järvinenf58b22f2008-11-24 21:14:43 -08001413 tp->retrans_out -= pcount;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001414 }
1415
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001416 return sacked;
Ilpo Järvinen9e10c47c2007-11-15 19:44:56 -08001417}
1418
Neal Cardwelldaef52b2012-02-12 18:37:10 +00001419/* Shift newly-SACKed bytes from this skb to the immediately previous
1420 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1421 */
Eric Dumazetf3319812017-10-05 22:21:26 -07001422static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev,
1423 struct sk_buff *skb,
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001424 struct tcp_sacktag_state *state,
1425 unsigned int pcount, int shifted, int mss,
1426 bool dup_sack)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001427{
1428 struct tcp_sock *tp = tcp_sk(sk);
Neal Cardwelldaef52b2012-02-12 18:37:10 +00001429 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1430 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001431
1432 BUG_ON(!pcount);
1433
Neal Cardwell4c90d3b2012-02-26 10:06:19 +00001434 /* Adjust counters and hints for the newly sacked sequence
1435 * range but discard the return value since prev is already
1436 * marked. We must tag the range first because the seq
1437 * advancement below implicitly advances
1438 * tcp_highest_sack_seq() when skb is highest_sack.
1439 */
1440 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
Yuchung Cheng59c9af42013-07-22 16:20:47 -07001441 start_seq, end_seq, dup_sack, pcount,
Eric Dumazet2fd66ff2018-09-21 08:51:47 -07001442 tcp_skb_timestamp_us(skb));
Yuchung Chengb9f64822016-09-19 23:39:14 -04001443 tcp_rate_skb_delivered(sk, skb, state->rate);
Neal Cardwell4c90d3b2012-02-26 10:06:19 +00001444
1445 if (skb == tp->lost_skb_hint)
Neal Cardwell0af2a0d2012-02-13 20:22:08 +00001446 tp->lost_cnt_hint += pcount;
1447
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001448 TCP_SKB_CB(prev)->end_seq += shifted;
1449 TCP_SKB_CB(skb)->seq += shifted;
1450
Eric Dumazetcd7d8492014-09-24 04:11:22 -07001451 tcp_skb_pcount_add(prev, pcount);
Eric Dumazet3b4929f2019-05-17 17:17:22 -07001452 WARN_ON_ONCE(tcp_skb_pcount(skb) < pcount);
Eric Dumazetcd7d8492014-09-24 04:11:22 -07001453 tcp_skb_pcount_add(skb, -pcount);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001454
1455 /* When we're adding to gso_segs == 1, gso_size will be zero,
1456 * in theory this shouldn't be necessary but as long as DSACK
1457 * code can come after this skb later on it's better to keep
1458 * setting gso_size to something.
1459 */
Eric Dumazetf69ad292015-06-11 09:15:18 -07001460 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1461 TCP_SKB_CB(prev)->tcp_gso_size = mss;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001462
1463 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
Eric Dumazet51466a72015-06-11 09:15:16 -07001464 if (tcp_skb_pcount(skb) <= 1)
Eric Dumazetf69ad292015-06-11 09:15:18 -07001465 TCP_SKB_CB(skb)->tcp_gso_size = 0;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001466
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001467 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1468 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1469
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001470 if (skb->len > 0) {
1471 BUG_ON(!tcp_skb_pcount(skb));
Eric Dumazetc10d9312016-04-29 14:16:47 -07001472 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001473 return false;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001474 }
1475
1476 /* Whole SKB was eaten :-) */
1477
Ilpo Järvinen92ee76b2008-11-24 21:26:56 -08001478 if (skb == tp->retransmit_skb_hint)
1479 tp->retransmit_skb_hint = prev;
Ilpo Järvinen92ee76b2008-11-24 21:26:56 -08001480 if (skb == tp->lost_skb_hint) {
1481 tp->lost_skb_hint = prev;
1482 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1483 }
1484
Eric Dumazet5e8a402f2013-10-04 10:31:41 -07001485 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
Martin KaFai Laua643b5d2016-04-25 14:44:49 -07001486 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
Eric Dumazet5e8a402f2013-10-04 10:31:41 -07001487 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1488 TCP_SKB_CB(prev)->end_seq++;
1489
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001490 if (skb == tcp_highest_sack(sk))
1491 tcp_advance_highest_sack(sk, skb);
1492
Martin KaFai Laucfea5a62016-04-19 22:39:29 -07001493 tcp_skb_collapse_tstamp(prev, skb);
Eric Dumazet9a568de2017-05-16 14:00:14 -07001494 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp))
1495 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0;
Yuchung Chengb9f64822016-09-19 23:39:14 -04001496
Eric Dumazet75c119a2017-10-05 22:21:27 -07001497 tcp_rtx_queue_unlink_and_free(skb, sk);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001498
Eric Dumazetc10d9312016-04-29 14:16:47 -07001499 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
Ilpo Järvinen111cc8b2008-11-24 21:27:22 -08001500
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001501 return true;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001502}
1503
1504/* I wish gso_size would have a bit more sane initialization than
1505 * something-or-zero which complicates things
1506 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001507static int tcp_skb_seglen(const struct sk_buff *skb)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001508{
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001509 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001510}
1511
1512/* Shifting pages past head area doesn't work */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001513static int skb_can_shift(const struct sk_buff *skb)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001514{
1515 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1516}
1517
Eric Dumazet3b4929f2019-05-17 17:17:22 -07001518int tcp_skb_shift(struct sk_buff *to, struct sk_buff *from,
1519 int pcount, int shiftlen)
1520{
1521 /* TCP min gso_size is 8 bytes (TCP_MIN_GSO_SIZE)
1522 * Since TCP_SKB_CB(skb)->tcp_gso_segs is 16 bits, we need
1523 * to make sure not storing more than 65535 * 8 bytes per skb,
1524 * even if current MSS is bigger.
1525 */
1526 if (unlikely(to->len + shiftlen >= 65535 * TCP_MIN_GSO_SIZE))
1527 return 0;
1528 if (unlikely(tcp_skb_pcount(to) + pcount > 65535))
1529 return 0;
1530 return skb_shift(to, from, shiftlen);
1531}
1532
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001533/* Try collapsing SACK blocks spanning across multiple skbs to a single
1534 * skb.
1535 */
1536static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001537 struct tcp_sacktag_state *state,
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001538 u32 start_seq, u32 end_seq,
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001539 bool dup_sack)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001540{
1541 struct tcp_sock *tp = tcp_sk(sk);
1542 struct sk_buff *prev;
1543 int mss;
1544 int pcount = 0;
1545 int len;
1546 int in_sack;
1547
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001548 /* Normally R but no L won't result in plain S */
1549 if (!dup_sack &&
Ilpo Järvinen9969ca52008-12-05 22:41:06 -08001550 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001551 goto fallback;
1552 if (!skb_can_shift(skb))
1553 goto fallback;
1554 /* This frame is about to be dropped (was ACKed). */
1555 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1556 goto fallback;
1557
1558 /* Can only happen with delayed DSACK + discard craziness */
Eric Dumazet75c119a2017-10-05 22:21:27 -07001559 prev = skb_rb_prev(skb);
1560 if (!prev)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001561 goto fallback;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001562
1563 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1564 goto fallback;
1565
Mat Martineau85712482020-01-09 07:59:20 -08001566 if (!tcp_skb_can_collapse(prev, skb))
Martin KaFai Laua643b5d2016-04-25 14:44:49 -07001567 goto fallback;
1568
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001569 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1570 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1571
1572 if (in_sack) {
1573 len = skb->len;
1574 pcount = tcp_skb_pcount(skb);
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001575 mss = tcp_skb_seglen(skb);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001576
1577 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1578 * drop this restriction as unnecessary
1579 */
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001580 if (mss != tcp_skb_seglen(prev))
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001581 goto fallback;
1582 } else {
1583 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1584 goto noop;
1585 /* CHECKME: This is non-MSS split case only?, this will
1586 * cause skipped skbs due to advancing loop btw, original
1587 * has that feature too
1588 */
1589 if (tcp_skb_pcount(skb) <= 1)
1590 goto noop;
1591
1592 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1593 if (!in_sack) {
1594 /* TODO: head merge to next could be attempted here
1595 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1596 * though it might not be worth of the additional hassle
1597 *
1598 * ...we can probably just fallback to what was done
1599 * previously. We could try merging non-SACKed ones
1600 * as well but it probably isn't going to buy off
1601 * because later SACKs might again split them, and
1602 * it would make skb timestamp tracking considerably
1603 * harder problem.
1604 */
1605 goto fallback;
1606 }
1607
1608 len = end_seq - TCP_SKB_CB(skb)->seq;
1609 BUG_ON(len < 0);
1610 BUG_ON(len > skb->len);
1611
1612 /* MSS boundaries should be honoured or else pcount will
1613 * severely break even though it makes things bit trickier.
1614 * Optimize common case to avoid most of the divides
1615 */
1616 mss = tcp_skb_mss(skb);
1617
1618 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1619 * drop this restriction as unnecessary
1620 */
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001621 if (mss != tcp_skb_seglen(prev))
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001622 goto fallback;
1623
1624 if (len == mss) {
1625 pcount = 1;
1626 } else if (len < mss) {
1627 goto noop;
1628 } else {
1629 pcount = len / mss;
1630 len = pcount * mss;
1631 }
1632 }
1633
Neal Cardwell4648dc92012-03-05 19:35:04 +00001634 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1635 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1636 goto fallback;
1637
Eric Dumazet3b4929f2019-05-17 17:17:22 -07001638 if (!tcp_skb_shift(prev, skb, pcount, len))
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001639 goto fallback;
Eric Dumazetf3319812017-10-05 22:21:26 -07001640 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack))
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001641 goto out;
1642
1643 /* Hole filled allows collapsing with the next as well, this is very
1644 * useful when hole on every nth skb pattern happens
1645 */
Eric Dumazet75c119a2017-10-05 22:21:27 -07001646 skb = skb_rb_next(prev);
1647 if (!skb)
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001648 goto out;
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001649
Ilpo Järvinenf0bc52f2008-12-05 22:40:47 -08001650 if (!skb_can_shift(skb) ||
Ilpo Järvinenf0bc52f2008-12-05 22:40:47 -08001651 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
Ilpo Järvinen775ffab2008-12-05 22:41:26 -08001652 (mss != tcp_skb_seglen(skb)))
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001653 goto out;
1654
1655 len = skb->len;
Eric Dumazet3b4929f2019-05-17 17:17:22 -07001656 pcount = tcp_skb_pcount(skb);
1657 if (tcp_skb_shift(prev, skb, pcount, len))
1658 tcp_shifted_skb(sk, prev, skb, state, pcount,
Eric Dumazetf3319812017-10-05 22:21:26 -07001659 len, mss, 0);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001660
1661out:
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001662 return prev;
1663
1664noop:
1665 return skb;
1666
1667fallback:
Eric Dumazetc10d9312016-04-29 14:16:47 -07001668 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001669 return NULL;
1670}
1671
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001672static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1673 struct tcp_sack_block *next_dup,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001674 struct tcp_sacktag_state *state,
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001675 u32 start_seq, u32 end_seq,
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001676 bool dup_sack_in)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001677{
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001678 struct tcp_sock *tp = tcp_sk(sk);
1679 struct sk_buff *tmp;
1680
Eric Dumazet75c119a2017-10-05 22:21:27 -07001681 skb_rbtree_walk_from(skb) {
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001682 int in_sack = 0;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001683 bool dup_sack = dup_sack_in;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001684
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001685 /* queue is in-order => we can short-circuit the walk early */
1686 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1687 break;
1688
Ian Morris00db4122015-04-03 09:17:27 +01001689 if (next_dup &&
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001690 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1691 in_sack = tcp_match_skb_to_sack(sk, skb,
1692 next_dup->start_seq,
1693 next_dup->end_seq);
1694 if (in_sack > 0)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001695 dup_sack = true;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001696 }
1697
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001698 /* skb reference here is a bit tricky to get right, since
1699 * shifting can eat and free both this skb and the next,
1700 * so not even _safe variant of the loop is enough.
1701 */
1702 if (in_sack <= 0) {
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001703 tmp = tcp_shift_skb_data(sk, skb, state,
1704 start_seq, end_seq, dup_sack);
Ian Morris00db4122015-04-03 09:17:27 +01001705 if (tmp) {
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001706 if (tmp != skb) {
1707 skb = tmp;
1708 continue;
1709 }
1710
1711 in_sack = 0;
1712 } else {
1713 in_sack = tcp_match_skb_to_sack(sk, skb,
1714 start_seq,
1715 end_seq);
1716 }
1717 }
1718
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001719 if (unlikely(in_sack < 0))
1720 break;
1721
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001722 if (in_sack) {
Neal Cardwellcc9a6722012-02-12 18:37:09 +00001723 TCP_SKB_CB(skb)->sacked =
1724 tcp_sacktag_one(sk,
1725 state,
1726 TCP_SKB_CB(skb)->sacked,
1727 TCP_SKB_CB(skb)->seq,
1728 TCP_SKB_CB(skb)->end_seq,
1729 dup_sack,
Yuchung Cheng59c9af42013-07-22 16:20:47 -07001730 tcp_skb_pcount(skb),
Eric Dumazet2fd66ff2018-09-21 08:51:47 -07001731 tcp_skb_timestamp_us(skb));
Yuchung Chengb9f64822016-09-19 23:39:14 -04001732 tcp_rate_skb_delivered(sk, skb, state->rate);
Eric Dumazete2080072017-10-04 12:59:58 -07001733 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1734 list_del_init(&skb->tcp_tsorted_anchor);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001735
Ilpo Järvinen832d11c2008-11-24 21:20:15 -08001736 if (!before(TCP_SKB_CB(skb)->seq,
1737 tcp_highest_sack_seq(tp)))
1738 tcp_advance_highest_sack(sk, skb);
1739 }
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001740 }
1741 return skb;
1742}
1743
Taehee Yoo4bfabc42019-02-25 18:42:33 +09001744static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk, u32 seq)
Eric Dumazet75c119a2017-10-05 22:21:27 -07001745{
1746 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node;
1747 struct sk_buff *skb;
Eric Dumazet75c119a2017-10-05 22:21:27 -07001748
1749 while (*p) {
1750 parent = *p;
1751 skb = rb_to_skb(parent);
1752 if (before(seq, TCP_SKB_CB(skb)->seq)) {
1753 p = &parent->rb_left;
1754 continue;
1755 }
1756 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) {
1757 p = &parent->rb_right;
1758 continue;
1759 }
Eric Dumazet75c119a2017-10-05 22:21:27 -07001760 return skb;
1761 }
1762 return NULL;
1763}
1764
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001765static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001766 u32 skip_to_seq)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001767{
Eric Dumazet75c119a2017-10-05 22:21:27 -07001768 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq))
1769 return skb;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001770
Taehee Yoo4bfabc42019-02-25 18:42:33 +09001771 return tcp_sacktag_bsearch(sk, skip_to_seq);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001772}
1773
1774static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1775 struct sock *sk,
1776 struct tcp_sack_block *next_dup,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001777 struct tcp_sacktag_state *state,
1778 u32 skip_to_seq)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001779{
Ian Morris51456b22015-04-03 09:17:26 +01001780 if (!next_dup)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001781 return skb;
1782
1783 if (before(next_dup->start_seq, skip_to_seq)) {
Taehee Yoo4bfabc42019-02-25 18:42:33 +09001784 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq);
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001785 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1786 next_dup->start_seq, next_dup->end_seq,
1787 1);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001788 }
1789
1790 return skb;
1791}
1792
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001793static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001794{
1795 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1796}
1797
Linus Torvalds1da177e2005-04-16 15:20:36 -07001798static int
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001799tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001800 u32 prior_snd_una, struct tcp_sacktag_state *state)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001801{
1802 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001803 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1804 TCP_SKB_CB(ack_skb)->sacked);
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001805 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
Adam Langley4389dde2008-07-19 00:07:02 -07001806 struct tcp_sack_block sp[TCP_NUM_SACKS];
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001807 struct tcp_sack_block *cache;
1808 struct sk_buff *skb;
Adam Langley4389dde2008-07-19 00:07:02 -07001809 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001810 int used_sacks;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001811 bool found_dup_sack = false;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001812 int i, j;
Baruch Evenfda03fb2007-02-04 23:35:57 -08001813 int first_sack_index;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001814
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001815 state->flag = 0;
Yuchung Cheng737ff312017-11-08 13:01:27 -08001816 state->reord = tp->snd_nxt;
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001817
Yuchung Cheng737ff312017-11-08 13:01:27 -08001818 if (!tp->sacked_out)
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001819 tcp_highest_sack_reset(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001820
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07001821 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
Priyaranjan Jhaa71d77e2020-07-16 12:12:34 -07001822 num_sacks, prior_snd_una, state);
Baruch Even6f746512007-02-04 23:36:42 -08001823
1824 /* Eliminate too old ACKs, but take into
1825 * account more or less fresh ones, they can
1826 * contain valid SACK info.
1827 */
1828 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1829 return 0;
1830
Ilpo Järvinen96a2d412007-11-14 15:47:18 -08001831 if (!tp->packets_out)
1832 goto out;
1833
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001834 used_sacks = 0;
1835 first_sack_index = 0;
1836 for (i = 0; i < num_sacks; i++) {
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001837 bool dup_sack = !i && found_dup_sack;
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001838
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07001839 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1840 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001841
1842 if (!tcp_is_sackblock_valid(tp, dup_sack,
1843 sp[used_sacks].start_seq,
1844 sp[used_sacks].end_seq)) {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001845 int mib_idx;
1846
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001847 if (dup_sack) {
1848 if (!tp->undo_marker)
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001849 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001850 else
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001851 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001852 } else {
1853 /* Don't count olds caused by ACK reordering */
1854 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1855 !after(sp[used_sacks].end_seq, tp->snd_una))
1856 continue;
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001857 mib_idx = LINUX_MIB_TCPSACKDISCARD;
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001858 }
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07001859
Eric Dumazetc10d9312016-04-29 14:16:47 -07001860 NET_INC_STATS(sock_net(sk), mib_idx);
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001861 if (i == 0)
1862 first_sack_index = -1;
1863 continue;
1864 }
1865
1866 /* Ignore very old stuff early */
Pengcheng Yangc9655002019-12-30 17:54:41 +08001867 if (!after(sp[used_sacks].end_seq, prior_snd_una)) {
1868 if (i == 0)
1869 first_sack_index = -1;
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001870 continue;
Pengcheng Yangc9655002019-12-30 17:54:41 +08001871 }
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001872
1873 used_sacks++;
1874 }
1875
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001876 /* order SACK blocks to allow in order walk of the retrans queue */
1877 for (i = used_sacks - 1; i > 0; i--) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001878 for (j = 0; j < i; j++) {
1879 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
Ilpo Järvinena0bffff2009-03-21 13:36:17 -07001880 swap(sp[j], sp[j + 1]);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001881
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001882 /* Track where the first SACK block goes to */
1883 if (j == first_sack_index)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001884 first_sack_index = j + 1;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001885 }
1886 }
1887 }
1888
Eric Dumazet75c119a2017-10-05 22:21:27 -07001889 state->mss_now = tcp_current_mss(sk);
Eric Dumazet75c119a2017-10-05 22:21:27 -07001890 skb = NULL;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001891 i = 0;
1892
1893 if (!tp->sacked_out) {
1894 /* It's already past, so skip checking against it */
1895 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1896 } else {
1897 cache = tp->recv_sack_cache;
1898 /* Skip empty blocks in at head of the cache */
1899 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1900 !cache->end_seq)
1901 cache++;
Baruch Evenfda03fb2007-02-04 23:35:57 -08001902 }
1903
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001904 while (i < used_sacks) {
Ilpo Järvinenfd6dad62007-11-15 19:49:47 -08001905 u32 start_seq = sp[i].start_seq;
1906 u32 end_seq = sp[i].end_seq;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001907 bool dup_sack = (found_dup_sack && (i == first_sack_index));
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001908 struct tcp_sack_block *next_dup = NULL;
Ilpo Järvinene56d6cd2007-11-01 00:09:37 -07001909
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001910 if (found_dup_sack && ((i + 1) == first_sack_index))
1911 next_dup = &sp[i + 1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001912
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001913 /* Skip too early cached blocks */
1914 while (tcp_sack_cache_ok(tp, cache) &&
1915 !before(start_seq, cache->end_seq))
1916 cache++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001917
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001918 /* Can skip some work by looking recv_sack_cache? */
1919 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1920 after(end_seq, cache->start_seq)) {
David S. Millerfe067e82007-03-07 12:12:44 -08001921
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001922 /* Head todo? */
1923 if (before(start_seq, cache->start_seq)) {
Taehee Yoo4bfabc42019-02-25 18:42:33 +09001924 skb = tcp_sacktag_skip(skb, sk, start_seq);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001925 skb = tcp_sacktag_walk(skb, sk, next_dup,
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001926 state,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001927 start_seq,
1928 cache->start_seq,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001929 dup_sack);
Baruch Evenfda03fb2007-02-04 23:35:57 -08001930 }
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001931
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001932 /* Rest of the block already fully processed? */
Ilpo Järvinen20de20b2007-11-16 16:17:05 -08001933 if (!after(end_seq, cache->end_seq))
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001934 goto advance_sp;
Ilpo Järvinen20de20b2007-11-16 16:17:05 -08001935
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001936 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001937 state,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001938 cache->end_seq);
Ilpo Järvinene56d6cd2007-11-01 00:09:37 -07001939
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001940 /* ...tail remains todo... */
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001941 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
Ilpo Järvinen20de20b2007-11-16 16:17:05 -08001942 /* ...but better entrypoint exists! */
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001943 skb = tcp_highest_sack(sk);
Ian Morris51456b22015-04-03 09:17:26 +01001944 if (!skb)
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001945 break;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001946 cache++;
1947 goto walk;
1948 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001949
Taehee Yoo4bfabc42019-02-25 18:42:33 +09001950 skb = tcp_sacktag_skip(skb, sk, cache->end_seq);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001951 /* Check overlap against next cached too (past this one already) */
1952 cache++;
1953 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001954 }
Ilpo Järvinenfbd52eb2007-11-10 21:24:19 -08001955
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001956 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1957 skb = tcp_highest_sack(sk);
Ian Morris51456b22015-04-03 09:17:26 +01001958 if (!skb)
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001959 break;
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001960 }
Taehee Yoo4bfabc42019-02-25 18:42:33 +09001961 skb = tcp_sacktag_skip(skb, sk, start_seq);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001962
1963walk:
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001964 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
Ilpo Järvinena1197f52008-12-05 22:42:22 -08001965 start_seq, end_seq, dup_sack);
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001966
1967advance_sp:
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001968 i++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001969 }
1970
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001971 /* Clear the head of the cache sack blocks so we can skip it next time */
1972 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1973 tp->recv_sack_cache[i].start_seq = 0;
1974 tp->recv_sack_cache[i].end_seq = 0;
1975 }
1976 for (j = 0; j < used_sacks; j++)
1977 tp->recv_sack_cache[i++] = sp[j];
1978
Yuchung Cheng737ff312017-11-08 13:01:27 -08001979 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker)
1980 tcp_check_sack_reordering(sk, state->reord, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001981
Yuchung Cheng9dac8832015-04-29 11:28:30 -07001982 tcp_verify_left_out(tp);
Ilpo Järvinen96a2d412007-11-14 15:47:18 -08001983out:
1984
Linus Torvalds1da177e2005-04-16 15:20:36 -07001985#if FASTRETRANS_DEBUG > 0
Ilpo Järvinen547b7922008-07-25 21:43:18 -07001986 WARN_ON((int)tp->sacked_out < 0);
1987 WARN_ON((int)tp->lost_out < 0);
1988 WARN_ON((int)tp->retrans_out < 0);
1989 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001990#endif
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02001991 return state->flag;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001992}
1993
Ilpo Järvinen882beba2008-04-07 22:33:07 -07001994/* Limits sacked_out so that sum with lost_out isn't ever larger than
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001995 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
Ilpo Järvinen30935cf2007-02-21 23:01:36 -08001996 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001997static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001998{
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07001999 u32 holes;
2000
2001 holes = max(tp->lost_out, 1U);
2002 holes = min(holes, tp->packets_out);
2003
2004 if ((tp->sacked_out + holes) > tp->packets_out) {
2005 tp->sacked_out = tp->packets_out - holes;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002006 return true;
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07002007 }
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002008 return false;
Ilpo Järvinen882beba2008-04-07 22:33:07 -07002009}
2010
2011/* If we receive more dupacks than we expected counting segments
2012 * in assumption of absent reordering, interpret this as reordering.
2013 * The only another reason could be bug in receiver TCP.
2014 */
2015static void tcp_check_reno_reordering(struct sock *sk, const int addend)
2016{
2017 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng737ff312017-11-08 13:01:27 -08002018
2019 if (!tcp_limit_reno_sacked(tp))
2020 return;
2021
2022 tp->reordering = min_t(u32, tp->packets_out + addend,
2023 sock_net(sk)->ipv4.sysctl_tcp_max_reordering);
Wei Wang7ec65372018-07-31 17:46:24 -07002024 tp->reord_seen++;
Yuchung Cheng737ff312017-11-08 13:01:27 -08002025 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER);
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07002026}
2027
2028/* Emulate SACKs for SACKless connection: account for a new dupack. */
2029
Yousuk Seungc634e342020-06-26 21:05:33 -07002030static void tcp_add_reno_sack(struct sock *sk, int num_dupack, bool ece_ack)
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07002031{
Eric Dumazet19119f22018-11-27 14:42:01 -08002032 if (num_dupack) {
2033 struct tcp_sock *tp = tcp_sk(sk);
2034 u32 prior_sacked = tp->sacked_out;
2035 s32 delivered;
Yuchung Chengddf1af62016-02-02 10:33:06 -08002036
Eric Dumazet19119f22018-11-27 14:42:01 -08002037 tp->sacked_out += num_dupack;
2038 tcp_check_reno_reordering(sk, 0);
2039 delivered = tp->sacked_out - prior_sacked;
2040 if (delivered > 0)
Yousuk Seung082d4fa2020-06-26 21:05:35 -07002041 tcp_count_delivered(tp, delivered, ece_ack);
Eric Dumazet19119f22018-11-27 14:42:01 -08002042 tcp_verify_left_out(tp);
2043 }
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07002044}
2045
2046/* Account for ACK, ACKing some data in Reno Recovery phase. */
2047
Yousuk Seungc634e342020-06-26 21:05:33 -07002048static void tcp_remove_reno_sacks(struct sock *sk, int acked, bool ece_ack)
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07002049{
2050 struct tcp_sock *tp = tcp_sk(sk);
2051
2052 if (acked > 0) {
2053 /* One ACK acked hole. The rest eat duplicate ACKs. */
Yousuk Seung082d4fa2020-06-26 21:05:35 -07002054 tcp_count_delivered(tp, max_t(int, acked - tp->sacked_out, 1),
2055 ece_ack);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002056 if (acked - 1 >= tp->sacked_out)
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07002057 tp->sacked_out = 0;
2058 else
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002059 tp->sacked_out -= acked - 1;
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07002060 }
2061 tcp_check_reno_reordering(sk, acked);
Ilpo Järvinen005903b2007-08-09 14:44:16 +03002062 tcp_verify_left_out(tp);
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07002063}
2064
2065static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
2066{
2067 tp->sacked_out = 0;
Ilpo Järvinen4ddf6672007-05-27 01:52:00 -07002068}
2069
Yuchung Cheng989e04c2014-08-22 14:15:22 -07002070void tcp_clear_retrans(struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002071{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002072 tp->retrans_out = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002073 tp->lost_out = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002074 tp->undo_marker = 0;
Yuchung Cheng6e08d5e2014-07-02 12:07:16 -07002075 tp->undo_retrans = -1;
Ilpo Järvinen4cd82992007-10-11 17:34:57 -07002076 tp->sacked_out = 0;
2077}
2078
Yuchung Cheng989e04c2014-08-22 14:15:22 -07002079static inline void tcp_init_undo(struct tcp_sock *tp)
2080{
2081 tp->undo_marker = tp->snd_una;
2082 /* Retransmission still in flight may cause DSACKs later. */
2083 tp->undo_retrans = tp->retrans_out ? : -1;
2084}
2085
Yuchung Chengb8fef652018-05-16 16:40:16 -07002086static bool tcp_is_rack(const struct sock *sk)
2087{
2088 return sock_net(sk)->ipv4.sysctl_tcp_recovery & TCP_RACK_LOSS_DETECTION;
2089}
2090
Yuchung Cheng2ad55f52018-05-16 16:40:14 -07002091/* If we detect SACK reneging, forget all SACK information
Linus Torvalds1da177e2005-04-16 15:20:36 -07002092 * and reset tags completely, otherwise preserve SACKs. If receiver
2093 * dropped its ofo queue, we will know this due to reneging detection.
2094 */
Yuchung Cheng2ad55f52018-05-16 16:40:14 -07002095static void tcp_timeout_mark_lost(struct sock *sk)
2096{
2097 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng56f8c5d2018-05-16 16:40:17 -07002098 struct sk_buff *skb, *head;
Yuchung Cheng2ad55f52018-05-16 16:40:14 -07002099 bool is_reneg; /* is receiver reneging on SACKs? */
2100
Yuchung Cheng56f8c5d2018-05-16 16:40:17 -07002101 head = tcp_rtx_queue_head(sk);
2102 is_reneg = head && (TCP_SKB_CB(head)->sacked & TCPCB_SACKED_ACKED);
Yuchung Cheng2ad55f52018-05-16 16:40:14 -07002103 if (is_reneg) {
2104 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2105 tp->sacked_out = 0;
2106 /* Mark SACK reneging until we recover from this loss event. */
2107 tp->is_sack_reneg = 1;
2108 } else if (tcp_is_reno(tp)) {
2109 tcp_reset_reno_sack(tp);
2110 }
2111
Yuchung Cheng56f8c5d2018-05-16 16:40:17 -07002112 skb = head;
Yuchung Cheng2ad55f52018-05-16 16:40:14 -07002113 skb_rbtree_walk_from(skb) {
2114 if (is_reneg)
2115 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
Yuchung Cheng56f8c5d2018-05-16 16:40:17 -07002116 else if (tcp_is_rack(sk) && skb != head &&
2117 tcp_rack_skb_timeout(tp, skb, 0) > 0)
2118 continue; /* Don't mark recently sent ones lost yet */
Yuchung Cheng2ad55f52018-05-16 16:40:14 -07002119 tcp_mark_skb_lost(sk, skb);
2120 }
2121 tcp_verify_left_out(tp);
2122 tcp_clear_all_retrans_hints(tp);
2123}
2124
2125/* Enter Loss state. */
Neal Cardwell5ae344c2014-08-04 19:12:29 -04002126void tcp_enter_loss(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002127{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002128 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002129 struct tcp_sock *tp = tcp_sk(sk);
Nikolay Borisov1043e252016-02-03 09:46:52 +02002130 struct net *net = sock_net(sk);
Yuchung Chengcc663f42017-04-07 11:42:05 -07002131 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002132
Yuchung Chengc77d62f2018-05-16 16:40:15 -07002133 tcp_timeout_mark_lost(sk);
2134
Linus Torvalds1da177e2005-04-16 15:20:36 -07002135 /* Reduce ssthresh if it has not yet been made inside this window. */
Yuchung Chenge33099f2013-03-20 13:33:00 +00002136 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
2137 !after(tp->high_seq, tp->snd_una) ||
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002138 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2139 tp->prior_ssthresh = tcp_current_ssthresh(sk);
Yuchung Cheng4faf7832017-08-03 20:38:51 -07002140 tp->prior_cwnd = tp->snd_cwnd;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002141 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2142 tcp_ca_event(sk, CA_EVENT_LOSS);
Yuchung Cheng989e04c2014-08-22 14:15:22 -07002143 tcp_init_undo(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002144 }
Yuchung Cheng56f8c5d2018-05-16 16:40:17 -07002145 tp->snd_cwnd = tcp_packets_in_flight(tp) + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002146 tp->snd_cwnd_cnt = 0;
Eric Dumazetc2203cf2017-05-16 14:00:04 -07002147 tp->snd_cwnd_stamp = tcp_jiffies32;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002148
Yuchung Cheng74c181d2013-08-12 16:41:25 -07002149 /* Timeout in disordered state after receiving substantial DUPACKs
2150 * suggests that the degree of reordering is over-estimated.
2151 */
2152 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
Nikolay Borisov1043e252016-02-03 09:46:52 +02002153 tp->sacked_out >= net->ipv4.sysctl_tcp_reordering)
Yuchung Cheng74c181d2013-08-12 16:41:25 -07002154 tp->reordering = min_t(unsigned int, tp->reordering,
Nikolay Borisov1043e252016-02-03 09:46:52 +02002155 net->ipv4.sysctl_tcp_reordering);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002156 tcp_set_ca_state(sk, TCP_CA_Loss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002157 tp->high_seq = tp->snd_nxt;
Florian Westphal735d3832014-09-29 13:08:30 +02002158 tcp_ecn_queue_cwr(tp);
Yuchung Chenge33099f2013-03-20 13:33:00 +00002159
Yuchung Chengcc663f42017-04-07 11:42:05 -07002160 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
2161 * loss recovery is underway except recurring timeout(s) on
2162 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
Yuchung Chenge33099f2013-03-20 13:33:00 +00002163 */
Eric Dumazetaf9b69a2017-10-26 21:55:10 -07002164 tp->frto = net->ipv4.sysctl_tcp_frto &&
Yuchung Chengcc663f42017-04-07 11:42:05 -07002165 (new_recovery || icsk->icsk_retransmits) &&
2166 !inet_csk(sk)->icsk_mtup.probe_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002167}
2168
Ilpo Järvinencadbd032007-12-31 04:49:21 -08002169/* If ACK arrived pointing to a remembered SACK, it means that our
2170 * remembered SACKs do not reflect real state of receiver i.e.
2171 * receiver _host_ is heavily congested (or buggy).
2172 *
Neal Cardwell5ae344c2014-08-04 19:12:29 -04002173 * To avoid big spurious retransmission bursts due to transient SACK
2174 * scoreboard oddities that look like reneging, we give the receiver a
2175 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2176 * restore sanity to the SACK scoreboard. If the apparent reneging
2177 * persists until this RTO then we'll clear the SACK scoreboard.
Ilpo Järvinencadbd032007-12-31 04:49:21 -08002178 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002179static bool tcp_check_sack_reneging(struct sock *sk, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002180{
Ilpo Järvinencadbd032007-12-31 04:49:21 -08002181 if (flag & FLAG_SACK_RENEGING) {
Neal Cardwell5ae344c2014-08-04 19:12:29 -04002182 struct tcp_sock *tp = tcp_sk(sk);
2183 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2184 msecs_to_jiffies(10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002185
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002186 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
Neal Cardwell5ae344c2014-08-04 19:12:29 -04002187 delay, TCP_RTO_MAX);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002188 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002189 }
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002190 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002191}
2192
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002193/* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2194 * counter when SACK is enabled (without SACK, sacked_out is used for
2195 * that purpose).
2196 *
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002197 * With reordering, holes may still be in flight, so RFC3517 recovery
2198 * uses pure sacked_out (total number of SACKed segments) even though
2199 * it violates the RFC that uses duplicate ACKs, often these are equal
2200 * but when e.g. out-of-window ACKs or packet duplication occurs,
2201 * they differ. Since neither occurs due to loss, TCP should really
2202 * ignore them.
2203 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04002204static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002205{
Yuchung Cheng713bafe2017-11-08 13:01:26 -08002206 return tp->sacked_out + 1;
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002207}
2208
Yuchung Cheng713bafe2017-11-08 13:01:26 -08002209/* Linux NewReno/SACK/ECN state machine.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002210 * --------------------------------------
2211 *
2212 * "Open" Normal state, no dubious events, fast path.
2213 * "Disorder" In all the respects it is "Open",
2214 * but requires a bit more attention. It is entered when
2215 * we see some SACKs or dupacks. It is split of "Open"
2216 * mainly to move some processing from fast path to slow one.
2217 * "CWR" CWND was reduced due to some Congestion Notification event.
2218 * It can be ECN, ICMP source quench, local device congestion.
2219 * "Recovery" CWND was reduced, we are fast-retransmitting.
2220 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2221 *
2222 * tcp_fastretrans_alert() is entered:
2223 * - each incoming ACK, if state is not "Open"
2224 * - when arrived ACK is unusual, namely:
2225 * * SACK
2226 * * Duplicate ACK.
2227 * * ECN ECE.
2228 *
2229 * Counting packets in flight is pretty simple.
2230 *
2231 * in_flight = packets_out - left_out + retrans_out
2232 *
2233 * packets_out is SND.NXT-SND.UNA counted in packets.
2234 *
2235 * retrans_out is number of retransmitted segments.
2236 *
2237 * left_out is number of segments left network, but not ACKed yet.
2238 *
2239 * left_out = sacked_out + lost_out
2240 *
2241 * sacked_out: Packets, which arrived to receiver out of order
2242 * and hence not ACKed. With SACKs this number is simply
2243 * amount of SACKed data. Even without SACKs
2244 * it is easy to give pretty reliable estimate of this number,
2245 * counting duplicate ACKs.
2246 *
2247 * lost_out: Packets lost by network. TCP has no explicit
2248 * "loss notification" feedback from network (for now).
2249 * It means that this number can be only _guessed_.
2250 * Actually, it is the heuristics to predict lossage that
2251 * distinguishes different algorithms.
2252 *
2253 * F.e. after RTO, when all the queue is considered as lost,
2254 * lost_out = packets_out and in_flight = retrans_out.
2255 *
Yuchung Chenga0370b32017-01-12 22:11:36 -08002256 * Essentially, we have now a few algorithms detecting
Linus Torvalds1da177e2005-04-16 15:20:36 -07002257 * lost packets.
2258 *
Yuchung Chenga0370b32017-01-12 22:11:36 -08002259 * If the receiver supports SACK:
2260 *
2261 * RFC6675/3517: It is the conventional algorithm. A packet is
2262 * considered lost if the number of higher sequence packets
2263 * SACKed is greater than or equal the DUPACK thoreshold
2264 * (reordering). This is implemented in tcp_mark_head_lost and
2265 * tcp_update_scoreboard.
2266 *
2267 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2268 * (2017-) that checks timing instead of counting DUPACKs.
2269 * Essentially a packet is considered lost if it's not S/ACKed
2270 * after RTT + reordering_window, where both metrics are
2271 * dynamically measured and adjusted. This is implemented in
2272 * tcp_rack_mark_lost.
2273 *
Yuchung Chenga0370b32017-01-12 22:11:36 -08002274 * If the receiver does not support SACK:
2275 *
2276 * NewReno (RFC6582): in Recovery we assume that one segment
Linus Torvalds1da177e2005-04-16 15:20:36 -07002277 * is lost (classic Reno). While we are in Recovery and
2278 * a partial ACK arrives, we assume that one more packet
2279 * is lost (NewReno). This heuristics are the same in NewReno
2280 * and SACK.
2281 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002282 * Really tricky (and requiring careful tuning) part of algorithm
2283 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2284 * The first determines the moment _when_ we should reduce CWND and,
2285 * hence, slow down forward transmission. In fact, it determines the moment
2286 * when we decide that hole is caused by loss, rather than by a reorder.
2287 *
2288 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2289 * holes, caused by lost packets.
2290 *
2291 * And the most logically complicated part of algorithm is undo
2292 * heuristics. We detect false retransmits due to both too early
2293 * fast retransmit (reordering) and underestimated RTO, analyzing
2294 * timestamps and D-SACKs. When we detect that some segments were
2295 * retransmitted by mistake and CWND reduction was wrong, we undo
2296 * window reduction and abort recovery phase. This logic is hidden
2297 * inside several functions named tcp_try_undo_<something>.
2298 */
2299
2300/* This function decides, when we should leave Disordered state
2301 * and enter Recovery phase, reducing congestion window.
2302 *
2303 * Main question: may we further continue forward transmission
2304 * with the same cwnd?
2305 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002306static bool tcp_time_to_recover(struct sock *sk, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002307{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002308 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002309
2310 /* Trick#1: The loss is proven. */
2311 if (tp->lost_out)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002312 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002313
2314 /* Not-A-Trick#2 : Classic rule... */
Yuchung Chengb38a51f2018-05-16 16:40:11 -07002315 if (!tcp_is_rack(sk) && tcp_dupack_heuristics(tp) > tp->reordering)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002316 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002317
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002318 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002319}
2320
Yuchung Cheng974c1232012-01-19 14:42:21 +00002321/* Detect loss in event "A" above by marking head of queue up as lost.
zhang kai636ef282020-05-07 11:08:30 +08002322 * For RFC3517 SACK, a segment is considered lost if it
Yuchung Cheng974c1232012-01-19 14:42:21 +00002323 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2324 * the maximum SACKed segments to pass before reaching this limit.
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002325 */
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002326static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002327{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002328 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002329 struct sk_buff *skb;
zhang kai636ef282020-05-07 11:08:30 +08002330 int cnt;
Yuchung Cheng974c1232012-01-19 14:42:21 +00002331 /* Use SACK to deduce losses of new sequences sent during recovery */
zhang kai636ef282020-05-07 11:08:30 +08002332 const u32 loss_high = tp->snd_nxt;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002333
Ilpo Järvinen547b7922008-07-25 21:43:18 -07002334 WARN_ON(packets > tp->packets_out);
Eric Dumazet5e76ee42017-10-05 22:21:24 -07002335 skb = tp->lost_skb_hint;
2336 if (skb) {
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002337 /* Head already handled? */
Eric Dumazet5e76ee42017-10-05 22:21:24 -07002338 if (mark_head && after(TCP_SKB_CB(skb)->seq, tp->snd_una))
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002339 return;
Eric Dumazet5e76ee42017-10-05 22:21:24 -07002340 cnt = tp->lost_cnt_hint;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002341 } else {
Eric Dumazet75c119a2017-10-05 22:21:27 -07002342 skb = tcp_rtx_queue_head(sk);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002343 cnt = 0;
2344 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002345
Eric Dumazet75c119a2017-10-05 22:21:27 -07002346 skb_rbtree_walk_from(skb) {
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002347 /* TODO: do this better */
2348 /* this is not the most efficient way to do this... */
2349 tp->lost_skb_hint = skb;
2350 tp->lost_cnt_hint = cnt;
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002351
Yuchung Cheng974c1232012-01-19 14:42:21 +00002352 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
Ilpo Järvinenc137f3d2008-04-07 22:32:38 -07002353 break;
2354
zhang kai636ef282020-05-07 11:08:30 +08002355 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002356 cnt += tcp_skb_pcount(skb);
2357
zhang kai636ef282020-05-07 11:08:30 +08002358 if (cnt > packets)
2359 break;
Ilpo Järvinenc137f3d2008-04-07 22:32:38 -07002360
Yuchung Cheng534a2102020-09-25 10:04:31 -07002361 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_LOST))
2362 tcp_mark_skb_lost(sk, skb);
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002363
2364 if (mark_head)
2365 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002366 }
Ilpo Järvinen005903b2007-08-09 14:44:16 +03002367 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002368}
2369
2370/* Account newly detected lost packet(s) */
2371
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002372static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002373{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002374 struct tcp_sock *tp = tcp_sk(sk);
2375
Yuchung Cheng6ac06ec2018-05-16 16:40:12 -07002376 if (tcp_is_sack(tp)) {
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08002377 int sacked_upto = tp->sacked_out - tp->reordering;
Ilpo Järvinen1fdb9362010-10-14 01:42:30 +00002378 if (sacked_upto >= 0)
2379 tcp_mark_head_lost(sk, sacked_upto, 0);
2380 else if (fast_rexmit)
2381 tcp_mark_head_lost(sk, 1, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002382 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002383}
2384
Yuchung Cheng77c63122015-10-16 21:57:44 -07002385static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2386{
2387 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2388 before(tp->rx_opt.rcv_tsecr, when);
2389}
2390
Yuchung Cheng659a8ad2015-10-16 21:57:46 -07002391/* skb is spurious retransmitted if the returned timestamp echo
2392 * reply is prior to the skb transmission time
2393 */
2394static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2395 const struct sk_buff *skb)
2396{
2397 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2398 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb));
2399}
2400
Linus Torvalds1da177e2005-04-16 15:20:36 -07002401/* Nothing was retransmitted or returned timestamp is less
2402 * than timestamp of the first retransmission.
2403 */
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00002404static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002405{
Yuchung Chengbc9f38c2019-04-29 15:46:13 -07002406 return tp->retrans_stamp &&
Yuchung Cheng77c63122015-10-16 21:57:44 -07002407 tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002408}
2409
2410/* Undo procedures. */
2411
Marcelo Leitner1f37bf82014-11-04 17:15:08 -02002412/* We can clear retrans_stamp when there are no retransmissions in the
2413 * window. It would seem that it is trivially available for us in
2414 * tp->retrans_out, however, that kind of assumptions doesn't consider
2415 * what will happen if errors occur when sending retransmission for the
2416 * second time. ...It could the that such segment has only
2417 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2418 * the head skb is enough except for some reneging corner cases that
2419 * are not worth the effort.
2420 *
2421 * Main reason for all this complexity is the fact that connection dying
2422 * time now depends on the validity of the retrans_stamp, in particular,
2423 * that successive retransmissions of a segment must not advance
2424 * retrans_stamp under any conditions.
2425 */
2426static bool tcp_any_retrans_done(const struct sock *sk)
2427{
2428 const struct tcp_sock *tp = tcp_sk(sk);
2429 struct sk_buff *skb;
2430
2431 if (tp->retrans_out)
2432 return true;
2433
Eric Dumazet75c119a2017-10-05 22:21:27 -07002434 skb = tcp_rtx_queue_head(sk);
Marcelo Leitner1f37bf82014-11-04 17:15:08 -02002435 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2436 return true;
2437
2438 return false;
2439}
2440
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002441static void DBGUNDO(struct sock *sk, const char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002442{
Joe Perches39347882017-09-10 19:02:25 -07002443#if FASTRETRANS_DEBUG > 1
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002444 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002445 struct inet_sock *inet = inet_sk(sk);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002446
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002447 if (sk->sk_family == AF_INET) {
Joe Perches91df42b2012-05-15 14:11:54 +00002448 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2449 msg,
2450 &inet->inet_daddr, ntohs(inet->inet_dport),
2451 tp->snd_cwnd, tcp_left_out(tp),
2452 tp->snd_ssthresh, tp->prior_ssthresh,
2453 tp->packets_out);
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002454 }
Eric Dumazetdfd56b82011-12-10 09:48:31 +00002455#if IS_ENABLED(CONFIG_IPV6)
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002456 else if (sk->sk_family == AF_INET6) {
Joe Perches91df42b2012-05-15 14:11:54 +00002457 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2458 msg,
Eric Dumazet019b1c92016-09-22 17:54:00 -07002459 &sk->sk_v6_daddr, ntohs(inet->inet_dport),
Joe Perches91df42b2012-05-15 14:11:54 +00002460 tp->snd_cwnd, tcp_left_out(tp),
2461 tp->snd_ssthresh, tp->prior_ssthresh,
2462 tp->packets_out);
YOSHIFUJI Hideaki569508c2008-04-14 04:09:36 -07002463 }
2464#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002465#endif
Joe Perches39347882017-09-10 19:02:25 -07002466}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002467
Yuchung Cheng7026b912013-05-29 14:20:13 +00002468static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002469{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002470 struct tcp_sock *tp = tcp_sk(sk);
2471
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002472 if (unmark_loss) {
2473 struct sk_buff *skb;
2474
Eric Dumazet75c119a2017-10-05 22:21:27 -07002475 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002476 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2477 }
2478 tp->lost_out = 0;
2479 tcp_clear_all_retrans_hints(tp);
2480 }
2481
Linus Torvalds1da177e2005-04-16 15:20:36 -07002482 if (tp->prior_ssthresh) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002483 const struct inet_connection_sock *icsk = inet_csk(sk);
2484
Florian Westphale9799182016-11-21 14:18:38 +01002485 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002486
Yuchung Cheng7026b912013-05-29 14:20:13 +00002487 if (tp->prior_ssthresh > tp->snd_ssthresh) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002488 tp->snd_ssthresh = tp->prior_ssthresh;
Florian Westphal735d3832014-09-29 13:08:30 +02002489 tcp_ecn_withdraw_cwr(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002490 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002491 }
Eric Dumazetc2203cf2017-05-16 14:00:04 -07002492 tp->snd_cwnd_stamp = tcp_jiffies32;
Yuchung Cheng7026b912013-05-29 14:20:13 +00002493 tp->undo_marker = 0;
Yuchung Chengcd1fc852017-12-07 11:33:31 -08002494 tp->rack.advanced = 1; /* Force RACK to re-exam losses */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002495}
2496
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00002497static inline bool tcp_may_undo(const struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002498{
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002499 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002500}
2501
2502/* People celebrate: "We love our President!" */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002503static bool tcp_try_undo_recovery(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002504{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002505 struct tcp_sock *tp = tcp_sk(sk);
2506
Linus Torvalds1da177e2005-04-16 15:20:36 -07002507 if (tcp_may_undo(tp)) {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07002508 int mib_idx;
2509
Linus Torvalds1da177e2005-04-16 15:20:36 -07002510 /* Happy end! We did not retransmit anything
2511 * or our original transmission succeeded.
2512 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002513 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
Yuchung Cheng7026b912013-05-29 14:20:13 +00002514 tcp_undo_cwnd_reduction(sk, false);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002515 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07002516 mib_idx = LINUX_MIB_TCPLOSSUNDO;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002517 else
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07002518 mib_idx = LINUX_MIB_TCPFULLUNDO;
2519
Eric Dumazetc10d9312016-04-29 14:16:47 -07002520 NET_INC_STATS(sock_net(sk), mib_idx);
Priyaranjan Jha1f255692017-11-03 16:38:48 -07002521 } else if (tp->rack.reo_wnd_persist) {
2522 tp->rack.reo_wnd_persist--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002523 }
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002524 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002525 /* Hold old state until something *above* high_seq
2526 * is ACKed. For Reno it is MUST to prevent false
2527 * fast retransmits (RFC2582). SACK TCP is safe. */
Marcelo Leitner1f37bf82014-11-04 17:15:08 -02002528 if (!tcp_any_retrans_done(sk))
2529 tp->retrans_stamp = 0;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002530 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002531 }
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002532 tcp_set_ca_state(sk, TCP_CA_Open);
Yousuk Seungd4761752017-12-07 13:41:34 -08002533 tp->is_sack_reneg = 0;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002534 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002535}
2536
2537/* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
Yuchung Chengc7d9d6a2013-05-29 14:20:14 +00002538static bool tcp_try_undo_dsack(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002539{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002540 struct tcp_sock *tp = tcp_sk(sk);
2541
Linus Torvalds1da177e2005-04-16 15:20:36 -07002542 if (tp->undo_marker && !tp->undo_retrans) {
Priyaranjan Jha1f255692017-11-03 16:38:48 -07002543 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH,
2544 tp->rack.reo_wnd_persist + 1);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002545 DBGUNDO(sk, "D-SACK");
Yuchung Cheng7026b912013-05-29 14:20:13 +00002546 tcp_undo_cwnd_reduction(sk, false);
Eric Dumazetc10d9312016-04-29 14:16:47 -07002547 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
Yuchung Chengc7d9d6a2013-05-29 14:20:14 +00002548 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002549 }
Yuchung Chengc7d9d6a2013-05-29 14:20:14 +00002550 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002551}
2552
Yuchung Chenge33099f2013-03-20 13:33:00 +00002553/* Undo during loss recovery after partial ACK or using F-RTO. */
2554static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002555{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002556 struct tcp_sock *tp = tcp_sk(sk);
2557
Yuchung Chenge33099f2013-03-20 13:33:00 +00002558 if (frto_undo || tcp_may_undo(tp)) {
Yuchung Cheng7026b912013-05-29 14:20:13 +00002559 tcp_undo_cwnd_reduction(sk, true);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002560
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002561 DBGUNDO(sk, "partial loss");
Eric Dumazetc10d9312016-04-29 14:16:47 -07002562 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
Yuchung Chenge33099f2013-03-20 13:33:00 +00002563 if (frto_undo)
Eric Dumazetc10d9312016-04-29 14:16:47 -07002564 NET_INC_STATS(sock_net(sk),
Eric Dumazet02a1d6e2016-04-27 16:44:39 -07002565 LINUX_MIB_TCPSPURIOUSRTOS);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002566 inet_csk(sk)->icsk_retransmits = 0;
Yousuk Seungd4761752017-12-07 13:41:34 -08002567 if (frto_undo || tcp_is_sack(tp)) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002568 tcp_set_ca_state(sk, TCP_CA_Open);
Yousuk Seungd4761752017-12-07 13:41:34 -08002569 tp->is_sack_reneg = 0;
2570 }
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002571 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002572 }
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002573 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002574}
2575
Yuchung Cheng37598242015-07-01 14:11:15 -07002576/* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002577 * It computes the number of packets to send (sndcnt) based on packets newly
2578 * delivered:
2579 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2580 * cwnd reductions across a full RTT.
Yuchung Cheng37598242015-07-01 14:11:15 -07002581 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
Yuchung Cheng7e901ee2020-10-30 18:34:12 -07002582 * But when SND_UNA is acked without further losses,
Yuchung Cheng37598242015-07-01 14:11:15 -07002583 * slow starts cwnd up to ssthresh to speed up the recovery.
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002584 */
Christoph Paasch5ee2c942014-07-14 16:58:32 +02002585static void tcp_init_cwnd_reduction(struct sock *sk)
Yuchung Cheng684bad12012-09-02 17:38:04 +00002586{
2587 struct tcp_sock *tp = tcp_sk(sk);
2588
2589 tp->high_seq = tp->snd_nxt;
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00002590 tp->tlp_high_seq = 0;
Yuchung Cheng684bad12012-09-02 17:38:04 +00002591 tp->snd_cwnd_cnt = 0;
2592 tp->prior_cwnd = tp->snd_cwnd;
2593 tp->prr_delivered = 0;
2594 tp->prr_out = 0;
Christoph Paasch5ee2c942014-07-14 16:58:32 +02002595 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
Florian Westphal735d3832014-09-29 13:08:30 +02002596 tcp_ecn_queue_cwr(tp);
Yuchung Cheng684bad12012-09-02 17:38:04 +00002597}
2598
Yuchung Cheng7e901ee2020-10-30 18:34:12 -07002599void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int newly_lost, int flag)
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002600{
2601 struct tcp_sock *tp = tcp_sk(sk);
2602 int sndcnt = 0;
2603 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2604
Yuchung Cheng8b8a3212016-01-06 12:42:38 -08002605 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2606 return;
2607
Yuchung Cheng684bad12012-09-02 17:38:04 +00002608 tp->prr_delivered += newly_acked_sacked;
Yuchung Cheng37598242015-07-01 14:11:15 -07002609 if (delta < 0) {
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002610 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2611 tp->prior_cwnd - 1;
2612 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
Yuchung Cheng7e901ee2020-10-30 18:34:12 -07002613 } else if (flag & FLAG_SND_UNA_ADVANCED && !newly_lost) {
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002614 sndcnt = min_t(int, delta,
2615 max_t(int, tp->prr_delivered - tp->prr_out,
2616 newly_acked_sacked) + 1);
Yuchung Cheng37598242015-07-01 14:11:15 -07002617 } else {
2618 sndcnt = min(delta, newly_acked_sacked);
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002619 }
Yuchung Cheng31ba0c12016-02-02 10:33:05 -08002620 /* Force a fast retransmit upon entering fast recovery */
2621 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
Yuchung Chengfb4d3d12012-09-02 17:38:03 +00002622 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2623}
2624
Yuchung Cheng684bad12012-09-02 17:38:04 +00002625static inline void tcp_end_cwnd_reduction(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002626{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002627 struct tcp_sock *tp = tcp_sk(sk);
Nandita Dukkipatia262f0c2011-08-21 20:21:57 +00002628
Yuchung Chengc0402762016-09-19 23:39:21 -04002629 if (inet_csk(sk)->icsk_ca_ops->cong_control)
2630 return;
2631
Yuchung Cheng684bad12012-09-02 17:38:04 +00002632 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
Yuchung Chenged254972017-08-01 13:22:32 -07002633 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH &&
2634 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) {
Yuchung Cheng684bad12012-09-02 17:38:04 +00002635 tp->snd_cwnd = tp->snd_ssthresh;
Eric Dumazetc2203cf2017-05-16 14:00:04 -07002636 tp->snd_cwnd_stamp = tcp_jiffies32;
Yuchung Cheng67d41202011-03-14 10:57:03 +00002637 }
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002638 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002639}
2640
Yuchung Cheng684bad12012-09-02 17:38:04 +00002641/* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
Christoph Paasch5ee2c942014-07-14 16:58:32 +02002642void tcp_enter_cwr(struct sock *sk)
Yuchung Cheng09484d12012-09-02 17:38:02 +00002643{
2644 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng09484d12012-09-02 17:38:02 +00002645
2646 tp->prior_ssthresh = 0;
Yuchung Cheng684bad12012-09-02 17:38:04 +00002647 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
Yuchung Cheng09484d12012-09-02 17:38:02 +00002648 tp->undo_marker = 0;
Christoph Paasch5ee2c942014-07-14 16:58:32 +02002649 tcp_init_cwnd_reduction(sk);
Yuchung Cheng09484d12012-09-02 17:38:02 +00002650 tcp_set_ca_state(sk, TCP_CA_CWR);
2651 }
2652}
Kenneth Klette Jonassen7782ad82015-06-10 19:08:16 +02002653EXPORT_SYMBOL(tcp_enter_cwr);
Yuchung Cheng09484d12012-09-02 17:38:02 +00002654
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07002655static void tcp_try_keep_open(struct sock *sk)
2656{
2657 struct tcp_sock *tp = tcp_sk(sk);
2658 int state = TCP_CA_Open;
2659
Neal Cardwellf6982042011-11-16 08:58:04 +00002660 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07002661 state = TCP_CA_Disorder;
2662
2663 if (inet_csk(sk)->icsk_ca_state != state) {
2664 tcp_set_ca_state(sk, state);
2665 tp->high_seq = tp->snd_nxt;
2666 }
2667}
2668
Yuchung Cheng31ba0c12016-02-02 10:33:05 -08002669static void tcp_try_to_open(struct sock *sk, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002670{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002671 struct tcp_sock *tp = tcp_sk(sk);
2672
Ilpo Järvinen86426c22007-08-09 14:45:17 +03002673 tcp_verify_left_out(tp);
2674
Yuchung Cheng9b441902013-03-20 13:32:58 +00002675 if (!tcp_any_retrans_done(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002676 tp->retrans_stamp = 0;
2677
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002678 if (flag & FLAG_ECE)
Christoph Paasch5ee2c942014-07-14 16:58:32 +02002679 tcp_enter_cwr(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002680
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002681 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07002682 tcp_try_keep_open(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002683 }
2684}
2685
John Heffner5d424d52006-03-20 17:53:41 -08002686static void tcp_mtup_probe_failed(struct sock *sk)
2687{
2688 struct inet_connection_sock *icsk = inet_csk(sk);
2689
2690 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2691 icsk->icsk_mtup.probe_size = 0;
Eric Dumazetc10d9312016-04-29 14:16:47 -07002692 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
John Heffner5d424d52006-03-20 17:53:41 -08002693}
2694
Ilpo Järvinen72211e92009-03-14 14:23:04 +00002695static void tcp_mtup_probe_success(struct sock *sk)
John Heffner5d424d52006-03-20 17:53:41 -08002696{
2697 struct tcp_sock *tp = tcp_sk(sk);
2698 struct inet_connection_sock *icsk = inet_csk(sk);
2699
2700 /* FIXME: breaks with very large cwnd */
2701 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2702 tp->snd_cwnd = tp->snd_cwnd *
2703 tcp_mss_to_mtu(sk, tp->mss_cache) /
2704 icsk->icsk_mtup.probe_size;
2705 tp->snd_cwnd_cnt = 0;
Eric Dumazetc2203cf2017-05-16 14:00:04 -07002706 tp->snd_cwnd_stamp = tcp_jiffies32;
John Heffner9c6d5e52010-10-06 21:18:02 -07002707 tp->snd_ssthresh = tcp_current_ssthresh(sk);
John Heffner5d424d52006-03-20 17:53:41 -08002708
2709 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2710 icsk->icsk_mtup.probe_size = 0;
2711 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
Eric Dumazetc10d9312016-04-29 14:16:47 -07002712 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
John Heffner5d424d52006-03-20 17:53:41 -08002713}
2714
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08002715/* Do a simple retransmit without using the backoff mechanisms in
2716 * tcp_timer. This is used for path mtu discovery.
2717 * The socket is already locked here.
2718 */
2719void tcp_simple_retransmit(struct sock *sk)
2720{
2721 const struct inet_connection_sock *icsk = inet_csk(sk);
2722 struct tcp_sock *tp = tcp_sk(sk);
2723 struct sk_buff *skb;
Alexander Duyckc31b70c2020-12-12 12:31:24 -08002724 int mss;
2725
2726 /* A fastopen SYN request is stored as two separate packets within
2727 * the retransmit queue, this is done by tcp_send_syn_data().
2728 * As a result simply checking the MSS of the frames in the queue
2729 * will not work for the SYN packet.
2730 *
2731 * Us being here is an indication of a path MTU issue so we can
2732 * assume that the fastopen SYN was lost and just mark all the
2733 * frames in the retransmit queue as lost. We will use an MSS of
2734 * -1 to mark all frames as lost, otherwise compute the current MSS.
2735 */
2736 if (tp->syn_data && sk->sk_state == TCP_SYN_SENT)
2737 mss = -1;
2738 else
2739 mss = tcp_current_mss(sk);
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08002740
Eric Dumazet75c119a2017-10-05 22:21:27 -07002741 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
Yuchung Cheng68698972020-09-25 10:04:30 -07002742 if (tcp_skb_seglen(skb) > mss)
Yuchung Cheng179ac352020-09-25 10:04:28 -07002743 tcp_mark_skb_lost(sk, skb);
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08002744 }
2745
2746 tcp_clear_retrans_hints_partial(tp);
2747
Yuchung Cheng0eb96bf2017-11-07 15:33:43 -08002748 if (!tp->lost_out)
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08002749 return;
2750
2751 if (tcp_is_reno(tp))
2752 tcp_limit_reno_sacked(tp);
2753
2754 tcp_verify_left_out(tp);
2755
2756 /* Don't muck with the congestion window here.
2757 * Reason is that we do not increase amount of _data_
2758 * in network, but units changed and effective
2759 * cwnd/ssthresh really reduced now.
2760 */
2761 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2762 tp->high_seq = tp->snd_nxt;
2763 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2764 tp->prior_ssthresh = 0;
2765 tp->undo_marker = 0;
2766 tcp_set_ca_state(sk, TCP_CA_Loss);
2767 }
2768 tcp_xmit_retransmit_queue(sk);
2769}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00002770EXPORT_SYMBOL(tcp_simple_retransmit);
Ilpo Järvinene1aa6802008-11-24 21:11:55 -08002771
Yuchung Cheng57dde7f2017-01-12 22:11:33 -08002772void tcp_enter_recovery(struct sock *sk, bool ece_ack)
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002773{
2774 struct tcp_sock *tp = tcp_sk(sk);
2775 int mib_idx;
2776
2777 if (tcp_is_reno(tp))
2778 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2779 else
2780 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2781
Eric Dumazetc10d9312016-04-29 14:16:47 -07002782 NET_INC_STATS(sock_net(sk), mib_idx);
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002783
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002784 tp->prior_ssthresh = 0;
Yuchung Cheng989e04c2014-08-22 14:15:22 -07002785 tcp_init_undo(tp);
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002786
Yuchung Cheng291a00d2015-07-01 14:11:14 -07002787 if (!tcp_in_cwnd_reduction(sk)) {
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002788 if (!ece_ack)
2789 tp->prior_ssthresh = tcp_current_ssthresh(sk);
Christoph Paasch5ee2c942014-07-14 16:58:32 +02002790 tcp_init_cwnd_reduction(sk);
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002791 }
Yuchung Cheng1fbc3402012-05-02 13:30:02 +00002792 tcp_set_ca_state(sk, TCP_CA_Recovery);
2793}
2794
Yuchung Chengab42d9e2013-03-20 13:32:59 +00002795/* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2796 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2797 */
Eric Dumazet19119f22018-11-27 14:42:01 -08002798static void tcp_process_loss(struct sock *sk, int flag, int num_dupack,
Yuchung Chenge662ca42016-02-02 10:33:04 -08002799 int *rexmit)
Yuchung Chengab42d9e2013-03-20 13:32:59 +00002800{
Yuchung Chengab42d9e2013-03-20 13:32:59 +00002801 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Chenge33099f2013-03-20 13:33:00 +00002802 bool recovered = !before(tp->snd_una, tp->high_seq);
Yuchung Chengab42d9e2013-03-20 13:32:59 +00002803
Eric Dumazetd983ea62019-10-10 20:17:38 -07002804 if ((flag & FLAG_SND_UNA_ADVANCED || rcu_access_pointer(tp->fastopen_rsk)) &&
Yuchung Chengda34ac72015-05-18 12:31:44 -07002805 tcp_try_undo_loss(sk, false))
2806 return;
2807
Yuchung Cheng89fe18e2017-01-12 22:11:37 -08002808 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
Yuchung Chengfc68e172018-02-27 14:15:02 -08002809 /* Step 3.b. A timeout is spurious if not all data are
2810 * lost, i.e., never-retransmitted data are (s)acked.
2811 */
2812 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2813 tcp_try_undo_loss(sk, true))
2814 return;
2815
Yuchung Chengb7b0ed92015-05-18 12:31:45 -07002816 if (after(tp->snd_nxt, tp->high_seq)) {
Eric Dumazet19119f22018-11-27 14:42:01 -08002817 if (flag & FLAG_DATA_SACKED || num_dupack)
Yuchung Chengb7b0ed92015-05-18 12:31:45 -07002818 tp->frto = 0; /* Step 3.a. loss was real */
Yuchung Chenge33099f2013-03-20 13:33:00 +00002819 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2820 tp->high_seq = tp->snd_nxt;
Yuchung Chenge662ca42016-02-02 10:33:04 -08002821 /* Step 2.b. Try send new data (but deferred until cwnd
2822 * is updated in tcp_ack()). Otherwise fall back to
2823 * the conventional recovery.
2824 */
Eric Dumazet75c119a2017-10-05 22:21:27 -07002825 if (!tcp_write_queue_empty(sk) &&
Yuchung Chenge662ca42016-02-02 10:33:04 -08002826 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2827 *rexmit = REXMIT_NEW;
2828 return;
2829 }
Yuchung Chenge33099f2013-03-20 13:33:00 +00002830 tp->frto = 0;
2831 }
2832 }
2833
2834 if (recovered) {
2835 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
Yuchung Chengab42d9e2013-03-20 13:32:59 +00002836 tcp_try_undo_recovery(sk);
2837 return;
2838 }
Yuchung Chenge33099f2013-03-20 13:33:00 +00002839 if (tcp_is_reno(tp)) {
2840 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2841 * delivered. Lower inflight to clock out (re)tranmissions.
2842 */
Eric Dumazet19119f22018-11-27 14:42:01 -08002843 if (after(tp->snd_nxt, tp->high_seq) && num_dupack)
Yousuk Seungc634e342020-06-26 21:05:33 -07002844 tcp_add_reno_sack(sk, num_dupack, flag & FLAG_ECE);
Yuchung Chenge33099f2013-03-20 13:33:00 +00002845 else if (flag & FLAG_SND_UNA_ADVANCED)
2846 tcp_reset_reno_sack(tp);
2847 }
Yuchung Chenge662ca42016-02-02 10:33:04 -08002848 *rexmit = REXMIT_LOST;
Yuchung Chengab42d9e2013-03-20 13:32:59 +00002849}
2850
Yuchung Chenga29cb692021-06-02 17:51:21 -07002851static bool tcp_force_fast_retransmit(struct sock *sk)
2852{
2853 struct tcp_sock *tp = tcp_sk(sk);
2854
2855 return after(tcp_highest_sack_seq(tp),
2856 tp->snd_una + tp->reordering * tp->mss_cache);
2857}
2858
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002859/* Undo during fast recovery after partial ACK. */
Yuchung Chenga29cb692021-06-02 17:51:21 -07002860static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una,
2861 bool *do_lost)
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002862{
2863 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002864
Yuchung Cheng7026b912013-05-29 14:20:13 +00002865 if (tp->undo_marker && tcp_packet_delayed(tp)) {
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002866 /* Plain luck! Hole if filled with delayed
Yuchung Cheng737ff312017-11-08 13:01:27 -08002867 * packet, rather than with a retransmit. Check reordering.
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002868 */
Yuchung Cheng737ff312017-11-08 13:01:27 -08002869 tcp_check_sack_reordering(sk, prior_snd_una, 1);
Yuchung Cheng7026b912013-05-29 14:20:13 +00002870
2871 /* We are getting evidence that the reordering degree is higher
2872 * than we realized. If there are no retransmits out then we
2873 * can undo. Otherwise we clock out new packets but do not
2874 * mark more packets lost or retransmit more.
2875 */
Yuchung Cheng31ba0c12016-02-02 10:33:05 -08002876 if (tp->retrans_out)
Yuchung Cheng7026b912013-05-29 14:20:13 +00002877 return true;
Yuchung Cheng7026b912013-05-29 14:20:13 +00002878
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002879 if (!tcp_any_retrans_done(sk))
2880 tp->retrans_stamp = 0;
2881
Yuchung Cheng7026b912013-05-29 14:20:13 +00002882 DBGUNDO(sk, "partial recovery");
2883 tcp_undo_cwnd_reduction(sk, true);
Eric Dumazetc10d9312016-04-29 14:16:47 -07002884 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
Yuchung Cheng7026b912013-05-29 14:20:13 +00002885 tcp_try_keep_open(sk);
Yuchung Chenga29cb692021-06-02 17:51:21 -07002886 } else {
2887 /* Partial ACK arrived. Force fast retransmit. */
2888 *do_lost = tcp_force_fast_retransmit(sk);
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002889 }
Yuchung Cheng7026b912013-05-29 14:20:13 +00002890 return false;
Yuchung Cheng6a63df42013-05-29 14:20:12 +00002891}
2892
Yuchung Cheng6ac06ec2018-05-16 16:40:12 -07002893static void tcp_identify_packet_loss(struct sock *sk, int *ack_flag)
Yuchung Cheng98e36d42017-01-12 22:11:35 -08002894{
2895 struct tcp_sock *tp = tcp_sk(sk);
2896
Yuchung Cheng6ac06ec2018-05-16 16:40:12 -07002897 if (tcp_rtx_queue_empty(sk))
2898 return;
2899
2900 if (unlikely(tcp_is_reno(tp))) {
2901 tcp_newreno_mark_lost(sk, *ack_flag & FLAG_SND_UNA_ADVANCED);
2902 } else if (tcp_is_rack(sk)) {
Yuchung Cheng98e36d42017-01-12 22:11:35 -08002903 u32 prior_retrans = tp->retrans_out;
2904
Pengcheng Yang62d9f1a2021-01-24 13:07:14 +08002905 if (tcp_rack_mark_lost(sk))
2906 *ack_flag &= ~FLAG_SET_XMIT_TIMER;
Yuchung Cheng98e36d42017-01-12 22:11:35 -08002907 if (prior_retrans > tp->retrans_out)
2908 *ack_flag |= FLAG_LOST_RETRANS;
2909 }
2910}
2911
Linus Torvalds1da177e2005-04-16 15:20:36 -07002912/* Process an event, which can update packets-in-flight not trivially.
2913 * Main goal of this function is to calculate new estimate for left_out,
2914 * taking into account both packets sitting in receiver's buffer and
2915 * packets lost by network.
2916 *
Yuchung Cheng31ba0c12016-02-02 10:33:05 -08002917 * Besides that it updates the congestion state when packet loss or ECN
2918 * is detected. But it does not reduce the cwnd, it is done by the
2919 * congestion control later.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002920 *
2921 * It does _not_ decide what to send, it is made in function
2922 * tcp_xmit_retransmit_queue().
2923 */
Yuchung Cheng737ff312017-11-08 13:01:27 -08002924static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una,
Eric Dumazet19119f22018-11-27 14:42:01 -08002925 int num_dupack, int *ack_flag, int *rexmit)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002926{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002927 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002928 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng31ba0c12016-02-02 10:33:05 -08002929 int fast_rexmit = 0, flag = *ack_flag;
Yousuk Seungc634e342020-06-26 21:05:33 -07002930 bool ece_ack = flag & FLAG_ECE;
Eric Dumazet19119f22018-11-27 14:42:01 -08002931 bool do_lost = num_dupack || ((flag & FLAG_DATA_SACKED) &&
2932 tcp_force_fast_retransmit(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002933
Eric Dumazet8ba6dda2017-10-05 22:21:25 -07002934 if (!tp->packets_out && tp->sacked_out)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002935 tp->sacked_out = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002936
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002937 /* Now state machine starts.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002938 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
Yousuk Seungc634e342020-06-26 21:05:33 -07002939 if (ece_ack)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002940 tp->prior_ssthresh = 0;
2941
2942 /* B. In all the states check for reneging SACKs. */
Ilpo Järvinencadbd032007-12-31 04:49:21 -08002943 if (tcp_check_sack_reneging(sk, flag))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002944 return;
2945
Yuchung Cheng974c1232012-01-19 14:42:21 +00002946 /* C. Check consistency of the current state. */
Ilpo Järvinen005903b2007-08-09 14:44:16 +03002947 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002948
Yuchung Cheng974c1232012-01-19 14:42:21 +00002949 /* D. Check state exit conditions. State can be terminated
Linus Torvalds1da177e2005-04-16 15:20:36 -07002950 * when high_seq is ACKed. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002951 if (icsk->icsk_ca_state == TCP_CA_Open) {
Eric Dumazeta7abf3c2021-03-11 12:35:05 -08002952 WARN_ON(tp->retrans_out != 0 && !tp->syn_data);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002953 tp->retrans_stamp = 0;
2954 } else if (!before(tp->snd_una, tp->high_seq)) {
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002955 switch (icsk->icsk_ca_state) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002956 case TCP_CA_CWR:
2957 /* CWR is to be held something *above* high_seq
2958 * is ACKed for CWR bit to reach receiver. */
2959 if (tp->snd_una != tp->high_seq) {
Yuchung Cheng684bad12012-09-02 17:38:04 +00002960 tcp_end_cwnd_reduction(sk);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002961 tcp_set_ca_state(sk, TCP_CA_Open);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002962 }
2963 break;
2964
Linus Torvalds1da177e2005-04-16 15:20:36 -07002965 case TCP_CA_Recovery:
Ilpo Järvinene60402d2007-08-09 15:14:46 +03002966 if (tcp_is_reno(tp))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002967 tcp_reset_reno_sack(tp);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002968 if (tcp_try_undo_recovery(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002969 return;
Yuchung Cheng684bad12012-09-02 17:38:04 +00002970 tcp_end_cwnd_reduction(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002971 break;
2972 }
2973 }
2974
Yuchung Cheng974c1232012-01-19 14:42:21 +00002975 /* E. Process state. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002976 switch (icsk->icsk_ca_state) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002977 case TCP_CA_Recovery:
Ilpo Järvinen2e605292007-08-02 19:46:58 -07002978 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
Eric Dumazet19119f22018-11-27 14:42:01 -08002979 if (tcp_is_reno(tp))
Yousuk Seungc634e342020-06-26 21:05:33 -07002980 tcp_add_reno_sack(sk, num_dupack, ece_ack);
Yuchung Chenga29cb692021-06-02 17:51:21 -07002981 } else if (tcp_try_undo_partial(sk, prior_snd_una, &do_lost))
Yuchung Chengc7d9d6a2013-05-29 14:20:14 +00002982 return;
Yuchung Chenga29cb692021-06-02 17:51:21 -07002983
2984 if (tcp_try_undo_dsack(sk))
2985 tcp_try_keep_open(sk);
2986
Yuchung Cheng6ac06ec2018-05-16 16:40:12 -07002987 tcp_identify_packet_loss(sk, ack_flag);
Yuchung Chenga29cb692021-06-02 17:51:21 -07002988 if (icsk->icsk_ca_state != TCP_CA_Recovery) {
2989 if (!tcp_time_to_recover(sk, flag))
2990 return;
2991 /* Undo reverts the recovery state. If loss is evident,
2992 * starts a new recovery (e.g. reordering then loss);
2993 */
2994 tcp_enter_recovery(sk, ece_ack);
2995 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002996 break;
2997 case TCP_CA_Loss:
Eric Dumazet19119f22018-11-27 14:42:01 -08002998 tcp_process_loss(sk, flag, num_dupack, rexmit);
Yuchung Cheng6ac06ec2018-05-16 16:40:12 -07002999 tcp_identify_packet_loss(sk, ack_flag);
Yuchung Cheng98e36d42017-01-12 22:11:35 -08003000 if (!(icsk->icsk_ca_state == TCP_CA_Open ||
3001 (*ack_flag & FLAG_LOST_RETRANS)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003002 return;
Yuchung Cheng291a00d2015-07-01 14:11:14 -07003003 /* Change state if cwnd is undone or retransmits are lost */
Joe Perchesa8eceea2020-03-12 15:50:22 -07003004 fallthrough;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003005 default:
Ilpo Järvinene60402d2007-08-09 15:14:46 +03003006 if (tcp_is_reno(tp)) {
Ilpo Järvinen2e605292007-08-02 19:46:58 -07003007 if (flag & FLAG_SND_UNA_ADVANCED)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003008 tcp_reset_reno_sack(tp);
Yousuk Seungc634e342020-06-26 21:05:33 -07003009 tcp_add_reno_sack(sk, num_dupack, ece_ack);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003010 }
3011
Neal Cardwellf6982042011-11-16 08:58:04 +00003012 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003013 tcp_try_undo_dsack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003014
Yuchung Cheng6ac06ec2018-05-16 16:40:12 -07003015 tcp_identify_packet_loss(sk, ack_flag);
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00003016 if (!tcp_time_to_recover(sk, flag)) {
Yuchung Cheng31ba0c12016-02-02 10:33:05 -08003017 tcp_try_to_open(sk, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003018 return;
3019 }
3020
John Heffner5d424d52006-03-20 17:53:41 -08003021 /* MTU probe failure: don't reduce cwnd */
3022 if (icsk->icsk_ca_state < TCP_CA_CWR &&
3023 icsk->icsk_mtup.probe_size &&
John Heffner0e7b13682006-03-20 21:32:58 -08003024 tp->snd_una == tp->mtu_probe.probe_seq_start) {
John Heffner5d424d52006-03-20 17:53:41 -08003025 tcp_mtup_probe_failed(sk);
3026 /* Restores the reduction we did in tcp_mtup_probe() */
3027 tp->snd_cwnd++;
3028 tcp_simple_retransmit(sk);
3029 return;
3030 }
3031
Linus Torvalds1da177e2005-04-16 15:20:36 -07003032 /* Otherwise enter Recovery state */
Yousuk Seungc634e342020-06-26 21:05:33 -07003033 tcp_enter_recovery(sk, ece_ack);
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08003034 fast_rexmit = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003035 }
3036
Yuchung Chengb38a51f2018-05-16 16:40:11 -07003037 if (!tcp_is_rack(sk) && do_lost)
Ilpo Järvinen85cc3912007-11-15 19:39:31 -08003038 tcp_update_scoreboard(sk, fast_rexmit);
Yuchung Chenge662ca42016-02-02 10:33:04 -08003039 *rexmit = REXMIT_LOST;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003040}
3041
Yuchung Chengeb36be02018-01-17 12:11:00 -08003042static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag)
Yuchung Chengf6722582015-10-16 21:57:42 -07003043{
Eric Dumazetbd239702017-10-27 07:47:28 -07003044 u32 wlen = sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen * HZ;
Neal Cardwell64033892016-09-19 23:39:10 -04003045 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Chengf6722582015-10-16 21:57:42 -07003046
Yuchung Chengeb36be02018-01-17 12:11:00 -08003047 if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) {
3048 /* If the remote keeps returning delayed ACKs, eventually
3049 * the min filter would pick it up and overestimate the
3050 * prop. delay when it expires. Skip suspected delayed ACKs.
3051 */
3052 return;
3053 }
Eric Dumazetac9517f2017-05-16 14:00:13 -07003054 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32,
Neal Cardwell64033892016-09-19 23:39:10 -04003055 rtt_us ? : jiffies_to_usecs(1));
Yuchung Chengf6722582015-10-16 21:57:42 -07003056}
3057
Yuchung Cheng775e68a2017-05-31 11:30:53 -07003058static bool tcp_ack_update_rtt(struct sock *sk, const int flag,
3059 long seq_rtt_us, long sack_rtt_us,
3060 long ca_rtt_us, struct rate_sample *rs)
Ilpo Järvinen41834b72008-12-05 22:43:26 -08003061{
Yuchung Cheng5b08e472013-07-22 16:20:46 -07003062 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen41834b72008-12-05 22:43:26 -08003063
Yuchung Cheng5b08e472013-07-22 16:20:46 -07003064 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
3065 * broken middle-boxes or peers may corrupt TS-ECR fields. But
3066 * Karn's algorithm forbids taking RTT if some retransmitted data
3067 * is acked (RFC6298).
3068 */
Eric Dumazet740b0f12014-02-26 14:02:48 -08003069 if (seq_rtt_us < 0)
3070 seq_rtt_us = sack_rtt_us;
Yuchung Chenged084952013-07-22 16:20:48 -07003071
Linus Torvalds1da177e2005-04-16 15:20:36 -07003072 /* RTTM Rule: A TSecr value received in a segment is used to
3073 * update the averaged RTT measurement only if the segment
3074 * acknowledges some new data, i.e., only if it advances the
3075 * left edge of the send window.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003076 * See draft-ietf-tcplw-high-performance-00, section 3.3.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003077 */
Eric Dumazet740b0f12014-02-26 14:02:48 -08003078 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
Eric Dumazet9a568de2017-05-16 14:00:14 -07003079 flag & FLAG_ACKED) {
3080 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
Eric Dumazet9a568de2017-05-16 14:00:14 -07003081
Eric Dumazet9efdda42018-11-24 09:12:24 -08003082 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
Jianfeng Wang730e7002020-07-30 23:49:16 +00003083 if (!delta)
3084 delta = 1;
Eric Dumazet9efdda42018-11-24 09:12:24 -08003085 seq_rtt_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
3086 ca_rtt_us = seq_rtt_us;
3087 }
Eric Dumazet9a568de2017-05-16 14:00:14 -07003088 }
Yuchung Cheng775e68a2017-05-31 11:30:53 -07003089 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */
Eric Dumazet740b0f12014-02-26 14:02:48 -08003090 if (seq_rtt_us < 0)
Yuchung Chenged084952013-07-22 16:20:48 -07003091 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003092
Yuchung Chengf6722582015-10-16 21:57:42 -07003093 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
3094 * always taken together with ACK, SACK, or TS-opts. Any negative
3095 * values will be skipped with the seq_rtt_us < 0 check above.
3096 */
Yuchung Chengeb36be02018-01-17 12:11:00 -08003097 tcp_update_rtt_min(sk, ca_rtt_us, flag);
Eric Dumazet740b0f12014-02-26 14:02:48 -08003098 tcp_rtt_estimator(sk, seq_rtt_us);
Yuchung Cheng5b08e472013-07-22 16:20:46 -07003099 tcp_set_rto(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003100
Yuchung Cheng5b08e472013-07-22 16:20:46 -07003101 /* RFC6298: only reset backoff on valid RTT measurement. */
3102 inet_csk(sk)->icsk_backoff = 0;
Yuchung Chenged084952013-07-22 16:20:48 -07003103 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003104}
3105
Yuchung Cheng375fe022013-07-22 16:20:45 -07003106/* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
Yuchung Cheng0f1c28a2015-09-18 11:36:14 -07003107void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
Yuchung Cheng375fe022013-07-22 16:20:45 -07003108{
Yuchung Cheng775e68a2017-05-31 11:30:53 -07003109 struct rate_sample rs;
Yuchung Cheng0f1c28a2015-09-18 11:36:14 -07003110 long rtt_us = -1L;
Yuchung Cheng375fe022013-07-22 16:20:45 -07003111
Eric Dumazet9a568de2017-05-16 14:00:14 -07003112 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack)
3113 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack);
Yuchung Cheng0f1c28a2015-09-18 11:36:14 -07003114
Yuchung Cheng775e68a2017-05-31 11:30:53 -07003115 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs);
Yuchung Cheng375fe022013-07-22 16:20:45 -07003116}
3117
Yuchung Cheng0f1c28a2015-09-18 11:36:14 -07003118
Eric Dumazet24901552014-05-02 21:18:05 -07003119static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003120{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003121 const struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazet24901552014-05-02 21:18:05 -07003122
3123 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
Eric Dumazetc2203cf2017-05-16 14:00:04 -07003124 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003125}
3126
Linus Torvalds1da177e2005-04-16 15:20:36 -07003127/* Restart timer after forward progress on connection.
3128 * RFC2988 recommends to restart timer to now+rto.
3129 */
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00003130void tcp_rearm_rto(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003131{
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00003132 const struct inet_connection_sock *icsk = inet_csk(sk);
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00003133 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003134
Jerry Chu168a8f52012-08-31 12:29:13 +00003135 /* If the retrans timer is currently being used by Fast Open
3136 * for SYN-ACK retrans purpose, stay put.
3137 */
Eric Dumazetd983ea62019-10-10 20:17:38 -07003138 if (rcu_access_pointer(tp->fastopen_rsk))
Jerry Chu168a8f52012-08-31 12:29:13 +00003139 return;
3140
Linus Torvalds1da177e2005-04-16 15:20:36 -07003141 if (!tp->packets_out) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003142 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003143 } else {
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00003144 u32 rto = inet_csk(sk)->icsk_rto;
3145 /* Offset the time elapsed after installing regular RTO */
Yuchung Chengbec41a12017-01-12 22:11:39 -08003146 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00003147 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
Neal Cardwelle1a10ef2017-08-03 09:19:52 -04003148 s64 delta_us = tcp_rto_delta_us(sk);
Eric Dumazetb17b8a22017-05-18 09:15:58 -07003149 /* delta_us may not be positive if the socket is locked
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00003150 * when the retrans timer fires and is rescheduled.
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00003151 */
Neal Cardwellcdbeb632017-08-16 17:53:36 -04003152 rto = usecs_to_jiffies(max_t(int, delta_us, 1));
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00003153 }
Eric Dumazet3f80e082018-10-23 11:54:16 -07003154 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
Eric Dumazet8dc242ad2020-05-04 11:27:49 -07003155 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003156 }
Yuchung Cheng750ea2b2012-05-02 13:30:04 +00003157}
3158
Neal Cardwelldf92c832017-08-03 09:19:54 -04003159/* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
3160static void tcp_set_xmit_timer(struct sock *sk)
3161{
Neal Cardwelled66dfa2017-11-17 21:06:14 -05003162 if (!tcp_schedule_loss_probe(sk, true))
Neal Cardwelldf92c832017-08-03 09:19:54 -04003163 tcp_rearm_rto(sk);
3164}
3165
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003166/* If we get here, the whole TSO packet has not been acked. */
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003167static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003168{
3169 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003170 u32 packets_acked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003171
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003172 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003173
3174 packets_acked = tcp_skb_pcount(skb);
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003175 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003176 return 0;
3177 packets_acked -= tcp_skb_pcount(skb);
3178
3179 if (packets_acked) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003180 BUG_ON(tcp_skb_pcount(skb) == 0);
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003181 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003182 }
3183
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003184 return packets_acked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003185}
3186
Eric Dumazetad971f62014-10-11 15:17:29 -07003187static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
Yousuk Seunge7ed11e2021-01-20 12:41:55 -08003188 const struct sk_buff *ack_skb, u32 prior_snd_una)
Eric Dumazetad971f62014-10-11 15:17:29 -07003189{
3190 const struct skb_shared_info *shinfo;
3191
3192 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
Soheil Hassas Yeganeh6b084922016-04-02 23:08:08 -04003193 if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
Eric Dumazetad971f62014-10-11 15:17:29 -07003194 return;
3195
3196 shinfo = skb_shinfo(skb);
Soheil Hassas Yeganeh0a2cf202016-04-27 23:39:01 -04003197 if (!before(shinfo->tskey, prior_snd_una) &&
Eric Dumazete2080072017-10-04 12:59:58 -07003198 before(shinfo->tskey, tcp_sk(sk)->snd_una)) {
3199 tcp_skb_tsorted_save(skb) {
Yousuk Seunge7ed11e2021-01-20 12:41:55 -08003200 __skb_tstamp_tx(skb, ack_skb, NULL, sk, SCM_TSTAMP_ACK);
Eric Dumazete2080072017-10-04 12:59:58 -07003201 } tcp_skb_tsorted_restore(skb);
3202 }
Eric Dumazetad971f62014-10-11 15:17:29 -07003203}
3204
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003205/* Remove acknowledged frames from the retransmission queue. If our packet
3206 * is before the ack sequence we can discard it as it's confirmed to have
3207 * arrived at the other end.
3208 */
Yousuk Seunge7ed11e2021-01-20 12:41:55 -08003209static int tcp_clean_rtx_queue(struct sock *sk, const struct sk_buff *ack_skb,
3210 u32 prior_fack, u32 prior_snd_una,
Yousuk Seungc634e342020-06-26 21:05:33 -07003211 struct tcp_sacktag_state *sack, bool ece_ack)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003212{
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08003213 const struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazet9a568de2017-05-16 14:00:14 -07003214 u64 first_ackt, last_ackt;
Eric Dumazet740b0f12014-02-26 14:02:48 -08003215 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen90638a02008-09-20 21:25:52 -07003216 u32 prior_sacked = tp->sacked_out;
Yuchung Cheng737ff312017-11-08 13:01:27 -08003217 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */
Eric Dumazet75c119a2017-10-05 22:21:27 -07003218 struct sk_buff *skb, *next;
Eric Dumazet740b0f12014-02-26 14:02:48 -08003219 bool fully_acked = true;
Kenneth Klette Jonassen31231a82015-05-01 01:10:58 +02003220 long sack_rtt_us = -1L;
Eric Dumazet740b0f12014-02-26 14:02:48 -08003221 long seq_rtt_us = -1L;
Kenneth Klette Jonassen31231a82015-05-01 01:10:58 +02003222 long ca_rtt_us = -1L;
Eric Dumazet740b0f12014-02-26 14:02:48 -08003223 u32 pkts_acked = 0;
Lawrence Brakmo6f094b92016-06-08 21:16:44 -07003224 u32 last_in_flight = 0;
Yuchung Cheng2f715c12013-10-24 08:59:27 -07003225 bool rtt_update;
Eric Dumazet740b0f12014-02-26 14:02:48 -08003226 int flag = 0;
3227
Eric Dumazet9a568de2017-05-16 14:00:14 -07003228 first_ackt = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003229
Eric Dumazet75c119a2017-10-05 22:21:27 -07003230 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) {
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09003231 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
Yuchung Cheng737ff312017-11-08 13:01:27 -08003232 const u32 start_seq = scb->seq;
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003233 u8 sacked = scb->sacked;
Eric Dumazet740b0f12014-02-26 14:02:48 -08003234 u32 acked_pcount;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003235
Gavin McCullagh2072c222007-12-29 19:11:21 -08003236 /* Determine how many packets and what bytes were acked, tso and else */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003237 if (after(scb->end_seq, tp->snd_una)) {
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003238 if (tcp_skb_pcount(skb) == 1 ||
3239 !after(tp->snd_una, scb->seq))
3240 break;
3241
Ilpo Järvinen72018832007-12-30 04:37:55 -08003242 acked_pcount = tcp_tso_acked(sk, skb);
3243 if (!acked_pcount)
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003244 break;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00003245 fully_acked = false;
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003246 } else {
Ilpo Järvinen72018832007-12-30 04:37:55 -08003247 acked_pcount = tcp_skb_pcount(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003248 }
3249
Eric Dumazetad971f62014-10-11 15:17:29 -07003250 if (unlikely(sacked & TCPCB_RETRANS)) {
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003251 if (sacked & TCPCB_SACKED_RETRANS)
Ilpo Järvinen72018832007-12-30 04:37:55 -08003252 tp->retrans_out -= acked_pcount;
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003253 flag |= FLAG_RETRANS_DATA_ACKED;
Kenneth Klette Jonassen3d0d26c2015-04-11 02:17:49 +02003254 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
Eric Dumazet2fd66ff2018-09-21 08:51:47 -07003255 last_ackt = tcp_skb_timestamp_us(skb);
Eric Dumazet9a568de2017-05-16 14:00:14 -07003256 WARN_ON_ONCE(last_ackt == 0);
3257 if (!first_ackt)
Eric Dumazet740b0f12014-02-26 14:02:48 -08003258 first_ackt = last_ackt;
3259
Lawrence Brakmo6f094b92016-06-08 21:16:44 -07003260 last_in_flight = TCP_SKB_CB(skb)->tx.in_flight;
Yuchung Cheng737ff312017-11-08 13:01:27 -08003261 if (before(start_seq, reord))
3262 reord = start_seq;
Kenneth Klette Jonassen3d0d26c2015-04-11 02:17:49 +02003263 if (!after(scb->end_seq, tp->high_seq))
3264 flag |= FLAG_ORIG_SACK_ACKED;
Stephen Hemminger2d2abba2005-11-10 16:56:12 -08003265 }
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003266
Yuchung Chengddf1af62016-02-02 10:33:06 -08003267 if (sacked & TCPCB_SACKED_ACKED) {
Ilpo Järvinen72018832007-12-30 04:37:55 -08003268 tp->sacked_out -= acked_pcount;
Yuchung Chengddf1af62016-02-02 10:33:06 -08003269 } else if (tcp_is_sack(tp)) {
Yousuk Seung082d4fa2020-06-26 21:05:35 -07003270 tcp_count_delivered(tp, acked_pcount, ece_ack);
Yuchung Chengddf1af62016-02-02 10:33:06 -08003271 if (!tcp_skb_spurious_retrans(tp, skb))
Yuchung Cheng1d0833d2017-01-12 22:11:34 -08003272 tcp_rack_advance(tp, sacked, scb->end_seq,
Eric Dumazet2fd66ff2018-09-21 08:51:47 -07003273 tcp_skb_timestamp_us(skb));
Yuchung Chengddf1af62016-02-02 10:33:06 -08003274 }
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003275 if (sacked & TCPCB_LOST)
Ilpo Järvinen72018832007-12-30 04:37:55 -08003276 tp->lost_out -= acked_pcount;
Ilpo Järvinen89d478f2007-12-30 04:35:27 -08003277
Ilpo Järvinen72018832007-12-30 04:37:55 -08003278 tp->packets_out -= acked_pcount;
3279 pkts_acked += acked_pcount;
Yuchung Chengb9f64822016-09-19 23:39:14 -04003280 tcp_rate_skb_delivered(sk, skb, sack->rate);
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003281
Ilpo Järvinen009a2e32007-09-20 11:34:38 -07003282 /* Initial outgoing SYN's get put onto the write_queue
3283 * just like anything else we transmit. It is not
3284 * true data, and if we misinform our callers that
3285 * this ACK acks real data, we will erroneously exit
3286 * connection startup slow start one packet too
3287 * quickly. This is severely frowned upon behavior.
3288 */
Eric Dumazetad971f62014-10-11 15:17:29 -07003289 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
Ilpo Järvinen009a2e32007-09-20 11:34:38 -07003290 flag |= FLAG_DATA_ACKED;
3291 } else {
3292 flag |= FLAG_SYN_ACKED;
3293 tp->retrans_stamp = 0;
3294 }
3295
Ilpo Järvinen13fcf852007-10-09 01:28:45 -07003296 if (!fully_acked)
3297 break;
3298
Yousuk Seunge7ed11e2021-01-20 12:41:55 -08003299 tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una);
Yousuk Seungfdb7eb22020-06-26 21:05:32 -07003300
Eric Dumazet75c119a2017-10-05 22:21:27 -07003301 next = skb_rb_next(skb);
Eric Dumazetad971f62014-10-11 15:17:29 -07003302 if (unlikely(skb == tp->retransmit_skb_hint))
Ilpo Järvinenef9da472008-09-20 21:25:15 -07003303 tp->retransmit_skb_hint = NULL;
Eric Dumazetad971f62014-10-11 15:17:29 -07003304 if (unlikely(skb == tp->lost_skb_hint))
Ilpo Järvinen90638a02008-09-20 21:25:52 -07003305 tp->lost_skb_hint = NULL;
Eric Dumazet2bec4452020-01-22 21:03:00 -08003306 tcp_highest_sack_replace(sk, skb, next);
Eric Dumazet75c119a2017-10-05 22:21:27 -07003307 tcp_rtx_queue_unlink_and_free(skb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003308 }
3309
Francis Yan0f872302016-11-27 23:07:14 -08003310 if (!skb)
3311 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3312
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07003313 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3314 tp->snd_up = tp->snd_una;
3315
Yousuk Seungff91e922020-06-30 09:49:33 -07003316 if (skb) {
Yousuk Seunge7ed11e2021-01-20 12:41:55 -08003317 tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una);
Yousuk Seungff91e922020-06-30 09:49:33 -07003318 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
3319 flag |= FLAG_SACK_RENEGING;
3320 }
Ilpo Järvinencadbd032007-12-31 04:49:21 -08003321
Eric Dumazet9a568de2017-05-16 14:00:14 -07003322 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3323 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt);
3324 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt);
Yuchung Chengeb36be02018-01-17 12:11:00 -08003325
3326 if (pkts_acked == 1 && last_in_flight < tp->mss_cache &&
3327 last_in_flight && !prior_sacked && fully_acked &&
3328 sack->rate->prior_delivered + 1 == tp->delivered &&
3329 !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) {
3330 /* Conservatively mark a delayed ACK. It's typically
3331 * from a lone runt packet over the round trip to
3332 * a receiver w/o out-of-order or CE events.
3333 */
3334 flag |= FLAG_ACK_MAYBE_DELAYED;
3335 }
Kenneth Klette Jonassen31231a82015-05-01 01:10:58 +02003336 }
Eric Dumazet9a568de2017-05-16 14:00:14 -07003337 if (sack->first_sackt) {
3338 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt);
3339 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt);
Eric Dumazet740b0f12014-02-26 14:02:48 -08003340 }
Yuchung Chengf6722582015-10-16 21:57:42 -07003341 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
Yuchung Cheng775e68a2017-05-31 11:30:53 -07003342 ca_rtt_us, sack->rate);
Yuchung Chenged084952013-07-22 16:20:48 -07003343
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003344 if (flag & FLAG_ACKED) {
Neal Cardwelldf92c832017-08-03 09:19:54 -04003345 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
Ilpo Järvinen72211e92009-03-14 14:23:04 +00003346 if (unlikely(icsk->icsk_mtup.probe_size &&
3347 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3348 tcp_mtup_probe_success(sk);
3349 }
3350
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003351 if (tcp_is_reno(tp)) {
Yousuk Seungc634e342020-06-26 21:05:33 -07003352 tcp_remove_reno_sacks(sk, pkts_acked, ece_ack);
Ilpo Järvinen1236f222018-06-29 13:07:53 +03003353
3354 /* If any of the cumulatively ACKed segments was
3355 * retransmitted, non-SACK case cannot confirm that
3356 * progress was due to original transmission due to
3357 * lack of TCPCB_SACKED_ACKED bits even if some of
3358 * the packets may have been never retransmitted.
3359 */
3360 if (flag & FLAG_RETRANS_DATA_ACKED)
3361 flag &= ~FLAG_ORIG_SACK_ACKED;
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003362 } else {
Ilpo Järvinen59a08cb2009-02-28 04:44:28 +00003363 int delta;
3364
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003365 /* Non-retransmitted hole got filled? That's reordering */
Yuchung Cheng737ff312017-11-08 13:01:27 -08003366 if (before(reord, prior_fack))
3367 tcp_check_sack_reordering(sk, reord, 0);
Ilpo Järvinen90638a02008-09-20 21:25:52 -07003368
Yuchung Cheng713bafe2017-11-08 13:01:26 -08003369 delta = prior_sacked - tp->sacked_out;
Ilpo Järvinen59a08cb2009-02-28 04:44:28 +00003370 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003371 }
Eric Dumazet740b0f12014-02-26 14:02:48 -08003372 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
Eric Dumazet2fd66ff2018-09-21 08:51:47 -07003373 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp,
3374 tcp_skb_timestamp_us(skb))) {
Yuchung Cheng2f715c12013-10-24 08:59:27 -07003375 /* Do not re-arm RTO if the sack RTT is measured from data sent
3376 * after when the head was last (re)transmitted. Otherwise the
3377 * timeout may continue to extend in loss recovery.
3378 */
Neal Cardwelldf92c832017-08-03 09:19:54 -04003379 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003380 }
3381
Lawrence Brakmo756ee172016-05-11 10:02:13 -07003382 if (icsk->icsk_ca_ops->pkts_acked) {
3383 struct ack_sample sample = { .pkts_acked = pkts_acked,
Yuchung Cheng775e68a2017-05-31 11:30:53 -07003384 .rtt_us = sack->rate->rtt_us,
Lawrence Brakmo6f094b92016-06-08 21:16:44 -07003385 .in_flight = last_in_flight };
Lawrence Brakmo756ee172016-05-11 10:02:13 -07003386
3387 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3388 }
Kenneth Klette Jonassen138998f2015-05-01 01:10:59 +02003389
Linus Torvalds1da177e2005-04-16 15:20:36 -07003390#if FASTRETRANS_DEBUG > 0
Ilpo Järvinen547b7922008-07-25 21:43:18 -07003391 WARN_ON((int)tp->sacked_out < 0);
3392 WARN_ON((int)tp->lost_out < 0);
3393 WARN_ON((int)tp->retrans_out < 0);
Ilpo Järvinene60402d2007-08-09 15:14:46 +03003394 if (!tp->packets_out && tcp_is_sack(tp)) {
Stephen Hemmingercfcabdc2007-10-09 01:59:42 -07003395 icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003396 if (tp->lost_out) {
Joe Perches91df42b2012-05-15 14:11:54 +00003397 pr_debug("Leak l=%u %d\n",
3398 tp->lost_out, icsk->icsk_ca_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003399 tp->lost_out = 0;
3400 }
3401 if (tp->sacked_out) {
Joe Perches91df42b2012-05-15 14:11:54 +00003402 pr_debug("Leak s=%u %d\n",
3403 tp->sacked_out, icsk->icsk_ca_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003404 tp->sacked_out = 0;
3405 }
3406 if (tp->retrans_out) {
Joe Perches91df42b2012-05-15 14:11:54 +00003407 pr_debug("Leak r=%u %d\n",
3408 tp->retrans_out, icsk->icsk_ca_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003409 tp->retrans_out = 0;
3410 }
3411 }
3412#endif
Ilpo Järvinen7c46a032007-09-20 11:33:43 -07003413 return flag;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003414}
3415
3416static void tcp_ack_probe(struct sock *sk)
3417{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003418 struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazet75c119a2017-10-05 22:21:27 -07003419 struct sk_buff *head = tcp_send_head(sk);
3420 const struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003421
3422 /* Was it a usable window open? */
Eric Dumazet75c119a2017-10-05 22:21:27 -07003423 if (!head)
3424 return;
3425 if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003426 icsk->icsk_backoff = 0;
Enke Chen9d9b1ee2021-01-15 14:30:58 -08003427 icsk->icsk_probes_tstamp = 0;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003428 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003429 /* Socket must be waked up by subsequent tcp_data_snd_check().
3430 * This function is not for random using!
3431 */
3432 } else {
Eric Dumazet21c8fe92015-05-06 14:26:24 -07003433 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
Eric Dumazetfcdd1cf2014-09-22 13:19:44 -07003434
Enke Chen344db932021-01-22 11:13:06 -08003435 when = tcp_clamp_probe0_to_user_timeout(sk, when);
3436 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, when, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003437 }
3438}
3439
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00003440static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003441{
Eric Dumazeta02cec22010-09-22 20:43:57 +00003442 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3443 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003444}
3445
Yuchung Cheng0f7cc9a2013-08-21 17:29:23 -07003446/* Decide wheather to run the increase function of congestion control. */
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00003447static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003448{
Yuchung Cheng0f7cc9a2013-08-21 17:29:23 -07003449 /* If reordering is high then always grow cwnd whenever data is
3450 * delivered regardless of its ordering. Otherwise stay conservative
Yuchung Cheng16edfe72013-09-05 15:36:09 -07003451 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
Yuchung Cheng0f7cc9a2013-08-21 17:29:23 -07003452 * new SACK or ECE mark may first advance cwnd here and later reduce
3453 * cwnd in tcp_fastretrans_alert() based on more states.
3454 */
Nikolay Borisov1043e252016-02-03 09:46:52 +02003455 if (tcp_sk(sk)->reordering > sock_net(sk)->ipv4.sysctl_tcp_reordering)
Yuchung Cheng0f7cc9a2013-08-21 17:29:23 -07003456 return flag & FLAG_FORWARD_PROGRESS;
3457
Yuchung Cheng16edfe72013-09-05 15:36:09 -07003458 return flag & FLAG_DATA_ACKED;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003459}
3460
Yuchung Chengd452e6c2016-02-02 10:33:09 -08003461/* The "ultimate" congestion control function that aims to replace the rigid
3462 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3463 * It's called toward the end of processing an ACK with precise rate
3464 * information. All transmission or retransmission are delayed afterwards.
3465 */
3466static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
Yuchung Chengc0402762016-09-19 23:39:21 -04003467 int flag, const struct rate_sample *rs)
Yuchung Chengd452e6c2016-02-02 10:33:09 -08003468{
Yuchung Chengc0402762016-09-19 23:39:21 -04003469 const struct inet_connection_sock *icsk = inet_csk(sk);
3470
3471 if (icsk->icsk_ca_ops->cong_control) {
3472 icsk->icsk_ca_ops->cong_control(sk, rs);
3473 return;
3474 }
3475
Yuchung Chengd452e6c2016-02-02 10:33:09 -08003476 if (tcp_in_cwnd_reduction(sk)) {
3477 /* Reduce cwnd if state mandates */
Yuchung Cheng7e901ee2020-10-30 18:34:12 -07003478 tcp_cwnd_reduction(sk, acked_sacked, rs->losses, flag);
Yuchung Chengd452e6c2016-02-02 10:33:09 -08003479 } else if (tcp_may_raise_cwnd(sk, flag)) {
3480 /* Advance cwnd if state allows */
3481 tcp_cong_avoid(sk, ack, acked_sacked);
3482 }
3483 tcp_update_pacing_rate(sk);
3484}
3485
Linus Torvalds1da177e2005-04-16 15:20:36 -07003486/* Check that window update is acceptable.
3487 * The function assumes that snd_una<=ack<=snd_next.
3488 */
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00003489static inline bool tcp_may_update_window(const struct tcp_sock *tp,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003490 const u32 ack, const u32 ack_seq,
3491 const u32 nwin)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003492{
Eric Dumazeta02cec22010-09-22 20:43:57 +00003493 return after(ack, tp->snd_una) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07003494 after(ack_seq, tp->snd_wl1) ||
Eric Dumazeta02cec22010-09-22 20:43:57 +00003495 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003496}
3497
Eric Dumazet0df48c22015-04-28 15:28:17 -07003498/* If we update tp->snd_una, also update tp->bytes_acked */
3499static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3500{
3501 u32 delta = ack - tp->snd_una;
3502
Eric Dumazet46cc6e42016-05-03 16:56:03 -07003503 sock_owned_by_me((struct sock *)tp);
Eric Dumazet0df48c22015-04-28 15:28:17 -07003504 tp->bytes_acked += delta;
3505 tp->snd_una = ack;
3506}
3507
Eric Dumazetbdd1f9e2015-04-28 15:28:18 -07003508/* If we update tp->rcv_nxt, also update tp->bytes_received */
3509static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3510{
3511 u32 delta = seq - tp->rcv_nxt;
3512
Eric Dumazet46cc6e42016-05-03 16:56:03 -07003513 sock_owned_by_me((struct sock *)tp);
Eric Dumazetbdd1f9e2015-04-28 15:28:18 -07003514 tp->bytes_received += delta;
Eric Dumazetdba7d9b2019-10-10 20:17:39 -07003515 WRITE_ONCE(tp->rcv_nxt, seq);
Eric Dumazetbdd1f9e2015-04-28 15:28:18 -07003516}
3517
Linus Torvalds1da177e2005-04-16 15:20:36 -07003518/* Update our send window.
3519 *
3520 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3521 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3522 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003523static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003524 u32 ack_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003525{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07003526 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003527 int flag = 0;
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07003528 u32 nwin = ntohs(tcp_hdr(skb)->window);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003529
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07003530 if (likely(!tcp_hdr(skb)->syn))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003531 nwin <<= tp->rx_opt.snd_wscale;
3532
3533 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3534 flag |= FLAG_WIN_UPDATE;
Hantzis Fotisee7537b2009-03-02 22:42:02 -08003535 tcp_update_wl(tp, ack_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003536
3537 if (tp->snd_wnd != nwin) {
3538 tp->snd_wnd = nwin;
3539
Florian Westphal31770e32017-08-30 19:24:58 +02003540 /* Note, it is the only place, where
3541 * fast path is recovered for sending TCP.
3542 */
3543 tp->pred_flags = 0;
3544 tcp_fast_path_check(sk);
3545
Eric Dumazet75c119a2017-10-05 22:21:27 -07003546 if (!tcp_write_queue_empty(sk))
Eric Dumazet6f021c62015-08-21 12:30:00 -07003547 tcp_slow_start_after_idle_check(sk);
3548
Linus Torvalds1da177e2005-04-16 15:20:36 -07003549 if (nwin > tp->max_window) {
3550 tp->max_window = nwin;
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08003551 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003552 }
3553 }
3554 }
3555
Eric Dumazet0df48c22015-04-28 15:28:17 -07003556 tcp_snd_una_update(tp, ack);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003557
3558 return flag;
3559}
3560
Jason Baron083ae302016-07-14 11:38:40 -04003561static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3562 u32 *last_oow_ack_time)
3563{
3564 if (*last_oow_ack_time) {
Eric Dumazet594208a2017-05-16 14:00:10 -07003565 s32 elapsed = (s32)(tcp_jiffies32 - *last_oow_ack_time);
Jason Baron083ae302016-07-14 11:38:40 -04003566
Eric Dumazet4170ba62017-10-27 07:47:30 -07003567 if (0 <= elapsed && elapsed < net->ipv4.sysctl_tcp_invalid_ratelimit) {
Jason Baron083ae302016-07-14 11:38:40 -04003568 NET_INC_STATS(net, mib_idx);
3569 return true; /* rate-limited: don't send yet! */
3570 }
3571 }
3572
Eric Dumazet594208a2017-05-16 14:00:10 -07003573 *last_oow_ack_time = tcp_jiffies32;
Jason Baron083ae302016-07-14 11:38:40 -04003574
3575 return false; /* not rate-limited: go ahead, send dupack now! */
3576}
3577
Eric Dumazet7970ddc2015-03-16 21:06:20 -07003578/* Return true if we're currently rate-limiting out-of-window ACKs and
3579 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3580 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3581 * attacks that send repeated SYNs or ACKs for the same connection. To
3582 * do this, we do not send a duplicate SYNACK or ACK if the remote
3583 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3584 */
3585bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3586 int mib_idx, u32 *last_oow_ack_time)
3587{
3588 /* Data packets without SYNs are not likely part of an ACK loop. */
3589 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3590 !tcp_hdr(skb)->syn)
Jason Baron083ae302016-07-14 11:38:40 -04003591 return false;
Eric Dumazet7970ddc2015-03-16 21:06:20 -07003592
Jason Baron083ae302016-07-14 11:38:40 -04003593 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
Eric Dumazet7970ddc2015-03-16 21:06:20 -07003594}
3595
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003596/* RFC 5961 7 [ACK Throttling] */
Neal Cardwellf2b2c582015-02-06 16:04:40 -05003597static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003598{
3599 /* unprotected vars, we dont care of overwrites */
3600 static u32 challenge_timestamp;
3601 static unsigned int challenge_count;
Neal Cardwellf2b2c582015-02-06 16:04:40 -05003602 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazetb530b682017-10-27 07:47:26 -07003603 struct net *net = sock_net(sk);
Eric Dumazet75ff39c2016-07-10 10:04:02 +02003604 u32 count, now;
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003605
Neal Cardwellf2b2c582015-02-06 16:04:40 -05003606 /* First check our per-socket dupack rate limit. */
Eric Dumazetb530b682017-10-27 07:47:26 -07003607 if (__tcp_oow_rate_limited(net,
Jason Baron083ae302016-07-14 11:38:40 -04003608 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3609 &tp->last_oow_ack_time))
Neal Cardwellf2b2c582015-02-06 16:04:40 -05003610 return;
3611
Eric Dumazet75ff39c2016-07-10 10:04:02 +02003612 /* Then check host-wide RFC 5961 rate limit. */
Neal Cardwellf2b2c582015-02-06 16:04:40 -05003613 now = jiffies / HZ;
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003614 if (now != challenge_timestamp) {
Eric Dumazetb530b682017-10-27 07:47:26 -07003615 u32 ack_limit = net->ipv4.sysctl_tcp_challenge_ack_limit;
3616 u32 half = (ack_limit + 1) >> 1;
Eric Dumazet75ff39c2016-07-10 10:04:02 +02003617
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003618 challenge_timestamp = now;
Eric Dumazetb530b682017-10-27 07:47:26 -07003619 WRITE_ONCE(challenge_count, half + prandom_u32_max(ack_limit));
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003620 }
Eric Dumazet75ff39c2016-07-10 10:04:02 +02003621 count = READ_ONCE(challenge_count);
3622 if (count > 0) {
3623 WRITE_ONCE(challenge_count, count - 1);
Eric Dumazetb530b682017-10-27 07:47:26 -07003624 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK);
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003625 tcp_send_ack(sk);
3626 }
3627}
3628
Eric Dumazet12fb3dd2013-04-19 07:19:48 +00003629static void tcp_store_ts_recent(struct tcp_sock *tp)
3630{
3631 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
Arnd Bergmanncca9bab2018-07-11 12:16:12 +02003632 tp->rx_opt.ts_recent_stamp = ktime_get_seconds();
Eric Dumazet12fb3dd2013-04-19 07:19:48 +00003633}
3634
3635static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3636{
3637 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3638 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3639 * extra check below makes sure this can only happen
3640 * for pure ACK frames. -DaveM
3641 *
3642 * Not only, also it occurs for expired timestamps.
3643 */
3644
3645 if (tcp_paws_check(&tp->rx_opt, 0))
3646 tcp_store_ts_recent(tp);
3647 }
3648}
3649
Yuchung Cheng76be93f2020-07-23 12:00:06 -07003650/* This routine deals with acks during a TLP episode and ends an episode by
3651 * resetting tlp_high_seq. Ref: TLP algorithm in draft-ietf-tcpm-rack
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003652 */
3653static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3654{
3655 struct tcp_sock *tp = tcp_sk(sk);
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003656
Sébastien Barré08abdff2015-01-12 10:30:40 +01003657 if (before(ack, tp->tlp_high_seq))
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003658 return;
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003659
Yuchung Cheng76be93f2020-07-23 12:00:06 -07003660 if (!tp->tlp_retrans) {
3661 /* TLP of new data has been acknowledged */
3662 tp->tlp_high_seq = 0;
Yuchung Cheng63f367d2021-07-27 10:42:57 -04003663 } else if (flag & FLAG_DSACK_TLP) {
Sébastien Barré08abdff2015-01-12 10:30:40 +01003664 /* This DSACK means original and TLP probe arrived; no loss */
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003665 tp->tlp_high_seq = 0;
Sébastien Barré08abdff2015-01-12 10:30:40 +01003666 } else if (after(ack, tp->tlp_high_seq)) {
3667 /* ACK advances: there was a loss, so reduce cwnd. Reset
3668 * tlp_high_seq in tcp_init_cwnd_reduction()
3669 */
3670 tcp_init_cwnd_reduction(sk);
3671 tcp_set_ca_state(sk, TCP_CA_CWR);
3672 tcp_end_cwnd_reduction(sk);
3673 tcp_try_keep_open(sk);
Eric Dumazetc10d9312016-04-29 14:16:47 -07003674 NET_INC_STATS(sock_net(sk),
Eric Dumazet02a1d6e2016-04-27 16:44:39 -07003675 LINUX_MIB_TCPLOSSPROBERECOVERY);
Sébastien Barré08abdff2015-01-12 10:30:40 +01003676 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3677 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3678 /* Pure dupack: original and TLP probe arrived; no loss */
3679 tp->tlp_high_seq = 0;
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003680 }
3681}
3682
Florian Westphal7354c8c2014-09-26 22:37:34 +02003683static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3684{
3685 const struct inet_connection_sock *icsk = inet_csk(sk);
3686
3687 if (icsk->icsk_ca_ops->in_ack_event)
3688 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3689}
3690
Yuchung Chenge662ca42016-02-02 10:33:04 -08003691/* Congestion control has updated the cwnd already. So if we're in
3692 * loss recovery then now we do any new sends (for FRTO) or
3693 * retransmits (for CA_Loss or CA_recovery) that make sense.
3694 */
3695static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3696{
3697 struct tcp_sock *tp = tcp_sk(sk);
3698
Yuchung Chengbc9f38c2019-04-29 15:46:13 -07003699 if (rexmit == REXMIT_NONE || sk->sk_state == TCP_SYN_SENT)
Yuchung Chenge662ca42016-02-02 10:33:04 -08003700 return;
3701
Mao Wenand0e8bca2020-01-02 22:02:27 +08003702 if (unlikely(rexmit == REXMIT_NEW)) {
Yuchung Chenge662ca42016-02-02 10:33:04 -08003703 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3704 TCP_NAGLE_OFF);
3705 if (after(tp->snd_nxt, tp->high_seq))
3706 return;
3707 tp->frto = 0;
3708 }
3709 tcp_xmit_retransmit_queue(sk);
3710}
3711
Yuchung Chenga77fa012018-04-17 23:18:47 -07003712/* Returns the number of packets newly acked or sacked by the current ACK */
3713static u32 tcp_newly_delivered(struct sock *sk, u32 prior_delivered, int flag)
3714{
Yuchung Chengfeb5f2e2018-04-17 23:18:49 -07003715 const struct net *net = sock_net(sk);
Yuchung Chenga77fa012018-04-17 23:18:47 -07003716 struct tcp_sock *tp = tcp_sk(sk);
3717 u32 delivered;
3718
3719 delivered = tp->delivered - prior_delivered;
Yuchung Chengfeb5f2e2018-04-17 23:18:49 -07003720 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVERED, delivered);
Yousuk Seung082d4fa2020-06-26 21:05:35 -07003721 if (flag & FLAG_ECE)
Yuchung Chengfeb5f2e2018-04-17 23:18:49 -07003722 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVEREDCE, delivered);
Yousuk Seung082d4fa2020-06-26 21:05:35 -07003723
Yuchung Chenga77fa012018-04-17 23:18:47 -07003724 return delivered;
3725}
3726
Linus Torvalds1da177e2005-04-16 15:20:36 -07003727/* This routine deals with incoming acks, but not outgoing ones. */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003728static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003729{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003730 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003731 struct tcp_sock *tp = tcp_sk(sk);
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02003732 struct tcp_sacktag_state sack_state;
Yuchung Chengb9f64822016-09-19 23:39:14 -04003733 struct rate_sample rs = { .prior_delivered = 0 };
Linus Torvalds1da177e2005-04-16 15:20:36 -07003734 u32 prior_snd_una = tp->snd_una;
Yousuk Seungd4761752017-12-07 13:41:34 -08003735 bool is_sack_reneg = tp->is_sack_reneg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003736 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3737 u32 ack = TCP_SKB_CB(skb)->ack_seq;
Eric Dumazet19119f22018-11-27 14:42:01 -08003738 int num_dupack = 0;
Nandita Dukkipati35f079e2013-05-21 15:12:07 +00003739 int prior_packets = tp->packets_out;
Yuchung Chengb9f64822016-09-19 23:39:14 -04003740 u32 delivered = tp->delivered;
3741 u32 lost = tp->lost;
Yuchung Chenge662ca42016-02-02 10:33:04 -08003742 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
Yuchung Cheng737ff312017-11-08 13:01:27 -08003743 u32 prior_fack;
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02003744
Eric Dumazet9a568de2017-05-16 14:00:14 -07003745 sack_state.first_sackt = 0;
Yuchung Chengb9f64822016-09-19 23:39:14 -04003746 sack_state.rate = &rs;
Yousuk Seungf00394c2020-06-26 21:05:34 -07003747 sack_state.sack_delivered = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003748
Eric Dumazet75c119a2017-10-05 22:21:27 -07003749 /* We very likely will need to access rtx queue. */
3750 prefetch(sk->tcp_rtx_queue.rb_node);
Eric Dumazetad971f62014-10-11 15:17:29 -07003751
John Dykstra96e0bf42009-03-22 21:49:57 -07003752 /* If the ack is older than previous acks
Linus Torvalds1da177e2005-04-16 15:20:36 -07003753 * then we can probably ignore it.
3754 */
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003755 if (before(ack, prior_snd_una)) {
3756 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3757 if (before(ack, prior_snd_una - tp->max_window)) {
Eric Dumazetd0e1a1b2017-05-23 15:24:46 -07003758 if (!(flag & FLAG_NO_CHALLENGE_ACK))
3759 tcp_send_challenge_ack(sk, skb);
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003760 return -1;
3761 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003762 goto old_ack;
Eric Dumazet354e4aa2012-10-21 19:57:11 +00003763 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003764
John Dykstra96e0bf42009-03-22 21:49:57 -07003765 /* If the ack includes data we haven't sent yet, discard
3766 * this segment (RFC793 Section 3.9).
3767 */
3768 if (after(ack, tp->snd_nxt))
Yafang Shao9946b342019-02-25 18:33:48 +08003769 return -1;
John Dykstra96e0bf42009-03-22 21:49:57 -07003770
Neal Cardwell0c9ab092014-08-14 16:13:07 -04003771 if (after(ack, prior_snd_una)) {
Ilpo Järvinen2e605292007-08-02 19:46:58 -07003772 flag |= FLAG_SND_UNA_ADVANCED;
Neal Cardwell0c9ab092014-08-14 16:13:07 -04003773 icsk->icsk_retransmits = 0;
Ilya Lesokhin6dac1522018-04-30 10:16:10 +03003774
3775#if IS_ENABLED(CONFIG_TLS_DEVICE)
Jakub Kicinski494bc1d2019-05-08 16:46:14 -07003776 if (static_branch_unlikely(&clean_acked_data_enabled.key))
Ilya Lesokhin6dac1522018-04-30 10:16:10 +03003777 if (icsk->icsk_clean_acked)
3778 icsk->icsk_clean_acked(sk, ack);
3779#endif
Neal Cardwell0c9ab092014-08-14 16:13:07 -04003780 }
Ilpo Järvinen2e605292007-08-02 19:46:58 -07003781
Eric Dumazet50895b92017-11-14 21:02:19 -08003782 prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una;
Yuchung Chengb9f64822016-09-19 23:39:14 -04003783 rs.prior_in_flight = tcp_packets_in_flight(tp);
Ilpo Järvinenc7caf8d2007-11-10 21:22:18 -08003784
Eric Dumazet12fb3dd2013-04-19 07:19:48 +00003785 /* ts_recent update must be made after we are sure that the packet
3786 * is in window.
3787 */
3788 if (flag & FLAG_UPDATE_TS_RECENT)
3789 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3790
Yafang Shao5e13a0d2018-11-11 20:10:10 +08003791 if ((flag & (FLAG_SLOWPATH | FLAG_SND_UNA_ADVANCED)) ==
3792 FLAG_SND_UNA_ADVANCED) {
Florian Westphal31770e32017-08-30 19:24:58 +02003793 /* Window is constant, pure forward advance.
3794 * No more checks are required.
3795 * Note, we use the fact that SND.UNA>=SND.WL2.
3796 */
3797 tcp_update_wl(tp, ack_seq);
3798 tcp_snd_una_update(tp, ack);
3799 flag |= FLAG_WIN_UPDATE;
3800
3801 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3802
3803 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3804 } else {
Florian Westphalc1d2b4c2017-08-30 19:24:57 +02003805 u32 ack_ev_flags = CA_ACK_SLOWPATH;
Florian Westphal98900922014-09-26 22:37:35 +02003806
Florian Westphalc1d2b4c2017-08-30 19:24:57 +02003807 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3808 flag |= FLAG_DATA;
3809 else
3810 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003811
Florian Westphalc1d2b4c2017-08-30 19:24:57 +02003812 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003813
Florian Westphalc1d2b4c2017-08-30 19:24:57 +02003814 if (TCP_SKB_CB(skb)->sacked)
3815 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3816 &sack_state);
3817
3818 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3819 flag |= FLAG_ECE;
3820 ack_ev_flags |= CA_ACK_ECE;
3821 }
3822
Yousuk Seung082d4fa2020-06-26 21:05:35 -07003823 if (sack_state.sack_delivered)
3824 tcp_count_delivered(tp, sack_state.sack_delivered,
3825 flag & FLAG_ECE);
3826
Florian Westphalc1d2b4c2017-08-30 19:24:57 +02003827 if (flag & FLAG_WIN_UPDATE)
3828 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3829
3830 tcp_in_ack_event(sk, ack_ev_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003831 }
3832
Denis Kirjanov257028402020-06-25 14:51:06 +03003833 /* This is a deviation from RFC3168 since it states that:
3834 * "When the TCP data sender is ready to set the CWR bit after reducing
3835 * the congestion window, it SHOULD set the CWR bit only on the first
3836 * new data packet that it transmits."
3837 * We accept CWR on pure ACKs to be more robust
3838 * with widely-deployed TCP implementations that do this.
3839 */
3840 tcp_ecn_accept_cwr(sk, skb);
3841
Linus Torvalds1da177e2005-04-16 15:20:36 -07003842 /* We passed data and got it acked, remove any soft error
3843 * log. Something worked...
3844 */
3845 sk->sk_err_soft = 0;
David S. Miller4b53fb62008-07-23 16:38:45 -07003846 icsk->icsk_probes_out = 0;
Eric Dumazet70eabf02017-05-16 14:00:07 -07003847 tp->rcv_tstamp = tcp_jiffies32;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003848 if (!prior_packets)
3849 goto no_queue;
3850
Linus Torvalds1da177e2005-04-16 15:20:36 -07003851 /* See if we can take anything off of the retransmit queue. */
Yousuk Seunge7ed11e2021-01-20 12:41:55 -08003852 flag |= tcp_clean_rtx_queue(sk, skb, prior_fack, prior_snd_una,
3853 &sack_state, flag & FLAG_ECE);
Nandita Dukkipatia262f0c2011-08-21 20:21:57 +00003854
Priyaranjan Jha1f255692017-11-03 16:38:48 -07003855 tcp_rack_update_reo_wnd(sk, &rs);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003856
Neal Cardwelldf92c832017-08-03 09:19:54 -04003857 if (tp->tlp_high_seq)
3858 tcp_process_tlp_ack(sk, ack, flag);
Neal Cardwelldf92c832017-08-03 09:19:54 -04003859
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003860 if (tcp_ack_is_dubious(sk, flag)) {
Eric Dumazet19119f22018-11-27 14:42:01 -08003861 if (!(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP))) {
3862 num_dupack = 1;
3863 /* Consider if pure acks were aggregated in tcp_add_backlog() */
3864 if (!(flag & FLAG_DATA))
3865 num_dupack = max_t(u16, 1, skb_shinfo(skb)->gso_segs);
3866 }
3867 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
Yuchung Cheng737ff312017-11-08 13:01:27 -08003868 &rexmit);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003869 }
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003870
Pengcheng Yang62d9f1a2021-01-24 13:07:14 +08003871 /* If needed, reset TLP/RTO timer when RACK doesn't set. */
3872 if (flag & FLAG_SET_XMIT_TIMER)
3873 tcp_set_xmit_timer(sk);
3874
Julian Anastasovc3a2e832017-02-06 23:14:14 +02003875 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3876 sk_dst_confirm(sk);
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00003877
Yuchung Chenga77fa012018-04-17 23:18:47 -07003878 delivered = tcp_newly_delivered(sk, delivered, flag);
Yuchung Chengb9f64822016-09-19 23:39:14 -04003879 lost = tp->lost - lost; /* freshly marked lost */
Yuchung Chenge4286602018-01-17 12:11:01 -08003880 rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED);
Yousuk Seungd4761752017-12-07 13:41:34 -08003881 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate);
Yuchung Chengdeed7be2017-01-12 22:11:32 -08003882 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate);
Yuchung Chenge662ca42016-02-02 10:33:04 -08003883 tcp_xmit_recovery(sk, rexmit);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003884 return 1;
3885
3886no_queue:
Neal Cardwell5628adf2011-11-16 08:58:02 +00003887 /* If data was DSACKed, see if we can undo a cwnd reduction. */
Yuchung Chenga77fa012018-04-17 23:18:47 -07003888 if (flag & FLAG_DSACKING_ACK) {
Eric Dumazet19119f22018-11-27 14:42:01 -08003889 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
Yuchung Cheng737ff312017-11-08 13:01:27 -08003890 &rexmit);
Yuchung Chenga77fa012018-04-17 23:18:47 -07003891 tcp_newly_delivered(sk, delivered, flag);
3892 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003893 /* If this ack opens up a zero window, clear backoff. It was
3894 * being used to time the probes, and is probably far higher than
3895 * it needs to be for normal retransmission.
3896 */
Eric Dumazet75c119a2017-10-05 22:21:27 -07003897 tcp_ack_probe(sk);
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00003898
3899 if (tp->tlp_high_seq)
3900 tcp_process_tlp_ack(sk, ack, flag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003901 return 1;
3902
3903old_ack:
Neal Cardwelle95ae2f2011-11-16 08:58:03 +00003904 /* If data was SACKed, tag it and see if we should send more data.
3905 * If data was DSACKed, see if we can undo a cwnd reduction.
3906 */
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07003907 if (TCP_SKB_CB(skb)->sacked) {
Yuchung Cheng59c9af42013-07-22 16:20:47 -07003908 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
Kenneth Klette Jonassen196da972015-05-01 01:10:57 +02003909 &sack_state);
Eric Dumazet19119f22018-11-27 14:42:01 -08003910 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
Yuchung Cheng737ff312017-11-08 13:01:27 -08003911 &rexmit);
Yuchung Chenga77fa012018-04-17 23:18:47 -07003912 tcp_newly_delivered(sk, delivered, flag);
Yuchung Chenge662ca42016-02-02 10:33:04 -08003913 tcp_xmit_recovery(sk, rexmit);
Ilpo Järvinen8aca6cb2008-06-04 11:34:22 -07003914 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003915
Linus Torvalds1da177e2005-04-16 15:20:36 -07003916 return 0;
3917}
3918
Daniel Lee7f9b8382015-04-06 14:37:26 -07003919static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3920 bool syn, struct tcp_fastopen_cookie *foc,
3921 bool exp_opt)
3922{
3923 /* Valid only in SYN or SYN-ACK with an even length. */
3924 if (!foc || !syn || len < 0 || (len & 1))
3925 return;
3926
3927 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3928 len <= TCP_FASTOPEN_COOKIE_MAX)
3929 memcpy(foc->val, cookie, len);
3930 else if (len != 0)
3931 len = -1;
3932 foc->len = len;
3933 foc->exp = exp_opt;
3934}
3935
Martin KaFai Lau7656d682020-08-20 12:00:33 -07003936static bool smc_parse_options(const struct tcphdr *th,
Ursula Braun60e2a772017-10-25 11:01:45 +02003937 struct tcp_options_received *opt_rx,
3938 const unsigned char *ptr,
3939 int opsize)
3940{
3941#if IS_ENABLED(CONFIG_SMC)
3942 if (static_branch_unlikely(&tcp_have_smc)) {
3943 if (th->syn && !(opsize & 1) &&
3944 opsize >= TCPOLEN_EXP_SMC_BASE &&
Martin KaFai Lau7656d682020-08-20 12:00:33 -07003945 get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC) {
Ursula Braun60e2a772017-10-25 11:01:45 +02003946 opt_rx->smc_ok = 1;
Martin KaFai Lau7656d682020-08-20 12:00:33 -07003947 return true;
3948 }
Ursula Braun60e2a772017-10-25 11:01:45 +02003949 }
3950#endif
Martin KaFai Lau7656d682020-08-20 12:00:33 -07003951 return false;
Ursula Braun60e2a772017-10-25 11:01:45 +02003952}
3953
Petar Penkov9349d602019-07-29 09:59:14 -07003954/* Try to parse the MSS option from the TCP header. Return 0 on failure, clamped
3955 * value on success.
3956 */
3957static u16 tcp_parse_mss_option(const struct tcphdr *th, u16 user_mss)
3958{
3959 const unsigned char *ptr = (const unsigned char *)(th + 1);
3960 int length = (th->doff * 4) - sizeof(struct tcphdr);
3961 u16 mss = 0;
3962
3963 while (length > 0) {
3964 int opcode = *ptr++;
3965 int opsize;
3966
3967 switch (opcode) {
3968 case TCPOPT_EOL:
3969 return mss;
3970 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3971 length--;
3972 continue;
3973 default:
3974 if (length < 2)
3975 return mss;
3976 opsize = *ptr++;
3977 if (opsize < 2) /* "silly options" */
3978 return mss;
3979 if (opsize > length)
3980 return mss; /* fail on partial options */
3981 if (opcode == TCPOPT_MSS && opsize == TCPOLEN_MSS) {
3982 u16 in_mss = get_unaligned_be16(ptr);
3983
3984 if (in_mss) {
3985 if (user_mss && user_mss < in_mss)
3986 in_mss = user_mss;
3987 mss = in_mss;
3988 }
3989 }
3990 ptr += opsize - 2;
3991 length -= opsize;
3992 }
3993 }
3994 return mss;
3995}
3996
Linus Torvalds1da177e2005-04-16 15:20:36 -07003997/* Look for tcp options. Normally only called on SYN and SYNACK packets.
3998 * But, this can also be called on packets in the established flow when
3999 * the fast version below fails.
4000 */
Eric Dumazeteed29f12017-06-07 10:34:36 -07004001void tcp_parse_options(const struct net *net,
4002 const struct sk_buff *skb,
Christoph Paasch1a2c6182013-03-17 08:23:34 +00004003 struct tcp_options_received *opt_rx, int estab,
Yuchung Cheng2100c8d2012-07-19 06:43:05 +00004004 struct tcp_fastopen_cookie *foc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004005{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04004006 const unsigned char *ptr;
4007 const struct tcphdr *th = tcp_hdr(skb);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004008 int length = (th->doff * 4) - sizeof(struct tcphdr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004009
Eric Dumazetcf533ea2011-10-21 05:22:42 -04004010 ptr = (const unsigned char *)(th + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004011 opt_rx->saw_tstamp = 0;
Martin KaFai Lau7656d682020-08-20 12:00:33 -07004012 opt_rx->saw_unknown = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004013
Stephen Hemminger2de979b2007-03-08 20:45:19 -08004014 while (length > 0) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004015 int opcode = *ptr++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004016 int opsize;
4017
4018 switch (opcode) {
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08004019 case TCPOPT_EOL:
4020 return;
4021 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
4022 length--;
4023 continue;
4024 default:
Young Xiao9609dad2019-05-29 16:10:59 +08004025 if (length < 2)
4026 return;
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08004027 opsize = *ptr++;
4028 if (opsize < 2) /* "silly options" */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004029 return;
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08004030 if (opsize > length)
4031 return; /* don't parse partial options */
4032 switch (opcode) {
4033 case TCPOPT_MSS:
4034 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07004035 u16 in_mss = get_unaligned_be16(ptr);
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08004036 if (in_mss) {
4037 if (opt_rx->user_mss &&
4038 opt_rx->user_mss < in_mss)
4039 in_mss = opt_rx->user_mss;
4040 opt_rx->mss_clamp = in_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004041 }
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07004042 }
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08004043 break;
4044 case TCPOPT_WINDOW:
4045 if (opsize == TCPOLEN_WINDOW && th->syn &&
Eric Dumazet9bb37ef2017-06-07 10:34:38 -07004046 !estab && net->ipv4.sysctl_tcp_window_scaling) {
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08004047 __u8 snd_wscale = *(__u8 *)ptr;
4048 opt_rx->wscale_ok = 1;
Gao Feng589c49c2017-04-04 21:09:48 +08004049 if (snd_wscale > TCP_MAX_WSCALE) {
4050 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
Joe Perchese87cc472012-05-13 21:56:26 +00004051 __func__,
Gao Feng589c49c2017-04-04 21:09:48 +08004052 snd_wscale,
4053 TCP_MAX_WSCALE);
4054 snd_wscale = TCP_MAX_WSCALE;
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08004055 }
4056 opt_rx->snd_wscale = snd_wscale;
4057 }
4058 break;
4059 case TCPOPT_TIMESTAMP:
4060 if ((opsize == TCPOLEN_TIMESTAMP) &&
4061 ((estab && opt_rx->tstamp_ok) ||
Eric Dumazet5d2ed052017-06-07 10:34:39 -07004062 (!estab && net->ipv4.sysctl_tcp_timestamps))) {
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08004063 opt_rx->saw_tstamp = 1;
Harvey Harrisond3e2ce32008-05-02 16:26:16 -07004064 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
4065 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08004066 }
4067 break;
4068 case TCPOPT_SACK_PERM:
4069 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
Eric Dumazetf9301032017-06-07 10:34:37 -07004070 !estab && net->ipv4.sysctl_tcp_sack) {
Vijay Subramanianab562222011-12-20 13:23:24 +00004071 opt_rx->sack_ok = TCP_SACK_SEEN;
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08004072 tcp_sack_reset(opt_rx);
4073 }
4074 break;
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07004075
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08004076 case TCPOPT_SACK:
4077 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
4078 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
4079 opt_rx->sack_ok) {
4080 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
4081 }
4082 break;
4083#ifdef CONFIG_TCP_MD5SIG
4084 case TCPOPT_MD5SIG:
4085 /*
4086 * The MD5 Hash has already been
4087 * checked (see tcp_v{4,6}_do_rcv()).
4088 */
4089 break;
4090#endif
Daniel Lee7f9b8382015-04-06 14:37:26 -07004091 case TCPOPT_FASTOPEN:
4092 tcp_parse_fastopen_option(
4093 opsize - TCPOLEN_FASTOPEN_BASE,
4094 ptr, th->syn, foc, false);
4095 break;
4096
Yuchung Cheng2100c8d2012-07-19 06:43:05 +00004097 case TCPOPT_EXP:
4098 /* Fast Open option shares code 254 using a
Daniel Lee7f9b8382015-04-06 14:37:26 -07004099 * 16 bits magic number.
Yuchung Cheng2100c8d2012-07-19 06:43:05 +00004100 */
Daniel Lee7f9b8382015-04-06 14:37:26 -07004101 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
4102 get_unaligned_be16(ptr) ==
Martin KaFai Lau7656d682020-08-20 12:00:33 -07004103 TCPOPT_FASTOPEN_MAGIC) {
Daniel Lee7f9b8382015-04-06 14:37:26 -07004104 tcp_parse_fastopen_option(opsize -
4105 TCPOLEN_EXP_FASTOPEN_BASE,
4106 ptr + 2, th->syn, foc, true);
Martin KaFai Lau7656d682020-08-20 12:00:33 -07004107 break;
4108 }
4109
4110 if (smc_parse_options(th, opt_rx, ptr, opsize))
4111 break;
4112
4113 opt_rx->saw_unknown = 1;
Yuchung Cheng2100c8d2012-07-19 06:43:05 +00004114 break;
4115
Martin KaFai Lau7656d682020-08-20 12:00:33 -07004116 default:
4117 opt_rx->saw_unknown = 1;
Yuchung Cheng2100c8d2012-07-19 06:43:05 +00004118 }
Ilpo Järvinenf038ac82008-01-03 20:36:55 -08004119 ptr += opsize-2;
4120 length -= opsize;
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07004121 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004122 }
4123}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00004124EXPORT_SYMBOL(tcp_parse_options);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004125
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004126static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
Ilpo Järvinena4356b22008-08-23 05:12:29 -07004127{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04004128 const __be32 *ptr = (const __be32 *)(th + 1);
Ilpo Järvinena4356b22008-08-23 05:12:29 -07004129
4130 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4131 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
4132 tp->rx_opt.saw_tstamp = 1;
4133 ++ptr;
4134 tp->rx_opt.rcv_tsval = ntohl(*ptr);
4135 ++ptr;
Andrew Vagine3e12022013-08-27 12:21:55 +04004136 if (*ptr)
4137 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
4138 else
4139 tp->rx_opt.rcv_tsecr = 0;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004140 return true;
Ilpo Järvinena4356b22008-08-23 05:12:29 -07004141 }
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004142 return false;
Ilpo Järvinena4356b22008-08-23 05:12:29 -07004143}
4144
Linus Torvalds1da177e2005-04-16 15:20:36 -07004145/* Fast parse options. This hopes to only see timestamps.
4146 * If it is wrong it falls back on tcp_parse_options().
4147 */
Eric Dumazeteed29f12017-06-07 10:34:36 -07004148static bool tcp_fast_parse_options(const struct net *net,
4149 const struct sk_buff *skb,
Christoph Paasch1a2c6182013-03-17 08:23:34 +00004150 const struct tcphdr *th, struct tcp_sock *tp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004151{
William Allen Simpson4957faade2009-12-02 18:25:27 +00004152 /* In the spirit of fast parsing, compare doff directly to constant
4153 * values. Because equality is used, short doff can be ignored here.
4154 */
4155 if (th->doff == (sizeof(*th) / 4)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004156 tp->rx_opt.saw_tstamp = 0;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004157 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004158 } else if (tp->rx_opt.tstamp_ok &&
William Allen Simpson4957faade2009-12-02 18:25:27 +00004159 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
Ilpo Järvinena4356b22008-08-23 05:12:29 -07004160 if (tcp_parse_aligned_timestamp(tp, th))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004161 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004162 }
Andrey Vaginee684b62013-02-11 05:50:19 +00004163
Eric Dumazeteed29f12017-06-07 10:34:36 -07004164 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL);
Andrew Vagine3e12022013-08-27 12:21:55 +04004165 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
Andrey Vaginee684b62013-02-11 05:50:19 +00004166 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
4167
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004168 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004169}
4170
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09004171#ifdef CONFIG_TCP_MD5SIG
4172/*
4173 * Parse MD5 Signature option
4174 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04004175const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09004176{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04004177 int length = (th->doff << 2) - sizeof(*th);
4178 const u8 *ptr = (const u8 *)(th + 1);
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09004179
Jann Horn7e5a2062018-04-20 15:57:30 +02004180 /* If not enough data remaining, we can short cut */
4181 while (length >= TCPOLEN_MD5SIG) {
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09004182 int opcode = *ptr++;
4183 int opsize;
4184
Weilong Chena22318e2013-12-23 14:37:26 +08004185 switch (opcode) {
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09004186 case TCPOPT_EOL:
4187 return NULL;
4188 case TCPOPT_NOP:
4189 length--;
4190 continue;
4191 default:
4192 opsize = *ptr++;
4193 if (opsize < 2 || opsize > length)
4194 return NULL;
4195 if (opcode == TCPOPT_MD5SIG)
Dmitry Popovba78e2d2010-08-07 20:24:28 -07004196 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09004197 }
4198 ptr += opsize - 2;
4199 length -= opsize;
4200 }
4201 return NULL;
4202}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00004203EXPORT_SYMBOL(tcp_parse_md5sig_option);
YOSHIFUJI Hideaki7d5d5522008-04-17 12:29:53 +09004204#endif
4205
Linus Torvalds1da177e2005-04-16 15:20:36 -07004206/* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
4207 *
4208 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
4209 * it can pass through stack. So, the following predicate verifies that
4210 * this segment is not used for anything but congestion avoidance or
4211 * fast retransmit. Moreover, we even are able to eliminate most of such
4212 * second order effects, if we apply some small "replay" window (~RTO)
4213 * to timestamp space.
4214 *
4215 * All these measures still do not guarantee that we reject wrapped ACKs
4216 * on networks with high bandwidth, when sequence space is recycled fastly,
4217 * but it guarantees that such events will be very rare and do not affect
4218 * connection seriously. This doesn't look nice, but alas, PAWS is really
4219 * buggy extension.
4220 *
4221 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
4222 * states that events when retransmit arrives after original data are rare.
4223 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
4224 * the biggest problem on large power networks even with minor reordering.
4225 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
4226 * up to bandwidth of 18Gigabit/sec. 8) ]
4227 */
4228
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004229static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004230{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04004231 const struct tcp_sock *tp = tcp_sk(sk);
4232 const struct tcphdr *th = tcp_hdr(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004233 u32 seq = TCP_SKB_CB(skb)->seq;
4234 u32 ack = TCP_SKB_CB(skb)->ack_seq;
4235
4236 return (/* 1. Pure ACK with correct sequence number. */
4237 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
4238
4239 /* 2. ... and duplicate ACK. */
4240 ack == tp->snd_una &&
4241
4242 /* 3. ... and does not update window. */
4243 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4244
4245 /* 4. ... and sits in replay window. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004246 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004247}
4248
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00004249static inline bool tcp_paws_discard(const struct sock *sk,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004250 const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004251{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004252 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinenc887e6d2009-03-14 14:23:03 +00004253
4254 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4255 !tcp_disordered_ack(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004256}
4257
4258/* Check segment sequence number for validity.
4259 *
4260 * Segment controls are considered valid, if the segment
4261 * fits to the window after truncation to the window. Acceptability
4262 * of data (and SYN, FIN, of course) is checked separately.
4263 * See tcp_data_queue(), for example.
4264 *
4265 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4266 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4267 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4268 * (borrowed from freebsd)
4269 */
4270
Vijay Subramanian67b95bd2012-07-19 21:32:18 +00004271static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004272{
4273 return !before(end_seq, tp->rcv_wup) &&
4274 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4275}
4276
4277/* When we get a reset we do this. */
Florian Westphal049fe382020-12-10 14:25:03 -08004278void tcp_reset(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004279{
Song Liu59415212017-10-23 09:20:25 -07004280 trace_tcp_receive_reset(sk);
4281
Jianguo Wu6787b7e2021-07-09 17:20:49 -07004282 /* mptcp can't tell us to ignore reset pkts,
4283 * so just ignore the return value of mptcp_incoming_options().
4284 */
Florian Westphal049fe382020-12-10 14:25:03 -08004285 if (sk_is_mptcp(sk))
4286 mptcp_incoming_options(sk, skb);
4287
Linus Torvalds1da177e2005-04-16 15:20:36 -07004288 /* We want the right error as BSD sees it (and indeed as we do). */
4289 switch (sk->sk_state) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004290 case TCP_SYN_SENT:
4291 sk->sk_err = ECONNREFUSED;
4292 break;
4293 case TCP_CLOSE_WAIT:
4294 sk->sk_err = EPIPE;
4295 break;
4296 case TCP_CLOSE:
4297 return;
4298 default:
4299 sk->sk_err = ECONNRESET;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004300 }
Tom Marshalla4d25802010-09-20 15:42:05 -07004301 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4302 smp_wmb();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004303
Soheil Hassas Yeganeha27fd7a2018-02-27 18:32:18 -05004304 tcp_write_queue_purge(sk);
Eric Dumazet3d476262017-04-18 09:45:51 -07004305 tcp_done(sk);
4306
Linus Torvalds1da177e2005-04-16 15:20:36 -07004307 if (!sock_flag(sk, SOCK_DEAD))
Alexander Aringe3ae2362021-06-27 18:48:21 -04004308 sk_error_report(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004309}
4310
4311/*
4312 * Process the FIN bit. This now behaves as it is supposed to work
4313 * and the FIN takes effect when it is validly part of sequence
4314 * space. Not before when we get holes.
4315 *
4316 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4317 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4318 * TIME-WAIT)
4319 *
4320 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4321 * close and we go into CLOSING (and later onto TIME-WAIT)
4322 *
4323 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4324 */
Eric Dumazete3e17b72016-02-06 11:16:28 -08004325void tcp_fin(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004326{
4327 struct tcp_sock *tp = tcp_sk(sk);
4328
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07004329 inet_csk_schedule_ack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004330
4331 sk->sk_shutdown |= RCV_SHUTDOWN;
4332 sock_set_flag(sk, SOCK_DONE);
4333
4334 switch (sk->sk_state) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004335 case TCP_SYN_RECV:
4336 case TCP_ESTABLISHED:
4337 /* Move to CLOSE_WAIT */
4338 tcp_set_state(sk, TCP_CLOSE_WAIT);
Wei Wang31954cd2019-01-25 10:53:19 -08004339 inet_csk_enter_pingpong_mode(sk);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004340 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004341
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004342 case TCP_CLOSE_WAIT:
4343 case TCP_CLOSING:
4344 /* Received a retransmission of the FIN, do
4345 * nothing.
4346 */
4347 break;
4348 case TCP_LAST_ACK:
4349 /* RFC793: Remain in the LAST-ACK state. */
4350 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004351
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004352 case TCP_FIN_WAIT1:
4353 /* This case occurs when a simultaneous close
4354 * happens, we must ack the received FIN and
4355 * enter the CLOSING state.
4356 */
4357 tcp_send_ack(sk);
4358 tcp_set_state(sk, TCP_CLOSING);
4359 break;
4360 case TCP_FIN_WAIT2:
4361 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4362 tcp_send_ack(sk);
4363 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4364 break;
4365 default:
4366 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4367 * cases we should never reach this piece of code.
4368 */
Joe Perches058bd4d2012-03-11 18:36:11 +00004369 pr_err("%s: Impossible, sk->sk_state=%d\n",
Harvey Harrison0dc47872008-03-05 20:47:47 -08004370 __func__, sk->sk_state);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004371 break;
Stephen Hemminger3ff50b72007-04-20 17:09:22 -07004372 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004373
4374 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4375 * Probably, we should reset in this case. For now drop them.
4376 */
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004377 skb_rbtree_purge(&tp->out_of_order_queue);
Ilpo Järvinene60402d2007-08-09 15:14:46 +03004378 if (tcp_is_sack(tp))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004379 tcp_sack_reset(&tp->rx_opt);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08004380 sk_mem_reclaim(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004381
4382 if (!sock_flag(sk, SOCK_DEAD)) {
4383 sk->sk_state_change(sk);
4384
4385 /* Do not send POLL_HUP for half duplex close. */
4386 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4387 sk->sk_state == TCP_CLOSE)
Pavel Emelyanov8d8ad9d2007-11-26 20:10:50 +08004388 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004389 else
Pavel Emelyanov8d8ad9d2007-11-26 20:10:50 +08004390 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004391 }
4392}
4393
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004394static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004395 u32 end_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004396{
4397 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4398 if (before(seq, sp->start_seq))
4399 sp->start_seq = seq;
4400 if (after(end_seq, sp->end_seq))
4401 sp->end_seq = end_seq;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004402 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004403 }
Eric Dumazeta2a385d2012-05-16 23:15:34 +00004404 return false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004405}
4406
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004407static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004408{
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004409 struct tcp_sock *tp = tcp_sk(sk);
4410
Eric Dumazet6496f6b2017-10-26 21:55:07 -07004411 if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) {
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07004412 int mib_idx;
4413
Linus Torvalds1da177e2005-04-16 15:20:36 -07004414 if (before(seq, tp->rcv_nxt))
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07004415 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004416 else
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07004417 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4418
Eric Dumazetc10d9312016-04-29 14:16:47 -07004419 NET_INC_STATS(sock_net(sk), mib_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004420
4421 tp->rx_opt.dsack = 1;
4422 tp->duplicate_sack[0].start_seq = seq;
4423 tp->duplicate_sack[0].end_seq = end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004424 }
4425}
4426
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004427static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004428{
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004429 struct tcp_sock *tp = tcp_sk(sk);
4430
Linus Torvalds1da177e2005-04-16 15:20:36 -07004431 if (!tp->rx_opt.dsack)
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004432 tcp_dsack_set(sk, seq, end_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004433 else
4434 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4435}
4436
Yuchung Cheng77881742018-08-29 14:53:56 -07004437static void tcp_rcv_spurious_retrans(struct sock *sk, const struct sk_buff *skb)
4438{
4439 /* When the ACK path fails or drops most ACKs, the sender would
4440 * timeout and spuriously retransmit the same segment repeatedly.
4441 * The receiver remembers and reflects via DSACKs. Leverage the
4442 * DSACK state and change the txhash to re-route speculatively.
4443 */
Yuchung Cheng9c30ae82021-01-19 11:26:19 -08004444 if (TCP_SKB_CB(skb)->seq == tcp_sk(sk)->duplicate_sack[0].start_seq &&
4445 sk_rethink_txhash(sk))
Abdul Kabbani32efcc02020-01-24 16:34:02 -05004446 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDUPLICATEDATAREHASH);
Yuchung Cheng77881742018-08-29 14:53:56 -07004447}
4448
Eric Dumazetcf533ea2011-10-21 05:22:42 -04004449static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004450{
4451 struct tcp_sock *tp = tcp_sk(sk);
4452
4453 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4454 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
Eric Dumazetc10d9312016-04-29 14:16:47 -07004455 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
Eric Dumazet9a9c9b52018-05-21 15:08:56 -07004456 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004457
Eric Dumazet6496f6b2017-10-26 21:55:07 -07004458 if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004459 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4460
Yuchung Cheng77881742018-08-29 14:53:56 -07004461 tcp_rcv_spurious_retrans(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004462 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4463 end_seq = tp->rcv_nxt;
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004464 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004465 }
4466 }
4467
4468 tcp_send_ack(sk);
4469}
4470
4471/* These routines update the SACK block as out-of-order packets arrive or
4472 * in-order packets close up the sequence space.
4473 */
4474static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4475{
4476 int this_sack;
4477 struct tcp_sack_block *sp = &tp->selective_acks[0];
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004478 struct tcp_sack_block *swalk = sp + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004479
4480 /* See if the recent change to the first SACK eats into
4481 * or hits the sequence space of other SACK blocks, if so coalesce.
4482 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004483 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004484 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4485 int i;
4486
4487 /* Zap SWALK, by moving every further SACK up by one slot.
4488 * Decrease num_sacks.
4489 */
4490 tp->rx_opt.num_sacks--;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004491 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4492 sp[i] = sp[i + 1];
Linus Torvalds1da177e2005-04-16 15:20:36 -07004493 continue;
4494 }
Julia Lawall44797582020-10-11 12:34:56 +02004495 this_sack++;
4496 swalk++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004497 }
4498}
4499
Eric Dumazet2b195852020-04-30 10:35:41 -07004500static void tcp_sack_compress_send_ack(struct sock *sk)
4501{
4502 struct tcp_sock *tp = tcp_sk(sk);
4503
4504 if (!tp->compressed_ack)
4505 return;
4506
4507 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
4508 __sock_put(sk);
4509
4510 /* Since we have to send one ack finally,
4511 * substract one from tp->compressed_ack to keep
4512 * LINUX_MIB_TCPACKCOMPRESSED accurate.
4513 */
4514 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
4515 tp->compressed_ack - 1);
4516
4517 tp->compressed_ack = 0;
4518 tcp_send_ack(sk);
4519}
4520
Eric Dumazetccd06282020-04-30 10:35:42 -07004521/* Reasonable amount of sack blocks included in TCP SACK option
4522 * The max is 4, but this becomes 3 if TCP timestamps are there.
4523 * Given that SACK packets might be lost, be conservative and use 2.
4524 */
4525#define TCP_SACK_BLOCKS_EXPECTED 2
4526
Linus Torvalds1da177e2005-04-16 15:20:36 -07004527static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4528{
4529 struct tcp_sock *tp = tcp_sk(sk);
4530 struct tcp_sack_block *sp = &tp->selective_acks[0];
4531 int cur_sacks = tp->rx_opt.num_sacks;
4532 int this_sack;
4533
4534 if (!cur_sacks)
4535 goto new_sack;
4536
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004537 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004538 if (tcp_sack_extend(sp, seq, end_seq)) {
Eric Dumazetccd06282020-04-30 10:35:42 -07004539 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED)
4540 tcp_sack_compress_send_ack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004541 /* Rotate this_sack to the first one. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004542 for (; this_sack > 0; this_sack--, sp--)
Ilpo Järvinena0bffff2009-03-21 13:36:17 -07004543 swap(*sp, *(sp - 1));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004544 if (cur_sacks > 1)
4545 tcp_sack_maybe_coalesce(tp);
4546 return;
4547 }
4548 }
4549
Eric Dumazetccd06282020-04-30 10:35:42 -07004550 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED)
4551 tcp_sack_compress_send_ack(sk);
4552
Linus Torvalds1da177e2005-04-16 15:20:36 -07004553 /* Could not find an adjacent existing SACK, build a new one,
4554 * put it at the front, and shift everyone else down. We
4555 * always know there is at least one SACK present already here.
4556 *
4557 * If the sack array is full, forget about the last one.
4558 */
Adam Langley4389dde2008-07-19 00:07:02 -07004559 if (this_sack >= TCP_NUM_SACKS) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004560 this_sack--;
4561 tp->rx_opt.num_sacks--;
4562 sp--;
4563 }
Stephen Hemminger2de979b2007-03-08 20:45:19 -08004564 for (; this_sack > 0; this_sack--, sp--)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004565 *sp = *(sp - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004566
4567new_sack:
4568 /* Build the new head SACK, and we're done. */
4569 sp->start_seq = seq;
4570 sp->end_seq = end_seq;
4571 tp->rx_opt.num_sacks++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004572}
4573
4574/* RCV.NXT advances, some SACKs should be eaten. */
4575
4576static void tcp_sack_remove(struct tcp_sock *tp)
4577{
4578 struct tcp_sack_block *sp = &tp->selective_acks[0];
4579 int num_sacks = tp->rx_opt.num_sacks;
4580 int this_sack;
4581
4582 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004583 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004584 tp->rx_opt.num_sacks = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004585 return;
4586 }
4587
Ilpo Järvinen056834d2007-12-31 14:57:14 -08004588 for (this_sack = 0; this_sack < num_sacks;) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004589 /* Check if the start of the sack is covered by RCV.NXT. */
4590 if (!before(tp->rcv_nxt, sp->start_seq)) {
4591 int i;
4592
4593 /* RCV.NXT must cover all the block! */
Ilpo Järvinen547b7922008-07-25 21:43:18 -07004594 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004595
4596 /* Zap this SACK, by moving forward any other SACKS. */
Weilong Chena22318e2013-12-23 14:37:26 +08004597 for (i = this_sack+1; i < num_sacks; i++)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004598 tp->selective_acks[i-1] = tp->selective_acks[i];
4599 num_sacks--;
4600 continue;
4601 }
4602 this_sack++;
4603 sp++;
4604 }
Ilpo Järvinen5861f8e2009-03-14 14:23:01 +00004605 tp->rx_opt.num_sacks = num_sacks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004606}
4607
Eric Dumazet1402d362012-04-23 07:11:42 +00004608/**
4609 * tcp_try_coalesce - try to merge skb to prior one
4610 * @sk: socket
4611 * @to: prior buffer
4612 * @from: buffer to add in queue
Eric Dumazet923dd342012-05-02 07:55:58 +00004613 * @fragstolen: pointer to boolean
Eric Dumazet1402d362012-04-23 07:11:42 +00004614 *
4615 * Before queueing skb @from after @to, try to merge them
4616 * to reduce overall memory use and queue lengths, if cost is small.
4617 * Packets in ofo or receive queues can stay a long time.
4618 * Better try to coalesce them right now to avoid future collapses.
Eric Dumazet783c1752012-04-23 17:34:36 +00004619 * Returns true if caller should free @from instead of queueing it
Eric Dumazet1402d362012-04-23 07:11:42 +00004620 */
Eric Dumazet783c1752012-04-23 17:34:36 +00004621static bool tcp_try_coalesce(struct sock *sk,
4622 struct sk_buff *to,
Eric Dumazet329033f2012-04-27 00:38:33 +00004623 struct sk_buff *from,
4624 bool *fragstolen)
Eric Dumazet1402d362012-04-23 07:11:42 +00004625{
Eric Dumazetbad43ca2012-05-19 03:02:02 +00004626 int delta;
Eric Dumazet1402d362012-04-23 07:11:42 +00004627
Eric Dumazet329033f2012-04-27 00:38:33 +00004628 *fragstolen = false;
Alexander Duyck34a802a2012-05-02 21:19:09 +00004629
Eric Dumazet1ca7ee32012-05-23 17:51:37 +00004630 /* Its possible this segment overlaps with prior segment in queue */
4631 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4632 return false;
4633
Mat Martineau85712482020-01-09 07:59:20 -08004634 if (!mptcp_skb_can_collapse(to, from))
4635 return false;
4636
Boris Pismenny41ed9c02018-07-13 14:33:38 +03004637#ifdef CONFIG_TLS_DEVICE
4638 if (from->decrypted != to->decrypted)
4639 return false;
4640#endif
4641
Eric Dumazetbad43ca2012-05-19 03:02:02 +00004642 if (!skb_try_coalesce(to, from, fragstolen, &delta))
Eric Dumazet783c1752012-04-23 17:34:36 +00004643 return false;
Alexander Duyck34a802a2012-05-02 21:19:09 +00004644
Alexander Duyck34a802a2012-05-02 21:19:09 +00004645 atomic_add(delta, &sk->sk_rmem_alloc);
4646 sk_mem_charge(sk, delta);
Eric Dumazetc10d9312016-04-29 14:16:47 -07004647 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
Alexander Duyck34a802a2012-05-02 21:19:09 +00004648 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4649 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
Eric Dumazete93a0432014-09-15 04:19:52 -07004650 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
Mike Maloney98aaa912017-08-22 17:08:48 -04004651
4652 if (TCP_SKB_CB(from)->has_rxtstamp) {
4653 TCP_SKB_CB(to)->has_rxtstamp = true;
Eric Dumazetbffa72c2017-09-19 05:14:24 -07004654 to->tstamp = from->tstamp;
Stephen Malloncadf9df2018-11-20 19:15:02 +11004655 skb_hwtstamps(to)->hwtstamp = skb_hwtstamps(from)->hwtstamp;
Mike Maloney98aaa912017-08-22 17:08:48 -04004656 }
4657
Alexander Duyck34a802a2012-05-02 21:19:09 +00004658 return true;
Eric Dumazet1402d362012-04-23 07:11:42 +00004659}
4660
Eric Dumazet58152ec2018-07-23 09:28:21 -07004661static bool tcp_ooo_try_coalesce(struct sock *sk,
4662 struct sk_buff *to,
4663 struct sk_buff *from,
4664 bool *fragstolen)
4665{
4666 bool res = tcp_try_coalesce(sk, to, from, fragstolen);
4667
4668 /* In case tcp_drop() is called later, update to->gso_segs */
4669 if (res) {
4670 u32 gso_segs = max_t(u16, 1, skb_shinfo(to)->gso_segs) +
4671 max_t(u16, 1, skb_shinfo(from)->gso_segs);
4672
4673 skb_shinfo(to)->gso_segs = min_t(u32, gso_segs, 0xFFFF);
4674 }
4675 return res;
4676}
4677
Eric Dumazet532182c2016-04-01 08:52:19 -07004678static void tcp_drop(struct sock *sk, struct sk_buff *skb)
4679{
4680 sk_drops_add(sk, skb);
4681 __kfree_skb(skb);
4682}
4683
Linus Torvalds1da177e2005-04-16 15:20:36 -07004684/* This one checks to see if we can put data from the
4685 * out_of_order queue into the receive_queue.
4686 */
4687static void tcp_ofo_queue(struct sock *sk)
4688{
4689 struct tcp_sock *tp = tcp_sk(sk);
4690 __u32 dsack_high = tp->rcv_nxt;
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004691 bool fin, fragstolen, eaten;
Eric Dumazetbd1e75a2014-09-19 08:26:20 -07004692 struct sk_buff *skb, *tail;
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004693 struct rb_node *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004694
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004695 p = rb_first(&tp->out_of_order_queue);
4696 while (p) {
Eric Dumazet18a4c0e2017-10-05 22:21:21 -07004697 skb = rb_to_skb(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004698 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4699 break;
4700
4701 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4702 __u32 dsack = dsack_high;
4703 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4704 dsack_high = TCP_SKB_CB(skb)->end_seq;
4705 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4706 }
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004707 p = rb_next(p);
4708 rb_erase(&skb->rbnode, &tp->out_of_order_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004709
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004710 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
Eric Dumazet532182c2016-04-01 08:52:19 -07004711 tcp_drop(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004712 continue;
4713 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004714
Eric Dumazetbd1e75a2014-09-19 08:26:20 -07004715 tail = skb_peek_tail(&sk->sk_receive_queue);
Eric Dumazetbffa72c2017-09-19 05:14:24 -07004716 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
Eric Dumazetbdd1f9e2015-04-28 15:28:18 -07004717 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004718 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
Eric Dumazetbd1e75a2014-09-19 08:26:20 -07004719 if (!eaten)
4720 __skb_queue_tail(&sk->sk_receive_queue, skb);
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004721 else
Eric Dumazetbd1e75a2014-09-19 08:26:20 -07004722 kfree_skb_partial(skb, fragstolen);
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004723
4724 if (unlikely(fin)) {
4725 tcp_fin(sk);
4726 /* tcp_fin() purges tp->out_of_order_queue,
4727 * so we must end this loop right now.
4728 */
4729 break;
4730 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004731 }
4732}
4733
4734static bool tcp_prune_ofo_queue(struct sock *sk);
4735static int tcp_prune_queue(struct sock *sk);
4736
4737static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4738 unsigned int size)
4739{
4740 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4741 !sk_rmem_schedule(sk, skb, size)) {
4742
4743 if (tcp_prune_queue(sk) < 0)
4744 return -1;
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07004745
Eric Dumazet36a65032016-08-17 14:17:09 -07004746 while (!sk_rmem_schedule(sk, skb, size)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004747 if (!tcp_prune_ofo_queue(sk))
4748 return -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004749 }
4750 }
4751 return 0;
4752}
4753
Eric Dumazete86b29192012-03-18 11:06:44 +00004754static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4755{
4756 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet18a4c0e2017-10-05 22:21:21 -07004757 struct rb_node **p, *parent;
Eric Dumazete86b29192012-03-18 11:06:44 +00004758 struct sk_buff *skb1;
4759 u32 seq, end_seq;
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004760 bool fragstolen;
Eric Dumazete86b29192012-03-18 11:06:44 +00004761
Yousuk Seungf4c9f852018-06-04 15:29:51 -07004762 tcp_ecn_check_ce(sk, skb);
Eric Dumazete86b29192012-03-18 11:06:44 +00004763
Mel Gormanc76562b2012-07-31 16:44:41 -07004764 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
Eric Dumazetc10d9312016-04-29 14:16:47 -07004765 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
Eric Dumazetba3bb0e2020-06-30 13:51:28 -07004766 sk->sk_data_ready(sk);
Eric Dumazet532182c2016-04-01 08:52:19 -07004767 tcp_drop(sk, skb);
Eric Dumazete86b29192012-03-18 11:06:44 +00004768 return;
4769 }
4770
Florian Westphal31770e32017-08-30 19:24:58 +02004771 /* Disable header prediction. */
4772 tp->pred_flags = 0;
Eric Dumazete86b29192012-03-18 11:06:44 +00004773 inet_csk_schedule_ack(sk);
4774
Thomas Higdonf9af2db2019-09-13 23:23:34 +00004775 tp->rcv_ooopack += max_t(u16, 1, skb_shinfo(skb)->gso_segs);
Eric Dumazetc10d9312016-04-29 14:16:47 -07004776 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004777 seq = TCP_SKB_CB(skb)->seq;
4778 end_seq = TCP_SKB_CB(skb)->end_seq;
Eric Dumazete86b29192012-03-18 11:06:44 +00004779
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004780 p = &tp->out_of_order_queue.rb_node;
4781 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
Eric Dumazete86b29192012-03-18 11:06:44 +00004782 /* Initial out of order segment, build 1 SACK. */
4783 if (tcp_is_sack(tp)) {
4784 tp->rx_opt.num_sacks = 1;
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004785 tp->selective_acks[0].start_seq = seq;
4786 tp->selective_acks[0].end_seq = end_seq;
Eric Dumazete86b29192012-03-18 11:06:44 +00004787 }
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004788 rb_link_node(&skb->rbnode, NULL, p);
4789 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4790 tp->ooo_last_skb = skb;
Eric Dumazete86b29192012-03-18 11:06:44 +00004791 goto end;
4792 }
4793
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004794 /* In the typical case, we are adding an skb to the end of the list.
4795 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4796 */
Eric Dumazet58152ec2018-07-23 09:28:21 -07004797 if (tcp_ooo_try_coalesce(sk, tp->ooo_last_skb,
4798 skb, &fragstolen)) {
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004799coalesce_done:
Eric Dumazet66205122020-06-15 20:37:07 -07004800 /* For non sack flows, do not grow window to force DUPACK
4801 * and trigger fast retransmit.
4802 */
4803 if (tcp_is_sack(tp))
Eric Dumazet240bfd12021-07-21 03:15:28 -07004804 tcp_grow_window(sk, skb, true);
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004805 kfree_skb_partial(skb, fragstolen);
4806 skb = NULL;
4807 goto add_sack;
Eric Dumazete86b29192012-03-18 11:06:44 +00004808 }
Eric Dumazet2594a2a2016-09-09 14:22:45 -07004809 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4810 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
4811 parent = &tp->ooo_last_skb->rbnode;
4812 p = &parent->rb_right;
4813 goto insert;
Eric Dumazete86b29192012-03-18 11:06:44 +00004814 }
4815
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004816 /* Find place to insert this segment. Handle overlaps on the way. */
4817 parent = NULL;
4818 while (*p) {
4819 parent = *p;
Eric Dumazet18a4c0e2017-10-05 22:21:21 -07004820 skb1 = rb_to_skb(parent);
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004821 if (before(seq, TCP_SKB_CB(skb1)->seq)) {
4822 p = &parent->rb_left;
4823 continue;
Eric Dumazete86b29192012-03-18 11:06:44 +00004824 }
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004825 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4826 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4827 /* All the bits are present. Drop. */
4828 NET_INC_STATS(sock_net(sk),
4829 LINUX_MIB_TCPOFOMERGE);
Eric Dumazet8541b212018-07-23 09:28:20 -07004830 tcp_drop(sk, skb);
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004831 skb = NULL;
4832 tcp_dsack_set(sk, seq, end_seq);
4833 goto add_sack;
4834 }
4835 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4836 /* Partial overlap. */
4837 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
4838 } else {
4839 /* skb's seq == skb1's seq and skb covers skb1.
4840 * Replace skb1 with skb.
4841 */
4842 rb_replace_node(&skb1->rbnode, &skb->rbnode,
4843 &tp->out_of_order_queue);
4844 tcp_dsack_extend(sk,
4845 TCP_SKB_CB(skb1)->seq,
4846 TCP_SKB_CB(skb1)->end_seq);
4847 NET_INC_STATS(sock_net(sk),
4848 LINUX_MIB_TCPOFOMERGE);
Eric Dumazet8541b212018-07-23 09:28:20 -07004849 tcp_drop(sk, skb1);
Eric Dumazet76f0dcb2016-09-13 22:55:05 -07004850 goto merge_right;
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004851 }
Eric Dumazet58152ec2018-07-23 09:28:21 -07004852 } else if (tcp_ooo_try_coalesce(sk, skb1,
4853 skb, &fragstolen)) {
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004854 goto coalesce_done;
4855 }
4856 p = &parent->rb_right;
Eric Dumazete86b29192012-03-18 11:06:44 +00004857 }
Eric Dumazet2594a2a2016-09-09 14:22:45 -07004858insert:
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004859 /* Insert segment into RB tree. */
4860 rb_link_node(&skb->rbnode, parent, p);
4861 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
Eric Dumazete86b29192012-03-18 11:06:44 +00004862
Eric Dumazet76f0dcb2016-09-13 22:55:05 -07004863merge_right:
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004864 /* Remove other segments covered by skb. */
Eric Dumazet18a4c0e2017-10-05 22:21:21 -07004865 while ((skb1 = skb_rb_next(skb)) != NULL) {
Eric Dumazete86b29192012-03-18 11:06:44 +00004866 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4867 break;
4868 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4869 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4870 end_seq);
4871 break;
4872 }
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004873 rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
Eric Dumazete86b29192012-03-18 11:06:44 +00004874 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4875 TCP_SKB_CB(skb1)->end_seq);
Eric Dumazetc10d9312016-04-29 14:16:47 -07004876 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
Eric Dumazet532182c2016-04-01 08:52:19 -07004877 tcp_drop(sk, skb1);
Eric Dumazete86b29192012-03-18 11:06:44 +00004878 }
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004879 /* If there is no skb after us, we are the last_skb ! */
Eric Dumazet18a4c0e2017-10-05 22:21:21 -07004880 if (!skb1)
Yaogong Wang9f5afea2016-09-07 14:49:28 -07004881 tp->ooo_last_skb = skb;
Eric Dumazete86b29192012-03-18 11:06:44 +00004882
4883add_sack:
4884 if (tcp_is_sack(tp))
4885 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4886end:
Eric Dumazet4e4f1fc2013-09-06 10:35:58 -07004887 if (skb) {
Eric Dumazet66205122020-06-15 20:37:07 -07004888 /* For non sack flows, do not grow window to force DUPACK
4889 * and trigger fast retransmit.
4890 */
4891 if (tcp_is_sack(tp))
Eric Dumazet240bfd12021-07-21 03:15:28 -07004892 tcp_grow_window(sk, skb, false);
Eric Dumazet60b1af32017-01-24 14:57:36 -08004893 skb_condense(skb);
Eric Dumazete86b29192012-03-18 11:06:44 +00004894 skb_set_owner_r(skb, sk);
Eric Dumazet4e4f1fc2013-09-06 10:35:58 -07004895 }
Eric Dumazete86b29192012-03-18 11:06:44 +00004896}
4897
Eric Dumazete7395f1f2018-11-26 14:49:12 -08004898static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb,
4899 bool *fragstolen)
Eric Dumazetb081f852012-05-02 09:58:29 +00004900{
4901 int eaten;
4902 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4903
Eric Dumazetb081f852012-05-02 09:58:29 +00004904 eaten = (tail &&
Eric Dumazetbffa72c2017-09-19 05:14:24 -07004905 tcp_try_coalesce(sk, tail,
Mike Maloney98aaa912017-08-22 17:08:48 -04004906 skb, fragstolen)) ? 1 : 0;
Eric Dumazetbdd1f9e2015-04-28 15:28:18 -07004907 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
Eric Dumazetb081f852012-05-02 09:58:29 +00004908 if (!eaten) {
4909 __skb_queue_tail(&sk->sk_receive_queue, skb);
4910 skb_set_owner_r(skb, sk);
4911 }
4912 return eaten;
4913}
Eric Dumazete86b29192012-03-18 11:06:44 +00004914
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004915int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4916{
Eric Dumazetcb934712014-09-17 03:14:42 -07004917 struct sk_buff *skb;
Eric Dumazet5d4c9bf2015-11-18 21:03:33 -08004918 int err = -ENOMEM;
4919 int data_len = 0;
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004920 bool fragstolen;
4921
Pavel Emelyanovc454e612012-10-29 05:05:33 +00004922 if (size == 0)
4923 return 0;
4924
Eric Dumazet5d4c9bf2015-11-18 21:03:33 -08004925 if (size > PAGE_SIZE) {
4926 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4927
4928 data_len = npages << PAGE_SHIFT;
4929 size = data_len + (size & ~PAGE_MASK);
4930 }
4931 skb = alloc_skb_with_frags(size - data_len, data_len,
4932 PAGE_ALLOC_COSTLY_ORDER,
4933 &err, sk->sk_allocation);
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004934 if (!skb)
4935 goto err;
4936
Eric Dumazet5d4c9bf2015-11-18 21:03:33 -08004937 skb_put(skb, size - data_len);
4938 skb->data_len = data_len;
4939 skb->len = size;
4940
Yafang Shaoea5d0c32018-06-28 00:22:56 -04004941 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
4942 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
Mel Gormanc76562b2012-07-31 16:44:41 -07004943 goto err_free;
Yafang Shaoea5d0c32018-06-28 00:22:56 -04004944 }
Mel Gormanc76562b2012-07-31 16:44:41 -07004945
Eric Dumazet5d4c9bf2015-11-18 21:03:33 -08004946 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
4947 if (err)
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004948 goto err_free;
4949
4950 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4951 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4952 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4953
Eric Dumazete7395f1f2018-11-26 14:49:12 -08004954 if (tcp_queue_rcv(sk, skb, &fragstolen)) {
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004955 WARN_ON_ONCE(fragstolen); /* should not happen */
4956 __kfree_skb(skb);
4957 }
4958 return size;
4959
4960err_free:
4961 kfree_skb(skb);
4962err:
Eric Dumazet5d4c9bf2015-11-18 21:03:33 -08004963 return err;
4964
Pavel Emelyanov292e8d82012-05-10 01:49:41 +00004965}
4966
Eric Dumazet03f45c82018-04-16 10:33:37 -07004967void tcp_data_ready(struct sock *sk)
4968{
Eric Dumazet39354eb2021-02-13 06:26:34 -08004969 if (tcp_epollin_ready(sk, sk->sk_rcvlowat) || sock_flag(sk, SOCK_DONE))
Eric Dumazet05dc72a2021-02-12 15:22:14 -08004970 sk->sk_data_ready(sk);
Eric Dumazet03f45c82018-04-16 10:33:37 -07004971}
4972
Linus Torvalds1da177e2005-04-16 15:20:36 -07004973static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4974{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004975 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet5357f0b2017-08-01 07:02:44 -07004976 bool fragstolen;
4977 int eaten;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004978
Jianguo Wu6787b7e2021-07-09 17:20:49 -07004979 /* If a subflow has been reset, the packet should not continue
4980 * to be processed, drop the packet.
4981 */
4982 if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb)) {
4983 __kfree_skb(skb);
4984 return;
4985 }
Mat Martineau648ef4b2020-01-21 16:56:24 -08004986
Eric Dumazet532182c2016-04-01 08:52:19 -07004987 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
4988 __kfree_skb(skb);
4989 return;
4990 }
Eric Dumazetf84af322010-04-28 15:31:51 -07004991 skb_dst_drop(skb);
Peter Pan(潘卫平)155c6e12014-09-24 22:17:02 +08004992 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004993
Ilpo Järvinen5861f8e2009-03-14 14:23:01 +00004994 tp->rx_opt.dsack = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004995
4996 /* Queue data for delivery to the user.
4997 * Packets in sequence go to the receive queue.
4998 * Out of sequence packets to the out_of_order_queue.
4999 */
5000 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
Yafang Shaofb223502018-06-24 10:02:54 -04005001 if (tcp_receive_window(tp) == 0) {
5002 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005003 goto out_of_window;
Yafang Shaofb223502018-06-24 10:02:54 -04005004 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07005005
5006 /* Ok. In sequence. In window. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005007queue_and_out:
Eric Dumazet5357f0b2017-08-01 07:02:44 -07005008 if (skb_queue_len(&sk->sk_receive_queue) == 0)
5009 sk_forced_mem_schedule(sk, skb->truesize);
Yafang Shaoea5d0c32018-06-28 00:22:56 -04005010 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
5011 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
Eric Dumazetba3bb0e2020-06-30 13:51:28 -07005012 sk->sk_data_ready(sk);
Eric Dumazet5357f0b2017-08-01 07:02:44 -07005013 goto drop;
Yafang Shaoea5d0c32018-06-28 00:22:56 -04005014 }
Eric Dumazet5357f0b2017-08-01 07:02:44 -07005015
Eric Dumazete7395f1f2018-11-26 14:49:12 -08005016 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
Stephen Hemminger2de979b2007-03-08 20:45:19 -08005017 if (skb->len)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005018 tcp_event_data_recv(sk, skb);
Peter Pan(潘卫平)155c6e12014-09-24 22:17:02 +08005019 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
Eric Dumazet20c4cb72011-10-20 17:44:03 -04005020 tcp_fin(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005021
Yaogong Wang9f5afea2016-09-07 14:49:28 -07005022 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005023 tcp_ofo_queue(sk);
5024
Yuchung Cheng15bdd562018-08-09 09:38:11 -07005025 /* RFC5681. 4.2. SHOULD send immediate ACK, when
Linus Torvalds1da177e2005-04-16 15:20:36 -07005026 * gap in queue is filled.
5027 */
Yaogong Wang9f5afea2016-09-07 14:49:28 -07005028 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
Yuchung Cheng15bdd562018-08-09 09:38:11 -07005029 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005030 }
5031
5032 if (tp->rx_opt.num_sacks)
5033 tcp_sack_remove(tp);
5034
Florian Westphal31770e32017-08-30 19:24:58 +02005035 tcp_fast_path_check(sk);
5036
Eric Dumazet923dd342012-05-02 07:55:58 +00005037 if (eaten > 0)
5038 kfree_skb_partial(skb, fragstolen);
Eric Dumazet1d57f192012-09-17 12:51:39 +00005039 if (!sock_flag(sk, SOCK_DEAD))
Eric Dumazet03f45c82018-04-16 10:33:37 -07005040 tcp_data_ready(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005041 return;
5042 }
5043
5044 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
Yuchung Cheng77881742018-08-29 14:53:56 -07005045 tcp_rcv_spurious_retrans(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005046 /* A retransmit, 2nd most common case. Force an immediate ack. */
Eric Dumazetc10d9312016-04-29 14:16:47 -07005047 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07005048 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005049
5050out_of_window:
Eric Dumazet9a9c9b52018-05-21 15:08:56 -07005051 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005052 inet_csk_schedule_ack(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005053drop:
Eric Dumazet532182c2016-04-01 08:52:19 -07005054 tcp_drop(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005055 return;
5056 }
5057
5058 /* Out of window. F.e. zero window probe. */
5059 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
5060 goto out_of_window;
5061
Linus Torvalds1da177e2005-04-16 15:20:36 -07005062 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5063 /* Partial packet, seq < rcv_next < end_seq */
Pavel Emelyanov1ed83462008-07-16 20:29:51 -07005064 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005065
Linus Torvalds1da177e2005-04-16 15:20:36 -07005066 /* If window is closed, drop tail of packet. But after
5067 * remembering D-SACK for its head made in previous line.
5068 */
Yafang Shaofb223502018-06-24 10:02:54 -04005069 if (!tcp_receive_window(tp)) {
5070 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005071 goto out_of_window;
Yafang Shaofb223502018-06-24 10:02:54 -04005072 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07005073 goto queue_and_out;
5074 }
5075
Eric Dumazete86b29192012-03-18 11:06:44 +00005076 tcp_data_queue_ofo(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005077}
5078
Yaogong Wang9f5afea2016-09-07 14:49:28 -07005079static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
Ilpo Järvinen2cf46632008-08-23 05:11:41 -07005080{
Yaogong Wang9f5afea2016-09-07 14:49:28 -07005081 if (list)
5082 return !skb_queue_is_last(list, skb) ? skb->next : NULL;
David S. Miller91521942009-05-28 21:35:47 -07005083
Eric Dumazet18a4c0e2017-10-05 22:21:21 -07005084 return skb_rb_next(skb);
Yaogong Wang9f5afea2016-09-07 14:49:28 -07005085}
Ilpo Järvinen2cf46632008-08-23 05:11:41 -07005086
Yaogong Wang9f5afea2016-09-07 14:49:28 -07005087static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
5088 struct sk_buff_head *list,
5089 struct rb_root *root)
5090{
5091 struct sk_buff *next = tcp_skb_next(skb, list);
5092
5093 if (list)
5094 __skb_unlink(skb, list);
5095 else
5096 rb_erase(&skb->rbnode, root);
5097
Ilpo Järvinen2cf46632008-08-23 05:11:41 -07005098 __kfree_skb(skb);
Eric Dumazetc10d9312016-04-29 14:16:47 -07005099 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
Ilpo Järvinen2cf46632008-08-23 05:11:41 -07005100
5101 return next;
5102}
5103
Yaogong Wang9f5afea2016-09-07 14:49:28 -07005104/* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
Eric Dumazet75c119a2017-10-05 22:21:27 -07005105void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
Yaogong Wang9f5afea2016-09-07 14:49:28 -07005106{
5107 struct rb_node **p = &root->rb_node;
5108 struct rb_node *parent = NULL;
5109 struct sk_buff *skb1;
5110
5111 while (*p) {
5112 parent = *p;
Eric Dumazet18a4c0e2017-10-05 22:21:21 -07005113 skb1 = rb_to_skb(parent);
Yaogong Wang9f5afea2016-09-07 14:49:28 -07005114 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
5115 p = &parent->rb_left;
5116 else
5117 p = &parent->rb_right;
5118 }
5119 rb_link_node(&skb->rbnode, parent, p);
5120 rb_insert_color(&skb->rbnode, root);
5121}
5122
Linus Torvalds1da177e2005-04-16 15:20:36 -07005123/* Collapse contiguous sequence of skbs head..tail with
5124 * sequence numbers start..end.
David S. Miller91521942009-05-28 21:35:47 -07005125 *
Yaogong Wang9f5afea2016-09-07 14:49:28 -07005126 * If tail is NULL, this means until the end of the queue.
David S. Miller91521942009-05-28 21:35:47 -07005127 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07005128 * Segments with FIN/SYN are not collapsed (only because this
5129 * simplifies code)
5130 */
5131static void
Yaogong Wang9f5afea2016-09-07 14:49:28 -07005132tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
5133 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005134{
Yaogong Wang9f5afea2016-09-07 14:49:28 -07005135 struct sk_buff *skb = head, *n;
5136 struct sk_buff_head tmp;
David S. Miller91521942009-05-28 21:35:47 -07005137 bool end_of_skbs;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005138
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08005139 /* First, check that queue is collapsible and find
Yaogong Wang9f5afea2016-09-07 14:49:28 -07005140 * the point where collapsing can be useful.
5141 */
David S. Miller91521942009-05-28 21:35:47 -07005142restart:
Yaogong Wang9f5afea2016-09-07 14:49:28 -07005143 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
5144 n = tcp_skb_next(skb, list);
5145
Linus Torvalds1da177e2005-04-16 15:20:36 -07005146 /* No new bits? It is possible on ofo queue. */
5147 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
Yaogong Wang9f5afea2016-09-07 14:49:28 -07005148 skb = tcp_collapse_one(sk, skb, list, root);
David S. Miller91521942009-05-28 21:35:47 -07005149 if (!skb)
5150 break;
5151 goto restart;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005152 }
5153
5154 /* The first skb to collapse is:
5155 * - not SYN/FIN and
5156 * - bloated or contains data before "start" or
Mat Martineau85712482020-01-09 07:59:20 -08005157 * overlaps to the next one and mptcp allow collapsing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005158 */
Eric Dumazete11ecdd2014-09-15 04:19:51 -07005159 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
Eric Dumazet94f08932017-10-26 21:55:09 -07005160 (tcp_win_from_space(sk, skb->truesize) > skb->len ||
David S. Miller91521942009-05-28 21:35:47 -07005161 before(TCP_SKB_CB(skb)->seq, start))) {
5162 end_of_skbs = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005163 break;
David S. Miller91521942009-05-28 21:35:47 -07005164 }
5165
Mat Martineau85712482020-01-09 07:59:20 -08005166 if (n && n != tail && mptcp_skb_can_collapse(skb, n) &&
Yaogong Wang9f5afea2016-09-07 14:49:28 -07005167 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
5168 end_of_skbs = false;
5169 break;
David S. Miller91521942009-05-28 21:35:47 -07005170 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07005171
5172 /* Decided to skip this, advance start seq. */
5173 start = TCP_SKB_CB(skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005174 }
Eric Dumazete11ecdd2014-09-15 04:19:51 -07005175 if (end_of_skbs ||
5176 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005177 return;
5178
Yaogong Wang9f5afea2016-09-07 14:49:28 -07005179 __skb_queue_head_init(&tmp);
5180
Linus Torvalds1da177e2005-04-16 15:20:36 -07005181 while (before(start, end)) {
Eric Dumazetb3d6cb92014-09-15 04:19:53 -07005182 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005183 struct sk_buff *nskb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005184
Eric Dumazetb3d6cb92014-09-15 04:19:53 -07005185 nskb = alloc_skb(copy, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005186 if (!nskb)
Yaogong Wang9f5afea2016-09-07 14:49:28 -07005187 break;
Arnaldo Carvalho de Meloc51957d2007-03-10 12:47:22 -03005188
Linus Torvalds1da177e2005-04-16 15:20:36 -07005189 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
Boris Pismenny41ed9c02018-07-13 14:33:38 +03005190#ifdef CONFIG_TLS_DEVICE
5191 nskb->decrypted = skb->decrypted;
5192#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07005193 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
Yaogong Wang9f5afea2016-09-07 14:49:28 -07005194 if (list)
5195 __skb_queue_before(list, skb, nskb);
5196 else
5197 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
Hideo Aoki3ab224b2007-12-31 00:11:19 -08005198 skb_set_owner_r(nskb, sk);
Mat Martineau85712482020-01-09 07:59:20 -08005199 mptcp_skb_ext_move(nskb, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005200
5201 /* Copy data, releasing collapsed skbs. */
5202 while (copy > 0) {
5203 int offset = start - TCP_SKB_CB(skb)->seq;
5204 int size = TCP_SKB_CB(skb)->end_seq - start;
5205
Kris Katterjohn09a62662006-01-08 22:24:28 -08005206 BUG_ON(offset < 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005207 if (size > 0) {
5208 size = min(copy, size);
5209 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
5210 BUG();
5211 TCP_SKB_CB(nskb)->end_seq += size;
5212 copy -= size;
5213 start += size;
5214 }
5215 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
Yaogong Wang9f5afea2016-09-07 14:49:28 -07005216 skb = tcp_collapse_one(sk, skb, list, root);
David S. Miller91521942009-05-28 21:35:47 -07005217 if (!skb ||
5218 skb == tail ||
Mat Martineau85712482020-01-09 07:59:20 -08005219 !mptcp_skb_can_collapse(nskb, skb) ||
Eric Dumazete11ecdd2014-09-15 04:19:51 -07005220 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
Yaogong Wang9f5afea2016-09-07 14:49:28 -07005221 goto end;
Boris Pismenny41ed9c02018-07-13 14:33:38 +03005222#ifdef CONFIG_TLS_DEVICE
5223 if (skb->decrypted != nskb->decrypted)
5224 goto end;
5225#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07005226 }
5227 }
5228 }
Yaogong Wang9f5afea2016-09-07 14:49:28 -07005229end:
5230 skb_queue_walk_safe(&tmp, skb, n)
5231 tcp_rbtree_insert(root, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005232}
5233
5234/* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
5235 * and tcp_collapse() them until all the queue is collapsed.
5236 */
5237static void tcp_collapse_ofo_queue(struct sock *sk)
5238{
5239 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet3d4bf932018-07-23 09:28:19 -07005240 u32 range_truesize, sum_tiny = 0;
Yaogong Wang9f5afea2016-09-07 14:49:28 -07005241 struct sk_buff *skb, *head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005242 u32 start, end;
5243
Eric Dumazet18a4c0e2017-10-05 22:21:21 -07005244 skb = skb_rb_first(&tp->out_of_order_queue);
Yaogong Wang9f5afea2016-09-07 14:49:28 -07005245new_range:
5246 if (!skb) {
Eric Dumazet18a4c0e2017-10-05 22:21:21 -07005247 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005248 return;
Yaogong Wang9f5afea2016-09-07 14:49:28 -07005249 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07005250 start = TCP_SKB_CB(skb)->seq;
5251 end = TCP_SKB_CB(skb)->end_seq;
Eric Dumazet3d4bf932018-07-23 09:28:19 -07005252 range_truesize = skb->truesize;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005253
Yaogong Wang9f5afea2016-09-07 14:49:28 -07005254 for (head = skb;;) {
Eric Dumazet18a4c0e2017-10-05 22:21:21 -07005255 skb = skb_rb_next(skb);
David S. Miller91521942009-05-28 21:35:47 -07005256
Yaogong Wang9f5afea2016-09-07 14:49:28 -07005257 /* Range is terminated when we see a gap or when
5258 * we are at the queue end.
5259 */
David S. Miller91521942009-05-28 21:35:47 -07005260 if (!skb ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07005261 after(TCP_SKB_CB(skb)->seq, end) ||
5262 before(TCP_SKB_CB(skb)->end_seq, start)) {
Eric Dumazet3d4bf932018-07-23 09:28:19 -07005263 /* Do not attempt collapsing tiny skbs */
5264 if (range_truesize != head->truesize ||
5265 end - start >= SKB_WITH_OVERHEAD(SK_MEM_QUANTUM)) {
5266 tcp_collapse(sk, NULL, &tp->out_of_order_queue,
5267 head, skb, start, end);
5268 } else {
5269 sum_tiny += range_truesize;
5270 if (sum_tiny > sk->sk_rcvbuf >> 3)
5271 return;
5272 }
Yaogong Wang9f5afea2016-09-07 14:49:28 -07005273 goto new_range;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005274 }
Yaogong Wang9f5afea2016-09-07 14:49:28 -07005275
Eric Dumazet3d4bf932018-07-23 09:28:19 -07005276 range_truesize += skb->truesize;
Yaogong Wang9f5afea2016-09-07 14:49:28 -07005277 if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
5278 start = TCP_SKB_CB(skb)->seq;
5279 if (after(TCP_SKB_CB(skb)->end_seq, end))
5280 end = TCP_SKB_CB(skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005281 }
5282}
5283
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07005284/*
Eric Dumazet36a65032016-08-17 14:17:09 -07005285 * Clean the out-of-order queue to make room.
5286 * We drop high sequences packets to :
5287 * 1) Let a chance for holes to be filled.
5288 * 2) not add too big latencies if thousands of packets sit there.
5289 * (But if application shrinks SO_RCVBUF, we could still end up
5290 * freeing whole queue here)
Eric Dumazet72cd43b2018-07-23 09:28:17 -07005291 * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks.
Eric Dumazet36a65032016-08-17 14:17:09 -07005292 *
5293 * Return true if queue has shrunk.
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07005294 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00005295static bool tcp_prune_ofo_queue(struct sock *sk)
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07005296{
5297 struct tcp_sock *tp = tcp_sk(sk);
Yaogong Wang9f5afea2016-09-07 14:49:28 -07005298 struct rb_node *node, *prev;
Eric Dumazet72cd43b2018-07-23 09:28:17 -07005299 int goal;
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07005300
Yaogong Wang9f5afea2016-09-07 14:49:28 -07005301 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
Eric Dumazet36a65032016-08-17 14:17:09 -07005302 return false;
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07005303
Eric Dumazet36a65032016-08-17 14:17:09 -07005304 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
Eric Dumazet72cd43b2018-07-23 09:28:17 -07005305 goal = sk->sk_rcvbuf >> 3;
Yaogong Wang9f5afea2016-09-07 14:49:28 -07005306 node = &tp->ooo_last_skb->rbnode;
5307 do {
5308 prev = rb_prev(node);
5309 rb_erase(node, &tp->out_of_order_queue);
Eric Dumazet72cd43b2018-07-23 09:28:17 -07005310 goal -= rb_to_skb(node)->truesize;
Eric Dumazet18a4c0e2017-10-05 22:21:21 -07005311 tcp_drop(sk, rb_to_skb(node));
Eric Dumazet72cd43b2018-07-23 09:28:17 -07005312 if (!prev || goal <= 0) {
5313 sk_mem_reclaim(sk);
5314 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
5315 !tcp_under_memory_pressure(sk))
5316 break;
5317 goal = sk->sk_rcvbuf >> 3;
5318 }
Yaogong Wang9f5afea2016-09-07 14:49:28 -07005319 node = prev;
5320 } while (node);
Eric Dumazet18a4c0e2017-10-05 22:21:21 -07005321 tp->ooo_last_skb = rb_to_skb(prev);
Eric Dumazet36a65032016-08-17 14:17:09 -07005322
5323 /* Reset SACK state. A conforming SACK implementation will
5324 * do the same at a timeout based retransmit. When a connection
5325 * is in a sad state like this, we care only about integrity
5326 * of the connection not performance.
5327 */
5328 if (tp->rx_opt.sack_ok)
5329 tcp_sack_reset(&tp->rx_opt);
5330 return true;
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07005331}
5332
Linus Torvalds1da177e2005-04-16 15:20:36 -07005333/* Reduce allocated memory if we can, trying to get
5334 * the socket within its memory limits again.
5335 *
5336 * Return less than zero if we should start dropping frames
5337 * until the socket owning process reads some of the data
5338 * to stabilize the situation.
5339 */
5340static int tcp_prune_queue(struct sock *sk)
5341{
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005342 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005343
Eric Dumazetc10d9312016-04-29 14:16:47 -07005344 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005345
5346 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005347 tcp_clamp_window(sk);
Eric Dumazetb8da51e2015-05-15 12:39:27 -07005348 else if (tcp_under_memory_pressure(sk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005349 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
5350
Eric Dumazetf4a33132018-07-23 09:28:18 -07005351 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5352 return 0;
5353
Linus Torvalds1da177e2005-04-16 15:20:36 -07005354 tcp_collapse_ofo_queue(sk);
David S. Miller91521942009-05-28 21:35:47 -07005355 if (!skb_queue_empty(&sk->sk_receive_queue))
Yaogong Wang9f5afea2016-09-07 14:49:28 -07005356 tcp_collapse(sk, &sk->sk_receive_queue, NULL,
David S. Miller91521942009-05-28 21:35:47 -07005357 skb_peek(&sk->sk_receive_queue),
5358 NULL,
5359 tp->copied_seq, tp->rcv_nxt);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08005360 sk_mem_reclaim(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005361
5362 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5363 return 0;
5364
5365 /* Collapsing did not help, destructive actions follow.
5366 * This must not ever occur. */
5367
Vitaliy Gusevb000cd32008-04-15 00:33:38 -07005368 tcp_prune_ofo_queue(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005369
5370 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5371 return 0;
5372
5373 /* If we are really being abused, tell the caller to silently
5374 * drop receive data on the floor. It will get retransmitted
5375 * and hopefully then we'll have sufficient space.
5376 */
Eric Dumazetc10d9312016-04-29 14:16:47 -07005377 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005378
5379 /* Massive buffer overcommit. */
Florian Westphal31770e32017-08-30 19:24:58 +02005380 tp->pred_flags = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005381 return -1;
5382}
5383
Eric Dumazeta2a385d2012-05-16 23:15:34 +00005384static bool tcp_should_expand_sndbuf(const struct sock *sk)
David S. Miller0d9901d2005-07-05 15:21:10 -07005385{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04005386 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005387
David S. Miller0d9901d2005-07-05 15:21:10 -07005388 /* If the user specified a specific send buffer setting, do
5389 * not modify it.
5390 */
5391 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00005392 return false;
David S. Miller0d9901d2005-07-05 15:21:10 -07005393
5394 /* If we are under global TCP memory pressure, do not expand. */
Eric Dumazetb8da51e2015-05-15 12:39:27 -07005395 if (tcp_under_memory_pressure(sk))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00005396 return false;
David S. Miller0d9901d2005-07-05 15:21:10 -07005397
5398 /* If we are under soft global TCP memory pressure, do not expand. */
Glauber Costa180d8cd2011-12-11 21:47:02 +00005399 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00005400 return false;
David S. Miller0d9901d2005-07-05 15:21:10 -07005401
5402 /* If we filled the congestion window, do not expand. */
Neal Cardwell65148902015-02-20 13:33:16 -05005403 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00005404 return false;
David S. Miller0d9901d2005-07-05 15:21:10 -07005405
Eric Dumazeta2a385d2012-05-16 23:15:34 +00005406 return true;
David S. Miller0d9901d2005-07-05 15:21:10 -07005407}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005408
Linus Torvalds1da177e2005-04-16 15:20:36 -07005409static void tcp_new_space(struct sock *sk)
5410{
5411 struct tcp_sock *tp = tcp_sk(sk);
5412
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005413 if (tcp_should_expand_sndbuf(sk)) {
Eric Dumazet6ae70532013-10-01 10:23:44 -07005414 tcp_sndbuf_expand(sk);
Eric Dumazetc2203cf2017-05-16 14:00:04 -07005415 tp->snd_cwnd_stamp = tcp_jiffies32;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005416 }
5417
Eric Dumazet739b2ad2021-07-21 02:06:14 -07005418 INDIRECT_CALL_1(sk->sk_write_space, sk_stream_write_space, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005419}
5420
Stephen Hemminger40efc6f2006-01-03 16:03:49 -08005421static void tcp_check_space(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005422{
Eric Dumazet0cbe6a82020-09-14 03:20:27 -07005423 /* pairs with tcp_poll() */
5424 smp_mb();
5425 if (sk->sk_socket &&
5426 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
5427 tcp_new_space(sk);
5428 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5429 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005430 }
5431}
5432
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005433static inline void tcp_data_snd_check(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005434{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005435 tcp_push_pending_frames(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005436 tcp_check_space(sk);
5437}
5438
5439/*
5440 * Check if sending an ack is needed.
5441 */
5442static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5443{
5444 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet5d9f4262018-05-17 14:47:26 -07005445 unsigned long rtt, delay;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005446
5447 /* More than one full frame received... */
Joe Perches9d4fb272009-11-23 10:41:23 -08005448 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07005449 /* ... and right edge of window advances far enough.
Eric Dumazet796f82e2018-04-16 10:33:36 -07005450 * (tcp_recvmsg() will send ACK otherwise).
5451 * If application uses SO_RCVLOWAT, we want send ack now if
5452 * we have not received enough bytes to satisfy the condition.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005453 */
Eric Dumazet796f82e2018-04-16 10:33:36 -07005454 (tp->rcv_nxt - tp->copied_seq < sk->sk_rcvlowat ||
5455 __tcp_select_window(sk) >= tp->rcv_wnd)) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07005456 /* We ACK each frame or... */
Yuchung Cheng466466d2018-08-09 09:38:09 -07005457 tcp_in_quickack_mode(sk) ||
5458 /* Protocol state mandates a one-time immediate ACK */
5459 inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOW) {
Eric Dumazet5d9f4262018-05-17 14:47:26 -07005460send_now:
Linus Torvalds1da177e2005-04-16 15:20:36 -07005461 tcp_send_ack(sk);
Eric Dumazet5d9f4262018-05-17 14:47:26 -07005462 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005463 }
Eric Dumazet5d9f4262018-05-17 14:47:26 -07005464
5465 if (!ofo_possible || RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5466 tcp_send_delayed_ack(sk);
5467 return;
5468 }
5469
Eric Dumazet9c21d2f2018-05-17 14:47:29 -07005470 if (!tcp_is_sack(tp) ||
5471 tp->compressed_ack >= sock_net(sk)->ipv4.sysctl_tcp_comp_sack_nr)
Eric Dumazet5d9f4262018-05-17 14:47:26 -07005472 goto send_now;
Eric Dumazet86de5922018-11-20 05:53:59 -08005473
5474 if (tp->compressed_ack_rcv_nxt != tp->rcv_nxt) {
5475 tp->compressed_ack_rcv_nxt = tp->rcv_nxt;
Eric Dumazet2b195852020-04-30 10:35:41 -07005476 tp->dup_ack_counter = 0;
Eric Dumazet86de5922018-11-20 05:53:59 -08005477 }
Eric Dumazet2b195852020-04-30 10:35:41 -07005478 if (tp->dup_ack_counter < TCP_FASTRETRANS_THRESH) {
5479 tp->dup_ack_counter++;
Eric Dumazet86de5922018-11-20 05:53:59 -08005480 goto send_now;
Eric Dumazet2b195852020-04-30 10:35:41 -07005481 }
5482 tp->compressed_ack++;
Eric Dumazet5d9f4262018-05-17 14:47:26 -07005483 if (hrtimer_is_queued(&tp->compressed_ack_timer))
5484 return;
5485
Eric Dumazet6d82aa22018-05-17 14:47:28 -07005486 /* compress ack timer : 5 % of rtt, but no more than tcp_comp_sack_delay_ns */
Eric Dumazet5d9f4262018-05-17 14:47:26 -07005487
5488 rtt = tp->rcv_rtt_est.rtt_us;
5489 if (tp->srtt_us && tp->srtt_us < rtt)
5490 rtt = tp->srtt_us;
5491
Eric Dumazet6d82aa22018-05-17 14:47:28 -07005492 delay = min_t(unsigned long, sock_net(sk)->ipv4.sysctl_tcp_comp_sack_delay_ns,
Eric Dumazet5d9f4262018-05-17 14:47:26 -07005493 rtt * (NSEC_PER_USEC >> 3)/20);
5494 sock_hold(sk);
Eric Dumazeta70437c2020-04-30 10:35:43 -07005495 hrtimer_start_range_ns(&tp->compressed_ack_timer, ns_to_ktime(delay),
5496 sock_net(sk)->ipv4.sysctl_tcp_comp_sack_slack_ns,
5497 HRTIMER_MODE_REL_PINNED_SOFT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005498}
5499
Stephen Hemminger40efc6f2006-01-03 16:03:49 -08005500static inline void tcp_ack_snd_check(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005501{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005502 if (!inet_csk_ack_scheduled(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005503 /* We sent a data segment already. */
5504 return;
5505 }
5506 __tcp_ack_snd_check(sk, 1);
5507}
5508
5509/*
5510 * This routine is only called when we have urgent data
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08005511 * signaled. Its the 'slow' part of tcp_urg. It could be
Linus Torvalds1da177e2005-04-16 15:20:36 -07005512 * moved inline now as tcp_urg is only called from one
5513 * place. We handle URGent data wrong. We have to - as
5514 * BSD still doesn't use the correction from RFC961.
5515 * For 1003.1g we should support a new option TCP_STDURG to permit
5516 * either form (or just set the sysctl tcp_stdurg).
5517 */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005518
Eric Dumazetcf533ea2011-10-21 05:22:42 -04005519static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005520{
5521 struct tcp_sock *tp = tcp_sk(sk);
5522 u32 ptr = ntohs(th->urg_ptr);
5523
Eric Dumazet3f4c7c62017-10-26 21:55:01 -07005524 if (ptr && !sock_net(sk)->ipv4.sysctl_tcp_stdurg)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005525 ptr--;
5526 ptr += ntohl(th->seq);
5527
5528 /* Ignore urgent data that we've already seen and read. */
5529 if (after(tp->copied_seq, ptr))
5530 return;
5531
5532 /* Do not replay urg ptr.
5533 *
5534 * NOTE: interesting situation not covered by specs.
5535 * Misbehaving sender may send urg ptr, pointing to segment,
5536 * which we already have in ofo queue. We are not able to fetch
5537 * such data and will stay in TCP_URG_NOTYET until will be eaten
5538 * by recvmsg(). Seems, we are not obliged to handle such wicked
5539 * situations. But it is worth to think about possibility of some
5540 * DoSes using some hypothetical application level deadlock.
5541 */
5542 if (before(ptr, tp->rcv_nxt))
5543 return;
5544
5545 /* Do we already have a newer (or duplicate) urgent pointer? */
5546 if (tp->urg_data && !after(ptr, tp->urg_seq))
5547 return;
5548
5549 /* Tell the world about our new urgent pointer. */
5550 sk_send_sigurg(sk);
5551
5552 /* We may be adding urgent data when the last byte read was
5553 * urgent. To do this requires some care. We cannot just ignore
5554 * tp->copied_seq since we would read the last urgent byte again
5555 * as data, nor can we alter copied_seq until this data arrives
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08005556 * or we break the semantics of SIOCATMARK (and thus sockatmark())
Linus Torvalds1da177e2005-04-16 15:20:36 -07005557 *
5558 * NOTE. Double Dutch. Rendering to plain English: author of comment
5559 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5560 * and expect that both A and B disappear from stream. This is _wrong_.
5561 * Though this happens in BSD with high probability, this is occasional.
5562 * Any application relying on this is buggy. Note also, that fix "works"
5563 * only in this artificial test. Insert some normal data between A and B and we will
5564 * decline of BSD again. Verdict: it is better to remove to trap
5565 * buggy users.
5566 */
5567 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005568 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005569 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5570 tp->copied_seq++;
5571 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
David S. Miller8728b832005-08-09 19:25:21 -07005572 __skb_unlink(skb, &sk->sk_receive_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005573 __kfree_skb(skb);
5574 }
5575 }
5576
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005577 tp->urg_data = TCP_URG_NOTYET;
Eric Dumazetd9b55bf2019-10-10 20:17:43 -07005578 WRITE_ONCE(tp->urg_seq, ptr);
Florian Westphal31770e32017-08-30 19:24:58 +02005579
5580 /* Disable header prediction. */
5581 tp->pred_flags = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005582}
5583
5584/* This is the 'fast' part of urgent handling. */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04005585static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005586{
5587 struct tcp_sock *tp = tcp_sk(sk);
5588
5589 /* Check if we get a new urgent pointer - normally not. */
5590 if (th->urg)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08005591 tcp_check_urg(sk, th);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005592
5593 /* Do we wait for any urgent data? - normally not... */
5594 if (tp->urg_data == TCP_URG_NOTYET) {
5595 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5596 th->syn;
5597
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005598 /* Is the urgent pointer pointing into this packet? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005599 if (ptr < skb->len) {
5600 u8 tmp;
5601 if (skb_copy_bits(skb, ptr, &tmp, 1))
5602 BUG();
5603 tp->urg_data = TCP_URG_VALID | tmp;
5604 if (!sock_flag(sk, SOCK_DEAD))
David S. Miller676d2362014-04-11 16:15:36 -04005605 sk->sk_data_ready(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005606 }
5607 }
5608}
5609
Jason Baron0e40f4c2017-01-17 13:37:19 -05005610/* Accept RST for rcv_nxt - 1 after a FIN.
5611 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5612 * FIN is sent followed by a RST packet. The RST is sent with the same
5613 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5614 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5615 * ACKs on the closed socket. In addition middleboxes can drop either the
5616 * challenge ACK or a subsequent RST.
5617 */
5618static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb)
5619{
5620 struct tcp_sock *tp = tcp_sk(sk);
5621
5622 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) &&
5623 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK |
5624 TCPF_CLOSING));
5625}
5626
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005627/* Does PAWS and seqno based validation of an incoming segment, flags will
5628 * play significant role here.
5629 */
Eric Dumazet0c246042012-07-17 01:41:30 +00005630static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5631 const struct tcphdr *th, int syn_inerr)
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005632{
5633 struct tcp_sock *tp = tcp_sk(sk);
Pau Espin Pedrole00431b2016-06-07 16:30:34 +02005634 bool rst_seq_match = false;
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005635
5636 /* RFC1323: H1. Apply PAWS check first. */
Eric Dumazeteed29f12017-06-07 10:34:36 -07005637 if (tcp_fast_parse_options(sock_net(sk), skb, th, tp) &&
5638 tp->rx_opt.saw_tstamp &&
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005639 tcp_paws_discard(sk, skb)) {
5640 if (!th->rst) {
Eric Dumazetc10d9312016-04-29 14:16:47 -07005641 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
Neal Cardwellf2b2c582015-02-06 16:04:40 -05005642 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5643 LINUX_MIB_TCPACKSKIPPEDPAWS,
5644 &tp->last_oow_ack_time))
5645 tcp_send_dupack(sk, skb);
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005646 goto discard;
5647 }
5648 /* Reset is accepted even if it did not pass PAWS. */
5649 }
5650
5651 /* Step 1: check sequence number */
5652 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5653 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5654 * (RST) segments are validated by checking their SEQ-fields."
5655 * And page 69: "If an incoming segment is not acceptable,
5656 * an acknowledgment should be sent in reply (unless the RST
5657 * bit is set, if so drop the segment and return)".
5658 */
Eric Dumazete3715892012-07-17 12:29:30 +00005659 if (!th->rst) {
5660 if (th->syn)
5661 goto syn_challenge;
Neal Cardwellf2b2c582015-02-06 16:04:40 -05005662 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5663 LINUX_MIB_TCPACKSKIPPEDSEQ,
5664 &tp->last_oow_ack_time))
5665 tcp_send_dupack(sk, skb);
Jason Baron0e40f4c2017-01-17 13:37:19 -05005666 } else if (tcp_reset_check(sk, skb)) {
Florian Westphal049fe382020-12-10 14:25:03 -08005667 tcp_reset(sk, skb);
Eric Dumazete3715892012-07-17 12:29:30 +00005668 }
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005669 goto discard;
5670 }
5671
5672 /* Step 2: check RST bit */
5673 if (th->rst) {
Jason Baron0e40f4c2017-01-17 13:37:19 -05005674 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5675 * FIN and SACK too if available):
5676 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5677 * the right-most SACK block,
Pau Espin Pedrole00431b2016-06-07 16:30:34 +02005678 * then
Eric Dumazet282f23c2012-07-17 10:13:05 +02005679 * RESET the connection
5680 * else
5681 * Send a challenge ACK
5682 */
Jason Baron0e40f4c2017-01-17 13:37:19 -05005683 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt ||
5684 tcp_reset_check(sk, skb)) {
Pau Espin Pedrole00431b2016-06-07 16:30:34 +02005685 rst_seq_match = true;
5686 } else if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
5687 struct tcp_sack_block *sp = &tp->selective_acks[0];
5688 int max_sack = sp[0].end_seq;
5689 int this_sack;
5690
5691 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
5692 ++this_sack) {
5693 max_sack = after(sp[this_sack].end_seq,
5694 max_sack) ?
5695 sp[this_sack].end_seq : max_sack;
5696 }
5697
5698 if (TCP_SKB_CB(skb)->seq == max_sack)
5699 rst_seq_match = true;
5700 }
5701
5702 if (rst_seq_match)
Florian Westphal049fe382020-12-10 14:25:03 -08005703 tcp_reset(sk, skb);
Wei Wangcf1ef3f2017-04-20 14:45:46 -07005704 else {
5705 /* Disable TFO if RST is out-of-order
5706 * and no data has been received
5707 * for current active TFO socket
5708 */
5709 if (tp->syn_fastopen && !tp->data_segs_in &&
5710 sk->sk_state == TCP_ESTABLISHED)
Wei Wang46c2fa32017-04-20 14:45:47 -07005711 tcp_fastopen_active_disable(sk);
Neal Cardwellf2b2c582015-02-06 16:04:40 -05005712 tcp_send_challenge_ack(sk, skb);
Wei Wangcf1ef3f2017-04-20 14:45:46 -07005713 }
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005714 goto discard;
5715 }
5716
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005717 /* step 3: check security and precedence [ignored] */
5718
Eric Dumazet0c246042012-07-17 01:41:30 +00005719 /* step 4: Check for a SYN
Sowmini Varadhancd214532014-10-30 12:48:08 -04005720 * RFC 5961 4.2 : Send a challenge ack
Eric Dumazet0c246042012-07-17 01:41:30 +00005721 */
5722 if (th->syn) {
Eric Dumazete3715892012-07-17 12:29:30 +00005723syn_challenge:
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005724 if (syn_inerr)
Eric Dumazetc10d9312016-04-29 14:16:47 -07005725 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5726 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
Neal Cardwellf2b2c582015-02-06 16:04:40 -05005727 tcp_send_challenge_ack(sk, skb);
Eric Dumazet0c246042012-07-17 01:41:30 +00005728 goto discard;
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005729 }
5730
Martin KaFai Lau00d211a2020-08-20 12:00:46 -07005731 bpf_skops_parse_hdr(sk, skb);
5732
Eric Dumazet0c246042012-07-17 01:41:30 +00005733 return true;
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005734
5735discard:
Eric Dumazet532182c2016-04-01 08:52:19 -07005736 tcp_drop(sk, skb);
Eric Dumazet0c246042012-07-17 01:41:30 +00005737 return false;
Ilpo Järvinencbe2d122008-08-23 05:10:12 -07005738}
5739
Linus Torvalds1da177e2005-04-16 15:20:36 -07005740/*
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09005741 * TCP receive function for the ESTABLISHED state.
Florian Westphal31770e32017-08-30 19:24:58 +02005742 *
5743 * It is split into a fast path and a slow path. The fast path is
5744 * disabled when:
5745 * - A zero window was announced from us - zero window probing
5746 * is only handled properly in the slow path.
5747 * - Out of order segments arrived.
5748 * - Urgent data is expected.
5749 * - There is no buffer space left
5750 * - Unexpected TCP flags/window values/header lengths are received
5751 * (detected by checking the TCP header against pred_flags)
5752 * - Data is sent in both directions. Fast path only supports pure senders
5753 * or pure receivers (this means either the sequence number or the ack
5754 * value must stay constant)
5755 * - Unexpected TCP option.
5756 *
5757 * When these conditions are not satisfied it drops into a standard
5758 * receive procedure patterned after RFC793 to handle all cases.
5759 * The first three cases are guaranteed by proper pred_flags setting,
5760 * the rest is checked inline. Fast processing is turned on in
5761 * tcp_data_queue when everything is OK.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005762 */
Yafang Shao3d97d882018-05-29 23:27:31 +08005763void tcp_rcv_established(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005764{
Yafang Shao3d97d882018-05-29 23:27:31 +08005765 const struct tcphdr *th = (const struct tcphdr *)skb->data;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005766 struct tcp_sock *tp = tcp_sk(sk);
Yafang Shao3d97d882018-05-29 23:27:31 +08005767 unsigned int len = skb->len;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005768
Masami Hiramatsuc3fde1b2017-12-29 11:45:51 +09005769 /* TCP congestion window tracking */
5770 trace_tcp_probe(sk, skb);
5771
Eric Dumazet9a568de2017-05-16 14:00:14 -07005772 tcp_mstamp_refresh(tp);
Ian Morris51456b22015-04-03 09:17:26 +01005773 if (unlikely(!sk->sk_rx_dst))
Eric Dumazet5d299f32012-08-06 05:09:33 +00005774 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
Florian Westphal31770e32017-08-30 19:24:58 +02005775 /*
5776 * Header prediction.
5777 * The code loosely follows the one in the famous
5778 * "30 instruction TCP receive" Van Jacobson mail.
5779 *
5780 * Van's trick is to deposit buffers into socket queue
5781 * on a device interrupt, to call tcp_recv function
5782 * on the receive process context and checksum and copy
5783 * the buffer to user space. smart...
5784 *
5785 * Our current scheme is not silly either but we take the
5786 * extra cost of the net_bh soft interrupt processing...
5787 * We do checksum and copy also but from device to kernel.
5788 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005789
5790 tp->rx_opt.saw_tstamp = 0;
5791
Florian Westphal31770e32017-08-30 19:24:58 +02005792 /* pred_flags is 0xS?10 << 16 + snd_wnd
5793 * if header_prediction is to be made
5794 * 'S' will always be tp->tcp_header_len >> 2
5795 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5796 * turn it off (when there are holes in the receive
5797 * space for instance)
5798 * PSH flag is ignored.
5799 */
5800
5801 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5802 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5803 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5804 int tcp_header_len = tp->tcp_header_len;
5805
5806 /* Timestamp header prediction: tcp_header_len
5807 * is automatically equal to th->doff*4 due to pred_flags
5808 * match.
5809 */
5810
5811 /* Check timestamp */
5812 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5813 /* No? Slow path! */
5814 if (!tcp_parse_aligned_timestamp(tp, th))
5815 goto slow_path;
5816
5817 /* If PAWS failed, check it more carefully in slow path */
5818 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5819 goto slow_path;
5820
5821 /* DO NOT update ts_recent here, if checksum fails
5822 * and timestamp was corrupted part, it will result
5823 * in a hung connection since we will drop all
5824 * future packets due to the PAWS test.
5825 */
5826 }
5827
5828 if (len <= tcp_header_len) {
5829 /* Bulk data transfer: sender */
5830 if (len == tcp_header_len) {
5831 /* Predicted packet is in window by definition.
5832 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5833 * Hence, check seq<=rcv_wup reduces to:
5834 */
5835 if (tcp_header_len ==
5836 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5837 tp->rcv_nxt == tp->rcv_wup)
5838 tcp_store_ts_recent(tp);
5839
5840 /* We know that such packets are checksummed
5841 * on entry.
5842 */
5843 tcp_ack(sk, skb, 0);
5844 __kfree_skb(skb);
5845 tcp_data_snd_check(sk);
Wei Wang3f6c65d2018-06-19 21:42:50 -07005846 /* When receiving pure ack in fast path, update
5847 * last ts ecr directly instead of calling
5848 * tcp_rcv_rtt_measure_ts()
5849 */
5850 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
Florian Westphal31770e32017-08-30 19:24:58 +02005851 return;
5852 } else { /* Header too small */
5853 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5854 goto discard;
5855 }
5856 } else {
5857 int eaten = 0;
5858 bool fragstolen = false;
5859
5860 if (tcp_checksum_complete(skb))
5861 goto csum_error;
5862
5863 if ((int)skb->truesize > sk->sk_forward_alloc)
5864 goto step5;
5865
5866 /* Predicted packet is in window by definition.
5867 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5868 * Hence, check seq<=rcv_wup reduces to:
5869 */
5870 if (tcp_header_len ==
5871 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5872 tp->rcv_nxt == tp->rcv_wup)
5873 tcp_store_ts_recent(tp);
5874
5875 tcp_rcv_rtt_measure_ts(sk, skb);
5876
5877 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
5878
5879 /* Bulk data transfer: receiver */
Eric Dumazete7395f1f2018-11-26 14:49:12 -08005880 __skb_pull(skb, tcp_header_len);
5881 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
Florian Westphal31770e32017-08-30 19:24:58 +02005882
5883 tcp_event_data_recv(sk, skb);
5884
5885 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5886 /* Well, only one small jumplet in fast path... */
5887 tcp_ack(sk, skb, FLAG_DATA);
5888 tcp_data_snd_check(sk);
5889 if (!inet_csk_ack_scheduled(sk))
5890 goto no_ack;
Neal Cardwell18ded912020-10-22 10:33:31 -04005891 } else {
5892 tcp_update_wl(tp, TCP_SKB_CB(skb)->seq);
Florian Westphal31770e32017-08-30 19:24:58 +02005893 }
5894
5895 __tcp_ack_snd_check(sk, 0);
5896no_ack:
5897 if (eaten)
5898 kfree_skb_partial(skb, fragstolen);
Eric Dumazet03f45c82018-04-16 10:33:37 -07005899 tcp_data_ready(sk);
Florian Westphal31770e32017-08-30 19:24:58 +02005900 return;
5901 }
5902 }
5903
5904slow_path:
Eric Dumazetfb3477c2016-04-29 14:16:48 -07005905 if (len < (th->doff << 2) || tcp_checksum_complete(skb))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005906 goto csum_error;
5907
Calvin Owens0c228e82014-11-20 15:09:53 -08005908 if (!th->ack && !th->rst && !th->syn)
Eric Dumazetc3ae62af2012-12-26 12:44:34 +00005909 goto discard;
5910
Florian Westphal31770e32017-08-30 19:24:58 +02005911 /*
5912 * Standard slow path.
5913 */
5914
Eric Dumazet0c246042012-07-17 01:41:30 +00005915 if (!tcp_validate_incoming(sk, skb, th, 1))
Vijay Subramanianc995ae22013-09-03 12:23:22 -07005916 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005917
Florian Westphal31770e32017-08-30 19:24:58 +02005918step5:
5919 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
John Dykstra96e0bf42009-03-22 21:49:57 -07005920 goto discard;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005921
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07005922 tcp_rcv_rtt_measure_ts(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005923
5924 /* Process urgent data. */
5925 tcp_urg(sk, skb, th);
5926
5927 /* step 7: process the segment text */
5928 tcp_data_queue(sk, skb);
5929
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07005930 tcp_data_snd_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005931 tcp_ack_snd_check(sk);
Vijay Subramanianc995ae22013-09-03 12:23:22 -07005932 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005933
5934csum_error:
Jakub Kicinski709c0312021-05-14 13:04:25 -07005935 trace_tcp_bad_csum(skb);
Eric Dumazetc10d9312016-04-29 14:16:47 -07005936 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
5937 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005938
5939discard:
Eric Dumazet532182c2016-04-01 08:52:19 -07005940 tcp_drop(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005941}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00005942EXPORT_SYMBOL(tcp_rcv_established);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005943
Martin KaFai Lau72be0fe2020-08-20 12:00:39 -07005944void tcp_init_transfer(struct sock *sk, int bpf_op, struct sk_buff *skb)
Yuchung Cheng98fa6272019-04-29 15:46:20 -07005945{
5946 struct inet_connection_sock *icsk = inet_csk(sk);
5947 struct tcp_sock *tp = tcp_sk(sk);
5948
5949 tcp_mtup_init(sk);
5950 icsk->icsk_af_ops->rebuild_header(sk);
5951 tcp_init_metrics(sk);
5952
5953 /* Initialize the congestion window to start the transfer.
5954 * Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
5955 * retransmitted. In light of RFC6298 more aggressive 1sec
5956 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
5957 * retransmission has occurred.
5958 */
5959 if (tp->total_retrans > 1 && tp->undo_marker)
5960 tp->snd_cwnd = 1;
5961 else
5962 tp->snd_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
5963 tp->snd_cwnd_stamp = tcp_jiffies32;
5964
Martin KaFai Lau72be0fe2020-08-20 12:00:39 -07005965 bpf_skops_established(sk, bpf_op, skb);
Nguyen Dinh Phibe5d1b62021-07-06 07:19:12 +08005966 /* Initialize congestion control unless BPF initialized it already: */
Neal Cardwell8919a9b2020-09-10 15:35:32 -04005967 if (!icsk->icsk_ca_initialized)
5968 tcp_init_congestion_control(sk);
Yuchung Cheng98fa6272019-04-29 15:46:20 -07005969 tcp_init_buffer_space(sk);
5970}
5971
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005972void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5973{
5974 struct tcp_sock *tp = tcp_sk(sk);
5975 struct inet_connection_sock *icsk = inet_csk(sk);
5976
5977 tcp_set_state(sk, TCP_ESTABLISHED);
Eric Dumazet70eabf02017-05-16 14:00:07 -07005978 icsk->icsk_ack.lrcvtime = tcp_jiffies32;
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005979
Ian Morris00db4122015-04-03 09:17:27 +01005980 if (skb) {
Eric Dumazet5d299f32012-08-06 05:09:33 +00005981 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005982 security_inet_conn_established(sk, skb);
Amritha Nambiarc6345ce2018-06-29 21:26:57 -07005983 sk_mark_napi_id(sk, skb);
David S. Miller41063e92012-06-19 21:22:05 -07005984 }
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005985
Martin KaFai Lau72be0fe2020-08-20 12:00:39 -07005986 tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB, skb);
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005987
5988 /* Prevent spurious tcp_cwnd_restart() on first data
5989 * packet.
5990 */
Eric Dumazetd635fbe2017-05-16 14:00:03 -07005991 tp->lsndtime = tcp_jiffies32;
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005992
Pavel Emelyanov370816a2012-04-19 03:40:01 +00005993 if (sock_flag(sk, SOCK_KEEPOPEN))
5994 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
Florian Westphal31770e32017-08-30 19:24:58 +02005995
5996 if (!tp->rx_opt.snd_wscale)
5997 __tcp_fast_path_on(tp, tp->snd_wnd);
5998 else
5999 tp->pred_flags = 0;
Pavel Emelyanov370816a2012-04-19 03:40:01 +00006000}
6001
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00006002static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
6003 struct tcp_fastopen_cookie *cookie)
6004{
6005 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet75c119a2017-10-05 22:21:27 -07006006 struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL;
Daniel Lee2646c832015-04-06 14:37:27 -07006007 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
6008 bool syn_drop = false;
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00006009
6010 if (mss == tp->rx_opt.user_mss) {
6011 struct tcp_options_received opt;
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00006012
6013 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
6014 tcp_clear_options(&opt);
6015 opt.user_mss = opt.mss_clamp = 0;
Eric Dumazeteed29f12017-06-07 10:34:36 -07006016 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL);
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00006017 mss = opt.mss_clamp;
6018 }
6019
Daniel Lee2646c832015-04-06 14:37:27 -07006020 if (!tp->syn_fastopen) {
6021 /* Ignore an unsolicited cookie */
Yuchung Cheng67da22d2012-07-19 06:43:11 +00006022 cookie->len = -1;
Daniel Lee2646c832015-04-06 14:37:27 -07006023 } else if (tp->total_retrans) {
6024 /* SYN timed out and the SYN-ACK neither has a cookie nor
6025 * acknowledges data. Presumably the remote received only
6026 * the retransmitted (regular) SYNs: either the original
6027 * SYN-data or the corresponding SYN-ACK was dropped.
6028 */
6029 syn_drop = (cookie->len < 0 && data);
6030 } else if (cookie->len < 0 && !tp->syn_data) {
6031 /* We requested a cookie but didn't get it. If we did not use
6032 * the (old) exp opt format then try so next time (try_exp=1).
6033 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
6034 */
6035 try_exp = tp->syn_fastopen_exp ? 2 : 1;
6036 }
Yuchung Cheng67da22d2012-07-19 06:43:11 +00006037
Daniel Lee2646c832015-04-06 14:37:27 -07006038 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00006039
6040 if (data) { /* Retransmit unacked data in SYN */
Jason Baron48027472019-10-23 11:09:26 -04006041 if (tp->total_retrans)
6042 tp->fastopen_client_fail = TFO_SYN_RETRANSMITTED;
6043 else
6044 tp->fastopen_client_fail = TFO_DATA_NOT_ACKED;
Eric Dumazeta7abf3c2021-03-11 12:35:05 -08006045 skb_rbtree_walk_from(data)
6046 tcp_mark_skb_lost(sk, data);
6047 tcp_xmit_retransmit_queue(sk);
Eric Dumazetc10d9312016-04-29 14:16:47 -07006048 NET_INC_STATS(sock_net(sk),
Eric Dumazet02a1d6e2016-04-27 16:44:39 -07006049 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00006050 return true;
6051 }
Yuchung Cheng6f736012012-10-19 15:14:44 +00006052 tp->syn_data_acked = tp->syn_data;
Yuchung Chengbef57672018-04-17 23:18:46 -07006053 if (tp->syn_data_acked) {
6054 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
6055 /* SYN-data is counted as two separate packets in tcp_ack() */
6056 if (tp->delivered > 1)
6057 --tp->delivered;
6058 }
Eric Dumazet61d2bca2016-02-01 21:03:07 -08006059
6060 tcp_fastopen_add_skb(sk, synack);
6061
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00006062 return false;
6063}
6064
Ursula Braun60e2a772017-10-25 11:01:45 +02006065static void smc_check_reset_syn(struct tcp_sock *tp)
6066{
6067#if IS_ENABLED(CONFIG_SMC)
6068 if (static_branch_unlikely(&tcp_have_smc)) {
6069 if (tp->syn_smc && !tp->rx_opt.smc_ok)
6070 tp->syn_smc = 0;
6071 }
6072#endif
6073}
6074
Yuchung Cheng7c1f0812019-04-29 15:46:14 -07006075static void tcp_try_undo_spurious_syn(struct sock *sk)
6076{
6077 struct tcp_sock *tp = tcp_sk(sk);
6078 u32 syn_stamp;
6079
6080 /* undo_marker is set when SYN or SYNACK times out. The timeout is
6081 * spurious if the ACK's timestamp option echo value matches the
6082 * original SYN timestamp.
6083 */
6084 syn_stamp = tp->retrans_stamp;
6085 if (tp->undo_marker && syn_stamp && tp->rx_opt.saw_tstamp &&
6086 syn_stamp == tp->rx_opt.rcv_tsecr)
6087 tp->undo_marker = 0;
6088}
6089
Linus Torvalds1da177e2005-04-16 15:20:36 -07006090static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
Eric Dumazetbda07a62015-09-29 07:42:40 -07006091 const struct tcphdr *th)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006092{
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08006093 struct inet_connection_sock *icsk = inet_csk(sk);
William Allen Simpson4957faade2009-12-02 18:25:27 +00006094 struct tcp_sock *tp = tcp_sk(sk);
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00006095 struct tcp_fastopen_cookie foc = { .len = -1 };
William Allen Simpson4957faade2009-12-02 18:25:27 +00006096 int saved_clamp = tp->rx_opt.mss_clamp;
Eric Dumazet0f9fa832017-04-18 09:45:52 -07006097 bool fastopen_fail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006098
Eric Dumazeteed29f12017-06-07 10:34:36 -07006099 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc);
Andrew Vagine3e12022013-08-27 12:21:55 +04006100 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
Andrey Vaginee684b62013-02-11 05:50:19 +00006101 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006102
6103 if (th->ack) {
6104 /* rfc793:
6105 * "If the state is SYN-SENT then
6106 * first check the ACK bit
6107 * If the ACK bit is set
6108 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
6109 * a reset (unless the RST bit is set, if so drop
6110 * the segment and return)"
Linus Torvalds1da177e2005-04-16 15:20:36 -07006111 */
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00006112 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
SeongJae Park9603d472020-02-02 03:38:26 +00006113 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
6114 /* Previous FIN/ACK or RST/ACK might be ignored. */
6115 if (icsk->icsk_retransmits == 0)
6116 inet_csk_reset_xmit_timer(sk,
6117 ICSK_TIME_RETRANS,
6118 TCP_TIMEOUT_MIN, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006119 goto reset_and_undo;
SeongJae Park9603d472020-02-02 03:38:26 +00006120 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07006121
6122 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
6123 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
Eric Dumazet9a568de2017-05-16 14:00:14 -07006124 tcp_time_stamp(tp))) {
Eric Dumazetc10d9312016-04-29 14:16:47 -07006125 NET_INC_STATS(sock_net(sk),
Eric Dumazet02a1d6e2016-04-27 16:44:39 -07006126 LINUX_MIB_PAWSACTIVEREJECTED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006127 goto reset_and_undo;
6128 }
6129
6130 /* Now ACK is acceptable.
6131 *
6132 * "If the RST bit is set
6133 * If the ACK was acceptable then signal the user "error:
6134 * connection reset", drop the segment, enter CLOSED state,
6135 * delete TCB, and return."
6136 */
6137
6138 if (th->rst) {
Florian Westphal049fe382020-12-10 14:25:03 -08006139 tcp_reset(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006140 goto discard;
6141 }
6142
6143 /* rfc793:
6144 * "fifth, if neither of the SYN or RST bits is set then
6145 * drop the segment and return."
6146 *
6147 * See note below!
6148 * --ANK(990513)
6149 */
6150 if (!th->syn)
6151 goto discard_and_undo;
6152
6153 /* rfc793:
6154 * "If the SYN bit is on ...
6155 * are acceptable then ...
6156 * (our SYN has been ACKed), change the connection
6157 * state to ESTABLISHED..."
6158 */
6159
Florian Westphal735d3832014-09-29 13:08:30 +02006160 tcp_ecn_rcv_synack(tp, th);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006161
Razvan Ghitulete1b3a6922012-08-14 16:30:20 +03006162 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
Yuchung Cheng7c1f0812019-04-29 15:46:14 -07006163 tcp_try_undo_spurious_syn(sk);
Florian Westphal31770e32017-08-30 19:24:58 +02006164 tcp_ack(sk, skb, FLAG_SLOWPATH);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006165
6166 /* Ok.. it's good. Set up sequence numbers and
6167 * move to established.
6168 */
Eric Dumazetdba7d9b2019-10-10 20:17:39 -07006169 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006170 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
6171
6172 /* RFC1323: The window in SYN & SYN/ACK segments is
6173 * never scaled.
6174 */
6175 tp->snd_wnd = ntohs(th->window);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006176
6177 if (!tp->rx_opt.wscale_ok) {
6178 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
6179 tp->window_clamp = min(tp->window_clamp, 65535U);
6180 }
6181
6182 if (tp->rx_opt.saw_tstamp) {
6183 tp->rx_opt.tstamp_ok = 1;
6184 tp->tcp_header_len =
6185 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
6186 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6187 tcp_store_ts_recent(tp);
6188 } else {
6189 tp->tcp_header_len = sizeof(struct tcphdr);
6190 }
6191
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08006192 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006193 tcp_initialize_rcv_mss(sk);
6194
6195 /* Remember, tcp_poll() does not lock socket!
6196 * Change state from SYN-SENT only after copied_seq
6197 * is initialized. */
Eric Dumazet7db48e92019-10-10 20:17:40 -07006198 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
William Allen Simpson4957faade2009-12-02 18:25:27 +00006199
Ursula Braun60e2a772017-10-25 11:01:45 +02006200 smc_check_reset_syn(tp);
6201
Ralf Baechlee16aa202006-12-07 00:11:33 -08006202 smp_mb();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006203
Pavel Emelyanov370816a2012-04-19 03:40:01 +00006204 tcp_finish_connect(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006205
Eric Dumazet0f9fa832017-04-18 09:45:52 -07006206 fastopen_fail = (tp->syn_fastopen || tp->syn_data) &&
6207 tcp_rcv_fastopen_synack(sk, skb, &foc);
Yuchung Cheng8e4178c2012-07-19 06:43:08 +00006208
Eric Dumazet0f9fa832017-04-18 09:45:52 -07006209 if (!sock_flag(sk, SOCK_DEAD)) {
6210 sk->sk_state_change(sk);
6211 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6212 }
6213 if (fastopen_fail)
6214 return -1;
Arnaldo Carvalho de Melo295f7322005-08-09 20:11:56 -07006215 if (sk->sk_write_pending ||
6216 icsk->icsk_accept_queue.rskq_defer_accept ||
Wei Wang31954cd2019-01-25 10:53:19 -08006217 inet_csk_in_pingpong_mode(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07006218 /* Save one ACK. Data will be ready after
6219 * several ticks, if write_pending is set.
6220 *
6221 * It may be deleted, but with this feature tcpdumps
6222 * look so _wonderfully_ clever, that I was not able
6223 * to stand against the temptation 8) --ANK
6224 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07006225 inet_csk_schedule_ack(sk);
Eric Dumazet9a9c9b52018-05-21 15:08:56 -07006226 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07006227 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
6228 TCP_DELACK_MAX, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006229
6230discard:
Eric Dumazet532182c2016-04-01 08:52:19 -07006231 tcp_drop(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006232 return 0;
6233 } else {
6234 tcp_send_ack(sk);
6235 }
6236 return -1;
6237 }
6238
6239 /* No ACK in the segment */
6240
6241 if (th->rst) {
6242 /* rfc793:
6243 * "If the RST bit is set
6244 *
6245 * Otherwise (no ACK) drop the segment and return."
6246 */
6247
6248 goto discard_and_undo;
6249 }
6250
6251 /* PAWS check. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08006252 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
Ilpo Järvinenc887e6d2009-03-14 14:23:03 +00006253 tcp_paws_reject(&tp->rx_opt, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07006254 goto discard_and_undo;
6255
6256 if (th->syn) {
6257 /* We see SYN without ACK. It is attempt of
6258 * simultaneous connect with crossed SYNs.
6259 * Particularly, it can be connect to self.
6260 */
6261 tcp_set_state(sk, TCP_SYN_RECV);
6262
6263 if (tp->rx_opt.saw_tstamp) {
6264 tp->rx_opt.tstamp_ok = 1;
6265 tcp_store_ts_recent(tp);
6266 tp->tcp_header_len =
6267 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
6268 } else {
6269 tp->tcp_header_len = sizeof(struct tcphdr);
6270 }
6271
Eric Dumazetdba7d9b2019-10-10 20:17:39 -07006272 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
Eric Dumazet7db48e92019-10-10 20:17:40 -07006273 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006274 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
6275
6276 /* RFC1323: The window in SYN & SYN/ACK segments is
6277 * never scaled.
6278 */
6279 tp->snd_wnd = ntohs(th->window);
6280 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
6281 tp->max_window = tp->snd_wnd;
6282
Florian Westphal735d3832014-09-29 13:08:30 +02006283 tcp_ecn_rcv_syn(tp, th);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006284
John Heffner5d424d52006-03-20 17:53:41 -08006285 tcp_mtup_init(sk);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08006286 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006287 tcp_initialize_rcv_mss(sk);
6288
Linus Torvalds1da177e2005-04-16 15:20:36 -07006289 tcp_send_synack(sk);
6290#if 0
6291 /* Note, we could accept data and URG from this segment.
Jerry Chu168a8f52012-08-31 12:29:13 +00006292 * There are no obstacles to make this (except that we must
6293 * either change tcp_recvmsg() to prevent it from returning data
6294 * before 3WHS completes per RFC793, or employ TCP Fast Open).
Linus Torvalds1da177e2005-04-16 15:20:36 -07006295 *
6296 * However, if we ignore data in ACKless segments sometimes,
6297 * we have no reasons to accept it sometimes.
6298 * Also, seems the code doing it in step6 of tcp_rcv_state_process
6299 * is not flawless. So, discard packet for sanity.
6300 * Uncomment this return to process the data.
6301 */
6302 return -1;
6303#else
6304 goto discard;
6305#endif
6306 }
6307 /* "fifth, if neither of the SYN or RST bits is set then
6308 * drop the segment and return."
6309 */
6310
6311discard_and_undo:
6312 tcp_clear_options(&tp->rx_opt);
6313 tp->rx_opt.mss_clamp = saved_clamp;
6314 goto discard;
6315
6316reset_and_undo:
6317 tcp_clear_options(&tp->rx_opt);
6318 tp->rx_opt.mss_clamp = saved_clamp;
6319 return 1;
6320}
6321
Yuchung Cheng6b94b1c2019-04-29 15:46:19 -07006322static void tcp_rcv_synrecv_state_fastopen(struct sock *sk)
6323{
Eric Dumazetd983ea62019-10-10 20:17:38 -07006324 struct request_sock *req;
6325
Neal Cardwelldad8cea2020-02-22 11:21:15 -05006326 /* If we are still handling the SYNACK RTO, see if timestamp ECR allows
6327 * undo. If peer SACKs triggered fast recovery, we can't undo here.
6328 */
6329 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
6330 tcp_try_undo_loss(sk, false);
Yuchung Chengcd736d82019-05-13 10:32:05 -07006331
6332 /* Reset rtx states to prevent spurious retransmits_timed_out() */
6333 tcp_sk(sk)->retrans_stamp = 0;
Yuchung Cheng6b94b1c2019-04-29 15:46:19 -07006334 inet_csk(sk)->icsk_retransmits = 0;
6335
6336 /* Once we leave TCP_SYN_RECV or TCP_FIN_WAIT_1,
6337 * we no longer need req so release it.
6338 */
Eric Dumazetd983ea62019-10-10 20:17:38 -07006339 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk,
6340 lockdep_sock_is_held(sk));
6341 reqsk_fastopen_remove(sk, req, false);
Yuchung Cheng6b94b1c2019-04-29 15:46:19 -07006342
6343 /* Re-arm the timer because data may have been sent out.
6344 * This is similar to the regular data transmission case
6345 * when new data has just been ack'ed.
6346 *
6347 * (TFO) - we could try to be more aggressive and
6348 * retransmitting any data sooner based on when they
6349 * are sent out.
6350 */
6351 tcp_rearm_rto(sk);
6352}
6353
Linus Torvalds1da177e2005-04-16 15:20:36 -07006354/*
6355 * This function implements the receiving procedure of RFC 793 for
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09006356 * all states except ESTABLISHED and TIME_WAIT.
Linus Torvalds1da177e2005-04-16 15:20:36 -07006357 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
6358 * address independent.
6359 */
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09006360
Eric Dumazet72ab4a82015-09-29 07:42:41 -07006361int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006362{
6363 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -08006364 struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazet72ab4a82015-09-29 07:42:41 -07006365 const struct tcphdr *th = tcp_hdr(skb);
Jerry Chu168a8f52012-08-31 12:29:13 +00006366 struct request_sock *req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006367 int queued = 0;
Eric Dumazet1f6afc82013-05-24 15:03:54 +00006368 bool acceptable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006369
Linus Torvalds1da177e2005-04-16 15:20:36 -07006370 switch (sk->sk_state) {
6371 case TCP_CLOSE:
6372 goto discard;
6373
6374 case TCP_LISTEN:
Stephen Hemminger2de979b2007-03-08 20:45:19 -08006375 if (th->ack)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006376 return 1;
6377
Stephen Hemminger2de979b2007-03-08 20:45:19 -08006378 if (th->rst)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006379 goto discard;
6380
Stephen Hemminger2de979b2007-03-08 20:45:19 -08006381 if (th->syn) {
Eric Dumazetfdf5af02011-12-02 23:41:42 +00006382 if (th->fin)
6383 goto discard;
Eric Dumazet449809a2017-03-01 08:39:49 -08006384 /* It is possible that we process SYN packets from backlog,
Eric Dumazet1ad98e92018-10-01 15:02:26 -07006385 * so we need to make sure to disable BH and RCU right there.
Eric Dumazet449809a2017-03-01 08:39:49 -08006386 */
Eric Dumazet1ad98e92018-10-01 15:02:26 -07006387 rcu_read_lock();
Eric Dumazet449809a2017-03-01 08:39:49 -08006388 local_bh_disable();
6389 acceptable = icsk->icsk_af_ops->conn_request(sk, skb) >= 0;
6390 local_bh_enable();
Eric Dumazet1ad98e92018-10-01 15:02:26 -07006391 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006392
Eric Dumazet449809a2017-03-01 08:39:49 -08006393 if (!acceptable)
6394 return 1;
Eric Dumazet0aea76d2016-04-21 22:13:01 -07006395 consume_skb(skb);
Masayuki Nakagawafb7e2392007-01-23 20:15:06 -08006396 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006397 }
6398 goto discard;
6399
6400 case TCP_SYN_SENT:
Eric Dumazet8804b272016-04-13 22:05:40 -07006401 tp->rx_opt.saw_tstamp = 0;
Eric Dumazet9a568de2017-05-16 14:00:14 -07006402 tcp_mstamp_refresh(tp);
Eric Dumazetbda07a62015-09-29 07:42:40 -07006403 queued = tcp_rcv_synsent_state_process(sk, skb, th);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006404 if (queued >= 0)
6405 return queued;
6406
6407 /* Do step6 onward by hand. */
6408 tcp_urg(sk, skb, th);
6409 __kfree_skb(skb);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07006410 tcp_data_snd_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006411 return 0;
6412 }
6413
Eric Dumazet9a568de2017-05-16 14:00:14 -07006414 tcp_mstamp_refresh(tp);
Eric Dumazet8804b272016-04-13 22:05:40 -07006415 tp->rx_opt.saw_tstamp = 0;
Eric Dumazetd983ea62019-10-10 20:17:38 -07006416 req = rcu_dereference_protected(tp->fastopen_rsk,
6417 lockdep_sock_is_held(sk));
Ian Morris00db4122015-04-03 09:17:27 +01006418 if (req) {
Eric Dumazete0f97592018-02-13 06:14:12 -08006419 bool req_stolen;
6420
Jerry Chu37561f62012-10-22 11:26:36 +00006421 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
Jerry Chu168a8f52012-08-31 12:29:13 +00006422 sk->sk_state != TCP_FIN_WAIT1);
6423
Eric Dumazete0f97592018-02-13 06:14:12 -08006424 if (!tcp_check_req(sk, skb, req, true, &req_stolen))
Jerry Chu168a8f52012-08-31 12:29:13 +00006425 goto discard;
Neal Cardwelle69bebd2012-09-22 04:18:57 +00006426 }
Eric Dumazetc3ae62af2012-12-26 12:44:34 +00006427
Calvin Owens0c228e82014-11-20 15:09:53 -08006428 if (!th->ack && !th->rst && !th->syn)
Eric Dumazetc3ae62af2012-12-26 12:44:34 +00006429 goto discard;
6430
Neal Cardwelle69bebd2012-09-22 04:18:57 +00006431 if (!tcp_validate_incoming(sk, skb, th, 0))
Eric Dumazet0c246042012-07-17 01:41:30 +00006432 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006433
6434 /* step 5: check the ACK field */
Florian Westphal31770e32017-08-30 19:24:58 +02006435 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
6436 FLAG_UPDATE_TS_RECENT |
Eric Dumazetd0e1a1b2017-05-23 15:24:46 -07006437 FLAG_NO_CHALLENGE_ACK) > 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006438
Eric Dumazetd0e1a1b2017-05-23 15:24:46 -07006439 if (!acceptable) {
6440 if (sk->sk_state == TCP_SYN_RECV)
6441 return 1; /* send one RST */
6442 tcp_send_challenge_ack(sk, skb);
6443 goto discard;
6444 }
Eric Dumazet1f6afc82013-05-24 15:03:54 +00006445 switch (sk->sk_state) {
6446 case TCP_SYN_RECV:
Yuchung Chengbef57672018-04-17 23:18:46 -07006447 tp->delivered++; /* SYN-ACK delivery isn't tracked in tcp_ack */
Yuchung Cheng0f1c28a2015-09-18 11:36:14 -07006448 if (!tp->srtt_us)
6449 tcp_synack_rtt_meas(sk, req);
6450
Joe Perches61eb90032013-05-24 18:36:13 +00006451 if (req) {
Yuchung Cheng6b94b1c2019-04-29 15:46:19 -07006452 tcp_rcv_synrecv_state_fastopen(sk);
Joe Perches61eb90032013-05-24 18:36:13 +00006453 } else {
Yuchung Cheng336c39a2019-04-29 15:46:16 -07006454 tcp_try_undo_spurious_syn(sk);
6455 tp->retrans_stamp = 0;
Martin KaFai Lau72be0fe2020-08-20 12:00:39 -07006456 tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB,
6457 skb);
Eric Dumazet7db48e92019-10-10 20:17:40 -07006458 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
Eric Dumazet1f6afc82013-05-24 15:03:54 +00006459 }
Joe Perches61eb90032013-05-24 18:36:13 +00006460 smp_mb();
6461 tcp_set_state(sk, TCP_ESTABLISHED);
6462 sk->sk_state_change(sk);
6463
6464 /* Note, that this wakeup is only for marginal crossed SYN case.
6465 * Passively open sockets are not waked up, because
6466 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
6467 */
6468 if (sk->sk_socket)
6469 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6470
6471 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
6472 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
6473 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6474
6475 if (tp->rx_opt.tstamp_ok)
6476 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6477
Yuchung Chengc0402762016-09-19 23:39:21 -04006478 if (!inet_csk(sk)->icsk_ca_ops->cong_control)
6479 tcp_update_pacing_rate(sk);
Neal Cardwell02cf4eb2013-10-21 15:40:19 -04006480
Joe Perches61eb90032013-05-24 18:36:13 +00006481 /* Prevent spurious tcp_cwnd_restart() on first data packet */
Eric Dumazetd635fbe2017-05-16 14:00:03 -07006482 tp->lsndtime = tcp_jiffies32;
Joe Perches61eb90032013-05-24 18:36:13 +00006483
6484 tcp_initialize_rcv_mss(sk);
Florian Westphal31770e32017-08-30 19:24:58 +02006485 tcp_fast_path_on(tp);
Eric Dumazet1f6afc82013-05-24 15:03:54 +00006486 break;
6487
Joe Perchesc48b22d2013-05-24 18:06:58 +00006488 case TCP_FIN_WAIT1: {
Joe Perchesc48b22d2013-05-24 18:06:58 +00006489 int tmo;
6490
Yuchung Cheng6b94b1c2019-04-29 15:46:19 -07006491 if (req)
6492 tcp_rcv_synrecv_state_fastopen(sk);
6493
Joe Perchesc48b22d2013-05-24 18:06:58 +00006494 if (tp->snd_una != tp->write_seq)
6495 break;
Eric Dumazet1f6afc82013-05-24 15:03:54 +00006496
Joe Perchesc48b22d2013-05-24 18:06:58 +00006497 tcp_set_state(sk, TCP_FIN_WAIT2);
6498 sk->sk_shutdown |= SEND_SHUTDOWN;
Eric Dumazet1f6afc82013-05-24 15:03:54 +00006499
Julian Anastasovc3a2e832017-02-06 23:14:14 +02006500 sk_dst_confirm(sk);
Eric Dumazet1f6afc82013-05-24 15:03:54 +00006501
Joe Perchesc48b22d2013-05-24 18:06:58 +00006502 if (!sock_flag(sk, SOCK_DEAD)) {
6503 /* Wake up lingering close() */
6504 sk->sk_state_change(sk);
6505 break;
6506 }
Eric Dumazet1f6afc82013-05-24 15:03:54 +00006507
Wei Wangcf1ef3f2017-04-20 14:45:46 -07006508 if (tp->linger2 < 0) {
6509 tcp_done(sk);
6510 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6511 return 1;
6512 }
6513 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6514 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6515 /* Receive out of order FIN after close() */
6516 if (tp->syn_fastopen && th->fin)
Wei Wang46c2fa32017-04-20 14:45:47 -07006517 tcp_fastopen_active_disable(sk);
Joe Perchesc48b22d2013-05-24 18:06:58 +00006518 tcp_done(sk);
Eric Dumazetc10d9312016-04-29 14:16:47 -07006519 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
Joe Perchesc48b22d2013-05-24 18:06:58 +00006520 return 1;
6521 }
Eric Dumazet1f6afc82013-05-24 15:03:54 +00006522
Joe Perchesc48b22d2013-05-24 18:06:58 +00006523 tmo = tcp_fin_time(sk);
6524 if (tmo > TCP_TIMEWAIT_LEN) {
6525 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
6526 } else if (th->fin || sock_owned_by_user(sk)) {
6527 /* Bad case. We could lose such FIN otherwise.
6528 * It is not a big problem, but it looks confusing
6529 * and not so rare event. We still can lose it now,
6530 * if it spins in bh_lock_sock(), but it is really
6531 * marginal case.
6532 */
6533 inet_csk_reset_keepalive_timer(sk, tmo);
6534 } else {
6535 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6536 goto discard;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006537 }
Eric Dumazet1f6afc82013-05-24 15:03:54 +00006538 break;
Joe Perchesc48b22d2013-05-24 18:06:58 +00006539 }
Eric Dumazet1f6afc82013-05-24 15:03:54 +00006540
6541 case TCP_CLOSING:
6542 if (tp->snd_una == tp->write_seq) {
6543 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6544 goto discard;
6545 }
6546 break;
6547
6548 case TCP_LAST_ACK:
6549 if (tp->snd_una == tp->write_seq) {
6550 tcp_update_metrics(sk);
6551 tcp_done(sk);
6552 goto discard;
6553 }
6554 break;
Eric Dumazetc3ae62af2012-12-26 12:44:34 +00006555 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07006556
6557 /* step 6: check the URG bit */
6558 tcp_urg(sk, skb, th);
6559
6560 /* step 7: process the segment text */
6561 switch (sk->sk_state) {
6562 case TCP_CLOSE_WAIT:
6563 case TCP_CLOSING:
6564 case TCP_LAST_ACK:
Mat Martineau648ef4b2020-01-21 16:56:24 -08006565 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
Jianguo Wu6787b7e2021-07-09 17:20:49 -07006566 /* If a subflow has been reset, the packet should not
6567 * continue to be processed, drop the packet.
6568 */
6569 if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb))
6570 goto discard;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006571 break;
Mat Martineau648ef4b2020-01-21 16:56:24 -08006572 }
Joe Perchesa8eceea2020-03-12 15:50:22 -07006573 fallthrough;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006574 case TCP_FIN_WAIT1:
6575 case TCP_FIN_WAIT2:
6576 /* RFC 793 says to queue data in these states,
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09006577 * RFC 1122 says we MUST send a reset.
Linus Torvalds1da177e2005-04-16 15:20:36 -07006578 * BSD 4.4 also does reset.
6579 */
6580 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6581 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6582 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
Eric Dumazetc10d9312016-04-29 14:16:47 -07006583 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
Florian Westphal049fe382020-12-10 14:25:03 -08006584 tcp_reset(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006585 return 1;
6586 }
6587 }
Joe Perchesa8eceea2020-03-12 15:50:22 -07006588 fallthrough;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09006589 case TCP_ESTABLISHED:
Linus Torvalds1da177e2005-04-16 15:20:36 -07006590 tcp_data_queue(sk, skb);
6591 queued = 1;
6592 break;
6593 }
6594
6595 /* tcp_data could move socket to TIME-WAIT */
6596 if (sk->sk_state != TCP_CLOSE) {
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07006597 tcp_data_snd_check(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006598 tcp_ack_snd_check(sk);
6599 }
6600
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09006601 if (!queued) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07006602discard:
Eric Dumazet532182c2016-04-01 08:52:19 -07006603 tcp_drop(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006604 }
6605 return 0;
6606}
Linus Torvalds1da177e2005-04-16 15:20:36 -07006607EXPORT_SYMBOL(tcp_rcv_state_process);
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006608
6609static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6610{
6611 struct inet_request_sock *ireq = inet_rsk(req);
6612
6613 if (family == AF_INET)
Joe Perchesba7a46f2014-11-11 10:59:17 -08006614 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6615 &ireq->ir_rmt_addr, port);
Octavian Purdila4135ab82014-06-28 21:20:54 +03006616#if IS_ENABLED(CONFIG_IPV6)
6617 else if (family == AF_INET6)
Joe Perchesba7a46f2014-11-11 10:59:17 -08006618 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6619 &ireq->ir_v6_rmt_addr, port);
Octavian Purdila4135ab82014-06-28 21:20:54 +03006620#endif
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006621}
6622
Florian Westphald82bd122014-09-29 13:08:29 +02006623/* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6624 *
6625 * If we receive a SYN packet with these bits set, it means a
6626 * network is playing bad games with TOS bits. In order to
6627 * avoid possible false congestion notifications, we disable
stephen hemmingerf4e715c2014-10-29 16:05:06 -07006628 * TCP ECN negotiation.
Florian Westphald82bd122014-09-29 13:08:29 +02006629 *
6630 * Exception: tcp_ca wants ECN. This is required for DCTCP
Florian Westphal843c2fd2015-01-30 20:45:20 +01006631 * congestion control: Linux DCTCP asserts ECT on all packets,
6632 * including SYN, which is most optimal solution; however,
6633 * others, such as FreeBSD do not.
Tilmans, Olivier (Nokia - BE/Antwerp)f6fee162019-04-03 13:49:42 +00006634 *
6635 * Exception: At least one of the reserved bits of the TCP header (th->res1) is
6636 * set, indicating the use of a future TCP extension (such as AccECN). See
6637 * RFC8311 §4.3 which updates RFC3168 to allow the development of such
6638 * extensions.
Florian Westphald82bd122014-09-29 13:08:29 +02006639 */
6640static void tcp_ecn_create_request(struct request_sock *req,
6641 const struct sk_buff *skb,
Florian Westphalf7b3bec2014-11-03 17:35:03 +01006642 const struct sock *listen_sk,
6643 const struct dst_entry *dst)
Florian Westphald82bd122014-09-29 13:08:29 +02006644{
6645 const struct tcphdr *th = tcp_hdr(skb);
6646 const struct net *net = sock_net(listen_sk);
6647 bool th_ecn = th->ece && th->cwr;
Florian Westphal843c2fd2015-01-30 20:45:20 +01006648 bool ect, ecn_ok;
Daniel Borkmannc3a8d942015-08-31 15:58:47 +02006649 u32 ecn_ok_dst;
Florian Westphald82bd122014-09-29 13:08:29 +02006650
6651 if (!th_ecn)
6652 return;
6653
6654 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
Daniel Borkmannc3a8d942015-08-31 15:58:47 +02006655 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6656 ecn_ok = net->ipv4.sysctl_tcp_ecn || ecn_ok_dst;
Florian Westphald82bd122014-09-29 13:08:29 +02006657
Tilmans, Olivier (Nokia - BE/Antwerp)f6fee162019-04-03 13:49:42 +00006658 if (((!ect || th->res1) && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
Lawrence Brakmo91b5b212017-06-30 20:02:49 -07006659 (ecn_ok_dst & DST_FEATURE_ECN_CA) ||
6660 tcp_bpf_ca_needs_ecn((struct sock *)req))
Florian Westphald82bd122014-09-29 13:08:29 +02006661 inet_rsk(req)->ecn_ok = 1;
6662}
6663
Eric Dumazet1bfc4432015-03-16 21:06:19 -07006664static void tcp_openreq_init(struct request_sock *req,
6665 const struct tcp_options_received *rx_opt,
6666 struct sk_buff *skb, const struct sock *sk)
6667{
6668 struct inet_request_sock *ireq = inet_rsk(req);
6669
Eric Dumazeted53d0a2015-10-08 19:33:23 -07006670 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
Eric Dumazet1bfc4432015-03-16 21:06:19 -07006671 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6672 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
Yuchung Cheng9e450c12019-04-29 15:46:15 -07006673 tcp_rsk(req)->snt_synack = 0;
Eric Dumazet1bfc4432015-03-16 21:06:19 -07006674 tcp_rsk(req)->last_oow_ack_time = 0;
6675 req->mss = rx_opt->mss_clamp;
6676 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6677 ireq->tstamp_ok = rx_opt->tstamp_ok;
6678 ireq->sack_ok = rx_opt->sack_ok;
6679 ireq->snd_wscale = rx_opt->snd_wscale;
6680 ireq->wscale_ok = rx_opt->wscale_ok;
6681 ireq->acked = 0;
6682 ireq->ecn_ok = 0;
6683 ireq->ir_rmt_port = tcp_hdr(skb)->source;
6684 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6685 ireq->ir_mark = inet_request_mark(sk, skb);
Ursula Braun60e2a772017-10-25 11:01:45 +02006686#if IS_ENABLED(CONFIG_SMC)
6687 ireq->smc_ok = rx_opt->smc_ok;
6688#endif
Eric Dumazet1bfc4432015-03-16 21:06:19 -07006689}
6690
Eric Dumazete49bb332015-03-17 18:32:27 -07006691struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
Eric Dumazeta1a53442015-10-04 21:08:11 -07006692 struct sock *sk_listener,
6693 bool attach_listener)
Eric Dumazete49bb332015-03-17 18:32:27 -07006694{
Eric Dumazeta1a53442015-10-04 21:08:11 -07006695 struct request_sock *req = reqsk_alloc(ops, sk_listener,
6696 attach_listener);
Eric Dumazete49bb332015-03-17 18:32:27 -07006697
6698 if (req) {
6699 struct inet_request_sock *ireq = inet_rsk(req);
6700
Eric Dumazetc92e8c02017-10-20 09:04:13 -07006701 ireq->ireq_opt = NULL;
Huw Davies56ac42b2016-06-27 15:05:28 -04006702#if IS_ENABLED(CONFIG_IPV6)
6703 ireq->pktopts = NULL;
6704#endif
Yafang Shaoa06ac0d2018-04-24 23:07:45 +08006705 atomic64_set(&ireq->ir_cookie, 0);
Eric Dumazete49bb332015-03-17 18:32:27 -07006706 ireq->ireq_state = TCP_NEW_SYN_RECV;
6707 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
Eric Dumazet0144a812015-03-24 21:45:56 -07006708 ireq->ireq_family = sk_listener->sk_family;
Eric Dumazete49bb332015-03-17 18:32:27 -07006709 }
6710
6711 return req;
6712}
6713EXPORT_SYMBOL(inet_reqsk_alloc);
6714
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006715/*
6716 * Return true if a syncookie should be sent
6717 */
Petar Penkov96511272019-07-29 09:59:13 -07006718static bool tcp_syn_flood_action(const struct sock *sk, const char *proto)
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006719{
Eric Dumazet8d2675f2015-10-02 11:43:25 -07006720 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006721 const char *msg = "Dropping request";
6722 bool want_cookie = false;
Nikolay Borisov12ed8242016-02-03 09:46:51 +02006723 struct net *net = sock_net(sk);
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006724
6725#ifdef CONFIG_SYN_COOKIES
Nikolay Borisov12ed8242016-02-03 09:46:51 +02006726 if (net->ipv4.sysctl_tcp_syncookies) {
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006727 msg = "Sending cookies";
6728 want_cookie = true;
Eric Dumazet02a1d6e2016-04-27 16:44:39 -07006729 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006730 } else
6731#endif
Eric Dumazet02a1d6e2016-04-27 16:44:39 -07006732 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006733
Eric Dumazet8d2675f2015-10-02 11:43:25 -07006734 if (!queue->synflood_warned &&
Nikolay Borisov12ed8242016-02-03 09:46:51 +02006735 net->ipv4.sysctl_tcp_syncookies != 2 &&
Eric Dumazet8d2675f2015-10-02 11:43:25 -07006736 xchg(&queue->synflood_warned, 1) == 0)
Willem de Bruijn0297c1c2018-09-09 19:12:12 -04006737 net_info_ratelimited("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
Petar Penkov96511272019-07-29 09:59:13 -07006738 proto, sk->sk_num, msg);
Eric Dumazet2985aaa2015-09-29 07:42:51 -07006739
Eric Dumazet41d25fe2015-03-25 15:08:47 -07006740 return want_cookie;
6741}
6742
Eric Dumazetcd8ae852015-05-03 21:34:46 -07006743static void tcp_reqsk_record_syn(const struct sock *sk,
6744 struct request_sock *req,
6745 const struct sk_buff *skb)
6746{
6747 if (tcp_sk(sk)->save_syn) {
6748 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
Martin KaFai Lau70a217f2020-08-20 12:00:14 -07006749 struct saved_syn *saved_syn;
Martin KaFai Lau267cf9f2020-08-20 12:01:23 -07006750 u32 mac_hdrlen;
6751 void *base;
6752
6753 if (tcp_sk(sk)->save_syn == 2) { /* Save full header. */
6754 base = skb_mac_header(skb);
6755 mac_hdrlen = skb_mac_header_len(skb);
6756 len += mac_hdrlen;
6757 } else {
6758 base = skb_network_header(skb);
6759 mac_hdrlen = 0;
6760 }
Eric Dumazetcd8ae852015-05-03 21:34:46 -07006761
Martin KaFai Lau70a217f2020-08-20 12:00:14 -07006762 saved_syn = kmalloc(struct_size(saved_syn, data, len),
6763 GFP_ATOMIC);
6764 if (saved_syn) {
Martin KaFai Lau267cf9f2020-08-20 12:01:23 -07006765 saved_syn->mac_hdrlen = mac_hdrlen;
Martin KaFai Lau70a217f2020-08-20 12:00:14 -07006766 saved_syn->network_hdrlen = skb_network_header_len(skb);
6767 saved_syn->tcp_hdrlen = tcp_hdrlen(skb);
Martin KaFai Lau267cf9f2020-08-20 12:01:23 -07006768 memcpy(saved_syn->data, base, len);
Martin KaFai Lau70a217f2020-08-20 12:00:14 -07006769 req->saved_syn = saved_syn;
Eric Dumazetcd8ae852015-05-03 21:34:46 -07006770 }
6771 }
6772}
6773
Petar Penkov9349d602019-07-29 09:59:14 -07006774/* If a SYN cookie is required and supported, returns a clamped MSS value to be
6775 * used for SYN cookie generation.
6776 */
6777u16 tcp_get_syncookie_mss(struct request_sock_ops *rsk_ops,
6778 const struct tcp_request_sock_ops *af_ops,
6779 struct sock *sk, struct tcphdr *th)
6780{
6781 struct tcp_sock *tp = tcp_sk(sk);
6782 u16 mss;
6783
6784 if (sock_net(sk)->ipv4.sysctl_tcp_syncookies != 2 &&
6785 !inet_csk_reqsk_queue_is_full(sk))
6786 return 0;
6787
6788 if (!tcp_syn_flood_action(sk, rsk_ops->slab_name))
6789 return 0;
6790
6791 if (sk_acceptq_is_full(sk)) {
6792 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6793 return 0;
6794 }
6795
6796 mss = tcp_parse_mss_option(th, tp->rx_opt.user_mss);
6797 if (!mss)
6798 mss = af_ops->mss_clamp;
6799
6800 return mss;
6801}
6802EXPORT_SYMBOL_GPL(tcp_get_syncookie_mss);
6803
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006804int tcp_conn_request(struct request_sock_ops *rsk_ops,
6805 const struct tcp_request_sock_ops *af_ops,
6806 struct sock *sk, struct sk_buff *skb)
6807{
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006808 struct tcp_fastopen_cookie foc = { .len = -1 };
Eric Dumazet7c85af82015-09-24 17:16:05 -07006809 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6810 struct tcp_options_received tmp_opt;
6811 struct tcp_sock *tp = tcp_sk(sk);
Nikolay Borisov12ed8242016-02-03 09:46:51 +02006812 struct net *net = sock_net(sk);
Eric Dumazet7c85af82015-09-24 17:16:05 -07006813 struct sock *fastopen_sk = NULL;
Eric Dumazet7c85af82015-09-24 17:16:05 -07006814 struct request_sock *req;
6815 bool want_cookie = false;
Eric Dumazeteaa72dc2017-08-29 15:16:01 -07006816 struct dst_entry *dst;
Eric Dumazet7c85af82015-09-24 17:16:05 -07006817 struct flowi fl;
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006818
6819 /* TW buckets are converted to open requests without
6820 * limitations, they conserve resources and peer is
6821 * evidently real one.
6822 */
Nikolay Borisov12ed8242016-02-03 09:46:51 +02006823 if ((net->ipv4.sysctl_tcp_syncookies == 2 ||
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006824 inet_csk_reqsk_queue_is_full(sk)) && !isn) {
Petar Penkov96511272019-07-29 09:59:13 -07006825 want_cookie = tcp_syn_flood_action(sk, rsk_ops->slab_name);
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006826 if (!want_cookie)
6827 goto drop;
6828 }
6829
Eric Dumazet5ea8ea22016-10-26 09:27:57 -07006830 if (sk_acceptq_is_full(sk)) {
Eric Dumazetc10d9312016-04-29 14:16:47 -07006831 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006832 goto drop;
6833 }
6834
Eric Dumazeta1a53442015-10-04 21:08:11 -07006835 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006836 if (!req)
6837 goto drop;
6838
Florian Westphalf8ace8d2020-07-30 21:25:50 +02006839 req->syncookie = want_cookie;
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006840 tcp_rsk(req)->af_specific = af_ops;
Florian Westphal95a22ca2016-12-01 11:32:06 +01006841 tcp_rsk(req)->ts_off = 0;
Peter Krystadcec37a62020-01-21 16:56:18 -08006842#if IS_ENABLED(CONFIG_MPTCP)
6843 tcp_rsk(req)->is_mptcp = 0;
6844#endif
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006845
6846 tcp_clear_options(&tmp_opt);
6847 tmp_opt.mss_clamp = af_ops->mss_clamp;
6848 tmp_opt.user_mss = tp->rx_opt.user_mss;
Eric Dumazeteed29f12017-06-07 10:34:36 -07006849 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0,
6850 want_cookie ? NULL : &foc);
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006851
6852 if (want_cookie && !tmp_opt.saw_tstamp)
6853 tcp_clear_options(&tmp_opt);
6854
Hans Wippelbc58a1b2018-03-23 11:05:45 +01006855 if (IS_ENABLED(CONFIG_SMC) && want_cookie)
6856 tmp_opt.smc_ok = 0;
6857
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006858 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6859 tcp_openreq_init(req, &tmp_opt, skb, sk);
KOVACS Krisztian7a682572016-09-23 11:27:42 +02006860 inet_rsk(req)->no_srccheck = inet_sk(sk)->transparent;
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006861
Eric Dumazet16f86162015-03-13 15:51:10 -07006862 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
David Ahern6dd9a142015-12-16 13:20:44 -08006863 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
Eric Dumazet16f86162015-03-13 15:51:10 -07006864
Florian Westphal7ea851d2020-11-30 16:36:30 +01006865 dst = af_ops->route_req(sk, skb, &fl, req);
6866 if (!dst)
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006867 goto drop_and_free;
6868
Eric Dumazet84b114b2017-05-05 06:56:54 -07006869 if (tmp_opt.tstamp_ok)
Eric Dumazet5d2ed052017-06-07 10:34:39 -07006870 tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb);
Florian Westphal95a22ca2016-12-01 11:32:06 +01006871
Florian Westphalf7b3bec2014-11-03 17:35:03 +01006872 if (!want_cookie && !isn) {
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006873 /* Kill the following clause, if you dislike this way. */
Soheil Hassas Yeganeh4396e462017-03-15 16:30:46 -04006874 if (!net->ipv4.sysctl_tcp_syncookies &&
6875 (net->ipv4.sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6876 (net->ipv4.sysctl_max_syn_backlog >> 2)) &&
6877 !tcp_peer_is_proven(req, dst)) {
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006878 /* Without syncookies last quarter of
6879 * backlog is filled with destinations,
6880 * proven to be alive.
6881 * It means that we continue to communicate
6882 * to destinations, already remembered
6883 * to the moment of synflood.
6884 */
6885 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6886 rsk_ops->family);
6887 goto drop_and_release;
6888 }
6889
Eric Dumazet84b114b2017-05-05 06:56:54 -07006890 isn = af_ops->init_seq(skb);
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006891 }
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006892
Florian Westphalf7b3bec2014-11-03 17:35:03 +01006893 tcp_ecn_create_request(req, skb, sk, dst);
6894
6895 if (want_cookie) {
6896 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
Florian Westphalf7b3bec2014-11-03 17:35:03 +01006897 if (!tmp_opt.tstamp_ok)
6898 inet_rsk(req)->ecn_ok = 0;
6899 }
6900
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006901 tcp_rsk(req)->snt_isn = isn;
Eric Dumazet58d607d2015-09-15 15:24:20 -07006902 tcp_rsk(req)->txhash = net_tx_rndhash();
Wei Wange9b12ed2020-09-09 17:50:46 -07006903 tcp_rsk(req)->syn_tos = TCP_SKB_CB(skb)->ip_dsfield;
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006904 tcp_openreq_init_rwin(req, sk, dst);
Amritha Nambiarc6345ce2018-06-29 21:26:57 -07006905 sk_rx_queue_set(req_to_sk(req), skb);
Eric Dumazetca6fb062015-10-02 11:43:35 -07006906 if (!want_cookie) {
Eric Dumazetca6fb062015-10-02 11:43:35 -07006907 tcp_reqsk_record_syn(sk, req, skb);
Christoph Paasch71c02372017-10-23 13:22:23 -07006908 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
Eric Dumazetca6fb062015-10-02 11:43:35 -07006909 }
Eric Dumazet7c85af82015-09-24 17:16:05 -07006910 if (fastopen_sk) {
Eric Dumazetca6fb062015-10-02 11:43:35 -07006911 af_ops->send_synack(fastopen_sk, dst, &fl, req,
Martin KaFai Lau331fca42020-08-20 12:00:52 -07006912 &foc, TCP_SYNACK_FASTOPEN, skb);
Eric Dumazet7656d842015-10-04 21:08:07 -07006913 /* Add the child socket directly into the accept queue */
Guillaume Nault9d3e1362019-03-08 22:09:47 +01006914 if (!inet_csk_reqsk_queue_add(sk, req, fastopen_sk)) {
6915 reqsk_fastopen_remove(fastopen_sk, req, false);
6916 bh_unlock_sock(fastopen_sk);
6917 sock_put(fastopen_sk);
Guillaume Nault9403cf22019-03-19 16:05:44 +01006918 goto drop_and_free;
Guillaume Nault9d3e1362019-03-08 22:09:47 +01006919 }
Eric Dumazet7656d842015-10-04 21:08:07 -07006920 sk->sk_data_ready(sk);
6921 bh_unlock_sock(fastopen_sk);
Eric Dumazet7c85af82015-09-24 17:16:05 -07006922 sock_put(fastopen_sk);
6923 } else {
Eric Dumazet9439ce02015-03-17 18:32:29 -07006924 tcp_rsk(req)->tfo_listener = false;
Eric Dumazetca6fb062015-10-02 11:43:35 -07006925 if (!want_cookie)
Lawrence Brakmo8550f322017-06-30 20:02:42 -07006926 inet_csk_reqsk_queue_hash_add(sk, req,
6927 tcp_timeout_init((struct sock *)req));
Eric Dumazetb3d05142016-04-13 22:05:39 -07006928 af_ops->send_synack(sk, dst, &fl, req, &foc,
6929 !want_cookie ? TCP_SYNACK_NORMAL :
Martin KaFai Lau331fca42020-08-20 12:00:52 -07006930 TCP_SYNACK_COOKIE,
6931 skb);
Eric Dumazet9caad862016-04-01 08:52:20 -07006932 if (want_cookie) {
6933 reqsk_free(req);
6934 return 0;
6935 }
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006936 }
Eric Dumazetca6fb062015-10-02 11:43:35 -07006937 reqsk_put(req);
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006938 return 0;
6939
6940drop_and_release:
6941 dst_release(dst);
6942drop_and_free:
Guillaume Nault9403cf22019-03-19 16:05:44 +01006943 __reqsk_free(req);
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006944drop:
Eric Dumazet9caad862016-04-01 08:52:20 -07006945 tcp_listendrop(sk);
Octavian Purdila1fb6f152014-06-25 17:10:02 +03006946 return 0;
6947}
6948EXPORT_SYMBOL(tcp_conn_request);