xref: /openbmc/linux/net/ipv4/tcp_input.c (revision 9144f784f852f9a125cabe9927b986d909bfa439)
1  // SPDX-License-Identifier: GPL-2.0
2  /*
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   *
9   * Authors:	Ross Biro
10   *		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.
44   *		Andrey Savochkin:	Fix RTT measurements in the presence of
45   *					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.
52   *		Andi Kleen:		Add tcp_measure_rcv_mss to make
53   *					connections with MSS<min(MTU,ann. MSS)
54   *					work without delayed acks.
55   *		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
63   */
64  
65  #define pr_fmt(fmt) "TCP: " fmt
66  
67  #include <linux/mm.h>
68  #include <linux/slab.h>
69  #include <linux/module.h>
70  #include <linux/sysctl.h>
71  #include <linux/kernel.h>
72  #include <linux/prefetch.h>
73  #include <net/dst.h>
74  #include <net/tcp.h>
75  #include <net/inet_common.h>
76  #include <linux/ipsec.h>
77  #include <asm/unaligned.h>
78  #include <linux/errqueue.h>
79  #include <trace/events/tcp.h>
80  #include <linux/jump_label_ratelimit.h>
81  #include <net/busy_poll.h>
82  #include <net/mptcp.h>
83  
84  int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
85  
86  #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				*/
93  #define FLAG_LOST_RETRANS	0x80 /* This ACK marks some retransmission lost */
94  #define FLAG_SLOWPATH		0x100 /* Do not skip RFC checks for window update.*/
95  #define FLAG_ORIG_SACK_ACKED	0x200 /* Never retransmitted data are (s)acked	*/
96  #define FLAG_SND_UNA_ADVANCED	0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
97  #define FLAG_DSACKING_ACK	0x800 /* SACK blocks contained D-SACK info */
98  #define FLAG_SET_XMIT_TIMER	0x1000 /* Set TLP or RTO timer */
99  #define FLAG_SACK_RENEGING	0x2000 /* snd_una advanced to a sacked seq */
100  #define FLAG_UPDATE_TS_RECENT	0x4000 /* tcp_replace_ts_recent() */
101  #define FLAG_NO_CHALLENGE_ACK	0x8000 /* do not call tcp_send_challenge_ack()	*/
102  #define FLAG_ACK_MAYBE_DELAYED	0x10000 /* Likely a delayed ACK */
103  #define FLAG_DSACK_TLP		0x20000 /* DSACK for tail loss probe */
104  
105  #define FLAG_ACKED		(FLAG_DATA_ACKED|FLAG_SYN_ACKED)
106  #define FLAG_NOT_DUP		(FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
107  #define FLAG_CA_ALERT		(FLAG_DATA_SACKED|FLAG_ECE|FLAG_DSACKING_ACK)
108  #define FLAG_FORWARD_PROGRESS	(FLAG_ACKED|FLAG_DATA_SACKED)
109  
110  #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
111  #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
112  
113  #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  
117  #if IS_ENABLED(CONFIG_TLS_DEVICE)
118  static DEFINE_STATIC_KEY_DEFERRED_FALSE(clean_acked_data_enabled, HZ);
119  
clean_acked_data_enable(struct inet_connection_sock * icsk,void (* cad)(struct sock * sk,u32 ack_seq))120  void 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;
124  	static_branch_deferred_inc(&clean_acked_data_enabled);
125  }
126  EXPORT_SYMBOL_GPL(clean_acked_data_enable);
127  
clean_acked_data_disable(struct inet_connection_sock * icsk)128  void clean_acked_data_disable(struct inet_connection_sock *icsk)
129  {
130  	static_branch_slow_dec_deferred(&clean_acked_data_enabled);
131  	icsk->icsk_clean_acked = NULL;
132  }
133  EXPORT_SYMBOL_GPL(clean_acked_data_disable);
134  
clean_acked_data_flush(void)135  void clean_acked_data_flush(void)
136  {
137  	static_key_deferred_flush(&clean_acked_data_enabled);
138  }
139  EXPORT_SYMBOL_GPL(clean_acked_data_flush);
140  #endif
141  
142  #ifdef CONFIG_CGROUP_BPF
bpf_skops_parse_hdr(struct sock * sk,struct sk_buff * skb)143  static 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);
150  	struct bpf_sock_ops_kern sock_ops;
151  
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  
166  	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);
175  }
176  
bpf_skops_established(struct sock * sk,int bpf_op,struct sk_buff * skb)177  static 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;
188  	/* 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));
191  
192  	BPF_CGROUP_RUN_PROG_SOCK_OPS(&sock_ops);
193  }
194  #else
bpf_skops_parse_hdr(struct sock * sk,struct sk_buff * skb)195  static void bpf_skops_parse_hdr(struct sock *sk, struct sk_buff *skb)
196  {
197  }
198  
bpf_skops_established(struct sock * sk,int bpf_op,struct sk_buff * skb)199  static void bpf_skops_established(struct sock *sk, int bpf_op,
200  				  struct sk_buff *skb)
201  {
202  }
203  #endif
204  
tcp_gro_dev_warn(struct sock * sk,const struct sk_buff * skb,unsigned int len)205  static void tcp_gro_dev_warn(struct sock *sk, const struct sk_buff *skb,
206  			     unsigned int len)
207  {
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);
217  		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");
220  		rcu_read_unlock();
221  	}
222  }
223  
224  /* Adapt the MSS value used to make delayed ack decision to the
225   * real world.
226   */
tcp_measure_rcv_mss(struct sock * sk,const struct sk_buff * skb)227  static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
228  {
229  	struct inet_connection_sock *icsk = inet_csk(sk);
230  	const unsigned int lss = icsk->icsk_ack.last_seg_size;
231  	unsigned int len;
232  
233  	icsk->icsk_ack.last_seg_size = 0;
234  
235  	/* skb->len may jitter because of SACKs, even if peer
236  	 * sends good full-sized frames.
237  	 */
238  	len = skb_shinfo(skb)->gso_size ? : skb->len;
239  	if (len >= icsk->icsk_ack.rcv_mss) {
240  		/* Note: divides are still a bit expensive.
241  		 * For the moment, only adjust scaling_ratio
242  		 * when we update icsk_ack.rcv_mss.
243  		 */
244  		if (unlikely(len != icsk->icsk_ack.rcv_mss)) {
245  			u64 val = (u64)skb->len << TCP_RMEM_TO_WIN_SCALE;
246  			u8 old_ratio = tcp_sk(sk)->scaling_ratio;
247  
248  			do_div(val, skb->truesize);
249  			tcp_sk(sk)->scaling_ratio = val ? val : 1;
250  
251  			if (old_ratio != tcp_sk(sk)->scaling_ratio)
252  				WRITE_ONCE(tcp_sk(sk)->window_clamp,
253  					   tcp_win_from_space(sk, sk->sk_rcvbuf));
254  		}
255  		icsk->icsk_ack.rcv_mss = min_t(unsigned int, len,
256  					       tcp_sk(sk)->advmss);
257  		/* Account for possibly-removed options */
258  		if (unlikely(len > icsk->icsk_ack.rcv_mss +
259  				   MAX_TCP_OPTION_SPACE))
260  			tcp_gro_dev_warn(sk, skb, len);
261  		/* If the skb has a len of exactly 1*MSS and has the PSH bit
262  		 * set then it is likely the end of an application write. So
263  		 * more data may not be arriving soon, and yet the data sender
264  		 * may be waiting for an ACK if cwnd-bound or using TX zero
265  		 * copy. So we set ICSK_ACK_PUSHED here so that
266  		 * tcp_cleanup_rbuf() will send an ACK immediately if the app
267  		 * reads all of the data and is not ping-pong. If len > MSS
268  		 * then this logic does not matter (and does not hurt) because
269  		 * tcp_cleanup_rbuf() will always ACK immediately if the app
270  		 * reads data and there is more than an MSS of unACKed data.
271  		 */
272  		if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_PSH)
273  			icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
274  	} else {
275  		/* Otherwise, we make more careful check taking into account,
276  		 * that SACKs block is variable.
277  		 *
278  		 * "len" is invariant segment length, including TCP header.
279  		 */
280  		len += skb->data - skb_transport_header(skb);
281  		if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
282  		    /* If PSH is not set, packet should be
283  		     * full sized, provided peer TCP is not badly broken.
284  		     * This observation (if it is correct 8)) allows
285  		     * to handle super-low mtu links fairly.
286  		     */
287  		    (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
288  		     !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
289  			/* Subtract also invariant (if peer is RFC compliant),
290  			 * tcp header plus fixed timestamp option length.
291  			 * Resulting "len" is MSS free of SACK jitter.
292  			 */
293  			len -= tcp_sk(sk)->tcp_header_len;
294  			icsk->icsk_ack.last_seg_size = len;
295  			if (len == lss) {
296  				icsk->icsk_ack.rcv_mss = len;
297  				return;
298  			}
299  		}
300  		if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
301  			icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
302  		icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
303  	}
304  }
305  
tcp_incr_quickack(struct sock * sk,unsigned int max_quickacks)306  static void tcp_incr_quickack(struct sock *sk, unsigned int max_quickacks)
307  {
308  	struct inet_connection_sock *icsk = inet_csk(sk);
309  	unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
310  
311  	if (quickacks == 0)
312  		quickacks = 2;
313  	quickacks = min(quickacks, max_quickacks);
314  	if (quickacks > icsk->icsk_ack.quick)
315  		icsk->icsk_ack.quick = quickacks;
316  }
317  
tcp_enter_quickack_mode(struct sock * sk,unsigned int max_quickacks)318  static void tcp_enter_quickack_mode(struct sock *sk, unsigned int max_quickacks)
319  {
320  	struct inet_connection_sock *icsk = inet_csk(sk);
321  
322  	tcp_incr_quickack(sk, max_quickacks);
323  	inet_csk_exit_pingpong_mode(sk);
324  	icsk->icsk_ack.ato = TCP_ATO_MIN;
325  }
326  
327  /* Send ACKs quickly, if "quick" count is not exhausted
328   * and the session is not interactive.
329   */
330  
tcp_in_quickack_mode(struct sock * sk)331  static bool tcp_in_quickack_mode(struct sock *sk)
332  {
333  	const struct inet_connection_sock *icsk = inet_csk(sk);
334  	const struct dst_entry *dst = __sk_dst_get(sk);
335  
336  	return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
337  		(icsk->icsk_ack.quick && !inet_csk_in_pingpong_mode(sk));
338  }
339  
tcp_ecn_queue_cwr(struct tcp_sock * tp)340  static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
341  {
342  	if (tp->ecn_flags & TCP_ECN_OK)
343  		tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
344  }
345  
tcp_ecn_accept_cwr(struct sock * sk,const struct sk_buff * skb)346  static void tcp_ecn_accept_cwr(struct sock *sk, const struct sk_buff *skb)
347  {
348  	if (tcp_hdr(skb)->cwr) {
349  		tcp_sk(sk)->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
350  
351  		/* If the sender is telling us it has entered CWR, then its
352  		 * cwnd may be very low (even just 1 packet), so we should ACK
353  		 * immediately.
354  		 */
355  		if (TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq)
356  			inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
357  	}
358  }
359  
tcp_ecn_withdraw_cwr(struct tcp_sock * tp)360  static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
361  {
362  	tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
363  }
364  
__tcp_ecn_check_ce(struct sock * sk,const struct sk_buff * skb)365  static void __tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
366  {
367  	struct tcp_sock *tp = tcp_sk(sk);
368  
369  	switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
370  	case INET_ECN_NOT_ECT:
371  		/* Funny extension: if ECT is not set on a segment,
372  		 * and we already seen ECT on a previous segment,
373  		 * it is probably a retransmit.
374  		 */
375  		if (tp->ecn_flags & TCP_ECN_SEEN)
376  			tcp_enter_quickack_mode(sk, 2);
377  		break;
378  	case INET_ECN_CE:
379  		if (tcp_ca_needs_ecn(sk))
380  			tcp_ca_event(sk, CA_EVENT_ECN_IS_CE);
381  
382  		if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
383  			/* Better not delay acks, sender can have a very low cwnd */
384  			tcp_enter_quickack_mode(sk, 2);
385  			tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
386  		}
387  		tp->ecn_flags |= TCP_ECN_SEEN;
388  		break;
389  	default:
390  		if (tcp_ca_needs_ecn(sk))
391  			tcp_ca_event(sk, CA_EVENT_ECN_NO_CE);
392  		tp->ecn_flags |= TCP_ECN_SEEN;
393  		break;
394  	}
395  }
396  
tcp_ecn_check_ce(struct sock * sk,const struct sk_buff * skb)397  static void tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
398  {
399  	if (tcp_sk(sk)->ecn_flags & TCP_ECN_OK)
400  		__tcp_ecn_check_ce(sk, skb);
401  }
402  
tcp_ecn_rcv_synack(struct tcp_sock * tp,const struct tcphdr * th)403  static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
404  {
405  	if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
406  		tp->ecn_flags &= ~TCP_ECN_OK;
407  }
408  
tcp_ecn_rcv_syn(struct tcp_sock * tp,const struct tcphdr * th)409  static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
410  {
411  	if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
412  		tp->ecn_flags &= ~TCP_ECN_OK;
413  }
414  
tcp_ecn_rcv_ecn_echo(const struct tcp_sock * tp,const struct tcphdr * th)415  static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
416  {
417  	if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
418  		return true;
419  	return false;
420  }
421  
422  /* Buffer size and advertised window tuning.
423   *
424   * 1. Tuning sk->sk_sndbuf, when connection enters established state.
425   */
426  
tcp_sndbuf_expand(struct sock * sk)427  static void tcp_sndbuf_expand(struct sock *sk)
428  {
429  	const struct tcp_sock *tp = tcp_sk(sk);
430  	const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
431  	int sndmem, per_mss;
432  	u32 nr_segs;
433  
434  	/* Worst case is non GSO/TSO : each frame consumes one skb
435  	 * and skb->head is kmalloced using power of two area of memory
436  	 */
437  	per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
438  		  MAX_TCP_HEADER +
439  		  SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
440  
441  	per_mss = roundup_pow_of_two(per_mss) +
442  		  SKB_DATA_ALIGN(sizeof(struct sk_buff));
443  
444  	nr_segs = max_t(u32, TCP_INIT_CWND, tcp_snd_cwnd(tp));
445  	nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
446  
447  	/* Fast Recovery (RFC 5681 3.2) :
448  	 * Cubic needs 1.7 factor, rounded to 2 to include
449  	 * extra cushion (application might react slowly to EPOLLOUT)
450  	 */
451  	sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2;
452  	sndmem *= nr_segs * per_mss;
453  
454  	if (sk->sk_sndbuf < sndmem)
455  		WRITE_ONCE(sk->sk_sndbuf,
456  			   min(sndmem, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[2])));
457  }
458  
459  /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
460   *
461   * All tcp_full_space() is split to two parts: "network" buffer, allocated
462   * forward and advertised in receiver window (tp->rcv_wnd) and
463   * "application buffer", required to isolate scheduling/application
464   * latencies from network.
465   * window_clamp is maximal advertised window. It can be less than
466   * tcp_full_space(), in this case tcp_full_space() - window_clamp
467   * is reserved for "application" buffer. The less window_clamp is
468   * the smoother our behaviour from viewpoint of network, but the lower
469   * throughput and the higher sensitivity of the connection to losses. 8)
470   *
471   * rcv_ssthresh is more strict window_clamp used at "slow start"
472   * phase to predict further behaviour of this connection.
473   * It is used for two goals:
474   * - to enforce header prediction at sender, even when application
475   *   requires some significant "application buffer". It is check #1.
476   * - to prevent pruning of receive queue because of misprediction
477   *   of receiver window. Check #2.
478   *
479   * The scheme does not work when sender sends good segments opening
480   * window and then starts to feed us spaghetti. But it should work
481   * in common situations. Otherwise, we have to rely on queue collapsing.
482   */
483  
484  /* Slow part of check#2. */
__tcp_grow_window(const struct sock * sk,const struct sk_buff * skb,unsigned int skbtruesize)485  static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb,
486  			     unsigned int skbtruesize)
487  {
488  	const struct tcp_sock *tp = tcp_sk(sk);
489  	/* Optimize this! */
490  	int truesize = tcp_win_from_space(sk, skbtruesize) >> 1;
491  	int window = tcp_win_from_space(sk, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2])) >> 1;
492  
493  	while (tp->rcv_ssthresh <= window) {
494  		if (truesize <= skb->len)
495  			return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
496  
497  		truesize >>= 1;
498  		window >>= 1;
499  	}
500  	return 0;
501  }
502  
503  /* Even if skb appears to have a bad len/truesize ratio, TCP coalescing
504   * can play nice with us, as sk_buff and skb->head might be either
505   * freed or shared with up to MAX_SKB_FRAGS segments.
506   * Only give a boost to drivers using page frag(s) to hold the frame(s),
507   * and if no payload was pulled in skb->head before reaching us.
508   */
truesize_adjust(bool adjust,const struct sk_buff * skb)509  static u32 truesize_adjust(bool adjust, const struct sk_buff *skb)
510  {
511  	u32 truesize = skb->truesize;
512  
513  	if (adjust && !skb_headlen(skb)) {
514  		truesize -= SKB_TRUESIZE(skb_end_offset(skb));
515  		/* paranoid check, some drivers might be buggy */
516  		if (unlikely((int)truesize < (int)skb->len))
517  			truesize = skb->truesize;
518  	}
519  	return truesize;
520  }
521  
tcp_grow_window(struct sock * sk,const struct sk_buff * skb,bool adjust)522  static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb,
523  			    bool adjust)
524  {
525  	struct tcp_sock *tp = tcp_sk(sk);
526  	int room;
527  
528  	room = min_t(int, tp->window_clamp, tcp_space(sk)) - tp->rcv_ssthresh;
529  
530  	if (room <= 0)
531  		return;
532  
533  	/* Check #1 */
534  	if (!tcp_under_memory_pressure(sk)) {
535  		unsigned int truesize = truesize_adjust(adjust, skb);
536  		int incr;
537  
538  		/* Check #2. Increase window, if skb with such overhead
539  		 * will fit to rcvbuf in future.
540  		 */
541  		if (tcp_win_from_space(sk, truesize) <= skb->len)
542  			incr = 2 * tp->advmss;
543  		else
544  			incr = __tcp_grow_window(sk, skb, truesize);
545  
546  		if (incr) {
547  			incr = max_t(int, incr, 2 * skb->len);
548  			tp->rcv_ssthresh += min(room, incr);
549  			inet_csk(sk)->icsk_ack.quick |= 1;
550  		}
551  	} else {
552  		/* Under pressure:
553  		 * Adjust rcv_ssthresh according to reserved mem
554  		 */
555  		tcp_adjust_rcv_ssthresh(sk);
556  	}
557  }
558  
559  /* 3. Try to fixup all. It is made immediately after connection enters
560   *    established state.
561   */
tcp_init_buffer_space(struct sock * sk)562  static void tcp_init_buffer_space(struct sock *sk)
563  {
564  	int tcp_app_win = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_app_win);
565  	struct tcp_sock *tp = tcp_sk(sk);
566  	int maxwin;
567  
568  	if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
569  		tcp_sndbuf_expand(sk);
570  
571  	tcp_mstamp_refresh(tp);
572  	tp->rcvq_space.time = tp->tcp_mstamp;
573  	tp->rcvq_space.seq = tp->copied_seq;
574  
575  	maxwin = tcp_full_space(sk);
576  
577  	if (tp->window_clamp >= maxwin) {
578  		WRITE_ONCE(tp->window_clamp, maxwin);
579  
580  		if (tcp_app_win && maxwin > 4 * tp->advmss)
581  			WRITE_ONCE(tp->window_clamp,
582  				   max(maxwin - (maxwin >> tcp_app_win),
583  				       4 * tp->advmss));
584  	}
585  
586  	/* Force reservation of one segment. */
587  	if (tcp_app_win &&
588  	    tp->window_clamp > 2 * tp->advmss &&
589  	    tp->window_clamp + tp->advmss > maxwin)
590  		WRITE_ONCE(tp->window_clamp,
591  			   max(2 * tp->advmss, maxwin - tp->advmss));
592  
593  	tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
594  	tp->snd_cwnd_stamp = tcp_jiffies32;
595  	tp->rcvq_space.space = min3(tp->rcv_ssthresh, tp->rcv_wnd,
596  				    (u32)TCP_INIT_CWND * tp->advmss);
597  }
598  
599  /* 4. Recalculate window clamp after socket hit its memory bounds. */
tcp_clamp_window(struct sock * sk)600  static void tcp_clamp_window(struct sock *sk)
601  {
602  	struct tcp_sock *tp = tcp_sk(sk);
603  	struct inet_connection_sock *icsk = inet_csk(sk);
604  	struct net *net = sock_net(sk);
605  	int rmem2;
606  
607  	icsk->icsk_ack.quick = 0;
608  	rmem2 = READ_ONCE(net->ipv4.sysctl_tcp_rmem[2]);
609  
610  	if (sk->sk_rcvbuf < rmem2 &&
611  	    !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
612  	    !tcp_under_memory_pressure(sk) &&
613  	    sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
614  		WRITE_ONCE(sk->sk_rcvbuf,
615  			   min(atomic_read(&sk->sk_rmem_alloc), rmem2));
616  	}
617  	if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
618  		tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
619  }
620  
621  /* Initialize RCV_MSS value.
622   * RCV_MSS is an our guess about MSS used by the peer.
623   * We haven't any direct information about the MSS.
624   * It's better to underestimate the RCV_MSS rather than overestimate.
625   * Overestimations make us ACKing less frequently than needed.
626   * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
627   */
tcp_initialize_rcv_mss(struct sock * sk)628  void tcp_initialize_rcv_mss(struct sock *sk)
629  {
630  	const struct tcp_sock *tp = tcp_sk(sk);
631  	unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
632  
633  	hint = min(hint, tp->rcv_wnd / 2);
634  	hint = min(hint, TCP_MSS_DEFAULT);
635  	hint = max(hint, TCP_MIN_MSS);
636  
637  	inet_csk(sk)->icsk_ack.rcv_mss = hint;
638  }
639  EXPORT_SYMBOL(tcp_initialize_rcv_mss);
640  
641  /* Receiver "autotuning" code.
642   *
643   * The algorithm for RTT estimation w/o timestamps is based on
644   * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
645   * <https://public.lanl.gov/radiant/pubs.html#DRS>
646   *
647   * More detail on this code can be found at
648   * <http://staff.psc.edu/jheffner/>,
649   * though this reference is out of date.  A new paper
650   * is pending.
651   */
tcp_rcv_rtt_update(struct tcp_sock * tp,u32 sample,int win_dep)652  static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
653  {
654  	u32 new_sample = tp->rcv_rtt_est.rtt_us;
655  	long m = sample;
656  
657  	if (new_sample != 0) {
658  		/* If we sample in larger samples in the non-timestamp
659  		 * case, we could grossly overestimate the RTT especially
660  		 * with chatty applications or bulk transfer apps which
661  		 * are stalled on filesystem I/O.
662  		 *
663  		 * Also, since we are only going for a minimum in the
664  		 * non-timestamp case, we do not smooth things out
665  		 * else with timestamps disabled convergence takes too
666  		 * long.
667  		 */
668  		if (!win_dep) {
669  			m -= (new_sample >> 3);
670  			new_sample += m;
671  		} else {
672  			m <<= 3;
673  			if (m < new_sample)
674  				new_sample = m;
675  		}
676  	} else {
677  		/* No previous measure. */
678  		new_sample = m << 3;
679  	}
680  
681  	tp->rcv_rtt_est.rtt_us = new_sample;
682  }
683  
tcp_rcv_rtt_measure(struct tcp_sock * tp)684  static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
685  {
686  	u32 delta_us;
687  
688  	if (tp->rcv_rtt_est.time == 0)
689  		goto new_measure;
690  	if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
691  		return;
692  	delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time);
693  	if (!delta_us)
694  		delta_us = 1;
695  	tcp_rcv_rtt_update(tp, delta_us, 1);
696  
697  new_measure:
698  	tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
699  	tp->rcv_rtt_est.time = tp->tcp_mstamp;
700  }
701  
tcp_rcv_rtt_measure_ts(struct sock * sk,const struct sk_buff * skb)702  static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
703  					  const struct sk_buff *skb)
704  {
705  	struct tcp_sock *tp = tcp_sk(sk);
706  
707  	if (tp->rx_opt.rcv_tsecr == tp->rcv_rtt_last_tsecr)
708  		return;
709  	tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
710  
711  	if (TCP_SKB_CB(skb)->end_seq -
712  	    TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss) {
713  		u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
714  		u32 delta_us;
715  
716  		if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
717  			if (!delta)
718  				delta = 1;
719  			delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
720  			tcp_rcv_rtt_update(tp, delta_us, 0);
721  		}
722  	}
723  }
724  
725  /*
726   * This function should be called every time data is copied to user space.
727   * It calculates the appropriate TCP receive buffer space.
728   */
tcp_rcv_space_adjust(struct sock * sk)729  void tcp_rcv_space_adjust(struct sock *sk)
730  {
731  	struct tcp_sock *tp = tcp_sk(sk);
732  	u32 copied;
733  	int time;
734  
735  	trace_tcp_rcv_space_adjust(sk);
736  
737  	tcp_mstamp_refresh(tp);
738  	time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time);
739  	if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0)
740  		return;
741  
742  	/* Number of bytes copied to user in last RTT */
743  	copied = tp->copied_seq - tp->rcvq_space.seq;
744  	if (copied <= tp->rcvq_space.space)
745  		goto new_measure;
746  
747  	/* A bit of theory :
748  	 * copied = bytes received in previous RTT, our base window
749  	 * To cope with packet losses, we need a 2x factor
750  	 * To cope with slow start, and sender growing its cwin by 100 %
751  	 * every RTT, we need a 4x factor, because the ACK we are sending
752  	 * now is for the next RTT, not the current one :
753  	 * <prev RTT . ><current RTT .. ><next RTT .... >
754  	 */
755  
756  	if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) &&
757  	    !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
758  		u64 rcvwin, grow;
759  		int rcvbuf;
760  
761  		/* minimal window to cope with packet losses, assuming
762  		 * steady state. Add some cushion because of small variations.
763  		 */
764  		rcvwin = ((u64)copied << 1) + 16 * tp->advmss;
765  
766  		/* Accommodate for sender rate increase (eg. slow start) */
767  		grow = rcvwin * (copied - tp->rcvq_space.space);
768  		do_div(grow, tp->rcvq_space.space);
769  		rcvwin += (grow << 1);
770  
771  		rcvbuf = min_t(u64, tcp_space_from_win(sk, rcvwin),
772  			       READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
773  		if (rcvbuf > sk->sk_rcvbuf) {
774  			WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
775  
776  			/* Make the window clamp follow along.  */
777  			WRITE_ONCE(tp->window_clamp,
778  				   tcp_win_from_space(sk, rcvbuf));
779  		}
780  	}
781  	tp->rcvq_space.space = copied;
782  
783  new_measure:
784  	tp->rcvq_space.seq = tp->copied_seq;
785  	tp->rcvq_space.time = tp->tcp_mstamp;
786  }
787  
788  /* There is something which you must keep in mind when you analyze the
789   * behavior of the tp->ato delayed ack timeout interval.  When a
790   * connection starts up, we want to ack as quickly as possible.  The
791   * problem is that "good" TCP's do slow start at the beginning of data
792   * transmission.  The means that until we send the first few ACK's the
793   * sender will sit on his end and only queue most of his data, because
794   * he can only send snd_cwnd unacked packets at any given time.  For
795   * each ACK we send, he increments snd_cwnd and transmits more of his
796   * queue.  -DaveM
797   */
tcp_event_data_recv(struct sock * sk,struct sk_buff * skb)798  static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
799  {
800  	struct tcp_sock *tp = tcp_sk(sk);
801  	struct inet_connection_sock *icsk = inet_csk(sk);
802  	u32 now;
803  
804  	inet_csk_schedule_ack(sk);
805  
806  	tcp_measure_rcv_mss(sk, skb);
807  
808  	tcp_rcv_rtt_measure(tp);
809  
810  	now = tcp_jiffies32;
811  
812  	if (!icsk->icsk_ack.ato) {
813  		/* The _first_ data packet received, initialize
814  		 * delayed ACK engine.
815  		 */
816  		tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
817  		icsk->icsk_ack.ato = TCP_ATO_MIN;
818  	} else {
819  		int m = now - icsk->icsk_ack.lrcvtime;
820  
821  		if (m <= TCP_ATO_MIN / 2) {
822  			/* The fastest case is the first. */
823  			icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
824  		} else if (m < icsk->icsk_ack.ato) {
825  			icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
826  			if (icsk->icsk_ack.ato > icsk->icsk_rto)
827  				icsk->icsk_ack.ato = icsk->icsk_rto;
828  		} else if (m > icsk->icsk_rto) {
829  			/* Too long gap. Apparently sender failed to
830  			 * restart window, so that we send ACKs quickly.
831  			 */
832  			tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
833  		}
834  	}
835  	icsk->icsk_ack.lrcvtime = now;
836  
837  	tcp_ecn_check_ce(sk, skb);
838  
839  	if (skb->len >= 128)
840  		tcp_grow_window(sk, skb, true);
841  }
842  
843  /* Called to compute a smoothed rtt estimate. The data fed to this
844   * routine either comes from timestamps, or from segments that were
845   * known _not_ to have been retransmitted [see Karn/Partridge
846   * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
847   * piece by Van Jacobson.
848   * NOTE: the next three routines used to be one big routine.
849   * To save cycles in the RFC 1323 implementation it was better to break
850   * it up into three procedures. -- erics
851   */
tcp_rtt_estimator(struct sock * sk,long mrtt_us)852  static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
853  {
854  	struct tcp_sock *tp = tcp_sk(sk);
855  	long m = mrtt_us; /* RTT */
856  	u32 srtt = tp->srtt_us;
857  
858  	/*	The following amusing code comes from Jacobson's
859  	 *	article in SIGCOMM '88.  Note that rtt and mdev
860  	 *	are scaled versions of rtt and mean deviation.
861  	 *	This is designed to be as fast as possible
862  	 *	m stands for "measurement".
863  	 *
864  	 *	On a 1990 paper the rto value is changed to:
865  	 *	RTO = rtt + 4 * mdev
866  	 *
867  	 * Funny. This algorithm seems to be very broken.
868  	 * These formulae increase RTO, when it should be decreased, increase
869  	 * too slowly, when it should be increased quickly, decrease too quickly
870  	 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
871  	 * does not matter how to _calculate_ it. Seems, it was trap
872  	 * that VJ failed to avoid. 8)
873  	 */
874  	if (srtt != 0) {
875  		m -= (srtt >> 3);	/* m is now error in rtt est */
876  		srtt += m;		/* rtt = 7/8 rtt + 1/8 new */
877  		if (m < 0) {
878  			m = -m;		/* m is now abs(error) */
879  			m -= (tp->mdev_us >> 2);   /* similar update on mdev */
880  			/* This is similar to one of Eifel findings.
881  			 * Eifel blocks mdev updates when rtt decreases.
882  			 * This solution is a bit different: we use finer gain
883  			 * for mdev in this case (alpha*beta).
884  			 * Like Eifel it also prevents growth of rto,
885  			 * but also it limits too fast rto decreases,
886  			 * happening in pure Eifel.
887  			 */
888  			if (m > 0)
889  				m >>= 3;
890  		} else {
891  			m -= (tp->mdev_us >> 2);   /* similar update on mdev */
892  		}
893  		tp->mdev_us += m;		/* mdev = 3/4 mdev + 1/4 new */
894  		if (tp->mdev_us > tp->mdev_max_us) {
895  			tp->mdev_max_us = tp->mdev_us;
896  			if (tp->mdev_max_us > tp->rttvar_us)
897  				tp->rttvar_us = tp->mdev_max_us;
898  		}
899  		if (after(tp->snd_una, tp->rtt_seq)) {
900  			if (tp->mdev_max_us < tp->rttvar_us)
901  				tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
902  			tp->rtt_seq = tp->snd_nxt;
903  			tp->mdev_max_us = tcp_rto_min_us(sk);
904  
905  			tcp_bpf_rtt(sk);
906  		}
907  	} else {
908  		/* no previous measure. */
909  		srtt = m << 3;		/* take the measured time to be rtt */
910  		tp->mdev_us = m << 1;	/* make sure rto = 3*rtt */
911  		tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
912  		tp->mdev_max_us = tp->rttvar_us;
913  		tp->rtt_seq = tp->snd_nxt;
914  
915  		tcp_bpf_rtt(sk);
916  	}
917  	tp->srtt_us = max(1U, srtt);
918  }
919  
tcp_update_pacing_rate(struct sock * sk)920  static void tcp_update_pacing_rate(struct sock *sk)
921  {
922  	const struct tcp_sock *tp = tcp_sk(sk);
923  	u64 rate;
924  
925  	/* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
926  	rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
927  
928  	/* current rate is (cwnd * mss) / srtt
929  	 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
930  	 * In Congestion Avoidance phase, set it to 120 % the current rate.
931  	 *
932  	 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
933  	 *	 If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
934  	 *	 end of slow start and should slow down.
935  	 */
936  	if (tcp_snd_cwnd(tp) < tp->snd_ssthresh / 2)
937  		rate *= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio);
938  	else
939  		rate *= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio);
940  
941  	rate *= max(tcp_snd_cwnd(tp), tp->packets_out);
942  
943  	if (likely(tp->srtt_us))
944  		do_div(rate, tp->srtt_us);
945  
946  	/* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate
947  	 * without any lock. We want to make sure compiler wont store
948  	 * intermediate values in this location.
949  	 */
950  	WRITE_ONCE(sk->sk_pacing_rate, min_t(u64, rate,
951  					     sk->sk_max_pacing_rate));
952  }
953  
954  /* Calculate rto without backoff.  This is the second half of Van Jacobson's
955   * routine referred to above.
956   */
tcp_set_rto(struct sock * sk)957  static void tcp_set_rto(struct sock *sk)
958  {
959  	const struct tcp_sock *tp = tcp_sk(sk);
960  	/* Old crap is replaced with new one. 8)
961  	 *
962  	 * More seriously:
963  	 * 1. If rtt variance happened to be less 50msec, it is hallucination.
964  	 *    It cannot be less due to utterly erratic ACK generation made
965  	 *    at least by solaris and freebsd. "Erratic ACKs" has _nothing_
966  	 *    to do with delayed acks, because at cwnd>2 true delack timeout
967  	 *    is invisible. Actually, Linux-2.4 also generates erratic
968  	 *    ACKs in some circumstances.
969  	 */
970  	inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
971  
972  	/* 2. Fixups made earlier cannot be right.
973  	 *    If we do not estimate RTO correctly without them,
974  	 *    all the algo is pure shit and should be replaced
975  	 *    with correct one. It is exactly, which we pretend to do.
976  	 */
977  
978  	/* NOTE: clamping at TCP_RTO_MIN is not required, current algo
979  	 * guarantees that rto is higher.
980  	 */
981  	tcp_bound_rto(sk);
982  }
983  
tcp_init_cwnd(const struct tcp_sock * tp,const struct dst_entry * dst)984  __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
985  {
986  	__u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
987  
988  	if (!cwnd)
989  		cwnd = TCP_INIT_CWND;
990  	return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
991  }
992  
993  struct tcp_sacktag_state {
994  	/* Timestamps for earliest and latest never-retransmitted segment
995  	 * that was SACKed. RTO needs the earliest RTT to stay conservative,
996  	 * but congestion control should still get an accurate delay signal.
997  	 */
998  	u64	first_sackt;
999  	u64	last_sackt;
1000  	u32	reord;
1001  	u32	sack_delivered;
1002  	int	flag;
1003  	unsigned int mss_now;
1004  	struct rate_sample *rate;
1005  };
1006  
1007  /* Take a notice that peer is sending D-SACKs. Skip update of data delivery
1008   * and spurious retransmission information if this DSACK is unlikely caused by
1009   * sender's action:
1010   * - DSACKed sequence range is larger than maximum receiver's window.
1011   * - Total no. of DSACKed segments exceed the total no. of retransmitted segs.
1012   */
tcp_dsack_seen(struct tcp_sock * tp,u32 start_seq,u32 end_seq,struct tcp_sacktag_state * state)1013  static u32 tcp_dsack_seen(struct tcp_sock *tp, u32 start_seq,
1014  			  u32 end_seq, struct tcp_sacktag_state *state)
1015  {
1016  	u32 seq_len, dup_segs = 1;
1017  
1018  	if (!before(start_seq, end_seq))
1019  		return 0;
1020  
1021  	seq_len = end_seq - start_seq;
1022  	/* Dubious DSACK: DSACKed range greater than maximum advertised rwnd */
1023  	if (seq_len > tp->max_window)
1024  		return 0;
1025  	if (seq_len > tp->mss_cache)
1026  		dup_segs = DIV_ROUND_UP(seq_len, tp->mss_cache);
1027  	else if (tp->tlp_high_seq && tp->tlp_high_seq == end_seq)
1028  		state->flag |= FLAG_DSACK_TLP;
1029  
1030  	tp->dsack_dups += dup_segs;
1031  	/* Skip the DSACK if dup segs weren't retransmitted by sender */
1032  	if (tp->dsack_dups > tp->total_retrans)
1033  		return 0;
1034  
1035  	tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
1036  	/* We increase the RACK ordering window in rounds where we receive
1037  	 * DSACKs that may have been due to reordering causing RACK to trigger
1038  	 * a spurious fast recovery. Thus RACK ignores DSACKs that happen
1039  	 * without having seen reordering, or that match TLP probes (TLP
1040  	 * is timer-driven, not triggered by RACK).
1041  	 */
1042  	if (tp->reord_seen && !(state->flag & FLAG_DSACK_TLP))
1043  		tp->rack.dsack_seen = 1;
1044  
1045  	state->flag |= FLAG_DSACKING_ACK;
1046  	/* A spurious retransmission is delivered */
1047  	state->sack_delivered += dup_segs;
1048  
1049  	return dup_segs;
1050  }
1051  
1052  /* It's reordering when higher sequence was delivered (i.e. sacked) before
1053   * some lower never-retransmitted sequence ("low_seq"). The maximum reordering
1054   * distance is approximated in full-mss packet distance ("reordering").
1055   */
tcp_check_sack_reordering(struct sock * sk,const u32 low_seq,const int ts)1056  static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq,
1057  				      const int ts)
1058  {
1059  	struct tcp_sock *tp = tcp_sk(sk);
1060  	const u32 mss = tp->mss_cache;
1061  	u32 fack, metric;
1062  
1063  	fack = tcp_highest_sack_seq(tp);
1064  	if (!before(low_seq, fack))
1065  		return;
1066  
1067  	metric = fack - low_seq;
1068  	if ((metric > tp->reordering * mss) && mss) {
1069  #if FASTRETRANS_DEBUG > 1
1070  		pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
1071  			 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
1072  			 tp->reordering,
1073  			 0,
1074  			 tp->sacked_out,
1075  			 tp->undo_marker ? tp->undo_retrans : 0);
1076  #endif
1077  		tp->reordering = min_t(u32, (metric + mss - 1) / mss,
1078  				       READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering));
1079  	}
1080  
1081  	/* This exciting event is worth to be remembered. 8) */
1082  	tp->reord_seen++;
1083  	NET_INC_STATS(sock_net(sk),
1084  		      ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER);
1085  }
1086  
1087   /* This must be called before lost_out or retrans_out are updated
1088    * on a new loss, because we want to know if all skbs previously
1089    * known to be lost have already been retransmitted, indicating
1090    * that this newly lost skb is our next skb to retransmit.
1091    */
tcp_verify_retransmit_hint(struct tcp_sock * tp,struct sk_buff * skb)1092  static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
1093  {
1094  	if ((!tp->retransmit_skb_hint && tp->retrans_out >= tp->lost_out) ||
1095  	    (tp->retransmit_skb_hint &&
1096  	     before(TCP_SKB_CB(skb)->seq,
1097  		    TCP_SKB_CB(tp->retransmit_skb_hint)->seq)))
1098  		tp->retransmit_skb_hint = skb;
1099  }
1100  
1101  /* Sum the number of packets on the wire we have marked as lost, and
1102   * notify the congestion control module that the given skb was marked lost.
1103   */
tcp_notify_skb_loss_event(struct tcp_sock * tp,const struct sk_buff * skb)1104  static void tcp_notify_skb_loss_event(struct tcp_sock *tp, const struct sk_buff *skb)
1105  {
1106  	tp->lost += tcp_skb_pcount(skb);
1107  }
1108  
tcp_mark_skb_lost(struct sock * sk,struct sk_buff * skb)1109  void tcp_mark_skb_lost(struct sock *sk, struct sk_buff *skb)
1110  {
1111  	__u8 sacked = TCP_SKB_CB(skb)->sacked;
1112  	struct tcp_sock *tp = tcp_sk(sk);
1113  
1114  	if (sacked & TCPCB_SACKED_ACKED)
1115  		return;
1116  
1117  	tcp_verify_retransmit_hint(tp, skb);
1118  	if (sacked & TCPCB_LOST) {
1119  		if (sacked & TCPCB_SACKED_RETRANS) {
1120  			/* Account for retransmits that are lost again */
1121  			TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1122  			tp->retrans_out -= tcp_skb_pcount(skb);
1123  			NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT,
1124  				      tcp_skb_pcount(skb));
1125  			tcp_notify_skb_loss_event(tp, skb);
1126  		}
1127  	} else {
1128  		tp->lost_out += tcp_skb_pcount(skb);
1129  		TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1130  		tcp_notify_skb_loss_event(tp, skb);
1131  	}
1132  }
1133  
1134  /* Updates the delivered and delivered_ce counts */
tcp_count_delivered(struct tcp_sock * tp,u32 delivered,bool ece_ack)1135  static void tcp_count_delivered(struct tcp_sock *tp, u32 delivered,
1136  				bool ece_ack)
1137  {
1138  	tp->delivered += delivered;
1139  	if (ece_ack)
1140  		tp->delivered_ce += delivered;
1141  }
1142  
1143  /* This procedure tags the retransmission queue when SACKs arrive.
1144   *
1145   * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1146   * Packets in queue with these bits set are counted in variables
1147   * sacked_out, retrans_out and lost_out, correspondingly.
1148   *
1149   * Valid combinations are:
1150   * Tag  InFlight	Description
1151   * 0	1		- orig segment is in flight.
1152   * S	0		- nothing flies, orig reached receiver.
1153   * L	0		- nothing flies, orig lost by net.
1154   * R	2		- both orig and retransmit are in flight.
1155   * L|R	1		- orig is lost, retransmit is in flight.
1156   * S|R  1		- orig reached receiver, retrans is still in flight.
1157   * (L|S|R is logically valid, it could occur when L|R is sacked,
1158   *  but it is equivalent to plain S and code short-curcuits it to S.
1159   *  L|S is logically invalid, it would mean -1 packet in flight 8))
1160   *
1161   * These 6 states form finite state machine, controlled by the following events:
1162   * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1163   * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1164   * 3. Loss detection event of two flavors:
1165   *	A. Scoreboard estimator decided the packet is lost.
1166   *	   A'. Reno "three dupacks" marks head of queue lost.
1167   *	B. SACK arrives sacking SND.NXT at the moment, when the
1168   *	   segment was retransmitted.
1169   * 4. D-SACK added new rule: D-SACK changes any tag to S.
1170   *
1171   * It is pleasant to note, that state diagram turns out to be commutative,
1172   * so that we are allowed not to be bothered by order of our actions,
1173   * when multiple events arrive simultaneously. (see the function below).
1174   *
1175   * Reordering detection.
1176   * --------------------
1177   * Reordering metric is maximal distance, which a packet can be displaced
1178   * in packet stream. With SACKs we can estimate it:
1179   *
1180   * 1. SACK fills old hole and the corresponding segment was not
1181   *    ever retransmitted -> reordering. Alas, we cannot use it
1182   *    when segment was retransmitted.
1183   * 2. The last flaw is solved with D-SACK. D-SACK arrives
1184   *    for retransmitted and already SACKed segment -> reordering..
1185   * Both of these heuristics are not used in Loss state, when we cannot
1186   * account for retransmits accurately.
1187   *
1188   * SACK block validation.
1189   * ----------------------
1190   *
1191   * SACK block range validation checks that the received SACK block fits to
1192   * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1193   * Note that SND.UNA is not included to the range though being valid because
1194   * it means that the receiver is rather inconsistent with itself reporting
1195   * SACK reneging when it should advance SND.UNA. Such SACK block this is
1196   * perfectly valid, however, in light of RFC2018 which explicitly states
1197   * that "SACK block MUST reflect the newest segment.  Even if the newest
1198   * segment is going to be discarded ...", not that it looks very clever
1199   * in case of head skb. Due to potentional receiver driven attacks, we
1200   * choose to avoid immediate execution of a walk in write queue due to
1201   * reneging and defer head skb's loss recovery to standard loss recovery
1202   * procedure that will eventually trigger (nothing forbids us doing this).
1203   *
1204   * Implements also blockage to start_seq wrap-around. Problem lies in the
1205   * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1206   * there's no guarantee that it will be before snd_nxt (n). The problem
1207   * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1208   * wrap (s_w):
1209   *
1210   *         <- outs wnd ->                          <- wrapzone ->
1211   *         u     e      n                         u_w   e_w  s n_w
1212   *         |     |      |                          |     |   |  |
1213   * |<------------+------+----- TCP seqno space --------------+---------->|
1214   * ...-- <2^31 ->|                                           |<--------...
1215   * ...---- >2^31 ------>|                                    |<--------...
1216   *
1217   * Current code wouldn't be vulnerable but it's better still to discard such
1218   * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1219   * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1220   * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1221   * equal to the ideal case (infinite seqno space without wrap caused issues).
1222   *
1223   * With D-SACK the lower bound is extended to cover sequence space below
1224   * SND.UNA down to undo_marker, which is the last point of interest. Yet
1225   * again, D-SACK block must not to go across snd_una (for the same reason as
1226   * for the normal SACK blocks, explained above). But there all simplicity
1227   * ends, TCP might receive valid D-SACKs below that. As long as they reside
1228   * fully below undo_marker they do not affect behavior in anyway and can
1229   * therefore be safely ignored. In rare cases (which are more or less
1230   * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1231   * fragmentation and packet reordering past skb's retransmission. To consider
1232   * them correctly, the acceptable range must be extended even more though
1233   * the exact amount is rather hard to quantify. However, tp->max_window can
1234   * be used as an exaggerated estimate.
1235   */
tcp_is_sackblock_valid(struct tcp_sock * tp,bool is_dsack,u32 start_seq,u32 end_seq)1236  static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1237  				   u32 start_seq, u32 end_seq)
1238  {
1239  	/* Too far in future, or reversed (interpretation is ambiguous) */
1240  	if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1241  		return false;
1242  
1243  	/* Nasty start_seq wrap-around check (see comments above) */
1244  	if (!before(start_seq, tp->snd_nxt))
1245  		return false;
1246  
1247  	/* In outstanding window? ...This is valid exit for D-SACKs too.
1248  	 * start_seq == snd_una is non-sensical (see comments above)
1249  	 */
1250  	if (after(start_seq, tp->snd_una))
1251  		return true;
1252  
1253  	if (!is_dsack || !tp->undo_marker)
1254  		return false;
1255  
1256  	/* ...Then it's D-SACK, and must reside below snd_una completely */
1257  	if (after(end_seq, tp->snd_una))
1258  		return false;
1259  
1260  	if (!before(start_seq, tp->undo_marker))
1261  		return true;
1262  
1263  	/* Too old */
1264  	if (!after(end_seq, tp->undo_marker))
1265  		return false;
1266  
1267  	/* Undo_marker boundary crossing (overestimates a lot). Known already:
1268  	 *   start_seq < undo_marker and end_seq >= undo_marker.
1269  	 */
1270  	return !before(start_seq, end_seq - tp->max_window);
1271  }
1272  
tcp_check_dsack(struct sock * sk,const struct sk_buff * ack_skb,struct tcp_sack_block_wire * sp,int num_sacks,u32 prior_snd_una,struct tcp_sacktag_state * state)1273  static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1274  			    struct tcp_sack_block_wire *sp, int num_sacks,
1275  			    u32 prior_snd_una, struct tcp_sacktag_state *state)
1276  {
1277  	struct tcp_sock *tp = tcp_sk(sk);
1278  	u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1279  	u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1280  	u32 dup_segs;
1281  
1282  	if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1283  		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1284  	} else if (num_sacks > 1) {
1285  		u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1286  		u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1287  
1288  		if (after(end_seq_0, end_seq_1) || before(start_seq_0, start_seq_1))
1289  			return false;
1290  		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKOFORECV);
1291  	} else {
1292  		return false;
1293  	}
1294  
1295  	dup_segs = tcp_dsack_seen(tp, start_seq_0, end_seq_0, state);
1296  	if (!dup_segs) {	/* Skip dubious DSACK */
1297  		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKIGNOREDDUBIOUS);
1298  		return false;
1299  	}
1300  
1301  	NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECVSEGS, dup_segs);
1302  
1303  	/* D-SACK for already forgotten data... Do dumb counting. */
1304  	if (tp->undo_marker && tp->undo_retrans > 0 &&
1305  	    !after(end_seq_0, prior_snd_una) &&
1306  	    after(end_seq_0, tp->undo_marker))
1307  		tp->undo_retrans = max_t(int, 0, tp->undo_retrans - dup_segs);
1308  
1309  	return true;
1310  }
1311  
1312  /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1313   * the incoming SACK may not exactly match but we can find smaller MSS
1314   * aligned portion of it that matches. Therefore we might need to fragment
1315   * which may fail and creates some hassle (caller must handle error case
1316   * returns).
1317   *
1318   * FIXME: this could be merged to shift decision code
1319   */
tcp_match_skb_to_sack(struct sock * sk,struct sk_buff * skb,u32 start_seq,u32 end_seq)1320  static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1321  				  u32 start_seq, u32 end_seq)
1322  {
1323  	int err;
1324  	bool in_sack;
1325  	unsigned int pkt_len;
1326  	unsigned int mss;
1327  
1328  	in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1329  		  !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1330  
1331  	if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1332  	    after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1333  		mss = tcp_skb_mss(skb);
1334  		in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1335  
1336  		if (!in_sack) {
1337  			pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1338  			if (pkt_len < mss)
1339  				pkt_len = mss;
1340  		} else {
1341  			pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1342  			if (pkt_len < mss)
1343  				return -EINVAL;
1344  		}
1345  
1346  		/* Round if necessary so that SACKs cover only full MSSes
1347  		 * and/or the remaining small portion (if present)
1348  		 */
1349  		if (pkt_len > mss) {
1350  			unsigned int new_len = (pkt_len / mss) * mss;
1351  			if (!in_sack && new_len < pkt_len)
1352  				new_len += mss;
1353  			pkt_len = new_len;
1354  		}
1355  
1356  		if (pkt_len >= skb->len && !in_sack)
1357  			return 0;
1358  
1359  		err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
1360  				   pkt_len, mss, GFP_ATOMIC);
1361  		if (err < 0)
1362  			return err;
1363  	}
1364  
1365  	return in_sack;
1366  }
1367  
1368  /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
tcp_sacktag_one(struct sock * sk,struct tcp_sacktag_state * state,u8 sacked,u32 start_seq,u32 end_seq,int dup_sack,int pcount,u64 xmit_time)1369  static u8 tcp_sacktag_one(struct sock *sk,
1370  			  struct tcp_sacktag_state *state, u8 sacked,
1371  			  u32 start_seq, u32 end_seq,
1372  			  int dup_sack, int pcount,
1373  			  u64 xmit_time)
1374  {
1375  	struct tcp_sock *tp = tcp_sk(sk);
1376  
1377  	/* Account D-SACK for retransmitted packet. */
1378  	if (dup_sack && (sacked & TCPCB_RETRANS)) {
1379  		if (tp->undo_marker && tp->undo_retrans > 0 &&
1380  		    after(end_seq, tp->undo_marker))
1381  			tp->undo_retrans = max_t(int, 0, tp->undo_retrans - pcount);
1382  		if ((sacked & TCPCB_SACKED_ACKED) &&
1383  		    before(start_seq, state->reord))
1384  				state->reord = start_seq;
1385  	}
1386  
1387  	/* Nothing to do; acked frame is about to be dropped (was ACKed). */
1388  	if (!after(end_seq, tp->snd_una))
1389  		return sacked;
1390  
1391  	if (!(sacked & TCPCB_SACKED_ACKED)) {
1392  		tcp_rack_advance(tp, sacked, end_seq, xmit_time);
1393  
1394  		if (sacked & TCPCB_SACKED_RETRANS) {
1395  			/* If the segment is not tagged as lost,
1396  			 * we do not clear RETRANS, believing
1397  			 * that retransmission is still in flight.
1398  			 */
1399  			if (sacked & TCPCB_LOST) {
1400  				sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1401  				tp->lost_out -= pcount;
1402  				tp->retrans_out -= pcount;
1403  			}
1404  		} else {
1405  			if (!(sacked & TCPCB_RETRANS)) {
1406  				/* New sack for not retransmitted frame,
1407  				 * which was in hole. It is reordering.
1408  				 */
1409  				if (before(start_seq,
1410  					   tcp_highest_sack_seq(tp)) &&
1411  				    before(start_seq, state->reord))
1412  					state->reord = start_seq;
1413  
1414  				if (!after(end_seq, tp->high_seq))
1415  					state->flag |= FLAG_ORIG_SACK_ACKED;
1416  				if (state->first_sackt == 0)
1417  					state->first_sackt = xmit_time;
1418  				state->last_sackt = xmit_time;
1419  			}
1420  
1421  			if (sacked & TCPCB_LOST) {
1422  				sacked &= ~TCPCB_LOST;
1423  				tp->lost_out -= pcount;
1424  			}
1425  		}
1426  
1427  		sacked |= TCPCB_SACKED_ACKED;
1428  		state->flag |= FLAG_DATA_SACKED;
1429  		tp->sacked_out += pcount;
1430  		/* Out-of-order packets delivered */
1431  		state->sack_delivered += pcount;
1432  
1433  		/* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1434  		if (tp->lost_skb_hint &&
1435  		    before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1436  			tp->lost_cnt_hint += pcount;
1437  	}
1438  
1439  	/* D-SACK. We can detect redundant retransmission in S|R and plain R
1440  	 * frames and clear it. undo_retrans is decreased above, L|R frames
1441  	 * are accounted above as well.
1442  	 */
1443  	if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1444  		sacked &= ~TCPCB_SACKED_RETRANS;
1445  		tp->retrans_out -= pcount;
1446  	}
1447  
1448  	return sacked;
1449  }
1450  
1451  /* Shift newly-SACKed bytes from this skb to the immediately previous
1452   * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1453   */
tcp_shifted_skb(struct sock * sk,struct sk_buff * prev,struct sk_buff * skb,struct tcp_sacktag_state * state,unsigned int pcount,int shifted,int mss,bool dup_sack)1454  static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev,
1455  			    struct sk_buff *skb,
1456  			    struct tcp_sacktag_state *state,
1457  			    unsigned int pcount, int shifted, int mss,
1458  			    bool dup_sack)
1459  {
1460  	struct tcp_sock *tp = tcp_sk(sk);
1461  	u32 start_seq = TCP_SKB_CB(skb)->seq;	/* start of newly-SACKed */
1462  	u32 end_seq = start_seq + shifted;	/* end of newly-SACKed */
1463  
1464  	BUG_ON(!pcount);
1465  
1466  	/* Adjust counters and hints for the newly sacked sequence
1467  	 * range but discard the return value since prev is already
1468  	 * marked. We must tag the range first because the seq
1469  	 * advancement below implicitly advances
1470  	 * tcp_highest_sack_seq() when skb is highest_sack.
1471  	 */
1472  	tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1473  			start_seq, end_seq, dup_sack, pcount,
1474  			tcp_skb_timestamp_us(skb));
1475  	tcp_rate_skb_delivered(sk, skb, state->rate);
1476  
1477  	if (skb == tp->lost_skb_hint)
1478  		tp->lost_cnt_hint += pcount;
1479  
1480  	TCP_SKB_CB(prev)->end_seq += shifted;
1481  	TCP_SKB_CB(skb)->seq += shifted;
1482  
1483  	tcp_skb_pcount_add(prev, pcount);
1484  	WARN_ON_ONCE(tcp_skb_pcount(skb) < pcount);
1485  	tcp_skb_pcount_add(skb, -pcount);
1486  
1487  	/* When we're adding to gso_segs == 1, gso_size will be zero,
1488  	 * in theory this shouldn't be necessary but as long as DSACK
1489  	 * code can come after this skb later on it's better to keep
1490  	 * setting gso_size to something.
1491  	 */
1492  	if (!TCP_SKB_CB(prev)->tcp_gso_size)
1493  		TCP_SKB_CB(prev)->tcp_gso_size = mss;
1494  
1495  	/* CHECKME: To clear or not to clear? Mimics normal skb currently */
1496  	if (tcp_skb_pcount(skb) <= 1)
1497  		TCP_SKB_CB(skb)->tcp_gso_size = 0;
1498  
1499  	/* Difference in this won't matter, both ACKed by the same cumul. ACK */
1500  	TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1501  
1502  	if (skb->len > 0) {
1503  		BUG_ON(!tcp_skb_pcount(skb));
1504  		NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1505  		return false;
1506  	}
1507  
1508  	/* Whole SKB was eaten :-) */
1509  
1510  	if (skb == tp->retransmit_skb_hint)
1511  		tp->retransmit_skb_hint = prev;
1512  	if (skb == tp->lost_skb_hint) {
1513  		tp->lost_skb_hint = prev;
1514  		tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1515  	}
1516  
1517  	TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1518  	TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
1519  	if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1520  		TCP_SKB_CB(prev)->end_seq++;
1521  
1522  	if (skb == tcp_highest_sack(sk))
1523  		tcp_advance_highest_sack(sk, skb);
1524  
1525  	tcp_skb_collapse_tstamp(prev, skb);
1526  	if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp))
1527  		TCP_SKB_CB(prev)->tx.delivered_mstamp = 0;
1528  
1529  	tcp_rtx_queue_unlink_and_free(skb, sk);
1530  
1531  	NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
1532  
1533  	return true;
1534  }
1535  
1536  /* I wish gso_size would have a bit more sane initialization than
1537   * something-or-zero which complicates things
1538   */
tcp_skb_seglen(const struct sk_buff * skb)1539  static int tcp_skb_seglen(const struct sk_buff *skb)
1540  {
1541  	return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1542  }
1543  
1544  /* Shifting pages past head area doesn't work */
skb_can_shift(const struct sk_buff * skb)1545  static int skb_can_shift(const struct sk_buff *skb)
1546  {
1547  	return !skb_headlen(skb) && skb_is_nonlinear(skb);
1548  }
1549  
tcp_skb_shift(struct sk_buff * to,struct sk_buff * from,int pcount,int shiftlen)1550  int tcp_skb_shift(struct sk_buff *to, struct sk_buff *from,
1551  		  int pcount, int shiftlen)
1552  {
1553  	/* TCP min gso_size is 8 bytes (TCP_MIN_GSO_SIZE)
1554  	 * Since TCP_SKB_CB(skb)->tcp_gso_segs is 16 bits, we need
1555  	 * to make sure not storing more than 65535 * 8 bytes per skb,
1556  	 * even if current MSS is bigger.
1557  	 */
1558  	if (unlikely(to->len + shiftlen >= 65535 * TCP_MIN_GSO_SIZE))
1559  		return 0;
1560  	if (unlikely(tcp_skb_pcount(to) + pcount > 65535))
1561  		return 0;
1562  	return skb_shift(to, from, shiftlen);
1563  }
1564  
1565  /* Try collapsing SACK blocks spanning across multiple skbs to a single
1566   * skb.
1567   */
tcp_shift_skb_data(struct sock * sk,struct sk_buff * skb,struct tcp_sacktag_state * state,u32 start_seq,u32 end_seq,bool dup_sack)1568  static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1569  					  struct tcp_sacktag_state *state,
1570  					  u32 start_seq, u32 end_seq,
1571  					  bool dup_sack)
1572  {
1573  	struct tcp_sock *tp = tcp_sk(sk);
1574  	struct sk_buff *prev;
1575  	int mss;
1576  	int pcount = 0;
1577  	int len;
1578  	int in_sack;
1579  
1580  	/* Normally R but no L won't result in plain S */
1581  	if (!dup_sack &&
1582  	    (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1583  		goto fallback;
1584  	if (!skb_can_shift(skb))
1585  		goto fallback;
1586  	/* This frame is about to be dropped (was ACKed). */
1587  	if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1588  		goto fallback;
1589  
1590  	/* Can only happen with delayed DSACK + discard craziness */
1591  	prev = skb_rb_prev(skb);
1592  	if (!prev)
1593  		goto fallback;
1594  
1595  	if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1596  		goto fallback;
1597  
1598  	if (!tcp_skb_can_collapse(prev, skb))
1599  		goto fallback;
1600  
1601  	in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1602  		  !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1603  
1604  	if (in_sack) {
1605  		len = skb->len;
1606  		pcount = tcp_skb_pcount(skb);
1607  		mss = tcp_skb_seglen(skb);
1608  
1609  		/* TODO: Fix DSACKs to not fragment already SACKed and we can
1610  		 * drop this restriction as unnecessary
1611  		 */
1612  		if (mss != tcp_skb_seglen(prev))
1613  			goto fallback;
1614  	} else {
1615  		if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1616  			goto noop;
1617  		/* CHECKME: This is non-MSS split case only?, this will
1618  		 * cause skipped skbs due to advancing loop btw, original
1619  		 * has that feature too
1620  		 */
1621  		if (tcp_skb_pcount(skb) <= 1)
1622  			goto noop;
1623  
1624  		in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1625  		if (!in_sack) {
1626  			/* TODO: head merge to next could be attempted here
1627  			 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1628  			 * though it might not be worth of the additional hassle
1629  			 *
1630  			 * ...we can probably just fallback to what was done
1631  			 * previously. We could try merging non-SACKed ones
1632  			 * as well but it probably isn't going to buy off
1633  			 * because later SACKs might again split them, and
1634  			 * it would make skb timestamp tracking considerably
1635  			 * harder problem.
1636  			 */
1637  			goto fallback;
1638  		}
1639  
1640  		len = end_seq - TCP_SKB_CB(skb)->seq;
1641  		BUG_ON(len < 0);
1642  		BUG_ON(len > skb->len);
1643  
1644  		/* MSS boundaries should be honoured or else pcount will
1645  		 * severely break even though it makes things bit trickier.
1646  		 * Optimize common case to avoid most of the divides
1647  		 */
1648  		mss = tcp_skb_mss(skb);
1649  
1650  		/* TODO: Fix DSACKs to not fragment already SACKed and we can
1651  		 * drop this restriction as unnecessary
1652  		 */
1653  		if (mss != tcp_skb_seglen(prev))
1654  			goto fallback;
1655  
1656  		if (len == mss) {
1657  			pcount = 1;
1658  		} else if (len < mss) {
1659  			goto noop;
1660  		} else {
1661  			pcount = len / mss;
1662  			len = pcount * mss;
1663  		}
1664  	}
1665  
1666  	/* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1667  	if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1668  		goto fallback;
1669  
1670  	if (!tcp_skb_shift(prev, skb, pcount, len))
1671  		goto fallback;
1672  	if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack))
1673  		goto out;
1674  
1675  	/* Hole filled allows collapsing with the next as well, this is very
1676  	 * useful when hole on every nth skb pattern happens
1677  	 */
1678  	skb = skb_rb_next(prev);
1679  	if (!skb)
1680  		goto out;
1681  
1682  	if (!skb_can_shift(skb) ||
1683  	    ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1684  	    (mss != tcp_skb_seglen(skb)))
1685  		goto out;
1686  
1687  	if (!tcp_skb_can_collapse(prev, skb))
1688  		goto out;
1689  	len = skb->len;
1690  	pcount = tcp_skb_pcount(skb);
1691  	if (tcp_skb_shift(prev, skb, pcount, len))
1692  		tcp_shifted_skb(sk, prev, skb, state, pcount,
1693  				len, mss, 0);
1694  
1695  out:
1696  	return prev;
1697  
1698  noop:
1699  	return skb;
1700  
1701  fallback:
1702  	NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1703  	return NULL;
1704  }
1705  
tcp_sacktag_walk(struct sk_buff * skb,struct sock * sk,struct tcp_sack_block * next_dup,struct tcp_sacktag_state * state,u32 start_seq,u32 end_seq,bool dup_sack_in)1706  static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1707  					struct tcp_sack_block *next_dup,
1708  					struct tcp_sacktag_state *state,
1709  					u32 start_seq, u32 end_seq,
1710  					bool dup_sack_in)
1711  {
1712  	struct tcp_sock *tp = tcp_sk(sk);
1713  	struct sk_buff *tmp;
1714  
1715  	skb_rbtree_walk_from(skb) {
1716  		int in_sack = 0;
1717  		bool dup_sack = dup_sack_in;
1718  
1719  		/* queue is in-order => we can short-circuit the walk early */
1720  		if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1721  			break;
1722  
1723  		if (next_dup  &&
1724  		    before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1725  			in_sack = tcp_match_skb_to_sack(sk, skb,
1726  							next_dup->start_seq,
1727  							next_dup->end_seq);
1728  			if (in_sack > 0)
1729  				dup_sack = true;
1730  		}
1731  
1732  		/* skb reference here is a bit tricky to get right, since
1733  		 * shifting can eat and free both this skb and the next,
1734  		 * so not even _safe variant of the loop is enough.
1735  		 */
1736  		if (in_sack <= 0) {
1737  			tmp = tcp_shift_skb_data(sk, skb, state,
1738  						 start_seq, end_seq, dup_sack);
1739  			if (tmp) {
1740  				if (tmp != skb) {
1741  					skb = tmp;
1742  					continue;
1743  				}
1744  
1745  				in_sack = 0;
1746  			} else {
1747  				in_sack = tcp_match_skb_to_sack(sk, skb,
1748  								start_seq,
1749  								end_seq);
1750  			}
1751  		}
1752  
1753  		if (unlikely(in_sack < 0))
1754  			break;
1755  
1756  		if (in_sack) {
1757  			TCP_SKB_CB(skb)->sacked =
1758  				tcp_sacktag_one(sk,
1759  						state,
1760  						TCP_SKB_CB(skb)->sacked,
1761  						TCP_SKB_CB(skb)->seq,
1762  						TCP_SKB_CB(skb)->end_seq,
1763  						dup_sack,
1764  						tcp_skb_pcount(skb),
1765  						tcp_skb_timestamp_us(skb));
1766  			tcp_rate_skb_delivered(sk, skb, state->rate);
1767  			if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1768  				list_del_init(&skb->tcp_tsorted_anchor);
1769  
1770  			if (!before(TCP_SKB_CB(skb)->seq,
1771  				    tcp_highest_sack_seq(tp)))
1772  				tcp_advance_highest_sack(sk, skb);
1773  		}
1774  	}
1775  	return skb;
1776  }
1777  
tcp_sacktag_bsearch(struct sock * sk,u32 seq)1778  static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk, u32 seq)
1779  {
1780  	struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node;
1781  	struct sk_buff *skb;
1782  
1783  	while (*p) {
1784  		parent = *p;
1785  		skb = rb_to_skb(parent);
1786  		if (before(seq, TCP_SKB_CB(skb)->seq)) {
1787  			p = &parent->rb_left;
1788  			continue;
1789  		}
1790  		if (!before(seq, TCP_SKB_CB(skb)->end_seq)) {
1791  			p = &parent->rb_right;
1792  			continue;
1793  		}
1794  		return skb;
1795  	}
1796  	return NULL;
1797  }
1798  
tcp_sacktag_skip(struct sk_buff * skb,struct sock * sk,u32 skip_to_seq)1799  static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1800  					u32 skip_to_seq)
1801  {
1802  	if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq))
1803  		return skb;
1804  
1805  	return tcp_sacktag_bsearch(sk, skip_to_seq);
1806  }
1807  
tcp_maybe_skipping_dsack(struct sk_buff * skb,struct sock * sk,struct tcp_sack_block * next_dup,struct tcp_sacktag_state * state,u32 skip_to_seq)1808  static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1809  						struct sock *sk,
1810  						struct tcp_sack_block *next_dup,
1811  						struct tcp_sacktag_state *state,
1812  						u32 skip_to_seq)
1813  {
1814  	if (!next_dup)
1815  		return skb;
1816  
1817  	if (before(next_dup->start_seq, skip_to_seq)) {
1818  		skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq);
1819  		skb = tcp_sacktag_walk(skb, sk, NULL, state,
1820  				       next_dup->start_seq, next_dup->end_seq,
1821  				       1);
1822  	}
1823  
1824  	return skb;
1825  }
1826  
tcp_sack_cache_ok(const struct tcp_sock * tp,const struct tcp_sack_block * cache)1827  static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1828  {
1829  	return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1830  }
1831  
1832  static int
tcp_sacktag_write_queue(struct sock * sk,const struct sk_buff * ack_skb,u32 prior_snd_una,struct tcp_sacktag_state * state)1833  tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1834  			u32 prior_snd_una, struct tcp_sacktag_state *state)
1835  {
1836  	struct tcp_sock *tp = tcp_sk(sk);
1837  	const unsigned char *ptr = (skb_transport_header(ack_skb) +
1838  				    TCP_SKB_CB(ack_skb)->sacked);
1839  	struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1840  	struct tcp_sack_block sp[TCP_NUM_SACKS];
1841  	struct tcp_sack_block *cache;
1842  	struct sk_buff *skb;
1843  	int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1844  	int used_sacks;
1845  	bool found_dup_sack = false;
1846  	int i, j;
1847  	int first_sack_index;
1848  
1849  	state->flag = 0;
1850  	state->reord = tp->snd_nxt;
1851  
1852  	if (!tp->sacked_out)
1853  		tcp_highest_sack_reset(sk);
1854  
1855  	found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1856  					 num_sacks, prior_snd_una, state);
1857  
1858  	/* Eliminate too old ACKs, but take into
1859  	 * account more or less fresh ones, they can
1860  	 * contain valid SACK info.
1861  	 */
1862  	if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1863  		return 0;
1864  
1865  	if (!tp->packets_out)
1866  		goto out;
1867  
1868  	used_sacks = 0;
1869  	first_sack_index = 0;
1870  	for (i = 0; i < num_sacks; i++) {
1871  		bool dup_sack = !i && found_dup_sack;
1872  
1873  		sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1874  		sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1875  
1876  		if (!tcp_is_sackblock_valid(tp, dup_sack,
1877  					    sp[used_sacks].start_seq,
1878  					    sp[used_sacks].end_seq)) {
1879  			int mib_idx;
1880  
1881  			if (dup_sack) {
1882  				if (!tp->undo_marker)
1883  					mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1884  				else
1885  					mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1886  			} else {
1887  				/* Don't count olds caused by ACK reordering */
1888  				if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1889  				    !after(sp[used_sacks].end_seq, tp->snd_una))
1890  					continue;
1891  				mib_idx = LINUX_MIB_TCPSACKDISCARD;
1892  			}
1893  
1894  			NET_INC_STATS(sock_net(sk), mib_idx);
1895  			if (i == 0)
1896  				first_sack_index = -1;
1897  			continue;
1898  		}
1899  
1900  		/* Ignore very old stuff early */
1901  		if (!after(sp[used_sacks].end_seq, prior_snd_una)) {
1902  			if (i == 0)
1903  				first_sack_index = -1;
1904  			continue;
1905  		}
1906  
1907  		used_sacks++;
1908  	}
1909  
1910  	/* order SACK blocks to allow in order walk of the retrans queue */
1911  	for (i = used_sacks - 1; i > 0; i--) {
1912  		for (j = 0; j < i; j++) {
1913  			if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1914  				swap(sp[j], sp[j + 1]);
1915  
1916  				/* Track where the first SACK block goes to */
1917  				if (j == first_sack_index)
1918  					first_sack_index = j + 1;
1919  			}
1920  		}
1921  	}
1922  
1923  	state->mss_now = tcp_current_mss(sk);
1924  	skb = NULL;
1925  	i = 0;
1926  
1927  	if (!tp->sacked_out) {
1928  		/* It's already past, so skip checking against it */
1929  		cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1930  	} else {
1931  		cache = tp->recv_sack_cache;
1932  		/* Skip empty blocks in at head of the cache */
1933  		while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1934  		       !cache->end_seq)
1935  			cache++;
1936  	}
1937  
1938  	while (i < used_sacks) {
1939  		u32 start_seq = sp[i].start_seq;
1940  		u32 end_seq = sp[i].end_seq;
1941  		bool dup_sack = (found_dup_sack && (i == first_sack_index));
1942  		struct tcp_sack_block *next_dup = NULL;
1943  
1944  		if (found_dup_sack && ((i + 1) == first_sack_index))
1945  			next_dup = &sp[i + 1];
1946  
1947  		/* Skip too early cached blocks */
1948  		while (tcp_sack_cache_ok(tp, cache) &&
1949  		       !before(start_seq, cache->end_seq))
1950  			cache++;
1951  
1952  		/* Can skip some work by looking recv_sack_cache? */
1953  		if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1954  		    after(end_seq, cache->start_seq)) {
1955  
1956  			/* Head todo? */
1957  			if (before(start_seq, cache->start_seq)) {
1958  				skb = tcp_sacktag_skip(skb, sk, start_seq);
1959  				skb = tcp_sacktag_walk(skb, sk, next_dup,
1960  						       state,
1961  						       start_seq,
1962  						       cache->start_seq,
1963  						       dup_sack);
1964  			}
1965  
1966  			/* Rest of the block already fully processed? */
1967  			if (!after(end_seq, cache->end_seq))
1968  				goto advance_sp;
1969  
1970  			skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1971  						       state,
1972  						       cache->end_seq);
1973  
1974  			/* ...tail remains todo... */
1975  			if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1976  				/* ...but better entrypoint exists! */
1977  				skb = tcp_highest_sack(sk);
1978  				if (!skb)
1979  					break;
1980  				cache++;
1981  				goto walk;
1982  			}
1983  
1984  			skb = tcp_sacktag_skip(skb, sk, cache->end_seq);
1985  			/* Check overlap against next cached too (past this one already) */
1986  			cache++;
1987  			continue;
1988  		}
1989  
1990  		if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1991  			skb = tcp_highest_sack(sk);
1992  			if (!skb)
1993  				break;
1994  		}
1995  		skb = tcp_sacktag_skip(skb, sk, start_seq);
1996  
1997  walk:
1998  		skb = tcp_sacktag_walk(skb, sk, next_dup, state,
1999  				       start_seq, end_seq, dup_sack);
2000  
2001  advance_sp:
2002  		i++;
2003  	}
2004  
2005  	/* Clear the head of the cache sack blocks so we can skip it next time */
2006  	for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
2007  		tp->recv_sack_cache[i].start_seq = 0;
2008  		tp->recv_sack_cache[i].end_seq = 0;
2009  	}
2010  	for (j = 0; j < used_sacks; j++)
2011  		tp->recv_sack_cache[i++] = sp[j];
2012  
2013  	if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker)
2014  		tcp_check_sack_reordering(sk, state->reord, 0);
2015  
2016  	tcp_verify_left_out(tp);
2017  out:
2018  
2019  #if FASTRETRANS_DEBUG > 0
2020  	WARN_ON((int)tp->sacked_out < 0);
2021  	WARN_ON((int)tp->lost_out < 0);
2022  	WARN_ON((int)tp->retrans_out < 0);
2023  	WARN_ON((int)tcp_packets_in_flight(tp) < 0);
2024  #endif
2025  	return state->flag;
2026  }
2027  
2028  /* Limits sacked_out so that sum with lost_out isn't ever larger than
2029   * packets_out. Returns false if sacked_out adjustement wasn't necessary.
2030   */
tcp_limit_reno_sacked(struct tcp_sock * tp)2031  static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
2032  {
2033  	u32 holes;
2034  
2035  	holes = max(tp->lost_out, 1U);
2036  	holes = min(holes, tp->packets_out);
2037  
2038  	if ((tp->sacked_out + holes) > tp->packets_out) {
2039  		tp->sacked_out = tp->packets_out - holes;
2040  		return true;
2041  	}
2042  	return false;
2043  }
2044  
2045  /* If we receive more dupacks than we expected counting segments
2046   * in assumption of absent reordering, interpret this as reordering.
2047   * The only another reason could be bug in receiver TCP.
2048   */
tcp_check_reno_reordering(struct sock * sk,const int addend)2049  static void tcp_check_reno_reordering(struct sock *sk, const int addend)
2050  {
2051  	struct tcp_sock *tp = tcp_sk(sk);
2052  
2053  	if (!tcp_limit_reno_sacked(tp))
2054  		return;
2055  
2056  	tp->reordering = min_t(u32, tp->packets_out + addend,
2057  			       READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering));
2058  	tp->reord_seen++;
2059  	NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER);
2060  }
2061  
2062  /* Emulate SACKs for SACKless connection: account for a new dupack. */
2063  
tcp_add_reno_sack(struct sock * sk,int num_dupack,bool ece_ack)2064  static void tcp_add_reno_sack(struct sock *sk, int num_dupack, bool ece_ack)
2065  {
2066  	if (num_dupack) {
2067  		struct tcp_sock *tp = tcp_sk(sk);
2068  		u32 prior_sacked = tp->sacked_out;
2069  		s32 delivered;
2070  
2071  		tp->sacked_out += num_dupack;
2072  		tcp_check_reno_reordering(sk, 0);
2073  		delivered = tp->sacked_out - prior_sacked;
2074  		if (delivered > 0)
2075  			tcp_count_delivered(tp, delivered, ece_ack);
2076  		tcp_verify_left_out(tp);
2077  	}
2078  }
2079  
2080  /* Account for ACK, ACKing some data in Reno Recovery phase. */
2081  
tcp_remove_reno_sacks(struct sock * sk,int acked,bool ece_ack)2082  static void tcp_remove_reno_sacks(struct sock *sk, int acked, bool ece_ack)
2083  {
2084  	struct tcp_sock *tp = tcp_sk(sk);
2085  
2086  	if (acked > 0) {
2087  		/* One ACK acked hole. The rest eat duplicate ACKs. */
2088  		tcp_count_delivered(tp, max_t(int, acked - tp->sacked_out, 1),
2089  				    ece_ack);
2090  		if (acked - 1 >= tp->sacked_out)
2091  			tp->sacked_out = 0;
2092  		else
2093  			tp->sacked_out -= acked - 1;
2094  	}
2095  	tcp_check_reno_reordering(sk, acked);
2096  	tcp_verify_left_out(tp);
2097  }
2098  
tcp_reset_reno_sack(struct tcp_sock * tp)2099  static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
2100  {
2101  	tp->sacked_out = 0;
2102  }
2103  
tcp_clear_retrans(struct tcp_sock * tp)2104  void tcp_clear_retrans(struct tcp_sock *tp)
2105  {
2106  	tp->retrans_out = 0;
2107  	tp->lost_out = 0;
2108  	tp->undo_marker = 0;
2109  	tp->undo_retrans = -1;
2110  	tp->sacked_out = 0;
2111  	tp->rto_stamp = 0;
2112  	tp->total_rto = 0;
2113  	tp->total_rto_recoveries = 0;
2114  	tp->total_rto_time = 0;
2115  }
2116  
tcp_init_undo(struct tcp_sock * tp)2117  static inline void tcp_init_undo(struct tcp_sock *tp)
2118  {
2119  	tp->undo_marker = tp->snd_una;
2120  
2121  	/* Retransmission still in flight may cause DSACKs later. */
2122  	/* First, account for regular retransmits in flight: */
2123  	tp->undo_retrans = tp->retrans_out;
2124  	/* Next, account for TLP retransmits in flight: */
2125  	if (tp->tlp_high_seq && tp->tlp_retrans)
2126  		tp->undo_retrans++;
2127  	/* Finally, avoid 0, because undo_retrans==0 means "can undo now": */
2128  	if (!tp->undo_retrans)
2129  		tp->undo_retrans = -1;
2130  }
2131  
tcp_is_rack(const struct sock * sk)2132  static bool tcp_is_rack(const struct sock *sk)
2133  {
2134  	return READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_recovery) &
2135  		TCP_RACK_LOSS_DETECTION;
2136  }
2137  
2138  /* If we detect SACK reneging, forget all SACK information
2139   * and reset tags completely, otherwise preserve SACKs. If receiver
2140   * dropped its ofo queue, we will know this due to reneging detection.
2141   */
tcp_timeout_mark_lost(struct sock * sk)2142  static void tcp_timeout_mark_lost(struct sock *sk)
2143  {
2144  	struct tcp_sock *tp = tcp_sk(sk);
2145  	struct sk_buff *skb, *head;
2146  	bool is_reneg;			/* is receiver reneging on SACKs? */
2147  
2148  	head = tcp_rtx_queue_head(sk);
2149  	is_reneg = head && (TCP_SKB_CB(head)->sacked & TCPCB_SACKED_ACKED);
2150  	if (is_reneg) {
2151  		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2152  		tp->sacked_out = 0;
2153  		/* Mark SACK reneging until we recover from this loss event. */
2154  		tp->is_sack_reneg = 1;
2155  	} else if (tcp_is_reno(tp)) {
2156  		tcp_reset_reno_sack(tp);
2157  	}
2158  
2159  	skb = head;
2160  	skb_rbtree_walk_from(skb) {
2161  		if (is_reneg)
2162  			TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2163  		else if (tcp_is_rack(sk) && skb != head &&
2164  			 tcp_rack_skb_timeout(tp, skb, 0) > 0)
2165  			continue; /* Don't mark recently sent ones lost yet */
2166  		tcp_mark_skb_lost(sk, skb);
2167  	}
2168  	tcp_verify_left_out(tp);
2169  	tcp_clear_all_retrans_hints(tp);
2170  }
2171  
2172  /* Enter Loss state. */
tcp_enter_loss(struct sock * sk)2173  void tcp_enter_loss(struct sock *sk)
2174  {
2175  	const struct inet_connection_sock *icsk = inet_csk(sk);
2176  	struct tcp_sock *tp = tcp_sk(sk);
2177  	struct net *net = sock_net(sk);
2178  	bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
2179  	u8 reordering;
2180  
2181  	tcp_timeout_mark_lost(sk);
2182  
2183  	/* Reduce ssthresh if it has not yet been made inside this window. */
2184  	if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
2185  	    !after(tp->high_seq, tp->snd_una) ||
2186  	    (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2187  		tp->prior_ssthresh = tcp_current_ssthresh(sk);
2188  		tp->prior_cwnd = tcp_snd_cwnd(tp);
2189  		tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2190  		tcp_ca_event(sk, CA_EVENT_LOSS);
2191  		tcp_init_undo(tp);
2192  	}
2193  	tcp_snd_cwnd_set(tp, tcp_packets_in_flight(tp) + 1);
2194  	tp->snd_cwnd_cnt   = 0;
2195  	tp->snd_cwnd_stamp = tcp_jiffies32;
2196  
2197  	/* Timeout in disordered state after receiving substantial DUPACKs
2198  	 * suggests that the degree of reordering is over-estimated.
2199  	 */
2200  	reordering = READ_ONCE(net->ipv4.sysctl_tcp_reordering);
2201  	if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
2202  	    tp->sacked_out >= reordering)
2203  		tp->reordering = min_t(unsigned int, tp->reordering,
2204  				       reordering);
2205  
2206  	tcp_set_ca_state(sk, TCP_CA_Loss);
2207  	tp->high_seq = tp->snd_nxt;
2208  	tp->tlp_high_seq = 0;
2209  	tcp_ecn_queue_cwr(tp);
2210  
2211  	/* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
2212  	 * loss recovery is underway except recurring timeout(s) on
2213  	 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
2214  	 */
2215  	tp->frto = READ_ONCE(net->ipv4.sysctl_tcp_frto) &&
2216  		   (new_recovery || icsk->icsk_retransmits) &&
2217  		   !inet_csk(sk)->icsk_mtup.probe_size;
2218  }
2219  
2220  /* If ACK arrived pointing to a remembered SACK, it means that our
2221   * remembered SACKs do not reflect real state of receiver i.e.
2222   * receiver _host_ is heavily congested (or buggy).
2223   *
2224   * To avoid big spurious retransmission bursts due to transient SACK
2225   * scoreboard oddities that look like reneging, we give the receiver a
2226   * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2227   * restore sanity to the SACK scoreboard. If the apparent reneging
2228   * persists until this RTO then we'll clear the SACK scoreboard.
2229   */
tcp_check_sack_reneging(struct sock * sk,int * ack_flag)2230  static bool tcp_check_sack_reneging(struct sock *sk, int *ack_flag)
2231  {
2232  	if (*ack_flag & FLAG_SACK_RENEGING &&
2233  	    *ack_flag & FLAG_SND_UNA_ADVANCED) {
2234  		struct tcp_sock *tp = tcp_sk(sk);
2235  		unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2236  					  msecs_to_jiffies(10));
2237  
2238  		inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2239  					  delay, TCP_RTO_MAX);
2240  		*ack_flag &= ~FLAG_SET_XMIT_TIMER;
2241  		return true;
2242  	}
2243  	return false;
2244  }
2245  
2246  /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2247   * counter when SACK is enabled (without SACK, sacked_out is used for
2248   * that purpose).
2249   *
2250   * With reordering, holes may still be in flight, so RFC3517 recovery
2251   * uses pure sacked_out (total number of SACKed segments) even though
2252   * it violates the RFC that uses duplicate ACKs, often these are equal
2253   * but when e.g. out-of-window ACKs or packet duplication occurs,
2254   * they differ. Since neither occurs due to loss, TCP should really
2255   * ignore them.
2256   */
tcp_dupack_heuristics(const struct tcp_sock * tp)2257  static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2258  {
2259  	return tp->sacked_out + 1;
2260  }
2261  
2262  /* Linux NewReno/SACK/ECN state machine.
2263   * --------------------------------------
2264   *
2265   * "Open"	Normal state, no dubious events, fast path.
2266   * "Disorder"   In all the respects it is "Open",
2267   *		but requires a bit more attention. It is entered when
2268   *		we see some SACKs or dupacks. It is split of "Open"
2269   *		mainly to move some processing from fast path to slow one.
2270   * "CWR"	CWND was reduced due to some Congestion Notification event.
2271   *		It can be ECN, ICMP source quench, local device congestion.
2272   * "Recovery"	CWND was reduced, we are fast-retransmitting.
2273   * "Loss"	CWND was reduced due to RTO timeout or SACK reneging.
2274   *
2275   * tcp_fastretrans_alert() is entered:
2276   * - each incoming ACK, if state is not "Open"
2277   * - when arrived ACK is unusual, namely:
2278   *	* SACK
2279   *	* Duplicate ACK.
2280   *	* ECN ECE.
2281   *
2282   * Counting packets in flight is pretty simple.
2283   *
2284   *	in_flight = packets_out - left_out + retrans_out
2285   *
2286   *	packets_out is SND.NXT-SND.UNA counted in packets.
2287   *
2288   *	retrans_out is number of retransmitted segments.
2289   *
2290   *	left_out is number of segments left network, but not ACKed yet.
2291   *
2292   *		left_out = sacked_out + lost_out
2293   *
2294   *     sacked_out: Packets, which arrived to receiver out of order
2295   *		   and hence not ACKed. With SACKs this number is simply
2296   *		   amount of SACKed data. Even without SACKs
2297   *		   it is easy to give pretty reliable estimate of this number,
2298   *		   counting duplicate ACKs.
2299   *
2300   *       lost_out: Packets lost by network. TCP has no explicit
2301   *		   "loss notification" feedback from network (for now).
2302   *		   It means that this number can be only _guessed_.
2303   *		   Actually, it is the heuristics to predict lossage that
2304   *		   distinguishes different algorithms.
2305   *
2306   *	F.e. after RTO, when all the queue is considered as lost,
2307   *	lost_out = packets_out and in_flight = retrans_out.
2308   *
2309   *		Essentially, we have now a few algorithms detecting
2310   *		lost packets.
2311   *
2312   *		If the receiver supports SACK:
2313   *
2314   *		RFC6675/3517: It is the conventional algorithm. A packet is
2315   *		considered lost if the number of higher sequence packets
2316   *		SACKed is greater than or equal the DUPACK thoreshold
2317   *		(reordering). This is implemented in tcp_mark_head_lost and
2318   *		tcp_update_scoreboard.
2319   *
2320   *		RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2321   *		(2017-) that checks timing instead of counting DUPACKs.
2322   *		Essentially a packet is considered lost if it's not S/ACKed
2323   *		after RTT + reordering_window, where both metrics are
2324   *		dynamically measured and adjusted. This is implemented in
2325   *		tcp_rack_mark_lost.
2326   *
2327   *		If the receiver does not support SACK:
2328   *
2329   *		NewReno (RFC6582): in Recovery we assume that one segment
2330   *		is lost (classic Reno). While we are in Recovery and
2331   *		a partial ACK arrives, we assume that one more packet
2332   *		is lost (NewReno). This heuristics are the same in NewReno
2333   *		and SACK.
2334   *
2335   * Really tricky (and requiring careful tuning) part of algorithm
2336   * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2337   * The first determines the moment _when_ we should reduce CWND and,
2338   * hence, slow down forward transmission. In fact, it determines the moment
2339   * when we decide that hole is caused by loss, rather than by a reorder.
2340   *
2341   * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2342   * holes, caused by lost packets.
2343   *
2344   * And the most logically complicated part of algorithm is undo
2345   * heuristics. We detect false retransmits due to both too early
2346   * fast retransmit (reordering) and underestimated RTO, analyzing
2347   * timestamps and D-SACKs. When we detect that some segments were
2348   * retransmitted by mistake and CWND reduction was wrong, we undo
2349   * window reduction and abort recovery phase. This logic is hidden
2350   * inside several functions named tcp_try_undo_<something>.
2351   */
2352  
2353  /* This function decides, when we should leave Disordered state
2354   * and enter Recovery phase, reducing congestion window.
2355   *
2356   * Main question: may we further continue forward transmission
2357   * with the same cwnd?
2358   */
tcp_time_to_recover(struct sock * sk,int flag)2359  static bool tcp_time_to_recover(struct sock *sk, int flag)
2360  {
2361  	struct tcp_sock *tp = tcp_sk(sk);
2362  
2363  	/* Trick#1: The loss is proven. */
2364  	if (tp->lost_out)
2365  		return true;
2366  
2367  	/* Not-A-Trick#2 : Classic rule... */
2368  	if (!tcp_is_rack(sk) && tcp_dupack_heuristics(tp) > tp->reordering)
2369  		return true;
2370  
2371  	return false;
2372  }
2373  
2374  /* Detect loss in event "A" above by marking head of queue up as lost.
2375   * For RFC3517 SACK, a segment is considered lost if it
2376   * has at least tp->reordering SACKed seqments above it; "packets" refers to
2377   * the maximum SACKed segments to pass before reaching this limit.
2378   */
tcp_mark_head_lost(struct sock * sk,int packets,int mark_head)2379  static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2380  {
2381  	struct tcp_sock *tp = tcp_sk(sk);
2382  	struct sk_buff *skb;
2383  	int cnt;
2384  	/* Use SACK to deduce losses of new sequences sent during recovery */
2385  	const u32 loss_high = tp->snd_nxt;
2386  
2387  	WARN_ON(packets > tp->packets_out);
2388  	skb = tp->lost_skb_hint;
2389  	if (skb) {
2390  		/* Head already handled? */
2391  		if (mark_head && after(TCP_SKB_CB(skb)->seq, tp->snd_una))
2392  			return;
2393  		cnt = tp->lost_cnt_hint;
2394  	} else {
2395  		skb = tcp_rtx_queue_head(sk);
2396  		cnt = 0;
2397  	}
2398  
2399  	skb_rbtree_walk_from(skb) {
2400  		/* TODO: do this better */
2401  		/* this is not the most efficient way to do this... */
2402  		tp->lost_skb_hint = skb;
2403  		tp->lost_cnt_hint = cnt;
2404  
2405  		if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2406  			break;
2407  
2408  		if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2409  			cnt += tcp_skb_pcount(skb);
2410  
2411  		if (cnt > packets)
2412  			break;
2413  
2414  		if (!(TCP_SKB_CB(skb)->sacked & TCPCB_LOST))
2415  			tcp_mark_skb_lost(sk, skb);
2416  
2417  		if (mark_head)
2418  			break;
2419  	}
2420  	tcp_verify_left_out(tp);
2421  }
2422  
2423  /* Account newly detected lost packet(s) */
2424  
tcp_update_scoreboard(struct sock * sk,int fast_rexmit)2425  static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2426  {
2427  	struct tcp_sock *tp = tcp_sk(sk);
2428  
2429  	if (tcp_is_sack(tp)) {
2430  		int sacked_upto = tp->sacked_out - tp->reordering;
2431  		if (sacked_upto >= 0)
2432  			tcp_mark_head_lost(sk, sacked_upto, 0);
2433  		else if (fast_rexmit)
2434  			tcp_mark_head_lost(sk, 1, 1);
2435  	}
2436  }
2437  
tcp_tsopt_ecr_before(const struct tcp_sock * tp,u32 when)2438  static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2439  {
2440  	return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2441  	       before(tp->rx_opt.rcv_tsecr, when);
2442  }
2443  
2444  /* skb is spurious retransmitted if the returned timestamp echo
2445   * reply is prior to the skb transmission time
2446   */
tcp_skb_spurious_retrans(const struct tcp_sock * tp,const struct sk_buff * skb)2447  static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2448  				     const struct sk_buff *skb)
2449  {
2450  	return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2451  	       tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb));
2452  }
2453  
2454  /* Nothing was retransmitted or returned timestamp is less
2455   * than timestamp of the first retransmission.
2456   */
tcp_packet_delayed(const struct tcp_sock * tp)2457  static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2458  {
2459  	const struct sock *sk = (const struct sock *)tp;
2460  
2461  	if (tp->retrans_stamp &&
2462  	    tcp_tsopt_ecr_before(tp, tp->retrans_stamp))
2463  		return true;  /* got echoed TS before first retransmission */
2464  
2465  	/* Check if nothing was retransmitted (retrans_stamp==0), which may
2466  	 * happen in fast recovery due to TSQ. But we ignore zero retrans_stamp
2467  	 * in TCP_SYN_SENT, since when we set FLAG_SYN_ACKED we also clear
2468  	 * retrans_stamp even if we had retransmitted the SYN.
2469  	 */
2470  	if (!tp->retrans_stamp &&	   /* no record of a retransmit/SYN? */
2471  	    sk->sk_state != TCP_SYN_SENT)  /* not the FLAG_SYN_ACKED case? */
2472  		return true;  /* nothing was retransmitted */
2473  
2474  	return false;
2475  }
2476  
2477  /* Undo procedures. */
2478  
2479  /* We can clear retrans_stamp when there are no retransmissions in the
2480   * window. It would seem that it is trivially available for us in
2481   * tp->retrans_out, however, that kind of assumptions doesn't consider
2482   * what will happen if errors occur when sending retransmission for the
2483   * second time. ...It could the that such segment has only
2484   * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2485   * the head skb is enough except for some reneging corner cases that
2486   * are not worth the effort.
2487   *
2488   * Main reason for all this complexity is the fact that connection dying
2489   * time now depends on the validity of the retrans_stamp, in particular,
2490   * that successive retransmissions of a segment must not advance
2491   * retrans_stamp under any conditions.
2492   */
tcp_any_retrans_done(const struct sock * sk)2493  static bool tcp_any_retrans_done(const struct sock *sk)
2494  {
2495  	const struct tcp_sock *tp = tcp_sk(sk);
2496  	struct sk_buff *skb;
2497  
2498  	if (tp->retrans_out)
2499  		return true;
2500  
2501  	skb = tcp_rtx_queue_head(sk);
2502  	if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2503  		return true;
2504  
2505  	return false;
2506  }
2507  
2508  /* If loss recovery is finished and there are no retransmits out in the
2509   * network, then we clear retrans_stamp so that upon the next loss recovery
2510   * retransmits_timed_out() and timestamp-undo are using the correct value.
2511   */
tcp_retrans_stamp_cleanup(struct sock * sk)2512  static void tcp_retrans_stamp_cleanup(struct sock *sk)
2513  {
2514  	if (!tcp_any_retrans_done(sk))
2515  		tcp_sk(sk)->retrans_stamp = 0;
2516  }
2517  
DBGUNDO(struct sock * sk,const char * msg)2518  static void DBGUNDO(struct sock *sk, const char *msg)
2519  {
2520  #if FASTRETRANS_DEBUG > 1
2521  	struct tcp_sock *tp = tcp_sk(sk);
2522  	struct inet_sock *inet = inet_sk(sk);
2523  
2524  	if (sk->sk_family == AF_INET) {
2525  		pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2526  			 msg,
2527  			 &inet->inet_daddr, ntohs(inet->inet_dport),
2528  			 tcp_snd_cwnd(tp), tcp_left_out(tp),
2529  			 tp->snd_ssthresh, tp->prior_ssthresh,
2530  			 tp->packets_out);
2531  	}
2532  #if IS_ENABLED(CONFIG_IPV6)
2533  	else if (sk->sk_family == AF_INET6) {
2534  		pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2535  			 msg,
2536  			 &sk->sk_v6_daddr, ntohs(inet->inet_dport),
2537  			 tcp_snd_cwnd(tp), tcp_left_out(tp),
2538  			 tp->snd_ssthresh, tp->prior_ssthresh,
2539  			 tp->packets_out);
2540  	}
2541  #endif
2542  #endif
2543  }
2544  
tcp_undo_cwnd_reduction(struct sock * sk,bool unmark_loss)2545  static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2546  {
2547  	struct tcp_sock *tp = tcp_sk(sk);
2548  
2549  	if (unmark_loss) {
2550  		struct sk_buff *skb;
2551  
2552  		skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2553  			TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2554  		}
2555  		tp->lost_out = 0;
2556  		tcp_clear_all_retrans_hints(tp);
2557  	}
2558  
2559  	if (tp->prior_ssthresh) {
2560  		const struct inet_connection_sock *icsk = inet_csk(sk);
2561  
2562  		tcp_snd_cwnd_set(tp, icsk->icsk_ca_ops->undo_cwnd(sk));
2563  
2564  		if (tp->prior_ssthresh > tp->snd_ssthresh) {
2565  			tp->snd_ssthresh = tp->prior_ssthresh;
2566  			tcp_ecn_withdraw_cwr(tp);
2567  		}
2568  	}
2569  	tp->snd_cwnd_stamp = tcp_jiffies32;
2570  	tp->undo_marker = 0;
2571  	tp->rack.advanced = 1; /* Force RACK to re-exam losses */
2572  }
2573  
tcp_may_undo(const struct tcp_sock * tp)2574  static inline bool tcp_may_undo(const struct tcp_sock *tp)
2575  {
2576  	return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2577  }
2578  
tcp_is_non_sack_preventing_reopen(struct sock * sk)2579  static bool tcp_is_non_sack_preventing_reopen(struct sock *sk)
2580  {
2581  	struct tcp_sock *tp = tcp_sk(sk);
2582  
2583  	if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2584  		/* Hold old state until something *above* high_seq
2585  		 * is ACKed. For Reno it is MUST to prevent false
2586  		 * fast retransmits (RFC2582). SACK TCP is safe. */
2587  		if (!tcp_any_retrans_done(sk))
2588  			tp->retrans_stamp = 0;
2589  		return true;
2590  	}
2591  	return false;
2592  }
2593  
2594  /* People celebrate: "We love our President!" */
tcp_try_undo_recovery(struct sock * sk)2595  static bool tcp_try_undo_recovery(struct sock *sk)
2596  {
2597  	struct tcp_sock *tp = tcp_sk(sk);
2598  
2599  	if (tcp_may_undo(tp)) {
2600  		int mib_idx;
2601  
2602  		/* Happy end! We did not retransmit anything
2603  		 * or our original transmission succeeded.
2604  		 */
2605  		DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2606  		tcp_undo_cwnd_reduction(sk, false);
2607  		if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2608  			mib_idx = LINUX_MIB_TCPLOSSUNDO;
2609  		else
2610  			mib_idx = LINUX_MIB_TCPFULLUNDO;
2611  
2612  		NET_INC_STATS(sock_net(sk), mib_idx);
2613  	} else if (tp->rack.reo_wnd_persist) {
2614  		tp->rack.reo_wnd_persist--;
2615  	}
2616  	if (tcp_is_non_sack_preventing_reopen(sk))
2617  		return true;
2618  	tcp_set_ca_state(sk, TCP_CA_Open);
2619  	tp->is_sack_reneg = 0;
2620  	return false;
2621  }
2622  
2623  /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
tcp_try_undo_dsack(struct sock * sk)2624  static bool tcp_try_undo_dsack(struct sock *sk)
2625  {
2626  	struct tcp_sock *tp = tcp_sk(sk);
2627  
2628  	if (tp->undo_marker && !tp->undo_retrans) {
2629  		tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH,
2630  					       tp->rack.reo_wnd_persist + 1);
2631  		DBGUNDO(sk, "D-SACK");
2632  		tcp_undo_cwnd_reduction(sk, false);
2633  		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2634  		return true;
2635  	}
2636  	return false;
2637  }
2638  
2639  /* Undo during loss recovery after partial ACK or using F-RTO. */
tcp_try_undo_loss(struct sock * sk,bool frto_undo)2640  static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2641  {
2642  	struct tcp_sock *tp = tcp_sk(sk);
2643  
2644  	if (frto_undo || tcp_may_undo(tp)) {
2645  		tcp_undo_cwnd_reduction(sk, true);
2646  
2647  		DBGUNDO(sk, "partial loss");
2648  		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2649  		if (frto_undo)
2650  			NET_INC_STATS(sock_net(sk),
2651  					LINUX_MIB_TCPSPURIOUSRTOS);
2652  		inet_csk(sk)->icsk_retransmits = 0;
2653  		if (tcp_is_non_sack_preventing_reopen(sk))
2654  			return true;
2655  		if (frto_undo || tcp_is_sack(tp)) {
2656  			tcp_set_ca_state(sk, TCP_CA_Open);
2657  			tp->is_sack_reneg = 0;
2658  		}
2659  		return true;
2660  	}
2661  	return false;
2662  }
2663  
2664  /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2665   * It computes the number of packets to send (sndcnt) based on packets newly
2666   * delivered:
2667   *   1) If the packets in flight is larger than ssthresh, PRR spreads the
2668   *	cwnd reductions across a full RTT.
2669   *   2) Otherwise PRR uses packet conservation to send as much as delivered.
2670   *      But when SND_UNA is acked without further losses,
2671   *      slow starts cwnd up to ssthresh to speed up the recovery.
2672   */
tcp_init_cwnd_reduction(struct sock * sk)2673  static void tcp_init_cwnd_reduction(struct sock *sk)
2674  {
2675  	struct tcp_sock *tp = tcp_sk(sk);
2676  
2677  	tp->high_seq = tp->snd_nxt;
2678  	tp->tlp_high_seq = 0;
2679  	tp->snd_cwnd_cnt = 0;
2680  	tp->prior_cwnd = tcp_snd_cwnd(tp);
2681  	tp->prr_delivered = 0;
2682  	tp->prr_out = 0;
2683  	tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2684  	tcp_ecn_queue_cwr(tp);
2685  }
2686  
tcp_cwnd_reduction(struct sock * sk,int newly_acked_sacked,int newly_lost,int flag)2687  void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int newly_lost, int flag)
2688  {
2689  	struct tcp_sock *tp = tcp_sk(sk);
2690  	int sndcnt = 0;
2691  	int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2692  
2693  	if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2694  		return;
2695  
2696  	tp->prr_delivered += newly_acked_sacked;
2697  	if (delta < 0) {
2698  		u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2699  			       tp->prior_cwnd - 1;
2700  		sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2701  	} else {
2702  		sndcnt = max_t(int, tp->prr_delivered - tp->prr_out,
2703  			       newly_acked_sacked);
2704  		if (flag & FLAG_SND_UNA_ADVANCED && !newly_lost)
2705  			sndcnt++;
2706  		sndcnt = min(delta, sndcnt);
2707  	}
2708  	/* Force a fast retransmit upon entering fast recovery */
2709  	sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2710  	tcp_snd_cwnd_set(tp, tcp_packets_in_flight(tp) + sndcnt);
2711  }
2712  
tcp_end_cwnd_reduction(struct sock * sk)2713  static inline void tcp_end_cwnd_reduction(struct sock *sk)
2714  {
2715  	struct tcp_sock *tp = tcp_sk(sk);
2716  
2717  	if (inet_csk(sk)->icsk_ca_ops->cong_control)
2718  		return;
2719  
2720  	/* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2721  	if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH &&
2722  	    (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) {
2723  		tcp_snd_cwnd_set(tp, tp->snd_ssthresh);
2724  		tp->snd_cwnd_stamp = tcp_jiffies32;
2725  	}
2726  	tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2727  }
2728  
2729  /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
tcp_enter_cwr(struct sock * sk)2730  void tcp_enter_cwr(struct sock *sk)
2731  {
2732  	struct tcp_sock *tp = tcp_sk(sk);
2733  
2734  	tp->prior_ssthresh = 0;
2735  	if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2736  		tp->undo_marker = 0;
2737  		tcp_init_cwnd_reduction(sk);
2738  		tcp_set_ca_state(sk, TCP_CA_CWR);
2739  	}
2740  }
2741  EXPORT_SYMBOL(tcp_enter_cwr);
2742  
tcp_try_keep_open(struct sock * sk)2743  static void tcp_try_keep_open(struct sock *sk)
2744  {
2745  	struct tcp_sock *tp = tcp_sk(sk);
2746  	int state = TCP_CA_Open;
2747  
2748  	if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2749  		state = TCP_CA_Disorder;
2750  
2751  	if (inet_csk(sk)->icsk_ca_state != state) {
2752  		tcp_set_ca_state(sk, state);
2753  		tp->high_seq = tp->snd_nxt;
2754  	}
2755  }
2756  
tcp_try_to_open(struct sock * sk,int flag)2757  static void tcp_try_to_open(struct sock *sk, int flag)
2758  {
2759  	struct tcp_sock *tp = tcp_sk(sk);
2760  
2761  	tcp_verify_left_out(tp);
2762  
2763  	if (!tcp_any_retrans_done(sk))
2764  		tp->retrans_stamp = 0;
2765  
2766  	if (flag & FLAG_ECE)
2767  		tcp_enter_cwr(sk);
2768  
2769  	if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2770  		tcp_try_keep_open(sk);
2771  	}
2772  }
2773  
tcp_mtup_probe_failed(struct sock * sk)2774  static void tcp_mtup_probe_failed(struct sock *sk)
2775  {
2776  	struct inet_connection_sock *icsk = inet_csk(sk);
2777  
2778  	icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2779  	icsk->icsk_mtup.probe_size = 0;
2780  	NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2781  }
2782  
tcp_mtup_probe_success(struct sock * sk)2783  static void tcp_mtup_probe_success(struct sock *sk)
2784  {
2785  	struct tcp_sock *tp = tcp_sk(sk);
2786  	struct inet_connection_sock *icsk = inet_csk(sk);
2787  	u64 val;
2788  
2789  	tp->prior_ssthresh = tcp_current_ssthresh(sk);
2790  
2791  	val = (u64)tcp_snd_cwnd(tp) * tcp_mss_to_mtu(sk, tp->mss_cache);
2792  	do_div(val, icsk->icsk_mtup.probe_size);
2793  	DEBUG_NET_WARN_ON_ONCE((u32)val != val);
2794  	tcp_snd_cwnd_set(tp, max_t(u32, 1U, val));
2795  
2796  	tp->snd_cwnd_cnt = 0;
2797  	tp->snd_cwnd_stamp = tcp_jiffies32;
2798  	tp->snd_ssthresh = tcp_current_ssthresh(sk);
2799  
2800  	icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2801  	icsk->icsk_mtup.probe_size = 0;
2802  	tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2803  	NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2804  }
2805  
2806  /* Sometimes we deduce that packets have been dropped due to reasons other than
2807   * congestion, like path MTU reductions or failed client TFO attempts. In these
2808   * cases we call this function to retransmit as many packets as cwnd allows,
2809   * without reducing cwnd. Given that retransmits will set retrans_stamp to a
2810   * non-zero value (and may do so in a later calling context due to TSQ), we
2811   * also enter CA_Loss so that we track when all retransmitted packets are ACKed
2812   * and clear retrans_stamp when that happens (to ensure later recurring RTOs
2813   * are using the correct retrans_stamp and don't declare ETIMEDOUT
2814   * prematurely).
2815   */
tcp_non_congestion_loss_retransmit(struct sock * sk)2816  static void tcp_non_congestion_loss_retransmit(struct sock *sk)
2817  {
2818  	const struct inet_connection_sock *icsk = inet_csk(sk);
2819  	struct tcp_sock *tp = tcp_sk(sk);
2820  
2821  	if (icsk->icsk_ca_state != TCP_CA_Loss) {
2822  		tp->high_seq = tp->snd_nxt;
2823  		tp->snd_ssthresh = tcp_current_ssthresh(sk);
2824  		tp->prior_ssthresh = 0;
2825  		tp->undo_marker = 0;
2826  		tcp_set_ca_state(sk, TCP_CA_Loss);
2827  	}
2828  	tcp_xmit_retransmit_queue(sk);
2829  }
2830  
2831  /* Do a simple retransmit without using the backoff mechanisms in
2832   * tcp_timer. This is used for path mtu discovery.
2833   * The socket is already locked here.
2834   */
tcp_simple_retransmit(struct sock * sk)2835  void tcp_simple_retransmit(struct sock *sk)
2836  {
2837  	struct tcp_sock *tp = tcp_sk(sk);
2838  	struct sk_buff *skb;
2839  	int mss;
2840  
2841  	/* A fastopen SYN request is stored as two separate packets within
2842  	 * the retransmit queue, this is done by tcp_send_syn_data().
2843  	 * As a result simply checking the MSS of the frames in the queue
2844  	 * will not work for the SYN packet.
2845  	 *
2846  	 * Us being here is an indication of a path MTU issue so we can
2847  	 * assume that the fastopen SYN was lost and just mark all the
2848  	 * frames in the retransmit queue as lost. We will use an MSS of
2849  	 * -1 to mark all frames as lost, otherwise compute the current MSS.
2850  	 */
2851  	if (tp->syn_data && sk->sk_state == TCP_SYN_SENT)
2852  		mss = -1;
2853  	else
2854  		mss = tcp_current_mss(sk);
2855  
2856  	skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2857  		if (tcp_skb_seglen(skb) > mss)
2858  			tcp_mark_skb_lost(sk, skb);
2859  	}
2860  
2861  	tcp_clear_retrans_hints_partial(tp);
2862  
2863  	if (!tp->lost_out)
2864  		return;
2865  
2866  	if (tcp_is_reno(tp))
2867  		tcp_limit_reno_sacked(tp);
2868  
2869  	tcp_verify_left_out(tp);
2870  
2871  	/* Don't muck with the congestion window here.
2872  	 * Reason is that we do not increase amount of _data_
2873  	 * in network, but units changed and effective
2874  	 * cwnd/ssthresh really reduced now.
2875  	 */
2876  	tcp_non_congestion_loss_retransmit(sk);
2877  }
2878  EXPORT_SYMBOL(tcp_simple_retransmit);
2879  
tcp_enter_recovery(struct sock * sk,bool ece_ack)2880  void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2881  {
2882  	struct tcp_sock *tp = tcp_sk(sk);
2883  	int mib_idx;
2884  
2885  	/* Start the clock with our fast retransmit, for undo and ETIMEDOUT. */
2886  	tcp_retrans_stamp_cleanup(sk);
2887  
2888  	if (tcp_is_reno(tp))
2889  		mib_idx = LINUX_MIB_TCPRENORECOVERY;
2890  	else
2891  		mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2892  
2893  	NET_INC_STATS(sock_net(sk), mib_idx);
2894  
2895  	tp->prior_ssthresh = 0;
2896  	tcp_init_undo(tp);
2897  
2898  	if (!tcp_in_cwnd_reduction(sk)) {
2899  		if (!ece_ack)
2900  			tp->prior_ssthresh = tcp_current_ssthresh(sk);
2901  		tcp_init_cwnd_reduction(sk);
2902  	}
2903  	tcp_set_ca_state(sk, TCP_CA_Recovery);
2904  }
2905  
tcp_update_rto_time(struct tcp_sock * tp)2906  static void tcp_update_rto_time(struct tcp_sock *tp)
2907  {
2908  	if (tp->rto_stamp) {
2909  		tp->total_rto_time += tcp_time_stamp(tp) - tp->rto_stamp;
2910  		tp->rto_stamp = 0;
2911  	}
2912  }
2913  
2914  /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2915   * recovered or spurious. Otherwise retransmits more on partial ACKs.
2916   */
tcp_process_loss(struct sock * sk,int flag,int num_dupack,int * rexmit)2917  static void tcp_process_loss(struct sock *sk, int flag, int num_dupack,
2918  			     int *rexmit)
2919  {
2920  	struct tcp_sock *tp = tcp_sk(sk);
2921  	bool recovered = !before(tp->snd_una, tp->high_seq);
2922  
2923  	if ((flag & FLAG_SND_UNA_ADVANCED || rcu_access_pointer(tp->fastopen_rsk)) &&
2924  	    tcp_try_undo_loss(sk, false))
2925  		return;
2926  
2927  	if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2928  		/* Step 3.b. A timeout is spurious if not all data are
2929  		 * lost, i.e., never-retransmitted data are (s)acked.
2930  		 */
2931  		if ((flag & FLAG_ORIG_SACK_ACKED) &&
2932  		    tcp_try_undo_loss(sk, true))
2933  			return;
2934  
2935  		if (after(tp->snd_nxt, tp->high_seq)) {
2936  			if (flag & FLAG_DATA_SACKED || num_dupack)
2937  				tp->frto = 0; /* Step 3.a. loss was real */
2938  		} else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2939  			tp->high_seq = tp->snd_nxt;
2940  			/* Step 2.b. Try send new data (but deferred until cwnd
2941  			 * is updated in tcp_ack()). Otherwise fall back to
2942  			 * the conventional recovery.
2943  			 */
2944  			if (!tcp_write_queue_empty(sk) &&
2945  			    after(tcp_wnd_end(tp), tp->snd_nxt)) {
2946  				*rexmit = REXMIT_NEW;
2947  				return;
2948  			}
2949  			tp->frto = 0;
2950  		}
2951  	}
2952  
2953  	if (recovered) {
2954  		/* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2955  		tcp_try_undo_recovery(sk);
2956  		return;
2957  	}
2958  	if (tcp_is_reno(tp)) {
2959  		/* A Reno DUPACK means new data in F-RTO step 2.b above are
2960  		 * delivered. Lower inflight to clock out (re)transmissions.
2961  		 */
2962  		if (after(tp->snd_nxt, tp->high_seq) && num_dupack)
2963  			tcp_add_reno_sack(sk, num_dupack, flag & FLAG_ECE);
2964  		else if (flag & FLAG_SND_UNA_ADVANCED)
2965  			tcp_reset_reno_sack(tp);
2966  	}
2967  	*rexmit = REXMIT_LOST;
2968  }
2969  
tcp_force_fast_retransmit(struct sock * sk)2970  static bool tcp_force_fast_retransmit(struct sock *sk)
2971  {
2972  	struct tcp_sock *tp = tcp_sk(sk);
2973  
2974  	return after(tcp_highest_sack_seq(tp),
2975  		     tp->snd_una + tp->reordering * tp->mss_cache);
2976  }
2977  
2978  /* Undo during fast recovery after partial ACK. */
tcp_try_undo_partial(struct sock * sk,u32 prior_snd_una,bool * do_lost)2979  static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una,
2980  				 bool *do_lost)
2981  {
2982  	struct tcp_sock *tp = tcp_sk(sk);
2983  
2984  	if (tp->undo_marker && tcp_packet_delayed(tp)) {
2985  		/* Plain luck! Hole if filled with delayed
2986  		 * packet, rather than with a retransmit. Check reordering.
2987  		 */
2988  		tcp_check_sack_reordering(sk, prior_snd_una, 1);
2989  
2990  		/* We are getting evidence that the reordering degree is higher
2991  		 * than we realized. If there are no retransmits out then we
2992  		 * can undo. Otherwise we clock out new packets but do not
2993  		 * mark more packets lost or retransmit more.
2994  		 */
2995  		if (tp->retrans_out)
2996  			return true;
2997  
2998  		if (!tcp_any_retrans_done(sk))
2999  			tp->retrans_stamp = 0;
3000  
3001  		DBGUNDO(sk, "partial recovery");
3002  		tcp_undo_cwnd_reduction(sk, true);
3003  		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
3004  		tcp_try_keep_open(sk);
3005  	} else {
3006  		/* Partial ACK arrived. Force fast retransmit. */
3007  		*do_lost = tcp_force_fast_retransmit(sk);
3008  	}
3009  	return false;
3010  }
3011  
tcp_identify_packet_loss(struct sock * sk,int * ack_flag)3012  static void tcp_identify_packet_loss(struct sock *sk, int *ack_flag)
3013  {
3014  	struct tcp_sock *tp = tcp_sk(sk);
3015  
3016  	if (tcp_rtx_queue_empty(sk))
3017  		return;
3018  
3019  	if (unlikely(tcp_is_reno(tp))) {
3020  		tcp_newreno_mark_lost(sk, *ack_flag & FLAG_SND_UNA_ADVANCED);
3021  	} else if (tcp_is_rack(sk)) {
3022  		u32 prior_retrans = tp->retrans_out;
3023  
3024  		if (tcp_rack_mark_lost(sk))
3025  			*ack_flag &= ~FLAG_SET_XMIT_TIMER;
3026  		if (prior_retrans > tp->retrans_out)
3027  			*ack_flag |= FLAG_LOST_RETRANS;
3028  	}
3029  }
3030  
3031  /* Process an event, which can update packets-in-flight not trivially.
3032   * Main goal of this function is to calculate new estimate for left_out,
3033   * taking into account both packets sitting in receiver's buffer and
3034   * packets lost by network.
3035   *
3036   * Besides that it updates the congestion state when packet loss or ECN
3037   * is detected. But it does not reduce the cwnd, it is done by the
3038   * congestion control later.
3039   *
3040   * It does _not_ decide what to send, it is made in function
3041   * tcp_xmit_retransmit_queue().
3042   */
tcp_fastretrans_alert(struct sock * sk,const u32 prior_snd_una,int num_dupack,int * ack_flag,int * rexmit)3043  static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una,
3044  				  int num_dupack, int *ack_flag, int *rexmit)
3045  {
3046  	struct inet_connection_sock *icsk = inet_csk(sk);
3047  	struct tcp_sock *tp = tcp_sk(sk);
3048  	int fast_rexmit = 0, flag = *ack_flag;
3049  	bool ece_ack = flag & FLAG_ECE;
3050  	bool do_lost = num_dupack || ((flag & FLAG_DATA_SACKED) &&
3051  				      tcp_force_fast_retransmit(sk));
3052  
3053  	if (!tp->packets_out && tp->sacked_out)
3054  		tp->sacked_out = 0;
3055  
3056  	/* Now state machine starts.
3057  	 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
3058  	if (ece_ack)
3059  		tp->prior_ssthresh = 0;
3060  
3061  	/* B. In all the states check for reneging SACKs. */
3062  	if (tcp_check_sack_reneging(sk, ack_flag))
3063  		return;
3064  
3065  	/* C. Check consistency of the current state. */
3066  	tcp_verify_left_out(tp);
3067  
3068  	/* D. Check state exit conditions. State can be terminated
3069  	 *    when high_seq is ACKed. */
3070  	if (icsk->icsk_ca_state == TCP_CA_Open) {
3071  		WARN_ON(tp->retrans_out != 0 && !tp->syn_data);
3072  		tp->retrans_stamp = 0;
3073  	} else if (!before(tp->snd_una, tp->high_seq)) {
3074  		switch (icsk->icsk_ca_state) {
3075  		case TCP_CA_CWR:
3076  			/* CWR is to be held something *above* high_seq
3077  			 * is ACKed for CWR bit to reach receiver. */
3078  			if (tp->snd_una != tp->high_seq) {
3079  				tcp_end_cwnd_reduction(sk);
3080  				tcp_set_ca_state(sk, TCP_CA_Open);
3081  			}
3082  			break;
3083  
3084  		case TCP_CA_Recovery:
3085  			if (tcp_is_reno(tp))
3086  				tcp_reset_reno_sack(tp);
3087  			if (tcp_try_undo_recovery(sk))
3088  				return;
3089  			tcp_end_cwnd_reduction(sk);
3090  			break;
3091  		}
3092  	}
3093  
3094  	/* E. Process state. */
3095  	switch (icsk->icsk_ca_state) {
3096  	case TCP_CA_Recovery:
3097  		if (!(flag & FLAG_SND_UNA_ADVANCED)) {
3098  			if (tcp_is_reno(tp))
3099  				tcp_add_reno_sack(sk, num_dupack, ece_ack);
3100  		} else if (tcp_try_undo_partial(sk, prior_snd_una, &do_lost))
3101  			return;
3102  
3103  		if (tcp_try_undo_dsack(sk))
3104  			tcp_try_to_open(sk, flag);
3105  
3106  		tcp_identify_packet_loss(sk, ack_flag);
3107  		if (icsk->icsk_ca_state != TCP_CA_Recovery) {
3108  			if (!tcp_time_to_recover(sk, flag))
3109  				return;
3110  			/* Undo reverts the recovery state. If loss is evident,
3111  			 * starts a new recovery (e.g. reordering then loss);
3112  			 */
3113  			tcp_enter_recovery(sk, ece_ack);
3114  		}
3115  		break;
3116  	case TCP_CA_Loss:
3117  		tcp_process_loss(sk, flag, num_dupack, rexmit);
3118  		if (icsk->icsk_ca_state != TCP_CA_Loss)
3119  			tcp_update_rto_time(tp);
3120  		tcp_identify_packet_loss(sk, ack_flag);
3121  		if (!(icsk->icsk_ca_state == TCP_CA_Open ||
3122  		      (*ack_flag & FLAG_LOST_RETRANS)))
3123  			return;
3124  		/* Change state if cwnd is undone or retransmits are lost */
3125  		fallthrough;
3126  	default:
3127  		if (tcp_is_reno(tp)) {
3128  			if (flag & FLAG_SND_UNA_ADVANCED)
3129  				tcp_reset_reno_sack(tp);
3130  			tcp_add_reno_sack(sk, num_dupack, ece_ack);
3131  		}
3132  
3133  		if (icsk->icsk_ca_state <= TCP_CA_Disorder)
3134  			tcp_try_undo_dsack(sk);
3135  
3136  		tcp_identify_packet_loss(sk, ack_flag);
3137  		if (!tcp_time_to_recover(sk, flag)) {
3138  			tcp_try_to_open(sk, flag);
3139  			return;
3140  		}
3141  
3142  		/* MTU probe failure: don't reduce cwnd */
3143  		if (icsk->icsk_ca_state < TCP_CA_CWR &&
3144  		    icsk->icsk_mtup.probe_size &&
3145  		    tp->snd_una == tp->mtu_probe.probe_seq_start) {
3146  			tcp_mtup_probe_failed(sk);
3147  			/* Restores the reduction we did in tcp_mtup_probe() */
3148  			tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) + 1);
3149  			tcp_simple_retransmit(sk);
3150  			return;
3151  		}
3152  
3153  		/* Otherwise enter Recovery state */
3154  		tcp_enter_recovery(sk, ece_ack);
3155  		fast_rexmit = 1;
3156  	}
3157  
3158  	if (!tcp_is_rack(sk) && do_lost)
3159  		tcp_update_scoreboard(sk, fast_rexmit);
3160  	*rexmit = REXMIT_LOST;
3161  }
3162  
tcp_update_rtt_min(struct sock * sk,u32 rtt_us,const int flag)3163  static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag)
3164  {
3165  	u32 wlen = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen) * HZ;
3166  	struct tcp_sock *tp = tcp_sk(sk);
3167  
3168  	if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) {
3169  		/* If the remote keeps returning delayed ACKs, eventually
3170  		 * the min filter would pick it up and overestimate the
3171  		 * prop. delay when it expires. Skip suspected delayed ACKs.
3172  		 */
3173  		return;
3174  	}
3175  	minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32,
3176  			   rtt_us ? : jiffies_to_usecs(1));
3177  }
3178  
tcp_ack_update_rtt(struct sock * sk,const int flag,long seq_rtt_us,long sack_rtt_us,long ca_rtt_us,struct rate_sample * rs)3179  static bool tcp_ack_update_rtt(struct sock *sk, const int flag,
3180  			       long seq_rtt_us, long sack_rtt_us,
3181  			       long ca_rtt_us, struct rate_sample *rs)
3182  {
3183  	const struct tcp_sock *tp = tcp_sk(sk);
3184  
3185  	/* Prefer RTT measured from ACK's timing to TS-ECR. This is because
3186  	 * broken middle-boxes or peers may corrupt TS-ECR fields. But
3187  	 * Karn's algorithm forbids taking RTT if some retransmitted data
3188  	 * is acked (RFC6298).
3189  	 */
3190  	if (seq_rtt_us < 0)
3191  		seq_rtt_us = sack_rtt_us;
3192  
3193  	/* RTTM Rule: A TSecr value received in a segment is used to
3194  	 * update the averaged RTT measurement only if the segment
3195  	 * acknowledges some new data, i.e., only if it advances the
3196  	 * left edge of the send window.
3197  	 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3198  	 */
3199  	if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
3200  	    flag & FLAG_ACKED) {
3201  		u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
3202  
3203  		if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
3204  			if (!delta)
3205  				delta = 1;
3206  			seq_rtt_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
3207  			ca_rtt_us = seq_rtt_us;
3208  		}
3209  	}
3210  	rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */
3211  	if (seq_rtt_us < 0)
3212  		return false;
3213  
3214  	/* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
3215  	 * always taken together with ACK, SACK, or TS-opts. Any negative
3216  	 * values will be skipped with the seq_rtt_us < 0 check above.
3217  	 */
3218  	tcp_update_rtt_min(sk, ca_rtt_us, flag);
3219  	tcp_rtt_estimator(sk, seq_rtt_us);
3220  	tcp_set_rto(sk);
3221  
3222  	/* RFC6298: only reset backoff on valid RTT measurement. */
3223  	inet_csk(sk)->icsk_backoff = 0;
3224  	return true;
3225  }
3226  
3227  /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
tcp_synack_rtt_meas(struct sock * sk,struct request_sock * req)3228  void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
3229  {
3230  	struct rate_sample rs;
3231  	long rtt_us = -1L;
3232  
3233  	if (req && !req->num_retrans && tcp_rsk(req)->snt_synack)
3234  		rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack);
3235  
3236  	tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs);
3237  }
3238  
3239  
tcp_cong_avoid(struct sock * sk,u32 ack,u32 acked)3240  static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
3241  {
3242  	const struct inet_connection_sock *icsk = inet_csk(sk);
3243  
3244  	icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
3245  	tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32;
3246  }
3247  
3248  /* Restart timer after forward progress on connection.
3249   * RFC2988 recommends to restart timer to now+rto.
3250   */
tcp_rearm_rto(struct sock * sk)3251  void tcp_rearm_rto(struct sock *sk)
3252  {
3253  	const struct inet_connection_sock *icsk = inet_csk(sk);
3254  	struct tcp_sock *tp = tcp_sk(sk);
3255  
3256  	/* If the retrans timer is currently being used by Fast Open
3257  	 * for SYN-ACK retrans purpose, stay put.
3258  	 */
3259  	if (rcu_access_pointer(tp->fastopen_rsk))
3260  		return;
3261  
3262  	if (!tp->packets_out) {
3263  		inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3264  	} else {
3265  		u32 rto = inet_csk(sk)->icsk_rto;
3266  		/* Offset the time elapsed after installing regular RTO */
3267  		if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
3268  		    icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
3269  			s64 delta_us = tcp_rto_delta_us(sk);
3270  			/* delta_us may not be positive if the socket is locked
3271  			 * when the retrans timer fires and is rescheduled.
3272  			 */
3273  			rto = usecs_to_jiffies(max_t(int, delta_us, 1));
3274  		}
3275  		tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3276  				     TCP_RTO_MAX);
3277  	}
3278  }
3279  
3280  /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
tcp_set_xmit_timer(struct sock * sk)3281  static void tcp_set_xmit_timer(struct sock *sk)
3282  {
3283  	if (!tcp_schedule_loss_probe(sk, true))
3284  		tcp_rearm_rto(sk);
3285  }
3286  
3287  /* If we get here, the whole TSO packet has not been acked. */
tcp_tso_acked(struct sock * sk,struct sk_buff * skb)3288  static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3289  {
3290  	struct tcp_sock *tp = tcp_sk(sk);
3291  	u32 packets_acked;
3292  
3293  	BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3294  
3295  	packets_acked = tcp_skb_pcount(skb);
3296  	if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3297  		return 0;
3298  	packets_acked -= tcp_skb_pcount(skb);
3299  
3300  	if (packets_acked) {
3301  		BUG_ON(tcp_skb_pcount(skb) == 0);
3302  		BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3303  	}
3304  
3305  	return packets_acked;
3306  }
3307  
tcp_ack_tstamp(struct sock * sk,struct sk_buff * skb,const struct sk_buff * ack_skb,u32 prior_snd_una)3308  static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3309  			   const struct sk_buff *ack_skb, u32 prior_snd_una)
3310  {
3311  	const struct skb_shared_info *shinfo;
3312  
3313  	/* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3314  	if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
3315  		return;
3316  
3317  	shinfo = skb_shinfo(skb);
3318  	if (!before(shinfo->tskey, prior_snd_una) &&
3319  	    before(shinfo->tskey, tcp_sk(sk)->snd_una)) {
3320  		tcp_skb_tsorted_save(skb) {
3321  			__skb_tstamp_tx(skb, ack_skb, NULL, sk, SCM_TSTAMP_ACK);
3322  		} tcp_skb_tsorted_restore(skb);
3323  	}
3324  }
3325  
3326  /* Remove acknowledged frames from the retransmission queue. If our packet
3327   * is before the ack sequence we can discard it as it's confirmed to have
3328   * arrived at the other end.
3329   */
tcp_clean_rtx_queue(struct sock * sk,const struct sk_buff * ack_skb,u32 prior_fack,u32 prior_snd_una,struct tcp_sacktag_state * sack,bool ece_ack)3330  static int tcp_clean_rtx_queue(struct sock *sk, const struct sk_buff *ack_skb,
3331  			       u32 prior_fack, u32 prior_snd_una,
3332  			       struct tcp_sacktag_state *sack, bool ece_ack)
3333  {
3334  	const struct inet_connection_sock *icsk = inet_csk(sk);
3335  	u64 first_ackt, last_ackt;
3336  	struct tcp_sock *tp = tcp_sk(sk);
3337  	u32 prior_sacked = tp->sacked_out;
3338  	u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */
3339  	struct sk_buff *skb, *next;
3340  	bool fully_acked = true;
3341  	long sack_rtt_us = -1L;
3342  	long seq_rtt_us = -1L;
3343  	long ca_rtt_us = -1L;
3344  	u32 pkts_acked = 0;
3345  	bool rtt_update;
3346  	int flag = 0;
3347  
3348  	first_ackt = 0;
3349  
3350  	for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) {
3351  		struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3352  		const u32 start_seq = scb->seq;
3353  		u8 sacked = scb->sacked;
3354  		u32 acked_pcount;
3355  
3356  		/* Determine how many packets and what bytes were acked, tso and else */
3357  		if (after(scb->end_seq, tp->snd_una)) {
3358  			if (tcp_skb_pcount(skb) == 1 ||
3359  			    !after(tp->snd_una, scb->seq))
3360  				break;
3361  
3362  			acked_pcount = tcp_tso_acked(sk, skb);
3363  			if (!acked_pcount)
3364  				break;
3365  			fully_acked = false;
3366  		} else {
3367  			acked_pcount = tcp_skb_pcount(skb);
3368  		}
3369  
3370  		if (unlikely(sacked & TCPCB_RETRANS)) {
3371  			if (sacked & TCPCB_SACKED_RETRANS)
3372  				tp->retrans_out -= acked_pcount;
3373  			flag |= FLAG_RETRANS_DATA_ACKED;
3374  		} else if (!(sacked & TCPCB_SACKED_ACKED)) {
3375  			last_ackt = tcp_skb_timestamp_us(skb);
3376  			WARN_ON_ONCE(last_ackt == 0);
3377  			if (!first_ackt)
3378  				first_ackt = last_ackt;
3379  
3380  			if (before(start_seq, reord))
3381  				reord = start_seq;
3382  			if (!after(scb->end_seq, tp->high_seq))
3383  				flag |= FLAG_ORIG_SACK_ACKED;
3384  		}
3385  
3386  		if (sacked & TCPCB_SACKED_ACKED) {
3387  			tp->sacked_out -= acked_pcount;
3388  		} else if (tcp_is_sack(tp)) {
3389  			tcp_count_delivered(tp, acked_pcount, ece_ack);
3390  			if (!tcp_skb_spurious_retrans(tp, skb))
3391  				tcp_rack_advance(tp, sacked, scb->end_seq,
3392  						 tcp_skb_timestamp_us(skb));
3393  		}
3394  		if (sacked & TCPCB_LOST)
3395  			tp->lost_out -= acked_pcount;
3396  
3397  		tp->packets_out -= acked_pcount;
3398  		pkts_acked += acked_pcount;
3399  		tcp_rate_skb_delivered(sk, skb, sack->rate);
3400  
3401  		/* Initial outgoing SYN's get put onto the write_queue
3402  		 * just like anything else we transmit.  It is not
3403  		 * true data, and if we misinform our callers that
3404  		 * this ACK acks real data, we will erroneously exit
3405  		 * connection startup slow start one packet too
3406  		 * quickly.  This is severely frowned upon behavior.
3407  		 */
3408  		if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3409  			flag |= FLAG_DATA_ACKED;
3410  		} else {
3411  			flag |= FLAG_SYN_ACKED;
3412  			tp->retrans_stamp = 0;
3413  		}
3414  
3415  		if (!fully_acked)
3416  			break;
3417  
3418  		tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una);
3419  
3420  		next = skb_rb_next(skb);
3421  		if (unlikely(skb == tp->retransmit_skb_hint))
3422  			tp->retransmit_skb_hint = NULL;
3423  		if (unlikely(skb == tp->lost_skb_hint))
3424  			tp->lost_skb_hint = NULL;
3425  		tcp_highest_sack_replace(sk, skb, next);
3426  		tcp_rtx_queue_unlink_and_free(skb, sk);
3427  	}
3428  
3429  	if (!skb)
3430  		tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3431  
3432  	if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3433  		tp->snd_up = tp->snd_una;
3434  
3435  	if (skb) {
3436  		tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una);
3437  		if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
3438  			flag |= FLAG_SACK_RENEGING;
3439  	}
3440  
3441  	if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3442  		seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt);
3443  		ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt);
3444  
3445  		if (pkts_acked == 1 && fully_acked && !prior_sacked &&
3446  		    (tp->snd_una - prior_snd_una) < tp->mss_cache &&
3447  		    sack->rate->prior_delivered + 1 == tp->delivered &&
3448  		    !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) {
3449  			/* Conservatively mark a delayed ACK. It's typically
3450  			 * from a lone runt packet over the round trip to
3451  			 * a receiver w/o out-of-order or CE events.
3452  			 */
3453  			flag |= FLAG_ACK_MAYBE_DELAYED;
3454  		}
3455  	}
3456  	if (sack->first_sackt) {
3457  		sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt);
3458  		ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt);
3459  	}
3460  	rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3461  					ca_rtt_us, sack->rate);
3462  
3463  	if (flag & FLAG_ACKED) {
3464  		flag |= FLAG_SET_XMIT_TIMER;  /* set TLP or RTO timer */
3465  		if (unlikely(icsk->icsk_mtup.probe_size &&
3466  			     !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3467  			tcp_mtup_probe_success(sk);
3468  		}
3469  
3470  		if (tcp_is_reno(tp)) {
3471  			tcp_remove_reno_sacks(sk, pkts_acked, ece_ack);
3472  
3473  			/* If any of the cumulatively ACKed segments was
3474  			 * retransmitted, non-SACK case cannot confirm that
3475  			 * progress was due to original transmission due to
3476  			 * lack of TCPCB_SACKED_ACKED bits even if some of
3477  			 * the packets may have been never retransmitted.
3478  			 */
3479  			if (flag & FLAG_RETRANS_DATA_ACKED)
3480  				flag &= ~FLAG_ORIG_SACK_ACKED;
3481  		} else {
3482  			int delta;
3483  
3484  			/* Non-retransmitted hole got filled? That's reordering */
3485  			if (before(reord, prior_fack))
3486  				tcp_check_sack_reordering(sk, reord, 0);
3487  
3488  			delta = prior_sacked - tp->sacked_out;
3489  			tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3490  		}
3491  	} else if (skb && rtt_update && sack_rtt_us >= 0 &&
3492  		   sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp,
3493  						    tcp_skb_timestamp_us(skb))) {
3494  		/* Do not re-arm RTO if the sack RTT is measured from data sent
3495  		 * after when the head was last (re)transmitted. Otherwise the
3496  		 * timeout may continue to extend in loss recovery.
3497  		 */
3498  		flag |= FLAG_SET_XMIT_TIMER;  /* set TLP or RTO timer */
3499  	}
3500  
3501  	if (icsk->icsk_ca_ops->pkts_acked) {
3502  		struct ack_sample sample = { .pkts_acked = pkts_acked,
3503  					     .rtt_us = sack->rate->rtt_us };
3504  
3505  		sample.in_flight = tp->mss_cache *
3506  			(tp->delivered - sack->rate->prior_delivered);
3507  		icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3508  	}
3509  
3510  #if FASTRETRANS_DEBUG > 0
3511  	WARN_ON((int)tp->sacked_out < 0);
3512  	WARN_ON((int)tp->lost_out < 0);
3513  	WARN_ON((int)tp->retrans_out < 0);
3514  	if (!tp->packets_out && tcp_is_sack(tp)) {
3515  		icsk = inet_csk(sk);
3516  		if (tp->lost_out) {
3517  			pr_debug("Leak l=%u %d\n",
3518  				 tp->lost_out, icsk->icsk_ca_state);
3519  			tp->lost_out = 0;
3520  		}
3521  		if (tp->sacked_out) {
3522  			pr_debug("Leak s=%u %d\n",
3523  				 tp->sacked_out, icsk->icsk_ca_state);
3524  			tp->sacked_out = 0;
3525  		}
3526  		if (tp->retrans_out) {
3527  			pr_debug("Leak r=%u %d\n",
3528  				 tp->retrans_out, icsk->icsk_ca_state);
3529  			tp->retrans_out = 0;
3530  		}
3531  	}
3532  #endif
3533  	return flag;
3534  }
3535  
tcp_ack_probe(struct sock * sk)3536  static void tcp_ack_probe(struct sock *sk)
3537  {
3538  	struct inet_connection_sock *icsk = inet_csk(sk);
3539  	struct sk_buff *head = tcp_send_head(sk);
3540  	const struct tcp_sock *tp = tcp_sk(sk);
3541  
3542  	/* Was it a usable window open? */
3543  	if (!head)
3544  		return;
3545  	if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) {
3546  		icsk->icsk_backoff = 0;
3547  		icsk->icsk_probes_tstamp = 0;
3548  		inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3549  		/* Socket must be waked up by subsequent tcp_data_snd_check().
3550  		 * This function is not for random using!
3551  		 */
3552  	} else {
3553  		unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3554  
3555  		when = tcp_clamp_probe0_to_user_timeout(sk, when);
3556  		tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, when, TCP_RTO_MAX);
3557  	}
3558  }
3559  
tcp_ack_is_dubious(const struct sock * sk,const int flag)3560  static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3561  {
3562  	return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3563  		inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3564  }
3565  
3566  /* Decide wheather to run the increase function of congestion control. */
tcp_may_raise_cwnd(const struct sock * sk,const int flag)3567  static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3568  {
3569  	/* If reordering is high then always grow cwnd whenever data is
3570  	 * delivered regardless of its ordering. Otherwise stay conservative
3571  	 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3572  	 * new SACK or ECE mark may first advance cwnd here and later reduce
3573  	 * cwnd in tcp_fastretrans_alert() based on more states.
3574  	 */
3575  	if (tcp_sk(sk)->reordering >
3576  	    READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reordering))
3577  		return flag & FLAG_FORWARD_PROGRESS;
3578  
3579  	return flag & FLAG_DATA_ACKED;
3580  }
3581  
3582  /* The "ultimate" congestion control function that aims to replace the rigid
3583   * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3584   * It's called toward the end of processing an ACK with precise rate
3585   * information. All transmission or retransmission are delayed afterwards.
3586   */
tcp_cong_control(struct sock * sk,u32 ack,u32 acked_sacked,int flag,const struct rate_sample * rs)3587  static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3588  			     int flag, const struct rate_sample *rs)
3589  {
3590  	const struct inet_connection_sock *icsk = inet_csk(sk);
3591  
3592  	if (icsk->icsk_ca_ops->cong_control) {
3593  		icsk->icsk_ca_ops->cong_control(sk, rs);
3594  		return;
3595  	}
3596  
3597  	if (tcp_in_cwnd_reduction(sk)) {
3598  		/* Reduce cwnd if state mandates */
3599  		tcp_cwnd_reduction(sk, acked_sacked, rs->losses, flag);
3600  	} else if (tcp_may_raise_cwnd(sk, flag)) {
3601  		/* Advance cwnd if state allows */
3602  		tcp_cong_avoid(sk, ack, acked_sacked);
3603  	}
3604  	tcp_update_pacing_rate(sk);
3605  }
3606  
3607  /* Check that window update is acceptable.
3608   * The function assumes that snd_una<=ack<=snd_next.
3609   */
tcp_may_update_window(const struct tcp_sock * tp,const u32 ack,const u32 ack_seq,const u32 nwin)3610  static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3611  					const u32 ack, const u32 ack_seq,
3612  					const u32 nwin)
3613  {
3614  	return	after(ack, tp->snd_una) ||
3615  		after(ack_seq, tp->snd_wl1) ||
3616  		(ack_seq == tp->snd_wl1 && (nwin > tp->snd_wnd || !nwin));
3617  }
3618  
3619  /* If we update tp->snd_una, also update tp->bytes_acked */
tcp_snd_una_update(struct tcp_sock * tp,u32 ack)3620  static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3621  {
3622  	u32 delta = ack - tp->snd_una;
3623  
3624  	sock_owned_by_me((struct sock *)tp);
3625  	tp->bytes_acked += delta;
3626  	tp->snd_una = ack;
3627  }
3628  
3629  /* If we update tp->rcv_nxt, also update tp->bytes_received */
tcp_rcv_nxt_update(struct tcp_sock * tp,u32 seq)3630  static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3631  {
3632  	u32 delta = seq - tp->rcv_nxt;
3633  
3634  	sock_owned_by_me((struct sock *)tp);
3635  	tp->bytes_received += delta;
3636  	WRITE_ONCE(tp->rcv_nxt, seq);
3637  }
3638  
3639  /* Update our send window.
3640   *
3641   * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3642   * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3643   */
tcp_ack_update_window(struct sock * sk,const struct sk_buff * skb,u32 ack,u32 ack_seq)3644  static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3645  				 u32 ack_seq)
3646  {
3647  	struct tcp_sock *tp = tcp_sk(sk);
3648  	int flag = 0;
3649  	u32 nwin = ntohs(tcp_hdr(skb)->window);
3650  
3651  	if (likely(!tcp_hdr(skb)->syn))
3652  		nwin <<= tp->rx_opt.snd_wscale;
3653  
3654  	if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3655  		flag |= FLAG_WIN_UPDATE;
3656  		tcp_update_wl(tp, ack_seq);
3657  
3658  		if (tp->snd_wnd != nwin) {
3659  			tp->snd_wnd = nwin;
3660  
3661  			/* Note, it is the only place, where
3662  			 * fast path is recovered for sending TCP.
3663  			 */
3664  			tp->pred_flags = 0;
3665  			tcp_fast_path_check(sk);
3666  
3667  			if (!tcp_write_queue_empty(sk))
3668  				tcp_slow_start_after_idle_check(sk);
3669  
3670  			if (nwin > tp->max_window) {
3671  				tp->max_window = nwin;
3672  				tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3673  			}
3674  		}
3675  	}
3676  
3677  	tcp_snd_una_update(tp, ack);
3678  
3679  	return flag;
3680  }
3681  
__tcp_oow_rate_limited(struct net * net,int mib_idx,u32 * last_oow_ack_time)3682  static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3683  				   u32 *last_oow_ack_time)
3684  {
3685  	/* Paired with the WRITE_ONCE() in this function. */
3686  	u32 val = READ_ONCE(*last_oow_ack_time);
3687  
3688  	if (val) {
3689  		s32 elapsed = (s32)(tcp_jiffies32 - val);
3690  
3691  		if (0 <= elapsed &&
3692  		    elapsed < READ_ONCE(net->ipv4.sysctl_tcp_invalid_ratelimit)) {
3693  			NET_INC_STATS(net, mib_idx);
3694  			return true;	/* rate-limited: don't send yet! */
3695  		}
3696  	}
3697  
3698  	/* Paired with the prior READ_ONCE() and with itself,
3699  	 * as we might be lockless.
3700  	 */
3701  	WRITE_ONCE(*last_oow_ack_time, tcp_jiffies32);
3702  
3703  	return false;	/* not rate-limited: go ahead, send dupack now! */
3704  }
3705  
3706  /* Return true if we're currently rate-limiting out-of-window ACKs and
3707   * thus shouldn't send a dupack right now. We rate-limit dupacks in
3708   * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3709   * attacks that send repeated SYNs or ACKs for the same connection. To
3710   * do this, we do not send a duplicate SYNACK or ACK if the remote
3711   * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3712   */
tcp_oow_rate_limited(struct net * net,const struct sk_buff * skb,int mib_idx,u32 * last_oow_ack_time)3713  bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3714  			  int mib_idx, u32 *last_oow_ack_time)
3715  {
3716  	/* Data packets without SYNs are not likely part of an ACK loop. */
3717  	if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3718  	    !tcp_hdr(skb)->syn)
3719  		return false;
3720  
3721  	return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3722  }
3723  
3724  /* RFC 5961 7 [ACK Throttling] */
tcp_send_challenge_ack(struct sock * sk)3725  static void tcp_send_challenge_ack(struct sock *sk)
3726  {
3727  	struct tcp_sock *tp = tcp_sk(sk);
3728  	struct net *net = sock_net(sk);
3729  	u32 count, now, ack_limit;
3730  
3731  	/* First check our per-socket dupack rate limit. */
3732  	if (__tcp_oow_rate_limited(net,
3733  				   LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3734  				   &tp->last_oow_ack_time))
3735  		return;
3736  
3737  	ack_limit = READ_ONCE(net->ipv4.sysctl_tcp_challenge_ack_limit);
3738  	if (ack_limit == INT_MAX)
3739  		goto send_ack;
3740  
3741  	/* Then check host-wide RFC 5961 rate limit. */
3742  	now = jiffies / HZ;
3743  	if (now != READ_ONCE(net->ipv4.tcp_challenge_timestamp)) {
3744  		u32 half = (ack_limit + 1) >> 1;
3745  
3746  		WRITE_ONCE(net->ipv4.tcp_challenge_timestamp, now);
3747  		WRITE_ONCE(net->ipv4.tcp_challenge_count,
3748  			   get_random_u32_inclusive(half, ack_limit + half - 1));
3749  	}
3750  	count = READ_ONCE(net->ipv4.tcp_challenge_count);
3751  	if (count > 0) {
3752  		WRITE_ONCE(net->ipv4.tcp_challenge_count, count - 1);
3753  send_ack:
3754  		NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK);
3755  		tcp_send_ack(sk);
3756  	}
3757  }
3758  
tcp_store_ts_recent(struct tcp_sock * tp)3759  static void tcp_store_ts_recent(struct tcp_sock *tp)
3760  {
3761  	tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3762  	tp->rx_opt.ts_recent_stamp = ktime_get_seconds();
3763  }
3764  
tcp_replace_ts_recent(struct tcp_sock * tp,u32 seq)3765  static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3766  {
3767  	if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3768  		/* PAWS bug workaround wrt. ACK frames, the PAWS discard
3769  		 * extra check below makes sure this can only happen
3770  		 * for pure ACK frames.  -DaveM
3771  		 *
3772  		 * Not only, also it occurs for expired timestamps.
3773  		 */
3774  
3775  		if (tcp_paws_check(&tp->rx_opt, 0))
3776  			tcp_store_ts_recent(tp);
3777  	}
3778  }
3779  
3780  /* This routine deals with acks during a TLP episode and ends an episode by
3781   * resetting tlp_high_seq. Ref: TLP algorithm in draft-ietf-tcpm-rack
3782   */
tcp_process_tlp_ack(struct sock * sk,u32 ack,int flag)3783  static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3784  {
3785  	struct tcp_sock *tp = tcp_sk(sk);
3786  
3787  	if (before(ack, tp->tlp_high_seq))
3788  		return;
3789  
3790  	if (!tp->tlp_retrans) {
3791  		/* TLP of new data has been acknowledged */
3792  		tp->tlp_high_seq = 0;
3793  	} else if (flag & FLAG_DSACK_TLP) {
3794  		/* This DSACK means original and TLP probe arrived; no loss */
3795  		tp->tlp_high_seq = 0;
3796  	} else if (after(ack, tp->tlp_high_seq)) {
3797  		/* ACK advances: there was a loss, so reduce cwnd. Reset
3798  		 * tlp_high_seq in tcp_init_cwnd_reduction()
3799  		 */
3800  		tcp_init_cwnd_reduction(sk);
3801  		tcp_set_ca_state(sk, TCP_CA_CWR);
3802  		tcp_end_cwnd_reduction(sk);
3803  		tcp_try_keep_open(sk);
3804  		NET_INC_STATS(sock_net(sk),
3805  				LINUX_MIB_TCPLOSSPROBERECOVERY);
3806  	} else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3807  			     FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3808  		/* Pure dupack: original and TLP probe arrived; no loss */
3809  		tp->tlp_high_seq = 0;
3810  	}
3811  }
3812  
tcp_in_ack_event(struct sock * sk,u32 flags)3813  static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3814  {
3815  	const struct inet_connection_sock *icsk = inet_csk(sk);
3816  
3817  	if (icsk->icsk_ca_ops->in_ack_event)
3818  		icsk->icsk_ca_ops->in_ack_event(sk, flags);
3819  }
3820  
3821  /* Congestion control has updated the cwnd already. So if we're in
3822   * loss recovery then now we do any new sends (for FRTO) or
3823   * retransmits (for CA_Loss or CA_recovery) that make sense.
3824   */
tcp_xmit_recovery(struct sock * sk,int rexmit)3825  static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3826  {
3827  	struct tcp_sock *tp = tcp_sk(sk);
3828  
3829  	if (rexmit == REXMIT_NONE || sk->sk_state == TCP_SYN_SENT)
3830  		return;
3831  
3832  	if (unlikely(rexmit == REXMIT_NEW)) {
3833  		__tcp_push_pending_frames(sk, tcp_current_mss(sk),
3834  					  TCP_NAGLE_OFF);
3835  		if (after(tp->snd_nxt, tp->high_seq))
3836  			return;
3837  		tp->frto = 0;
3838  	}
3839  	tcp_xmit_retransmit_queue(sk);
3840  }
3841  
3842  /* Returns the number of packets newly acked or sacked by the current ACK */
tcp_newly_delivered(struct sock * sk,u32 prior_delivered,int flag)3843  static u32 tcp_newly_delivered(struct sock *sk, u32 prior_delivered, int flag)
3844  {
3845  	const struct net *net = sock_net(sk);
3846  	struct tcp_sock *tp = tcp_sk(sk);
3847  	u32 delivered;
3848  
3849  	delivered = tp->delivered - prior_delivered;
3850  	NET_ADD_STATS(net, LINUX_MIB_TCPDELIVERED, delivered);
3851  	if (flag & FLAG_ECE)
3852  		NET_ADD_STATS(net, LINUX_MIB_TCPDELIVEREDCE, delivered);
3853  
3854  	return delivered;
3855  }
3856  
3857  /* This routine deals with incoming acks, but not outgoing ones. */
tcp_ack(struct sock * sk,const struct sk_buff * skb,int flag)3858  static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3859  {
3860  	struct inet_connection_sock *icsk = inet_csk(sk);
3861  	struct tcp_sock *tp = tcp_sk(sk);
3862  	struct tcp_sacktag_state sack_state;
3863  	struct rate_sample rs = { .prior_delivered = 0 };
3864  	u32 prior_snd_una = tp->snd_una;
3865  	bool is_sack_reneg = tp->is_sack_reneg;
3866  	u32 ack_seq = TCP_SKB_CB(skb)->seq;
3867  	u32 ack = TCP_SKB_CB(skb)->ack_seq;
3868  	int num_dupack = 0;
3869  	int prior_packets = tp->packets_out;
3870  	u32 delivered = tp->delivered;
3871  	u32 lost = tp->lost;
3872  	int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3873  	u32 prior_fack;
3874  
3875  	sack_state.first_sackt = 0;
3876  	sack_state.rate = &rs;
3877  	sack_state.sack_delivered = 0;
3878  
3879  	/* We very likely will need to access rtx queue. */
3880  	prefetch(sk->tcp_rtx_queue.rb_node);
3881  
3882  	/* If the ack is older than previous acks
3883  	 * then we can probably ignore it.
3884  	 */
3885  	if (before(ack, prior_snd_una)) {
3886  		u32 max_window;
3887  
3888  		/* do not accept ACK for bytes we never sent. */
3889  		max_window = min_t(u64, tp->max_window, tp->bytes_acked);
3890  		/* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3891  		if (before(ack, prior_snd_una - max_window)) {
3892  			if (!(flag & FLAG_NO_CHALLENGE_ACK))
3893  				tcp_send_challenge_ack(sk);
3894  			return -SKB_DROP_REASON_TCP_TOO_OLD_ACK;
3895  		}
3896  		goto old_ack;
3897  	}
3898  
3899  	/* If the ack includes data we haven't sent yet, discard
3900  	 * this segment (RFC793 Section 3.9).
3901  	 */
3902  	if (after(ack, tp->snd_nxt))
3903  		return -SKB_DROP_REASON_TCP_ACK_UNSENT_DATA;
3904  
3905  	if (after(ack, prior_snd_una)) {
3906  		flag |= FLAG_SND_UNA_ADVANCED;
3907  		icsk->icsk_retransmits = 0;
3908  
3909  #if IS_ENABLED(CONFIG_TLS_DEVICE)
3910  		if (static_branch_unlikely(&clean_acked_data_enabled.key))
3911  			if (icsk->icsk_clean_acked)
3912  				icsk->icsk_clean_acked(sk, ack);
3913  #endif
3914  	}
3915  
3916  	prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una;
3917  	rs.prior_in_flight = tcp_packets_in_flight(tp);
3918  
3919  	/* ts_recent update must be made after we are sure that the packet
3920  	 * is in window.
3921  	 */
3922  	if (flag & FLAG_UPDATE_TS_RECENT)
3923  		tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3924  
3925  	if ((flag & (FLAG_SLOWPATH | FLAG_SND_UNA_ADVANCED)) ==
3926  	    FLAG_SND_UNA_ADVANCED) {
3927  		/* Window is constant, pure forward advance.
3928  		 * No more checks are required.
3929  		 * Note, we use the fact that SND.UNA>=SND.WL2.
3930  		 */
3931  		tcp_update_wl(tp, ack_seq);
3932  		tcp_snd_una_update(tp, ack);
3933  		flag |= FLAG_WIN_UPDATE;
3934  
3935  		tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3936  
3937  		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3938  	} else {
3939  		u32 ack_ev_flags = CA_ACK_SLOWPATH;
3940  
3941  		if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3942  			flag |= FLAG_DATA;
3943  		else
3944  			NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3945  
3946  		flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3947  
3948  		if (TCP_SKB_CB(skb)->sacked)
3949  			flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3950  							&sack_state);
3951  
3952  		if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3953  			flag |= FLAG_ECE;
3954  			ack_ev_flags |= CA_ACK_ECE;
3955  		}
3956  
3957  		if (sack_state.sack_delivered)
3958  			tcp_count_delivered(tp, sack_state.sack_delivered,
3959  					    flag & FLAG_ECE);
3960  
3961  		if (flag & FLAG_WIN_UPDATE)
3962  			ack_ev_flags |= CA_ACK_WIN_UPDATE;
3963  
3964  		tcp_in_ack_event(sk, ack_ev_flags);
3965  	}
3966  
3967  	/* This is a deviation from RFC3168 since it states that:
3968  	 * "When the TCP data sender is ready to set the CWR bit after reducing
3969  	 * the congestion window, it SHOULD set the CWR bit only on the first
3970  	 * new data packet that it transmits."
3971  	 * We accept CWR on pure ACKs to be more robust
3972  	 * with widely-deployed TCP implementations that do this.
3973  	 */
3974  	tcp_ecn_accept_cwr(sk, skb);
3975  
3976  	/* We passed data and got it acked, remove any soft error
3977  	 * log. Something worked...
3978  	 */
3979  	WRITE_ONCE(sk->sk_err_soft, 0);
3980  	icsk->icsk_probes_out = 0;
3981  	tp->rcv_tstamp = tcp_jiffies32;
3982  	if (!prior_packets)
3983  		goto no_queue;
3984  
3985  	/* See if we can take anything off of the retransmit queue. */
3986  	flag |= tcp_clean_rtx_queue(sk, skb, prior_fack, prior_snd_una,
3987  				    &sack_state, flag & FLAG_ECE);
3988  
3989  	tcp_rack_update_reo_wnd(sk, &rs);
3990  
3991  	if (tp->tlp_high_seq)
3992  		tcp_process_tlp_ack(sk, ack, flag);
3993  
3994  	if (tcp_ack_is_dubious(sk, flag)) {
3995  		if (!(flag & (FLAG_SND_UNA_ADVANCED |
3996  			      FLAG_NOT_DUP | FLAG_DSACKING_ACK))) {
3997  			num_dupack = 1;
3998  			/* Consider if pure acks were aggregated in tcp_add_backlog() */
3999  			if (!(flag & FLAG_DATA))
4000  				num_dupack = max_t(u16, 1, skb_shinfo(skb)->gso_segs);
4001  		}
4002  		tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
4003  				      &rexmit);
4004  	}
4005  
4006  	/* If needed, reset TLP/RTO timer when RACK doesn't set. */
4007  	if (flag & FLAG_SET_XMIT_TIMER)
4008  		tcp_set_xmit_timer(sk);
4009  
4010  	if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
4011  		sk_dst_confirm(sk);
4012  
4013  	delivered = tcp_newly_delivered(sk, delivered, flag);
4014  	lost = tp->lost - lost;			/* freshly marked lost */
4015  	rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED);
4016  	tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate);
4017  	tcp_cong_control(sk, ack, delivered, flag, sack_state.rate);
4018  	tcp_xmit_recovery(sk, rexmit);
4019  	return 1;
4020  
4021  no_queue:
4022  	/* If data was DSACKed, see if we can undo a cwnd reduction. */
4023  	if (flag & FLAG_DSACKING_ACK) {
4024  		tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
4025  				      &rexmit);
4026  		tcp_newly_delivered(sk, delivered, flag);
4027  	}
4028  	/* If this ack opens up a zero window, clear backoff.  It was
4029  	 * being used to time the probes, and is probably far higher than
4030  	 * it needs to be for normal retransmission.
4031  	 */
4032  	tcp_ack_probe(sk);
4033  
4034  	if (tp->tlp_high_seq)
4035  		tcp_process_tlp_ack(sk, ack, flag);
4036  	return 1;
4037  
4038  old_ack:
4039  	/* If data was SACKed, tag it and see if we should send more data.
4040  	 * If data was DSACKed, see if we can undo a cwnd reduction.
4041  	 */
4042  	if (TCP_SKB_CB(skb)->sacked) {
4043  		flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
4044  						&sack_state);
4045  		tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
4046  				      &rexmit);
4047  		tcp_newly_delivered(sk, delivered, flag);
4048  		tcp_xmit_recovery(sk, rexmit);
4049  	}
4050  
4051  	return 0;
4052  }
4053  
tcp_parse_fastopen_option(int len,const unsigned char * cookie,bool syn,struct tcp_fastopen_cookie * foc,bool exp_opt)4054  static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
4055  				      bool syn, struct tcp_fastopen_cookie *foc,
4056  				      bool exp_opt)
4057  {
4058  	/* Valid only in SYN or SYN-ACK with an even length.  */
4059  	if (!foc || !syn || len < 0 || (len & 1))
4060  		return;
4061  
4062  	if (len >= TCP_FASTOPEN_COOKIE_MIN &&
4063  	    len <= TCP_FASTOPEN_COOKIE_MAX)
4064  		memcpy(foc->val, cookie, len);
4065  	else if (len != 0)
4066  		len = -1;
4067  	foc->len = len;
4068  	foc->exp = exp_opt;
4069  }
4070  
smc_parse_options(const struct tcphdr * th,struct tcp_options_received * opt_rx,const unsigned char * ptr,int opsize)4071  static bool smc_parse_options(const struct tcphdr *th,
4072  			      struct tcp_options_received *opt_rx,
4073  			      const unsigned char *ptr,
4074  			      int opsize)
4075  {
4076  #if IS_ENABLED(CONFIG_SMC)
4077  	if (static_branch_unlikely(&tcp_have_smc)) {
4078  		if (th->syn && !(opsize & 1) &&
4079  		    opsize >= TCPOLEN_EXP_SMC_BASE &&
4080  		    get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC) {
4081  			opt_rx->smc_ok = 1;
4082  			return true;
4083  		}
4084  	}
4085  #endif
4086  	return false;
4087  }
4088  
4089  /* Try to parse the MSS option from the TCP header. Return 0 on failure, clamped
4090   * value on success.
4091   */
tcp_parse_mss_option(const struct tcphdr * th,u16 user_mss)4092  u16 tcp_parse_mss_option(const struct tcphdr *th, u16 user_mss)
4093  {
4094  	const unsigned char *ptr = (const unsigned char *)(th + 1);
4095  	int length = (th->doff * 4) - sizeof(struct tcphdr);
4096  	u16 mss = 0;
4097  
4098  	while (length > 0) {
4099  		int opcode = *ptr++;
4100  		int opsize;
4101  
4102  		switch (opcode) {
4103  		case TCPOPT_EOL:
4104  			return mss;
4105  		case TCPOPT_NOP:	/* Ref: RFC 793 section 3.1 */
4106  			length--;
4107  			continue;
4108  		default:
4109  			if (length < 2)
4110  				return mss;
4111  			opsize = *ptr++;
4112  			if (opsize < 2) /* "silly options" */
4113  				return mss;
4114  			if (opsize > length)
4115  				return mss;	/* fail on partial options */
4116  			if (opcode == TCPOPT_MSS && opsize == TCPOLEN_MSS) {
4117  				u16 in_mss = get_unaligned_be16(ptr);
4118  
4119  				if (in_mss) {
4120  					if (user_mss && user_mss < in_mss)
4121  						in_mss = user_mss;
4122  					mss = in_mss;
4123  				}
4124  			}
4125  			ptr += opsize - 2;
4126  			length -= opsize;
4127  		}
4128  	}
4129  	return mss;
4130  }
4131  EXPORT_SYMBOL_GPL(tcp_parse_mss_option);
4132  
4133  /* Look for tcp options. Normally only called on SYN and SYNACK packets.
4134   * But, this can also be called on packets in the established flow when
4135   * the fast version below fails.
4136   */
tcp_parse_options(const struct net * net,const struct sk_buff * skb,struct tcp_options_received * opt_rx,int estab,struct tcp_fastopen_cookie * foc)4137  void tcp_parse_options(const struct net *net,
4138  		       const struct sk_buff *skb,
4139  		       struct tcp_options_received *opt_rx, int estab,
4140  		       struct tcp_fastopen_cookie *foc)
4141  {
4142  	const unsigned char *ptr;
4143  	const struct tcphdr *th = tcp_hdr(skb);
4144  	int length = (th->doff * 4) - sizeof(struct tcphdr);
4145  
4146  	ptr = (const unsigned char *)(th + 1);
4147  	opt_rx->saw_tstamp = 0;
4148  	opt_rx->saw_unknown = 0;
4149  
4150  	while (length > 0) {
4151  		int opcode = *ptr++;
4152  		int opsize;
4153  
4154  		switch (opcode) {
4155  		case TCPOPT_EOL:
4156  			return;
4157  		case TCPOPT_NOP:	/* Ref: RFC 793 section 3.1 */
4158  			length--;
4159  			continue;
4160  		default:
4161  			if (length < 2)
4162  				return;
4163  			opsize = *ptr++;
4164  			if (opsize < 2) /* "silly options" */
4165  				return;
4166  			if (opsize > length)
4167  				return;	/* don't parse partial options */
4168  			switch (opcode) {
4169  			case TCPOPT_MSS:
4170  				if (opsize == TCPOLEN_MSS && th->syn && !estab) {
4171  					u16 in_mss = get_unaligned_be16(ptr);
4172  					if (in_mss) {
4173  						if (opt_rx->user_mss &&
4174  						    opt_rx->user_mss < in_mss)
4175  							in_mss = opt_rx->user_mss;
4176  						opt_rx->mss_clamp = in_mss;
4177  					}
4178  				}
4179  				break;
4180  			case TCPOPT_WINDOW:
4181  				if (opsize == TCPOLEN_WINDOW && th->syn &&
4182  				    !estab && READ_ONCE(net->ipv4.sysctl_tcp_window_scaling)) {
4183  					__u8 snd_wscale = *(__u8 *)ptr;
4184  					opt_rx->wscale_ok = 1;
4185  					if (snd_wscale > TCP_MAX_WSCALE) {
4186  						net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
4187  								     __func__,
4188  								     snd_wscale,
4189  								     TCP_MAX_WSCALE);
4190  						snd_wscale = TCP_MAX_WSCALE;
4191  					}
4192  					opt_rx->snd_wscale = snd_wscale;
4193  				}
4194  				break;
4195  			case TCPOPT_TIMESTAMP:
4196  				if ((opsize == TCPOLEN_TIMESTAMP) &&
4197  				    ((estab && opt_rx->tstamp_ok) ||
4198  				     (!estab && READ_ONCE(net->ipv4.sysctl_tcp_timestamps)))) {
4199  					opt_rx->saw_tstamp = 1;
4200  					opt_rx->rcv_tsval = get_unaligned_be32(ptr);
4201  					opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
4202  				}
4203  				break;
4204  			case TCPOPT_SACK_PERM:
4205  				if (opsize == TCPOLEN_SACK_PERM && th->syn &&
4206  				    !estab && READ_ONCE(net->ipv4.sysctl_tcp_sack)) {
4207  					opt_rx->sack_ok = TCP_SACK_SEEN;
4208  					tcp_sack_reset(opt_rx);
4209  				}
4210  				break;
4211  
4212  			case TCPOPT_SACK:
4213  				if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
4214  				   !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
4215  				   opt_rx->sack_ok) {
4216  					TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
4217  				}
4218  				break;
4219  #ifdef CONFIG_TCP_MD5SIG
4220  			case TCPOPT_MD5SIG:
4221  				/* The MD5 Hash has already been
4222  				 * checked (see tcp_v{4,6}_rcv()).
4223  				 */
4224  				break;
4225  #endif
4226  			case TCPOPT_FASTOPEN:
4227  				tcp_parse_fastopen_option(
4228  					opsize - TCPOLEN_FASTOPEN_BASE,
4229  					ptr, th->syn, foc, false);
4230  				break;
4231  
4232  			case TCPOPT_EXP:
4233  				/* Fast Open option shares code 254 using a
4234  				 * 16 bits magic number.
4235  				 */
4236  				if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
4237  				    get_unaligned_be16(ptr) ==
4238  				    TCPOPT_FASTOPEN_MAGIC) {
4239  					tcp_parse_fastopen_option(opsize -
4240  						TCPOLEN_EXP_FASTOPEN_BASE,
4241  						ptr + 2, th->syn, foc, true);
4242  					break;
4243  				}
4244  
4245  				if (smc_parse_options(th, opt_rx, ptr, opsize))
4246  					break;
4247  
4248  				opt_rx->saw_unknown = 1;
4249  				break;
4250  
4251  			default:
4252  				opt_rx->saw_unknown = 1;
4253  			}
4254  			ptr += opsize-2;
4255  			length -= opsize;
4256  		}
4257  	}
4258  }
4259  EXPORT_SYMBOL(tcp_parse_options);
4260  
tcp_parse_aligned_timestamp(struct tcp_sock * tp,const struct tcphdr * th)4261  static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
4262  {
4263  	const __be32 *ptr = (const __be32 *)(th + 1);
4264  
4265  	if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4266  			  | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
4267  		tp->rx_opt.saw_tstamp = 1;
4268  		++ptr;
4269  		tp->rx_opt.rcv_tsval = ntohl(*ptr);
4270  		++ptr;
4271  		if (*ptr)
4272  			tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
4273  		else
4274  			tp->rx_opt.rcv_tsecr = 0;
4275  		return true;
4276  	}
4277  	return false;
4278  }
4279  
4280  /* Fast parse options. This hopes to only see timestamps.
4281   * If it is wrong it falls back on tcp_parse_options().
4282   */
tcp_fast_parse_options(const struct net * net,const struct sk_buff * skb,const struct tcphdr * th,struct tcp_sock * tp)4283  static bool tcp_fast_parse_options(const struct net *net,
4284  				   const struct sk_buff *skb,
4285  				   const struct tcphdr *th, struct tcp_sock *tp)
4286  {
4287  	/* In the spirit of fast parsing, compare doff directly to constant
4288  	 * values.  Because equality is used, short doff can be ignored here.
4289  	 */
4290  	if (th->doff == (sizeof(*th) / 4)) {
4291  		tp->rx_opt.saw_tstamp = 0;
4292  		return false;
4293  	} else if (tp->rx_opt.tstamp_ok &&
4294  		   th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
4295  		if (tcp_parse_aligned_timestamp(tp, th))
4296  			return true;
4297  	}
4298  
4299  	tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL);
4300  	if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
4301  		tp->rx_opt.rcv_tsecr -= tp->tsoffset;
4302  
4303  	return true;
4304  }
4305  
4306  #ifdef CONFIG_TCP_MD5SIG
4307  /*
4308   * Parse MD5 Signature option
4309   */
tcp_parse_md5sig_option(const struct tcphdr * th)4310  const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
4311  {
4312  	int length = (th->doff << 2) - sizeof(*th);
4313  	const u8 *ptr = (const u8 *)(th + 1);
4314  
4315  	/* If not enough data remaining, we can short cut */
4316  	while (length >= TCPOLEN_MD5SIG) {
4317  		int opcode = *ptr++;
4318  		int opsize;
4319  
4320  		switch (opcode) {
4321  		case TCPOPT_EOL:
4322  			return NULL;
4323  		case TCPOPT_NOP:
4324  			length--;
4325  			continue;
4326  		default:
4327  			opsize = *ptr++;
4328  			if (opsize < 2 || opsize > length)
4329  				return NULL;
4330  			if (opcode == TCPOPT_MD5SIG)
4331  				return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
4332  		}
4333  		ptr += opsize - 2;
4334  		length -= opsize;
4335  	}
4336  	return NULL;
4337  }
4338  EXPORT_SYMBOL(tcp_parse_md5sig_option);
4339  #endif
4340  
4341  /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
4342   *
4343   * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
4344   * it can pass through stack. So, the following predicate verifies that
4345   * this segment is not used for anything but congestion avoidance or
4346   * fast retransmit. Moreover, we even are able to eliminate most of such
4347   * second order effects, if we apply some small "replay" window (~RTO)
4348   * to timestamp space.
4349   *
4350   * All these measures still do not guarantee that we reject wrapped ACKs
4351   * on networks with high bandwidth, when sequence space is recycled fastly,
4352   * but it guarantees that such events will be very rare and do not affect
4353   * connection seriously. This doesn't look nice, but alas, PAWS is really
4354   * buggy extension.
4355   *
4356   * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
4357   * states that events when retransmit arrives after original data are rare.
4358   * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
4359   * the biggest problem on large power networks even with minor reordering.
4360   * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
4361   * up to bandwidth of 18Gigabit/sec. 8) ]
4362   */
4363  
tcp_disordered_ack(const struct sock * sk,const struct sk_buff * skb)4364  static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
4365  {
4366  	const struct tcp_sock *tp = tcp_sk(sk);
4367  	const struct tcphdr *th = tcp_hdr(skb);
4368  	u32 seq = TCP_SKB_CB(skb)->seq;
4369  	u32 ack = TCP_SKB_CB(skb)->ack_seq;
4370  
4371  	return (/* 1. Pure ACK with correct sequence number. */
4372  		(th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
4373  
4374  		/* 2. ... and duplicate ACK. */
4375  		ack == tp->snd_una &&
4376  
4377  		/* 3. ... and does not update window. */
4378  		!tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4379  
4380  		/* 4. ... and sits in replay window. */
4381  		(s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
4382  }
4383  
tcp_paws_discard(const struct sock * sk,const struct sk_buff * skb)4384  static inline bool tcp_paws_discard(const struct sock *sk,
4385  				   const struct sk_buff *skb)
4386  {
4387  	const struct tcp_sock *tp = tcp_sk(sk);
4388  
4389  	return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4390  	       !tcp_disordered_ack(sk, skb);
4391  }
4392  
4393  /* Check segment sequence number for validity.
4394   *
4395   * Segment controls are considered valid, if the segment
4396   * fits to the window after truncation to the window. Acceptability
4397   * of data (and SYN, FIN, of course) is checked separately.
4398   * See tcp_data_queue(), for example.
4399   *
4400   * Also, controls (RST is main one) are accepted using RCV.WUP instead
4401   * of RCV.NXT. Peer still did not advance his SND.UNA when we
4402   * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4403   * (borrowed from freebsd)
4404   */
4405  
tcp_sequence(const struct tcp_sock * tp,u32 seq,u32 end_seq)4406  static enum skb_drop_reason tcp_sequence(const struct tcp_sock *tp,
4407  					 u32 seq, u32 end_seq)
4408  {
4409  	if (before(end_seq, tp->rcv_wup))
4410  		return SKB_DROP_REASON_TCP_OLD_SEQUENCE;
4411  
4412  	if (after(seq, tp->rcv_nxt + tcp_receive_window(tp)))
4413  		return SKB_DROP_REASON_TCP_INVALID_SEQUENCE;
4414  
4415  	return SKB_NOT_DROPPED_YET;
4416  }
4417  
4418  
tcp_done_with_error(struct sock * sk,int err)4419  void tcp_done_with_error(struct sock *sk, int err)
4420  {
4421  	/* This barrier is coupled with smp_rmb() in tcp_poll() */
4422  	WRITE_ONCE(sk->sk_err, err);
4423  	smp_wmb();
4424  
4425  	tcp_write_queue_purge(sk);
4426  	tcp_done(sk);
4427  
4428  	if (!sock_flag(sk, SOCK_DEAD))
4429  		sk_error_report(sk);
4430  }
4431  EXPORT_SYMBOL(tcp_done_with_error);
4432  
4433  /* When we get a reset we do this. */
tcp_reset(struct sock * sk,struct sk_buff * skb)4434  void tcp_reset(struct sock *sk, struct sk_buff *skb)
4435  {
4436  	int err;
4437  
4438  	trace_tcp_receive_reset(sk);
4439  
4440  	/* mptcp can't tell us to ignore reset pkts,
4441  	 * so just ignore the return value of mptcp_incoming_options().
4442  	 */
4443  	if (sk_is_mptcp(sk))
4444  		mptcp_incoming_options(sk, skb);
4445  
4446  	/* We want the right error as BSD sees it (and indeed as we do). */
4447  	switch (sk->sk_state) {
4448  	case TCP_SYN_SENT:
4449  		err = ECONNREFUSED;
4450  		break;
4451  	case TCP_CLOSE_WAIT:
4452  		err = EPIPE;
4453  		break;
4454  	case TCP_CLOSE:
4455  		return;
4456  	default:
4457  		err = ECONNRESET;
4458  	}
4459  	tcp_done_with_error(sk, err);
4460  }
4461  
4462  /*
4463   * 	Process the FIN bit. This now behaves as it is supposed to work
4464   *	and the FIN takes effect when it is validly part of sequence
4465   *	space. Not before when we get holes.
4466   *
4467   *	If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4468   *	(and thence onto LAST-ACK and finally, CLOSE, we never enter
4469   *	TIME-WAIT)
4470   *
4471   *	If we are in FINWAIT-1, a received FIN indicates simultaneous
4472   *	close and we go into CLOSING (and later onto TIME-WAIT)
4473   *
4474   *	If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4475   */
tcp_fin(struct sock * sk)4476  void tcp_fin(struct sock *sk)
4477  {
4478  	struct tcp_sock *tp = tcp_sk(sk);
4479  
4480  	inet_csk_schedule_ack(sk);
4481  
4482  	WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN);
4483  	sock_set_flag(sk, SOCK_DONE);
4484  
4485  	switch (sk->sk_state) {
4486  	case TCP_SYN_RECV:
4487  	case TCP_ESTABLISHED:
4488  		/* Move to CLOSE_WAIT */
4489  		tcp_set_state(sk, TCP_CLOSE_WAIT);
4490  		inet_csk_enter_pingpong_mode(sk);
4491  		break;
4492  
4493  	case TCP_CLOSE_WAIT:
4494  	case TCP_CLOSING:
4495  		/* Received a retransmission of the FIN, do
4496  		 * nothing.
4497  		 */
4498  		break;
4499  	case TCP_LAST_ACK:
4500  		/* RFC793: Remain in the LAST-ACK state. */
4501  		break;
4502  
4503  	case TCP_FIN_WAIT1:
4504  		/* This case occurs when a simultaneous close
4505  		 * happens, we must ack the received FIN and
4506  		 * enter the CLOSING state.
4507  		 */
4508  		tcp_send_ack(sk);
4509  		tcp_set_state(sk, TCP_CLOSING);
4510  		break;
4511  	case TCP_FIN_WAIT2:
4512  		/* Received a FIN -- send ACK and enter TIME_WAIT. */
4513  		tcp_send_ack(sk);
4514  		tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4515  		break;
4516  	default:
4517  		/* Only TCP_LISTEN and TCP_CLOSE are left, in these
4518  		 * cases we should never reach this piece of code.
4519  		 */
4520  		pr_err("%s: Impossible, sk->sk_state=%d\n",
4521  		       __func__, sk->sk_state);
4522  		break;
4523  	}
4524  
4525  	/* It _is_ possible, that we have something out-of-order _after_ FIN.
4526  	 * Probably, we should reset in this case. For now drop them.
4527  	 */
4528  	skb_rbtree_purge(&tp->out_of_order_queue);
4529  	if (tcp_is_sack(tp))
4530  		tcp_sack_reset(&tp->rx_opt);
4531  
4532  	if (!sock_flag(sk, SOCK_DEAD)) {
4533  		sk->sk_state_change(sk);
4534  
4535  		/* Do not send POLL_HUP for half duplex close. */
4536  		if (sk->sk_shutdown == SHUTDOWN_MASK ||
4537  		    sk->sk_state == TCP_CLOSE)
4538  			sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4539  		else
4540  			sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4541  	}
4542  }
4543  
tcp_sack_extend(struct tcp_sack_block * sp,u32 seq,u32 end_seq)4544  static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4545  				  u32 end_seq)
4546  {
4547  	if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4548  		if (before(seq, sp->start_seq))
4549  			sp->start_seq = seq;
4550  		if (after(end_seq, sp->end_seq))
4551  			sp->end_seq = end_seq;
4552  		return true;
4553  	}
4554  	return false;
4555  }
4556  
tcp_dsack_set(struct sock * sk,u32 seq,u32 end_seq)4557  static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4558  {
4559  	struct tcp_sock *tp = tcp_sk(sk);
4560  
4561  	if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) {
4562  		int mib_idx;
4563  
4564  		if (before(seq, tp->rcv_nxt))
4565  			mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4566  		else
4567  			mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4568  
4569  		NET_INC_STATS(sock_net(sk), mib_idx);
4570  
4571  		tp->rx_opt.dsack = 1;
4572  		tp->duplicate_sack[0].start_seq = seq;
4573  		tp->duplicate_sack[0].end_seq = end_seq;
4574  	}
4575  }
4576  
tcp_dsack_extend(struct sock * sk,u32 seq,u32 end_seq)4577  static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4578  {
4579  	struct tcp_sock *tp = tcp_sk(sk);
4580  
4581  	if (!tp->rx_opt.dsack)
4582  		tcp_dsack_set(sk, seq, end_seq);
4583  	else
4584  		tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4585  }
4586  
tcp_rcv_spurious_retrans(struct sock * sk,const struct sk_buff * skb)4587  static void tcp_rcv_spurious_retrans(struct sock *sk, const struct sk_buff *skb)
4588  {
4589  	/* When the ACK path fails or drops most ACKs, the sender would
4590  	 * timeout and spuriously retransmit the same segment repeatedly.
4591  	 * The receiver remembers and reflects via DSACKs. Leverage the
4592  	 * DSACK state and change the txhash to re-route speculatively.
4593  	 */
4594  	if (TCP_SKB_CB(skb)->seq == tcp_sk(sk)->duplicate_sack[0].start_seq &&
4595  	    sk_rethink_txhash(sk))
4596  		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDUPLICATEDATAREHASH);
4597  }
4598  
tcp_send_dupack(struct sock * sk,const struct sk_buff * skb)4599  static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4600  {
4601  	struct tcp_sock *tp = tcp_sk(sk);
4602  
4603  	if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4604  	    before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4605  		NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4606  		tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4607  
4608  		if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) {
4609  			u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4610  
4611  			tcp_rcv_spurious_retrans(sk, skb);
4612  			if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4613  				end_seq = tp->rcv_nxt;
4614  			tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4615  		}
4616  	}
4617  
4618  	tcp_send_ack(sk);
4619  }
4620  
4621  /* These routines update the SACK block as out-of-order packets arrive or
4622   * in-order packets close up the sequence space.
4623   */
tcp_sack_maybe_coalesce(struct tcp_sock * tp)4624  static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4625  {
4626  	int this_sack;
4627  	struct tcp_sack_block *sp = &tp->selective_acks[0];
4628  	struct tcp_sack_block *swalk = sp + 1;
4629  
4630  	/* See if the recent change to the first SACK eats into
4631  	 * or hits the sequence space of other SACK blocks, if so coalesce.
4632  	 */
4633  	for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4634  		if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4635  			int i;
4636  
4637  			/* Zap SWALK, by moving every further SACK up by one slot.
4638  			 * Decrease num_sacks.
4639  			 */
4640  			tp->rx_opt.num_sacks--;
4641  			for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4642  				sp[i] = sp[i + 1];
4643  			continue;
4644  		}
4645  		this_sack++;
4646  		swalk++;
4647  	}
4648  }
4649  
tcp_sack_compress_send_ack(struct sock * sk)4650  void tcp_sack_compress_send_ack(struct sock *sk)
4651  {
4652  	struct tcp_sock *tp = tcp_sk(sk);
4653  
4654  	if (!tp->compressed_ack)
4655  		return;
4656  
4657  	if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
4658  		__sock_put(sk);
4659  
4660  	/* Since we have to send one ack finally,
4661  	 * substract one from tp->compressed_ack to keep
4662  	 * LINUX_MIB_TCPACKCOMPRESSED accurate.
4663  	 */
4664  	NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
4665  		      tp->compressed_ack - 1);
4666  
4667  	tp->compressed_ack = 0;
4668  	tcp_send_ack(sk);
4669  }
4670  
4671  /* Reasonable amount of sack blocks included in TCP SACK option
4672   * The max is 4, but this becomes 3 if TCP timestamps are there.
4673   * Given that SACK packets might be lost, be conservative and use 2.
4674   */
4675  #define TCP_SACK_BLOCKS_EXPECTED 2
4676  
tcp_sack_new_ofo_skb(struct sock * sk,u32 seq,u32 end_seq)4677  static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4678  {
4679  	struct tcp_sock *tp = tcp_sk(sk);
4680  	struct tcp_sack_block *sp = &tp->selective_acks[0];
4681  	int cur_sacks = tp->rx_opt.num_sacks;
4682  	int this_sack;
4683  
4684  	if (!cur_sacks)
4685  		goto new_sack;
4686  
4687  	for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4688  		if (tcp_sack_extend(sp, seq, end_seq)) {
4689  			if (this_sack >= TCP_SACK_BLOCKS_EXPECTED)
4690  				tcp_sack_compress_send_ack(sk);
4691  			/* Rotate this_sack to the first one. */
4692  			for (; this_sack > 0; this_sack--, sp--)
4693  				swap(*sp, *(sp - 1));
4694  			if (cur_sacks > 1)
4695  				tcp_sack_maybe_coalesce(tp);
4696  			return;
4697  		}
4698  	}
4699  
4700  	if (this_sack >= TCP_SACK_BLOCKS_EXPECTED)
4701  		tcp_sack_compress_send_ack(sk);
4702  
4703  	/* Could not find an adjacent existing SACK, build a new one,
4704  	 * put it at the front, and shift everyone else down.  We
4705  	 * always know there is at least one SACK present already here.
4706  	 *
4707  	 * If the sack array is full, forget about the last one.
4708  	 */
4709  	if (this_sack >= TCP_NUM_SACKS) {
4710  		this_sack--;
4711  		tp->rx_opt.num_sacks--;
4712  		sp--;
4713  	}
4714  	for (; this_sack > 0; this_sack--, sp--)
4715  		*sp = *(sp - 1);
4716  
4717  new_sack:
4718  	/* Build the new head SACK, and we're done. */
4719  	sp->start_seq = seq;
4720  	sp->end_seq = end_seq;
4721  	tp->rx_opt.num_sacks++;
4722  }
4723  
4724  /* RCV.NXT advances, some SACKs should be eaten. */
4725  
tcp_sack_remove(struct tcp_sock * tp)4726  static void tcp_sack_remove(struct tcp_sock *tp)
4727  {
4728  	struct tcp_sack_block *sp = &tp->selective_acks[0];
4729  	int num_sacks = tp->rx_opt.num_sacks;
4730  	int this_sack;
4731  
4732  	/* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4733  	if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4734  		tp->rx_opt.num_sacks = 0;
4735  		return;
4736  	}
4737  
4738  	for (this_sack = 0; this_sack < num_sacks;) {
4739  		/* Check if the start of the sack is covered by RCV.NXT. */
4740  		if (!before(tp->rcv_nxt, sp->start_seq)) {
4741  			int i;
4742  
4743  			/* RCV.NXT must cover all the block! */
4744  			WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4745  
4746  			/* Zap this SACK, by moving forward any other SACKS. */
4747  			for (i = this_sack+1; i < num_sacks; i++)
4748  				tp->selective_acks[i-1] = tp->selective_acks[i];
4749  			num_sacks--;
4750  			continue;
4751  		}
4752  		this_sack++;
4753  		sp++;
4754  	}
4755  	tp->rx_opt.num_sacks = num_sacks;
4756  }
4757  
4758  /**
4759   * tcp_try_coalesce - try to merge skb to prior one
4760   * @sk: socket
4761   * @to: prior buffer
4762   * @from: buffer to add in queue
4763   * @fragstolen: pointer to boolean
4764   *
4765   * Before queueing skb @from after @to, try to merge them
4766   * to reduce overall memory use and queue lengths, if cost is small.
4767   * Packets in ofo or receive queues can stay a long time.
4768   * Better try to coalesce them right now to avoid future collapses.
4769   * Returns true if caller should free @from instead of queueing it
4770   */
tcp_try_coalesce(struct sock * sk,struct sk_buff * to,struct sk_buff * from,bool * fragstolen)4771  static bool tcp_try_coalesce(struct sock *sk,
4772  			     struct sk_buff *to,
4773  			     struct sk_buff *from,
4774  			     bool *fragstolen)
4775  {
4776  	int delta;
4777  
4778  	*fragstolen = false;
4779  
4780  	/* Its possible this segment overlaps with prior segment in queue */
4781  	if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4782  		return false;
4783  
4784  	if (!mptcp_skb_can_collapse(to, from))
4785  		return false;
4786  
4787  #ifdef CONFIG_TLS_DEVICE
4788  	if (from->decrypted != to->decrypted)
4789  		return false;
4790  #endif
4791  
4792  	if (!skb_try_coalesce(to, from, fragstolen, &delta))
4793  		return false;
4794  
4795  	atomic_add(delta, &sk->sk_rmem_alloc);
4796  	sk_mem_charge(sk, delta);
4797  	NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4798  	TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4799  	TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4800  	TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4801  
4802  	if (TCP_SKB_CB(from)->has_rxtstamp) {
4803  		TCP_SKB_CB(to)->has_rxtstamp = true;
4804  		to->tstamp = from->tstamp;
4805  		skb_hwtstamps(to)->hwtstamp = skb_hwtstamps(from)->hwtstamp;
4806  	}
4807  
4808  	return true;
4809  }
4810  
tcp_ooo_try_coalesce(struct sock * sk,struct sk_buff * to,struct sk_buff * from,bool * fragstolen)4811  static bool tcp_ooo_try_coalesce(struct sock *sk,
4812  			     struct sk_buff *to,
4813  			     struct sk_buff *from,
4814  			     bool *fragstolen)
4815  {
4816  	bool res = tcp_try_coalesce(sk, to, from, fragstolen);
4817  
4818  	/* In case tcp_drop_reason() is called later, update to->gso_segs */
4819  	if (res) {
4820  		u32 gso_segs = max_t(u16, 1, skb_shinfo(to)->gso_segs) +
4821  			       max_t(u16, 1, skb_shinfo(from)->gso_segs);
4822  
4823  		skb_shinfo(to)->gso_segs = min_t(u32, gso_segs, 0xFFFF);
4824  	}
4825  	return res;
4826  }
4827  
tcp_drop_reason(struct sock * sk,struct sk_buff * skb,enum skb_drop_reason reason)4828  static void tcp_drop_reason(struct sock *sk, struct sk_buff *skb,
4829  			    enum skb_drop_reason reason)
4830  {
4831  	sk_drops_add(sk, skb);
4832  	kfree_skb_reason(skb, reason);
4833  }
4834  
4835  /* This one checks to see if we can put data from the
4836   * out_of_order queue into the receive_queue.
4837   */
tcp_ofo_queue(struct sock * sk)4838  static void tcp_ofo_queue(struct sock *sk)
4839  {
4840  	struct tcp_sock *tp = tcp_sk(sk);
4841  	__u32 dsack_high = tp->rcv_nxt;
4842  	bool fin, fragstolen, eaten;
4843  	struct sk_buff *skb, *tail;
4844  	struct rb_node *p;
4845  
4846  	p = rb_first(&tp->out_of_order_queue);
4847  	while (p) {
4848  		skb = rb_to_skb(p);
4849  		if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4850  			break;
4851  
4852  		if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4853  			__u32 dsack = dsack_high;
4854  			if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4855  				dsack_high = TCP_SKB_CB(skb)->end_seq;
4856  			tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4857  		}
4858  		p = rb_next(p);
4859  		rb_erase(&skb->rbnode, &tp->out_of_order_queue);
4860  
4861  		if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
4862  			tcp_drop_reason(sk, skb, SKB_DROP_REASON_TCP_OFO_DROP);
4863  			continue;
4864  		}
4865  
4866  		tail = skb_peek_tail(&sk->sk_receive_queue);
4867  		eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4868  		tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4869  		fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
4870  		if (!eaten)
4871  			__skb_queue_tail(&sk->sk_receive_queue, skb);
4872  		else
4873  			kfree_skb_partial(skb, fragstolen);
4874  
4875  		if (unlikely(fin)) {
4876  			tcp_fin(sk);
4877  			/* tcp_fin() purges tp->out_of_order_queue,
4878  			 * so we must end this loop right now.
4879  			 */
4880  			break;
4881  		}
4882  	}
4883  }
4884  
4885  static bool tcp_prune_ofo_queue(struct sock *sk, const struct sk_buff *in_skb);
4886  static int tcp_prune_queue(struct sock *sk, const struct sk_buff *in_skb);
4887  
tcp_try_rmem_schedule(struct sock * sk,struct sk_buff * skb,unsigned int size)4888  static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4889  				 unsigned int size)
4890  {
4891  	if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4892  	    !sk_rmem_schedule(sk, skb, size)) {
4893  
4894  		if (tcp_prune_queue(sk, skb) < 0)
4895  			return -1;
4896  
4897  		while (!sk_rmem_schedule(sk, skb, size)) {
4898  			if (!tcp_prune_ofo_queue(sk, skb))
4899  				return -1;
4900  		}
4901  	}
4902  	return 0;
4903  }
4904  
tcp_data_queue_ofo(struct sock * sk,struct sk_buff * skb)4905  static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4906  {
4907  	struct tcp_sock *tp = tcp_sk(sk);
4908  	struct rb_node **p, *parent;
4909  	struct sk_buff *skb1;
4910  	u32 seq, end_seq;
4911  	bool fragstolen;
4912  
4913  	tcp_ecn_check_ce(sk, skb);
4914  
4915  	if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4916  		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
4917  		sk->sk_data_ready(sk);
4918  		tcp_drop_reason(sk, skb, SKB_DROP_REASON_PROTO_MEM);
4919  		return;
4920  	}
4921  
4922  	/* Disable header prediction. */
4923  	tp->pred_flags = 0;
4924  	inet_csk_schedule_ack(sk);
4925  
4926  	tp->rcv_ooopack += max_t(u16, 1, skb_shinfo(skb)->gso_segs);
4927  	NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4928  	seq = TCP_SKB_CB(skb)->seq;
4929  	end_seq = TCP_SKB_CB(skb)->end_seq;
4930  
4931  	p = &tp->out_of_order_queue.rb_node;
4932  	if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4933  		/* Initial out of order segment, build 1 SACK. */
4934  		if (tcp_is_sack(tp)) {
4935  			tp->rx_opt.num_sacks = 1;
4936  			tp->selective_acks[0].start_seq = seq;
4937  			tp->selective_acks[0].end_seq = end_seq;
4938  		}
4939  		rb_link_node(&skb->rbnode, NULL, p);
4940  		rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4941  		tp->ooo_last_skb = skb;
4942  		goto end;
4943  	}
4944  
4945  	/* In the typical case, we are adding an skb to the end of the list.
4946  	 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4947  	 */
4948  	if (tcp_ooo_try_coalesce(sk, tp->ooo_last_skb,
4949  				 skb, &fragstolen)) {
4950  coalesce_done:
4951  		/* For non sack flows, do not grow window to force DUPACK
4952  		 * and trigger fast retransmit.
4953  		 */
4954  		if (tcp_is_sack(tp))
4955  			tcp_grow_window(sk, skb, true);
4956  		kfree_skb_partial(skb, fragstolen);
4957  		skb = NULL;
4958  		goto add_sack;
4959  	}
4960  	/* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4961  	if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
4962  		parent = &tp->ooo_last_skb->rbnode;
4963  		p = &parent->rb_right;
4964  		goto insert;
4965  	}
4966  
4967  	/* Find place to insert this segment. Handle overlaps on the way. */
4968  	parent = NULL;
4969  	while (*p) {
4970  		parent = *p;
4971  		skb1 = rb_to_skb(parent);
4972  		if (before(seq, TCP_SKB_CB(skb1)->seq)) {
4973  			p = &parent->rb_left;
4974  			continue;
4975  		}
4976  		if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4977  			if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4978  				/* All the bits are present. Drop. */
4979  				NET_INC_STATS(sock_net(sk),
4980  					      LINUX_MIB_TCPOFOMERGE);
4981  				tcp_drop_reason(sk, skb,
4982  						SKB_DROP_REASON_TCP_OFOMERGE);
4983  				skb = NULL;
4984  				tcp_dsack_set(sk, seq, end_seq);
4985  				goto add_sack;
4986  			}
4987  			if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4988  				/* Partial overlap. */
4989  				tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
4990  			} else {
4991  				/* skb's seq == skb1's seq and skb covers skb1.
4992  				 * Replace skb1 with skb.
4993  				 */
4994  				rb_replace_node(&skb1->rbnode, &skb->rbnode,
4995  						&tp->out_of_order_queue);
4996  				tcp_dsack_extend(sk,
4997  						 TCP_SKB_CB(skb1)->seq,
4998  						 TCP_SKB_CB(skb1)->end_seq);
4999  				NET_INC_STATS(sock_net(sk),
5000  					      LINUX_MIB_TCPOFOMERGE);
5001  				tcp_drop_reason(sk, skb1,
5002  						SKB_DROP_REASON_TCP_OFOMERGE);
5003  				goto merge_right;
5004  			}
5005  		} else if (tcp_ooo_try_coalesce(sk, skb1,
5006  						skb, &fragstolen)) {
5007  			goto coalesce_done;
5008  		}
5009  		p = &parent->rb_right;
5010  	}
5011  insert:
5012  	/* Insert segment into RB tree. */
5013  	rb_link_node(&skb->rbnode, parent, p);
5014  	rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
5015  
5016  merge_right:
5017  	/* Remove other segments covered by skb. */
5018  	while ((skb1 = skb_rb_next(skb)) != NULL) {
5019  		if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
5020  			break;
5021  		if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
5022  			tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
5023  					 end_seq);
5024  			break;
5025  		}
5026  		rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
5027  		tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
5028  				 TCP_SKB_CB(skb1)->end_seq);
5029  		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
5030  		tcp_drop_reason(sk, skb1, SKB_DROP_REASON_TCP_OFOMERGE);
5031  	}
5032  	/* If there is no skb after us, we are the last_skb ! */
5033  	if (!skb1)
5034  		tp->ooo_last_skb = skb;
5035  
5036  add_sack:
5037  	if (tcp_is_sack(tp))
5038  		tcp_sack_new_ofo_skb(sk, seq, end_seq);
5039  end:
5040  	if (skb) {
5041  		/* For non sack flows, do not grow window to force DUPACK
5042  		 * and trigger fast retransmit.
5043  		 */
5044  		if (tcp_is_sack(tp))
5045  			tcp_grow_window(sk, skb, false);
5046  		skb_condense(skb);
5047  		skb_set_owner_r(skb, sk);
5048  	}
5049  }
5050  
tcp_queue_rcv(struct sock * sk,struct sk_buff * skb,bool * fragstolen)5051  static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb,
5052  				      bool *fragstolen)
5053  {
5054  	int eaten;
5055  	struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
5056  
5057  	eaten = (tail &&
5058  		 tcp_try_coalesce(sk, tail,
5059  				  skb, fragstolen)) ? 1 : 0;
5060  	tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
5061  	if (!eaten) {
5062  		__skb_queue_tail(&sk->sk_receive_queue, skb);
5063  		skb_set_owner_r(skb, sk);
5064  	}
5065  	return eaten;
5066  }
5067  
tcp_send_rcvq(struct sock * sk,struct msghdr * msg,size_t size)5068  int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
5069  {
5070  	struct sk_buff *skb;
5071  	int err = -ENOMEM;
5072  	int data_len = 0;
5073  	bool fragstolen;
5074  
5075  	if (size == 0)
5076  		return 0;
5077  
5078  	if (size > PAGE_SIZE) {
5079  		int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
5080  
5081  		data_len = npages << PAGE_SHIFT;
5082  		size = data_len + (size & ~PAGE_MASK);
5083  	}
5084  	skb = alloc_skb_with_frags(size - data_len, data_len,
5085  				   PAGE_ALLOC_COSTLY_ORDER,
5086  				   &err, sk->sk_allocation);
5087  	if (!skb)
5088  		goto err;
5089  
5090  	skb_put(skb, size - data_len);
5091  	skb->data_len = data_len;
5092  	skb->len = size;
5093  
5094  	if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
5095  		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
5096  		goto err_free;
5097  	}
5098  
5099  	err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
5100  	if (err)
5101  		goto err_free;
5102  
5103  	TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
5104  	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
5105  	TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
5106  
5107  	if (tcp_queue_rcv(sk, skb, &fragstolen)) {
5108  		WARN_ON_ONCE(fragstolen); /* should not happen */
5109  		__kfree_skb(skb);
5110  	}
5111  	return size;
5112  
5113  err_free:
5114  	kfree_skb(skb);
5115  err:
5116  	return err;
5117  
5118  }
5119  
tcp_data_ready(struct sock * sk)5120  void tcp_data_ready(struct sock *sk)
5121  {
5122  	if (tcp_epollin_ready(sk, sk->sk_rcvlowat) || sock_flag(sk, SOCK_DONE))
5123  		sk->sk_data_ready(sk);
5124  }
5125  
tcp_data_queue(struct sock * sk,struct sk_buff * skb)5126  static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
5127  {
5128  	struct tcp_sock *tp = tcp_sk(sk);
5129  	enum skb_drop_reason reason;
5130  	bool fragstolen;
5131  	int eaten;
5132  
5133  	/* If a subflow has been reset, the packet should not continue
5134  	 * to be processed, drop the packet.
5135  	 */
5136  	if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb)) {
5137  		__kfree_skb(skb);
5138  		return;
5139  	}
5140  
5141  	if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
5142  		__kfree_skb(skb);
5143  		return;
5144  	}
5145  	skb_dst_drop(skb);
5146  	__skb_pull(skb, tcp_hdr(skb)->doff * 4);
5147  
5148  	reason = SKB_DROP_REASON_NOT_SPECIFIED;
5149  	tp->rx_opt.dsack = 0;
5150  
5151  	/*  Queue data for delivery to the user.
5152  	 *  Packets in sequence go to the receive queue.
5153  	 *  Out of sequence packets to the out_of_order_queue.
5154  	 */
5155  	if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
5156  		if (tcp_receive_window(tp) == 0) {
5157  			reason = SKB_DROP_REASON_TCP_ZEROWINDOW;
5158  			NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
5159  			goto out_of_window;
5160  		}
5161  
5162  		/* Ok. In sequence. In window. */
5163  queue_and_out:
5164  		if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
5165  			/* TODO: maybe ratelimit these WIN 0 ACK ? */
5166  			inet_csk(sk)->icsk_ack.pending |=
5167  					(ICSK_ACK_NOMEM | ICSK_ACK_NOW);
5168  			inet_csk_schedule_ack(sk);
5169  			sk->sk_data_ready(sk);
5170  
5171  			if (skb_queue_len(&sk->sk_receive_queue)) {
5172  				reason = SKB_DROP_REASON_PROTO_MEM;
5173  				NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
5174  				goto drop;
5175  			}
5176  			sk_forced_mem_schedule(sk, skb->truesize);
5177  		}
5178  
5179  		eaten = tcp_queue_rcv(sk, skb, &fragstolen);
5180  		if (skb->len)
5181  			tcp_event_data_recv(sk, skb);
5182  		if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
5183  			tcp_fin(sk);
5184  
5185  		if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5186  			tcp_ofo_queue(sk);
5187  
5188  			/* RFC5681. 4.2. SHOULD send immediate ACK, when
5189  			 * gap in queue is filled.
5190  			 */
5191  			if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
5192  				inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
5193  		}
5194  
5195  		if (tp->rx_opt.num_sacks)
5196  			tcp_sack_remove(tp);
5197  
5198  		tcp_fast_path_check(sk);
5199  
5200  		if (eaten > 0)
5201  			kfree_skb_partial(skb, fragstolen);
5202  		if (!sock_flag(sk, SOCK_DEAD))
5203  			tcp_data_ready(sk);
5204  		return;
5205  	}
5206  
5207  	if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
5208  		tcp_rcv_spurious_retrans(sk, skb);
5209  		/* A retransmit, 2nd most common case.  Force an immediate ack. */
5210  		reason = SKB_DROP_REASON_TCP_OLD_DATA;
5211  		NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
5212  		tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
5213  
5214  out_of_window:
5215  		tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
5216  		inet_csk_schedule_ack(sk);
5217  drop:
5218  		tcp_drop_reason(sk, skb, reason);
5219  		return;
5220  	}
5221  
5222  	/* Out of window. F.e. zero window probe. */
5223  	if (!before(TCP_SKB_CB(skb)->seq,
5224  		    tp->rcv_nxt + tcp_receive_window(tp))) {
5225  		reason = SKB_DROP_REASON_TCP_OVERWINDOW;
5226  		goto out_of_window;
5227  	}
5228  
5229  	if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5230  		/* Partial packet, seq < rcv_next < end_seq */
5231  		tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
5232  
5233  		/* If window is closed, drop tail of packet. But after
5234  		 * remembering D-SACK for its head made in previous line.
5235  		 */
5236  		if (!tcp_receive_window(tp)) {
5237  			reason = SKB_DROP_REASON_TCP_ZEROWINDOW;
5238  			NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
5239  			goto out_of_window;
5240  		}
5241  		goto queue_and_out;
5242  	}
5243  
5244  	tcp_data_queue_ofo(sk, skb);
5245  }
5246  
tcp_skb_next(struct sk_buff * skb,struct sk_buff_head * list)5247  static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
5248  {
5249  	if (list)
5250  		return !skb_queue_is_last(list, skb) ? skb->next : NULL;
5251  
5252  	return skb_rb_next(skb);
5253  }
5254  
tcp_collapse_one(struct sock * sk,struct sk_buff * skb,struct sk_buff_head * list,struct rb_root * root)5255  static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
5256  					struct sk_buff_head *list,
5257  					struct rb_root *root)
5258  {
5259  	struct sk_buff *next = tcp_skb_next(skb, list);
5260  
5261  	if (list)
5262  		__skb_unlink(skb, list);
5263  	else
5264  		rb_erase(&skb->rbnode, root);
5265  
5266  	__kfree_skb(skb);
5267  	NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
5268  
5269  	return next;
5270  }
5271  
5272  /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
tcp_rbtree_insert(struct rb_root * root,struct sk_buff * skb)5273  void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
5274  {
5275  	struct rb_node **p = &root->rb_node;
5276  	struct rb_node *parent = NULL;
5277  	struct sk_buff *skb1;
5278  
5279  	while (*p) {
5280  		parent = *p;
5281  		skb1 = rb_to_skb(parent);
5282  		if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
5283  			p = &parent->rb_left;
5284  		else
5285  			p = &parent->rb_right;
5286  	}
5287  	rb_link_node(&skb->rbnode, parent, p);
5288  	rb_insert_color(&skb->rbnode, root);
5289  }
5290  
5291  /* Collapse contiguous sequence of skbs head..tail with
5292   * sequence numbers start..end.
5293   *
5294   * If tail is NULL, this means until the end of the queue.
5295   *
5296   * Segments with FIN/SYN are not collapsed (only because this
5297   * simplifies code)
5298   */
5299  static void
tcp_collapse(struct sock * sk,struct sk_buff_head * list,struct rb_root * root,struct sk_buff * head,struct sk_buff * tail,u32 start,u32 end)5300  tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
5301  	     struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
5302  {
5303  	struct sk_buff *skb = head, *n;
5304  	struct sk_buff_head tmp;
5305  	bool end_of_skbs;
5306  
5307  	/* First, check that queue is collapsible and find
5308  	 * the point where collapsing can be useful.
5309  	 */
5310  restart:
5311  	for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
5312  		n = tcp_skb_next(skb, list);
5313  
5314  		/* No new bits? It is possible on ofo queue. */
5315  		if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
5316  			skb = tcp_collapse_one(sk, skb, list, root);
5317  			if (!skb)
5318  				break;
5319  			goto restart;
5320  		}
5321  
5322  		/* The first skb to collapse is:
5323  		 * - not SYN/FIN and
5324  		 * - bloated or contains data before "start" or
5325  		 *   overlaps to the next one and mptcp allow collapsing.
5326  		 */
5327  		if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
5328  		    (tcp_win_from_space(sk, skb->truesize) > skb->len ||
5329  		     before(TCP_SKB_CB(skb)->seq, start))) {
5330  			end_of_skbs = false;
5331  			break;
5332  		}
5333  
5334  		if (n && n != tail && mptcp_skb_can_collapse(skb, n) &&
5335  		    TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
5336  			end_of_skbs = false;
5337  			break;
5338  		}
5339  
5340  		/* Decided to skip this, advance start seq. */
5341  		start = TCP_SKB_CB(skb)->end_seq;
5342  	}
5343  	if (end_of_skbs ||
5344  	    (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
5345  		return;
5346  
5347  	__skb_queue_head_init(&tmp);
5348  
5349  	while (before(start, end)) {
5350  		int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
5351  		struct sk_buff *nskb;
5352  
5353  		nskb = alloc_skb(copy, GFP_ATOMIC);
5354  		if (!nskb)
5355  			break;
5356  
5357  		memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
5358  #ifdef CONFIG_TLS_DEVICE
5359  		nskb->decrypted = skb->decrypted;
5360  #endif
5361  		TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
5362  		if (list)
5363  			__skb_queue_before(list, skb, nskb);
5364  		else
5365  			__skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
5366  		skb_set_owner_r(nskb, sk);
5367  		mptcp_skb_ext_move(nskb, skb);
5368  
5369  		/* Copy data, releasing collapsed skbs. */
5370  		while (copy > 0) {
5371  			int offset = start - TCP_SKB_CB(skb)->seq;
5372  			int size = TCP_SKB_CB(skb)->end_seq - start;
5373  
5374  			BUG_ON(offset < 0);
5375  			if (size > 0) {
5376  				size = min(copy, size);
5377  				if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
5378  					BUG();
5379  				TCP_SKB_CB(nskb)->end_seq += size;
5380  				copy -= size;
5381  				start += size;
5382  			}
5383  			if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
5384  				skb = tcp_collapse_one(sk, skb, list, root);
5385  				if (!skb ||
5386  				    skb == tail ||
5387  				    !mptcp_skb_can_collapse(nskb, skb) ||
5388  				    (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
5389  					goto end;
5390  #ifdef CONFIG_TLS_DEVICE
5391  				if (skb->decrypted != nskb->decrypted)
5392  					goto end;
5393  #endif
5394  			}
5395  		}
5396  	}
5397  end:
5398  	skb_queue_walk_safe(&tmp, skb, n)
5399  		tcp_rbtree_insert(root, skb);
5400  }
5401  
5402  /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
5403   * and tcp_collapse() them until all the queue is collapsed.
5404   */
tcp_collapse_ofo_queue(struct sock * sk)5405  static void tcp_collapse_ofo_queue(struct sock *sk)
5406  {
5407  	struct tcp_sock *tp = tcp_sk(sk);
5408  	u32 range_truesize, sum_tiny = 0;
5409  	struct sk_buff *skb, *head;
5410  	u32 start, end;
5411  
5412  	skb = skb_rb_first(&tp->out_of_order_queue);
5413  new_range:
5414  	if (!skb) {
5415  		tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue);
5416  		return;
5417  	}
5418  	start = TCP_SKB_CB(skb)->seq;
5419  	end = TCP_SKB_CB(skb)->end_seq;
5420  	range_truesize = skb->truesize;
5421  
5422  	for (head = skb;;) {
5423  		skb = skb_rb_next(skb);
5424  
5425  		/* Range is terminated when we see a gap or when
5426  		 * we are at the queue end.
5427  		 */
5428  		if (!skb ||
5429  		    after(TCP_SKB_CB(skb)->seq, end) ||
5430  		    before(TCP_SKB_CB(skb)->end_seq, start)) {
5431  			/* Do not attempt collapsing tiny skbs */
5432  			if (range_truesize != head->truesize ||
5433  			    end - start >= SKB_WITH_OVERHEAD(PAGE_SIZE)) {
5434  				tcp_collapse(sk, NULL, &tp->out_of_order_queue,
5435  					     head, skb, start, end);
5436  			} else {
5437  				sum_tiny += range_truesize;
5438  				if (sum_tiny > sk->sk_rcvbuf >> 3)
5439  					return;
5440  			}
5441  			goto new_range;
5442  		}
5443  
5444  		range_truesize += skb->truesize;
5445  		if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
5446  			start = TCP_SKB_CB(skb)->seq;
5447  		if (after(TCP_SKB_CB(skb)->end_seq, end))
5448  			end = TCP_SKB_CB(skb)->end_seq;
5449  	}
5450  }
5451  
5452  /*
5453   * Clean the out-of-order queue to make room.
5454   * We drop high sequences packets to :
5455   * 1) Let a chance for holes to be filled.
5456   *    This means we do not drop packets from ooo queue if their sequence
5457   *    is before incoming packet sequence.
5458   * 2) not add too big latencies if thousands of packets sit there.
5459   *    (But if application shrinks SO_RCVBUF, we could still end up
5460   *     freeing whole queue here)
5461   * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks.
5462   *
5463   * Return true if queue has shrunk.
5464   */
tcp_prune_ofo_queue(struct sock * sk,const struct sk_buff * in_skb)5465  static bool tcp_prune_ofo_queue(struct sock *sk, const struct sk_buff *in_skb)
5466  {
5467  	struct tcp_sock *tp = tcp_sk(sk);
5468  	struct rb_node *node, *prev;
5469  	bool pruned = false;
5470  	int goal;
5471  
5472  	if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
5473  		return false;
5474  
5475  	goal = sk->sk_rcvbuf >> 3;
5476  	node = &tp->ooo_last_skb->rbnode;
5477  
5478  	do {
5479  		struct sk_buff *skb = rb_to_skb(node);
5480  
5481  		/* If incoming skb would land last in ofo queue, stop pruning. */
5482  		if (after(TCP_SKB_CB(in_skb)->seq, TCP_SKB_CB(skb)->seq))
5483  			break;
5484  		pruned = true;
5485  		prev = rb_prev(node);
5486  		rb_erase(node, &tp->out_of_order_queue);
5487  		goal -= skb->truesize;
5488  		tcp_drop_reason(sk, skb, SKB_DROP_REASON_TCP_OFO_QUEUE_PRUNE);
5489  		tp->ooo_last_skb = rb_to_skb(prev);
5490  		if (!prev || goal <= 0) {
5491  			if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
5492  			    !tcp_under_memory_pressure(sk))
5493  				break;
5494  			goal = sk->sk_rcvbuf >> 3;
5495  		}
5496  		node = prev;
5497  	} while (node);
5498  
5499  	if (pruned) {
5500  		NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
5501  		/* Reset SACK state.  A conforming SACK implementation will
5502  		 * do the same at a timeout based retransmit.  When a connection
5503  		 * is in a sad state like this, we care only about integrity
5504  		 * of the connection not performance.
5505  		 */
5506  		if (tp->rx_opt.sack_ok)
5507  			tcp_sack_reset(&tp->rx_opt);
5508  	}
5509  	return pruned;
5510  }
5511  
5512  /* Reduce allocated memory if we can, trying to get
5513   * the socket within its memory limits again.
5514   *
5515   * Return less than zero if we should start dropping frames
5516   * until the socket owning process reads some of the data
5517   * to stabilize the situation.
5518   */
tcp_prune_queue(struct sock * sk,const struct sk_buff * in_skb)5519  static int tcp_prune_queue(struct sock *sk, const struct sk_buff *in_skb)
5520  {
5521  	struct tcp_sock *tp = tcp_sk(sk);
5522  
5523  	NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
5524  
5525  	if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
5526  		tcp_clamp_window(sk);
5527  	else if (tcp_under_memory_pressure(sk))
5528  		tcp_adjust_rcv_ssthresh(sk);
5529  
5530  	if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5531  		return 0;
5532  
5533  	tcp_collapse_ofo_queue(sk);
5534  	if (!skb_queue_empty(&sk->sk_receive_queue))
5535  		tcp_collapse(sk, &sk->sk_receive_queue, NULL,
5536  			     skb_peek(&sk->sk_receive_queue),
5537  			     NULL,
5538  			     tp->copied_seq, tp->rcv_nxt);
5539  
5540  	if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5541  		return 0;
5542  
5543  	/* Collapsing did not help, destructive actions follow.
5544  	 * This must not ever occur. */
5545  
5546  	tcp_prune_ofo_queue(sk, in_skb);
5547  
5548  	if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5549  		return 0;
5550  
5551  	/* If we are really being abused, tell the caller to silently
5552  	 * drop receive data on the floor.  It will get retransmitted
5553  	 * and hopefully then we'll have sufficient space.
5554  	 */
5555  	NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
5556  
5557  	/* Massive buffer overcommit. */
5558  	tp->pred_flags = 0;
5559  	return -1;
5560  }
5561  
tcp_should_expand_sndbuf(struct sock * sk)5562  static bool tcp_should_expand_sndbuf(struct sock *sk)
5563  {
5564  	const struct tcp_sock *tp = tcp_sk(sk);
5565  
5566  	/* If the user specified a specific send buffer setting, do
5567  	 * not modify it.
5568  	 */
5569  	if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
5570  		return false;
5571  
5572  	/* If we are under global TCP memory pressure, do not expand.  */
5573  	if (tcp_under_memory_pressure(sk)) {
5574  		int unused_mem = sk_unused_reserved_mem(sk);
5575  
5576  		/* Adjust sndbuf according to reserved mem. But make sure
5577  		 * it never goes below SOCK_MIN_SNDBUF.
5578  		 * See sk_stream_moderate_sndbuf() for more details.
5579  		 */
5580  		if (unused_mem > SOCK_MIN_SNDBUF)
5581  			WRITE_ONCE(sk->sk_sndbuf, unused_mem);
5582  
5583  		return false;
5584  	}
5585  
5586  	/* If we are under soft global TCP memory pressure, do not expand.  */
5587  	if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
5588  		return false;
5589  
5590  	/* If we filled the congestion window, do not expand.  */
5591  	if (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp))
5592  		return false;
5593  
5594  	return true;
5595  }
5596  
tcp_new_space(struct sock * sk)5597  static void tcp_new_space(struct sock *sk)
5598  {
5599  	struct tcp_sock *tp = tcp_sk(sk);
5600  
5601  	if (tcp_should_expand_sndbuf(sk)) {
5602  		tcp_sndbuf_expand(sk);
5603  		tp->snd_cwnd_stamp = tcp_jiffies32;
5604  	}
5605  
5606  	INDIRECT_CALL_1(sk->sk_write_space, sk_stream_write_space, sk);
5607  }
5608  
5609  /* Caller made space either from:
5610   * 1) Freeing skbs in rtx queues (after tp->snd_una has advanced)
5611   * 2) Sent skbs from output queue (and thus advancing tp->snd_nxt)
5612   *
5613   * We might be able to generate EPOLLOUT to the application if:
5614   * 1) Space consumed in output/rtx queues is below sk->sk_sndbuf/2
5615   * 2) notsent amount (tp->write_seq - tp->snd_nxt) became
5616   *    small enough that tcp_stream_memory_free() decides it
5617   *    is time to generate EPOLLOUT.
5618   */
tcp_check_space(struct sock * sk)5619  void tcp_check_space(struct sock *sk)
5620  {
5621  	/* pairs with tcp_poll() */
5622  	smp_mb();
5623  	if (sk->sk_socket &&
5624  	    test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
5625  		tcp_new_space(sk);
5626  		if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5627  			tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
5628  	}
5629  }
5630  
tcp_data_snd_check(struct sock * sk)5631  static inline void tcp_data_snd_check(struct sock *sk)
5632  {
5633  	tcp_push_pending_frames(sk);
5634  	tcp_check_space(sk);
5635  }
5636  
5637  /*
5638   * Check if sending an ack is needed.
5639   */
__tcp_ack_snd_check(struct sock * sk,int ofo_possible)5640  static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5641  {
5642  	struct tcp_sock *tp = tcp_sk(sk);
5643  	unsigned long rtt, delay;
5644  
5645  	    /* More than one full frame received... */
5646  	if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5647  	     /* ... and right edge of window advances far enough.
5648  	      * (tcp_recvmsg() will send ACK otherwise).
5649  	      * If application uses SO_RCVLOWAT, we want send ack now if
5650  	      * we have not received enough bytes to satisfy the condition.
5651  	      */
5652  	    (tp->rcv_nxt - tp->copied_seq < sk->sk_rcvlowat ||
5653  	     __tcp_select_window(sk) >= tp->rcv_wnd)) ||
5654  	    /* We ACK each frame or... */
5655  	    tcp_in_quickack_mode(sk) ||
5656  	    /* Protocol state mandates a one-time immediate ACK */
5657  	    inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOW) {
5658  send_now:
5659  		tcp_send_ack(sk);
5660  		return;
5661  	}
5662  
5663  	if (!ofo_possible || RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5664  		tcp_send_delayed_ack(sk);
5665  		return;
5666  	}
5667  
5668  	if (!tcp_is_sack(tp) ||
5669  	    tp->compressed_ack >= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_nr))
5670  		goto send_now;
5671  
5672  	if (tp->compressed_ack_rcv_nxt != tp->rcv_nxt) {
5673  		tp->compressed_ack_rcv_nxt = tp->rcv_nxt;
5674  		tp->dup_ack_counter = 0;
5675  	}
5676  	if (tp->dup_ack_counter < TCP_FASTRETRANS_THRESH) {
5677  		tp->dup_ack_counter++;
5678  		goto send_now;
5679  	}
5680  	tp->compressed_ack++;
5681  	if (hrtimer_is_queued(&tp->compressed_ack_timer))
5682  		return;
5683  
5684  	/* compress ack timer : 5 % of rtt, but no more than tcp_comp_sack_delay_ns */
5685  
5686  	rtt = tp->rcv_rtt_est.rtt_us;
5687  	if (tp->srtt_us && tp->srtt_us < rtt)
5688  		rtt = tp->srtt_us;
5689  
5690  	delay = min_t(unsigned long,
5691  		      READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_delay_ns),
5692  		      rtt * (NSEC_PER_USEC >> 3)/20);
5693  	sock_hold(sk);
5694  	hrtimer_start_range_ns(&tp->compressed_ack_timer, ns_to_ktime(delay),
5695  			       READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_slack_ns),
5696  			       HRTIMER_MODE_REL_PINNED_SOFT);
5697  }
5698  
tcp_ack_snd_check(struct sock * sk)5699  static inline void tcp_ack_snd_check(struct sock *sk)
5700  {
5701  	if (!inet_csk_ack_scheduled(sk)) {
5702  		/* We sent a data segment already. */
5703  		return;
5704  	}
5705  	__tcp_ack_snd_check(sk, 1);
5706  }
5707  
5708  /*
5709   *	This routine is only called when we have urgent data
5710   *	signaled. Its the 'slow' part of tcp_urg. It could be
5711   *	moved inline now as tcp_urg is only called from one
5712   *	place. We handle URGent data wrong. We have to - as
5713   *	BSD still doesn't use the correction from RFC961.
5714   *	For 1003.1g we should support a new option TCP_STDURG to permit
5715   *	either form (or just set the sysctl tcp_stdurg).
5716   */
5717  
tcp_check_urg(struct sock * sk,const struct tcphdr * th)5718  static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5719  {
5720  	struct tcp_sock *tp = tcp_sk(sk);
5721  	u32 ptr = ntohs(th->urg_ptr);
5722  
5723  	if (ptr && !READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_stdurg))
5724  		ptr--;
5725  	ptr += ntohl(th->seq);
5726  
5727  	/* Ignore urgent data that we've already seen and read. */
5728  	if (after(tp->copied_seq, ptr))
5729  		return;
5730  
5731  	/* Do not replay urg ptr.
5732  	 *
5733  	 * NOTE: interesting situation not covered by specs.
5734  	 * Misbehaving sender may send urg ptr, pointing to segment,
5735  	 * which we already have in ofo queue. We are not able to fetch
5736  	 * such data and will stay in TCP_URG_NOTYET until will be eaten
5737  	 * by recvmsg(). Seems, we are not obliged to handle such wicked
5738  	 * situations. But it is worth to think about possibility of some
5739  	 * DoSes using some hypothetical application level deadlock.
5740  	 */
5741  	if (before(ptr, tp->rcv_nxt))
5742  		return;
5743  
5744  	/* Do we already have a newer (or duplicate) urgent pointer? */
5745  	if (tp->urg_data && !after(ptr, tp->urg_seq))
5746  		return;
5747  
5748  	/* Tell the world about our new urgent pointer. */
5749  	sk_send_sigurg(sk);
5750  
5751  	/* We may be adding urgent data when the last byte read was
5752  	 * urgent. To do this requires some care. We cannot just ignore
5753  	 * tp->copied_seq since we would read the last urgent byte again
5754  	 * as data, nor can we alter copied_seq until this data arrives
5755  	 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5756  	 *
5757  	 * NOTE. Double Dutch. Rendering to plain English: author of comment
5758  	 * above did something sort of 	send("A", MSG_OOB); send("B", MSG_OOB);
5759  	 * and expect that both A and B disappear from stream. This is _wrong_.
5760  	 * Though this happens in BSD with high probability, this is occasional.
5761  	 * Any application relying on this is buggy. Note also, that fix "works"
5762  	 * only in this artificial test. Insert some normal data between A and B and we will
5763  	 * decline of BSD again. Verdict: it is better to remove to trap
5764  	 * buggy users.
5765  	 */
5766  	if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5767  	    !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5768  		struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5769  		tp->copied_seq++;
5770  		if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5771  			__skb_unlink(skb, &sk->sk_receive_queue);
5772  			__kfree_skb(skb);
5773  		}
5774  	}
5775  
5776  	WRITE_ONCE(tp->urg_data, TCP_URG_NOTYET);
5777  	WRITE_ONCE(tp->urg_seq, ptr);
5778  
5779  	/* Disable header prediction. */
5780  	tp->pred_flags = 0;
5781  }
5782  
5783  /* This is the 'fast' part of urgent handling. */
tcp_urg(struct sock * sk,struct sk_buff * skb,const struct tcphdr * th)5784  static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5785  {
5786  	struct tcp_sock *tp = tcp_sk(sk);
5787  
5788  	/* Check if we get a new urgent pointer - normally not. */
5789  	if (unlikely(th->urg))
5790  		tcp_check_urg(sk, th);
5791  
5792  	/* Do we wait for any urgent data? - normally not... */
5793  	if (unlikely(tp->urg_data == TCP_URG_NOTYET)) {
5794  		u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5795  			  th->syn;
5796  
5797  		/* Is the urgent pointer pointing into this packet? */
5798  		if (ptr < skb->len) {
5799  			u8 tmp;
5800  			if (skb_copy_bits(skb, ptr, &tmp, 1))
5801  				BUG();
5802  			WRITE_ONCE(tp->urg_data, TCP_URG_VALID | tmp);
5803  			if (!sock_flag(sk, SOCK_DEAD))
5804  				sk->sk_data_ready(sk);
5805  		}
5806  	}
5807  }
5808  
5809  /* Accept RST for rcv_nxt - 1 after a FIN.
5810   * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5811   * FIN is sent followed by a RST packet. The RST is sent with the same
5812   * sequence number as the FIN, and thus according to RFC 5961 a challenge
5813   * ACK should be sent. However, Mac OSX rate limits replies to challenge
5814   * ACKs on the closed socket. In addition middleboxes can drop either the
5815   * challenge ACK or a subsequent RST.
5816   */
tcp_reset_check(const struct sock * sk,const struct sk_buff * skb)5817  static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb)
5818  {
5819  	const struct tcp_sock *tp = tcp_sk(sk);
5820  
5821  	return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) &&
5822  			(1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK |
5823  					       TCPF_CLOSING));
5824  }
5825  
5826  /* Does PAWS and seqno based validation of an incoming segment, flags will
5827   * play significant role here.
5828   */
tcp_validate_incoming(struct sock * sk,struct sk_buff * skb,const struct tcphdr * th,int syn_inerr)5829  static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5830  				  const struct tcphdr *th, int syn_inerr)
5831  {
5832  	struct tcp_sock *tp = tcp_sk(sk);
5833  	SKB_DR(reason);
5834  
5835  	/* RFC1323: H1. Apply PAWS check first. */
5836  	if (tcp_fast_parse_options(sock_net(sk), skb, th, tp) &&
5837  	    tp->rx_opt.saw_tstamp &&
5838  	    tcp_paws_discard(sk, skb)) {
5839  		if (!th->rst) {
5840  			if (unlikely(th->syn))
5841  				goto syn_challenge;
5842  			NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5843  			if (!tcp_oow_rate_limited(sock_net(sk), skb,
5844  						  LINUX_MIB_TCPACKSKIPPEDPAWS,
5845  						  &tp->last_oow_ack_time))
5846  				tcp_send_dupack(sk, skb);
5847  			SKB_DR_SET(reason, TCP_RFC7323_PAWS);
5848  			goto discard;
5849  		}
5850  		/* Reset is accepted even if it did not pass PAWS. */
5851  	}
5852  
5853  	/* Step 1: check sequence number */
5854  	reason = tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
5855  	if (reason) {
5856  		/* RFC793, page 37: "In all states except SYN-SENT, all reset
5857  		 * (RST) segments are validated by checking their SEQ-fields."
5858  		 * And page 69: "If an incoming segment is not acceptable,
5859  		 * an acknowledgment should be sent in reply (unless the RST
5860  		 * bit is set, if so drop the segment and return)".
5861  		 */
5862  		if (!th->rst) {
5863  			if (th->syn)
5864  				goto syn_challenge;
5865  			if (!tcp_oow_rate_limited(sock_net(sk), skb,
5866  						  LINUX_MIB_TCPACKSKIPPEDSEQ,
5867  						  &tp->last_oow_ack_time))
5868  				tcp_send_dupack(sk, skb);
5869  		} else if (tcp_reset_check(sk, skb)) {
5870  			goto reset;
5871  		}
5872  		goto discard;
5873  	}
5874  
5875  	/* Step 2: check RST bit */
5876  	if (th->rst) {
5877  		/* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5878  		 * FIN and SACK too if available):
5879  		 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5880  		 * the right-most SACK block,
5881  		 * then
5882  		 *     RESET the connection
5883  		 * else
5884  		 *     Send a challenge ACK
5885  		 */
5886  		if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt ||
5887  		    tcp_reset_check(sk, skb))
5888  			goto reset;
5889  
5890  		if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
5891  			struct tcp_sack_block *sp = &tp->selective_acks[0];
5892  			int max_sack = sp[0].end_seq;
5893  			int this_sack;
5894  
5895  			for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
5896  			     ++this_sack) {
5897  				max_sack = after(sp[this_sack].end_seq,
5898  						 max_sack) ?
5899  					sp[this_sack].end_seq : max_sack;
5900  			}
5901  
5902  			if (TCP_SKB_CB(skb)->seq == max_sack)
5903  				goto reset;
5904  		}
5905  
5906  		/* Disable TFO if RST is out-of-order
5907  		 * and no data has been received
5908  		 * for current active TFO socket
5909  		 */
5910  		if (tp->syn_fastopen && !tp->data_segs_in &&
5911  		    sk->sk_state == TCP_ESTABLISHED)
5912  			tcp_fastopen_active_disable(sk);
5913  		tcp_send_challenge_ack(sk);
5914  		SKB_DR_SET(reason, TCP_RESET);
5915  		goto discard;
5916  	}
5917  
5918  	/* step 3: check security and precedence [ignored] */
5919  
5920  	/* step 4: Check for a SYN
5921  	 * RFC 5961 4.2 : Send a challenge ack
5922  	 */
5923  	if (th->syn) {
5924  		if (sk->sk_state == TCP_SYN_RECV && sk->sk_socket && th->ack &&
5925  		    TCP_SKB_CB(skb)->seq + 1 == TCP_SKB_CB(skb)->end_seq &&
5926  		    TCP_SKB_CB(skb)->seq + 1 == tp->rcv_nxt &&
5927  		    TCP_SKB_CB(skb)->ack_seq == tp->snd_nxt)
5928  			goto pass;
5929  syn_challenge:
5930  		if (syn_inerr)
5931  			TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5932  		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5933  		tcp_send_challenge_ack(sk);
5934  		SKB_DR_SET(reason, TCP_INVALID_SYN);
5935  		goto discard;
5936  	}
5937  
5938  pass:
5939  	bpf_skops_parse_hdr(sk, skb);
5940  
5941  	return true;
5942  
5943  discard:
5944  	tcp_drop_reason(sk, skb, reason);
5945  	return false;
5946  
5947  reset:
5948  	tcp_reset(sk, skb);
5949  	__kfree_skb(skb);
5950  	return false;
5951  }
5952  
5953  /*
5954   *	TCP receive function for the ESTABLISHED state.
5955   *
5956   *	It is split into a fast path and a slow path. The fast path is
5957   * 	disabled when:
5958   *	- A zero window was announced from us - zero window probing
5959   *        is only handled properly in the slow path.
5960   *	- Out of order segments arrived.
5961   *	- Urgent data is expected.
5962   *	- There is no buffer space left
5963   *	- Unexpected TCP flags/window values/header lengths are received
5964   *	  (detected by checking the TCP header against pred_flags)
5965   *	- Data is sent in both directions. Fast path only supports pure senders
5966   *	  or pure receivers (this means either the sequence number or the ack
5967   *	  value must stay constant)
5968   *	- Unexpected TCP option.
5969   *
5970   *	When these conditions are not satisfied it drops into a standard
5971   *	receive procedure patterned after RFC793 to handle all cases.
5972   *	The first three cases are guaranteed by proper pred_flags setting,
5973   *	the rest is checked inline. Fast processing is turned on in
5974   *	tcp_data_queue when everything is OK.
5975   */
tcp_rcv_established(struct sock * sk,struct sk_buff * skb)5976  void tcp_rcv_established(struct sock *sk, struct sk_buff *skb)
5977  {
5978  	enum skb_drop_reason reason = SKB_DROP_REASON_NOT_SPECIFIED;
5979  	const struct tcphdr *th = (const struct tcphdr *)skb->data;
5980  	struct tcp_sock *tp = tcp_sk(sk);
5981  	unsigned int len = skb->len;
5982  
5983  	/* TCP congestion window tracking */
5984  	trace_tcp_probe(sk, skb);
5985  
5986  	tcp_mstamp_refresh(tp);
5987  	if (unlikely(!rcu_access_pointer(sk->sk_rx_dst)))
5988  		inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5989  	/*
5990  	 *	Header prediction.
5991  	 *	The code loosely follows the one in the famous
5992  	 *	"30 instruction TCP receive" Van Jacobson mail.
5993  	 *
5994  	 *	Van's trick is to deposit buffers into socket queue
5995  	 *	on a device interrupt, to call tcp_recv function
5996  	 *	on the receive process context and checksum and copy
5997  	 *	the buffer to user space. smart...
5998  	 *
5999  	 *	Our current scheme is not silly either but we take the
6000  	 *	extra cost of the net_bh soft interrupt processing...
6001  	 *	We do checksum and copy also but from device to kernel.
6002  	 */
6003  
6004  	tp->rx_opt.saw_tstamp = 0;
6005  
6006  	/*	pred_flags is 0xS?10 << 16 + snd_wnd
6007  	 *	if header_prediction is to be made
6008  	 *	'S' will always be tp->tcp_header_len >> 2
6009  	 *	'?' will be 0 for the fast path, otherwise pred_flags is 0 to
6010  	 *  turn it off	(when there are holes in the receive
6011  	 *	 space for instance)
6012  	 *	PSH flag is ignored.
6013  	 */
6014  
6015  	if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
6016  	    TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
6017  	    !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
6018  		int tcp_header_len = tp->tcp_header_len;
6019  
6020  		/* Timestamp header prediction: tcp_header_len
6021  		 * is automatically equal to th->doff*4 due to pred_flags
6022  		 * match.
6023  		 */
6024  
6025  		/* Check timestamp */
6026  		if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
6027  			/* No? Slow path! */
6028  			if (!tcp_parse_aligned_timestamp(tp, th))
6029  				goto slow_path;
6030  
6031  			/* If PAWS failed, check it more carefully in slow path */
6032  			if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
6033  				goto slow_path;
6034  
6035  			/* DO NOT update ts_recent here, if checksum fails
6036  			 * and timestamp was corrupted part, it will result
6037  			 * in a hung connection since we will drop all
6038  			 * future packets due to the PAWS test.
6039  			 */
6040  		}
6041  
6042  		if (len <= tcp_header_len) {
6043  			/* Bulk data transfer: sender */
6044  			if (len == tcp_header_len) {
6045  				/* Predicted packet is in window by definition.
6046  				 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
6047  				 * Hence, check seq<=rcv_wup reduces to:
6048  				 */
6049  				if (tcp_header_len ==
6050  				    (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
6051  				    tp->rcv_nxt == tp->rcv_wup)
6052  					tcp_store_ts_recent(tp);
6053  
6054  				/* We know that such packets are checksummed
6055  				 * on entry.
6056  				 */
6057  				tcp_ack(sk, skb, 0);
6058  				__kfree_skb(skb);
6059  				tcp_data_snd_check(sk);
6060  				/* When receiving pure ack in fast path, update
6061  				 * last ts ecr directly instead of calling
6062  				 * tcp_rcv_rtt_measure_ts()
6063  				 */
6064  				tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
6065  				return;
6066  			} else { /* Header too small */
6067  				reason = SKB_DROP_REASON_PKT_TOO_SMALL;
6068  				TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
6069  				goto discard;
6070  			}
6071  		} else {
6072  			int eaten = 0;
6073  			bool fragstolen = false;
6074  
6075  			if (tcp_checksum_complete(skb))
6076  				goto csum_error;
6077  
6078  			if ((int)skb->truesize > sk->sk_forward_alloc)
6079  				goto step5;
6080  
6081  			/* Predicted packet is in window by definition.
6082  			 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
6083  			 * Hence, check seq<=rcv_wup reduces to:
6084  			 */
6085  			if (tcp_header_len ==
6086  			    (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
6087  			    tp->rcv_nxt == tp->rcv_wup)
6088  				tcp_store_ts_recent(tp);
6089  
6090  			tcp_rcv_rtt_measure_ts(sk, skb);
6091  
6092  			NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
6093  
6094  			/* Bulk data transfer: receiver */
6095  			skb_dst_drop(skb);
6096  			__skb_pull(skb, tcp_header_len);
6097  			eaten = tcp_queue_rcv(sk, skb, &fragstolen);
6098  
6099  			tcp_event_data_recv(sk, skb);
6100  
6101  			if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
6102  				/* Well, only one small jumplet in fast path... */
6103  				tcp_ack(sk, skb, FLAG_DATA);
6104  				tcp_data_snd_check(sk);
6105  				if (!inet_csk_ack_scheduled(sk))
6106  					goto no_ack;
6107  			} else {
6108  				tcp_update_wl(tp, TCP_SKB_CB(skb)->seq);
6109  			}
6110  
6111  			__tcp_ack_snd_check(sk, 0);
6112  no_ack:
6113  			if (eaten)
6114  				kfree_skb_partial(skb, fragstolen);
6115  			tcp_data_ready(sk);
6116  			return;
6117  		}
6118  	}
6119  
6120  slow_path:
6121  	if (len < (th->doff << 2) || tcp_checksum_complete(skb))
6122  		goto csum_error;
6123  
6124  	if (!th->ack && !th->rst && !th->syn) {
6125  		reason = SKB_DROP_REASON_TCP_FLAGS;
6126  		goto discard;
6127  	}
6128  
6129  	/*
6130  	 *	Standard slow path.
6131  	 */
6132  
6133  	if (!tcp_validate_incoming(sk, skb, th, 1))
6134  		return;
6135  
6136  step5:
6137  	reason = tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT);
6138  	if ((int)reason < 0) {
6139  		reason = -reason;
6140  		goto discard;
6141  	}
6142  	tcp_rcv_rtt_measure_ts(sk, skb);
6143  
6144  	/* Process urgent data. */
6145  	tcp_urg(sk, skb, th);
6146  
6147  	/* step 7: process the segment text */
6148  	tcp_data_queue(sk, skb);
6149  
6150  	tcp_data_snd_check(sk);
6151  	tcp_ack_snd_check(sk);
6152  	return;
6153  
6154  csum_error:
6155  	reason = SKB_DROP_REASON_TCP_CSUM;
6156  	trace_tcp_bad_csum(skb);
6157  	TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
6158  	TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
6159  
6160  discard:
6161  	tcp_drop_reason(sk, skb, reason);
6162  }
6163  EXPORT_SYMBOL(tcp_rcv_established);
6164  
tcp_init_transfer(struct sock * sk,int bpf_op,struct sk_buff * skb)6165  void tcp_init_transfer(struct sock *sk, int bpf_op, struct sk_buff *skb)
6166  {
6167  	struct inet_connection_sock *icsk = inet_csk(sk);
6168  	struct tcp_sock *tp = tcp_sk(sk);
6169  
6170  	tcp_mtup_init(sk);
6171  	icsk->icsk_af_ops->rebuild_header(sk);
6172  	tcp_init_metrics(sk);
6173  
6174  	/* Initialize the congestion window to start the transfer.
6175  	 * Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
6176  	 * retransmitted. In light of RFC6298 more aggressive 1sec
6177  	 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
6178  	 * retransmission has occurred.
6179  	 */
6180  	if (tp->total_retrans > 1 && tp->undo_marker)
6181  		tcp_snd_cwnd_set(tp, 1);
6182  	else
6183  		tcp_snd_cwnd_set(tp, tcp_init_cwnd(tp, __sk_dst_get(sk)));
6184  	tp->snd_cwnd_stamp = tcp_jiffies32;
6185  
6186  	bpf_skops_established(sk, bpf_op, skb);
6187  	/* Initialize congestion control unless BPF initialized it already: */
6188  	if (!icsk->icsk_ca_initialized)
6189  		tcp_init_congestion_control(sk);
6190  	tcp_init_buffer_space(sk);
6191  }
6192  
tcp_finish_connect(struct sock * sk,struct sk_buff * skb)6193  void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
6194  {
6195  	struct tcp_sock *tp = tcp_sk(sk);
6196  	struct inet_connection_sock *icsk = inet_csk(sk);
6197  
6198  	tcp_set_state(sk, TCP_ESTABLISHED);
6199  	icsk->icsk_ack.lrcvtime = tcp_jiffies32;
6200  
6201  	if (skb) {
6202  		icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
6203  		security_inet_conn_established(sk, skb);
6204  		sk_mark_napi_id(sk, skb);
6205  	}
6206  
6207  	tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB, skb);
6208  
6209  	/* Prevent spurious tcp_cwnd_restart() on first data
6210  	 * packet.
6211  	 */
6212  	tp->lsndtime = tcp_jiffies32;
6213  
6214  	if (sock_flag(sk, SOCK_KEEPOPEN))
6215  		inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
6216  
6217  	if (!tp->rx_opt.snd_wscale)
6218  		__tcp_fast_path_on(tp, tp->snd_wnd);
6219  	else
6220  		tp->pred_flags = 0;
6221  }
6222  
tcp_rcv_fastopen_synack(struct sock * sk,struct sk_buff * synack,struct tcp_fastopen_cookie * cookie)6223  static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
6224  				    struct tcp_fastopen_cookie *cookie)
6225  {
6226  	struct tcp_sock *tp = tcp_sk(sk);
6227  	struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL;
6228  	u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
6229  	bool syn_drop = false;
6230  
6231  	if (mss == tp->rx_opt.user_mss) {
6232  		struct tcp_options_received opt;
6233  
6234  		/* Get original SYNACK MSS value if user MSS sets mss_clamp */
6235  		tcp_clear_options(&opt);
6236  		opt.user_mss = opt.mss_clamp = 0;
6237  		tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL);
6238  		mss = opt.mss_clamp;
6239  	}
6240  
6241  	if (!tp->syn_fastopen) {
6242  		/* Ignore an unsolicited cookie */
6243  		cookie->len = -1;
6244  	} else if (tp->total_retrans) {
6245  		/* SYN timed out and the SYN-ACK neither has a cookie nor
6246  		 * acknowledges data. Presumably the remote received only
6247  		 * the retransmitted (regular) SYNs: either the original
6248  		 * SYN-data or the corresponding SYN-ACK was dropped.
6249  		 */
6250  		syn_drop = (cookie->len < 0 && data);
6251  	} else if (cookie->len < 0 && !tp->syn_data) {
6252  		/* We requested a cookie but didn't get it. If we did not use
6253  		 * the (old) exp opt format then try so next time (try_exp=1).
6254  		 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
6255  		 */
6256  		try_exp = tp->syn_fastopen_exp ? 2 : 1;
6257  	}
6258  
6259  	tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
6260  
6261  	if (data) { /* Retransmit unacked data in SYN */
6262  		if (tp->total_retrans)
6263  			tp->fastopen_client_fail = TFO_SYN_RETRANSMITTED;
6264  		else
6265  			tp->fastopen_client_fail = TFO_DATA_NOT_ACKED;
6266  		skb_rbtree_walk_from(data)
6267  			 tcp_mark_skb_lost(sk, data);
6268  		tcp_non_congestion_loss_retransmit(sk);
6269  		NET_INC_STATS(sock_net(sk),
6270  				LINUX_MIB_TCPFASTOPENACTIVEFAIL);
6271  		return true;
6272  	}
6273  	tp->syn_data_acked = tp->syn_data;
6274  	if (tp->syn_data_acked) {
6275  		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
6276  		/* SYN-data is counted as two separate packets in tcp_ack() */
6277  		if (tp->delivered > 1)
6278  			--tp->delivered;
6279  	}
6280  
6281  	tcp_fastopen_add_skb(sk, synack);
6282  
6283  	return false;
6284  }
6285  
smc_check_reset_syn(struct tcp_sock * tp)6286  static void smc_check_reset_syn(struct tcp_sock *tp)
6287  {
6288  #if IS_ENABLED(CONFIG_SMC)
6289  	if (static_branch_unlikely(&tcp_have_smc)) {
6290  		if (tp->syn_smc && !tp->rx_opt.smc_ok)
6291  			tp->syn_smc = 0;
6292  	}
6293  #endif
6294  }
6295  
tcp_try_undo_spurious_syn(struct sock * sk)6296  static void tcp_try_undo_spurious_syn(struct sock *sk)
6297  {
6298  	struct tcp_sock *tp = tcp_sk(sk);
6299  	u32 syn_stamp;
6300  
6301  	/* undo_marker is set when SYN or SYNACK times out. The timeout is
6302  	 * spurious if the ACK's timestamp option echo value matches the
6303  	 * original SYN timestamp.
6304  	 */
6305  	syn_stamp = tp->retrans_stamp;
6306  	if (tp->undo_marker && syn_stamp && tp->rx_opt.saw_tstamp &&
6307  	    syn_stamp == tp->rx_opt.rcv_tsecr)
6308  		tp->undo_marker = 0;
6309  }
6310  
tcp_rcv_synsent_state_process(struct sock * sk,struct sk_buff * skb,const struct tcphdr * th)6311  static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
6312  					 const struct tcphdr *th)
6313  {
6314  	struct inet_connection_sock *icsk = inet_csk(sk);
6315  	struct tcp_sock *tp = tcp_sk(sk);
6316  	struct tcp_fastopen_cookie foc = { .len = -1 };
6317  	int saved_clamp = tp->rx_opt.mss_clamp;
6318  	bool fastopen_fail;
6319  	SKB_DR(reason);
6320  
6321  	tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc);
6322  	if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
6323  		tp->rx_opt.rcv_tsecr -= tp->tsoffset;
6324  
6325  	if (th->ack) {
6326  		/* rfc793:
6327  		 * "If the state is SYN-SENT then
6328  		 *    first check the ACK bit
6329  		 *      If the ACK bit is set
6330  		 *	  If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
6331  		 *        a reset (unless the RST bit is set, if so drop
6332  		 *        the segment and return)"
6333  		 */
6334  		if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
6335  		    after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
6336  			/* Previous FIN/ACK or RST/ACK might be ignored. */
6337  			if (icsk->icsk_retransmits == 0)
6338  				inet_csk_reset_xmit_timer(sk,
6339  						ICSK_TIME_RETRANS,
6340  						TCP_TIMEOUT_MIN, TCP_RTO_MAX);
6341  			goto reset_and_undo;
6342  		}
6343  
6344  		if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
6345  		    !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
6346  			     tcp_time_stamp(tp))) {
6347  			NET_INC_STATS(sock_net(sk),
6348  					LINUX_MIB_PAWSACTIVEREJECTED);
6349  			goto reset_and_undo;
6350  		}
6351  
6352  		/* Now ACK is acceptable.
6353  		 *
6354  		 * "If the RST bit is set
6355  		 *    If the ACK was acceptable then signal the user "error:
6356  		 *    connection reset", drop the segment, enter CLOSED state,
6357  		 *    delete TCB, and return."
6358  		 */
6359  
6360  		if (th->rst) {
6361  			tcp_reset(sk, skb);
6362  consume:
6363  			__kfree_skb(skb);
6364  			return 0;
6365  		}
6366  
6367  		/* rfc793:
6368  		 *   "fifth, if neither of the SYN or RST bits is set then
6369  		 *    drop the segment and return."
6370  		 *
6371  		 *    See note below!
6372  		 *                                        --ANK(990513)
6373  		 */
6374  		if (!th->syn) {
6375  			SKB_DR_SET(reason, TCP_FLAGS);
6376  			goto discard_and_undo;
6377  		}
6378  		/* rfc793:
6379  		 *   "If the SYN bit is on ...
6380  		 *    are acceptable then ...
6381  		 *    (our SYN has been ACKed), change the connection
6382  		 *    state to ESTABLISHED..."
6383  		 */
6384  
6385  		tcp_ecn_rcv_synack(tp, th);
6386  
6387  		tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6388  		tcp_try_undo_spurious_syn(sk);
6389  		tcp_ack(sk, skb, FLAG_SLOWPATH);
6390  
6391  		/* Ok.. it's good. Set up sequence numbers and
6392  		 * move to established.
6393  		 */
6394  		WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
6395  		tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
6396  
6397  		/* RFC1323: The window in SYN & SYN/ACK segments is
6398  		 * never scaled.
6399  		 */
6400  		tp->snd_wnd = ntohs(th->window);
6401  
6402  		if (!tp->rx_opt.wscale_ok) {
6403  			tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
6404  			WRITE_ONCE(tp->window_clamp,
6405  				   min(tp->window_clamp, 65535U));
6406  		}
6407  
6408  		if (tp->rx_opt.saw_tstamp) {
6409  			tp->rx_opt.tstamp_ok	   = 1;
6410  			tp->tcp_header_len =
6411  				sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
6412  			tp->advmss	    -= TCPOLEN_TSTAMP_ALIGNED;
6413  			tcp_store_ts_recent(tp);
6414  		} else {
6415  			tp->tcp_header_len = sizeof(struct tcphdr);
6416  		}
6417  
6418  		tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
6419  		tcp_initialize_rcv_mss(sk);
6420  
6421  		/* Remember, tcp_poll() does not lock socket!
6422  		 * Change state from SYN-SENT only after copied_seq
6423  		 * is initialized. */
6424  		WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6425  
6426  		smc_check_reset_syn(tp);
6427  
6428  		smp_mb();
6429  
6430  		tcp_finish_connect(sk, skb);
6431  
6432  		fastopen_fail = (tp->syn_fastopen || tp->syn_data) &&
6433  				tcp_rcv_fastopen_synack(sk, skb, &foc);
6434  
6435  		if (!sock_flag(sk, SOCK_DEAD)) {
6436  			sk->sk_state_change(sk);
6437  			sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6438  		}
6439  		if (fastopen_fail)
6440  			return -1;
6441  		if (sk->sk_write_pending ||
6442  		    READ_ONCE(icsk->icsk_accept_queue.rskq_defer_accept) ||
6443  		    inet_csk_in_pingpong_mode(sk)) {
6444  			/* Save one ACK. Data will be ready after
6445  			 * several ticks, if write_pending is set.
6446  			 *
6447  			 * It may be deleted, but with this feature tcpdumps
6448  			 * look so _wonderfully_ clever, that I was not able
6449  			 * to stand against the temptation 8)     --ANK
6450  			 */
6451  			inet_csk_schedule_ack(sk);
6452  			tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
6453  			inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
6454  						  TCP_DELACK_MAX, TCP_RTO_MAX);
6455  			goto consume;
6456  		}
6457  		tcp_send_ack(sk);
6458  		return -1;
6459  	}
6460  
6461  	/* No ACK in the segment */
6462  
6463  	if (th->rst) {
6464  		/* rfc793:
6465  		 * "If the RST bit is set
6466  		 *
6467  		 *      Otherwise (no ACK) drop the segment and return."
6468  		 */
6469  		SKB_DR_SET(reason, TCP_RESET);
6470  		goto discard_and_undo;
6471  	}
6472  
6473  	/* PAWS check. */
6474  	if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
6475  	    tcp_paws_reject(&tp->rx_opt, 0)) {
6476  		SKB_DR_SET(reason, TCP_RFC7323_PAWS);
6477  		goto discard_and_undo;
6478  	}
6479  	if (th->syn) {
6480  		/* We see SYN without ACK. It is attempt of
6481  		 * simultaneous connect with crossed SYNs.
6482  		 * Particularly, it can be connect to self.
6483  		 */
6484  		tcp_set_state(sk, TCP_SYN_RECV);
6485  
6486  		if (tp->rx_opt.saw_tstamp) {
6487  			tp->rx_opt.tstamp_ok = 1;
6488  			tcp_store_ts_recent(tp);
6489  			tp->tcp_header_len =
6490  				sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
6491  		} else {
6492  			tp->tcp_header_len = sizeof(struct tcphdr);
6493  		}
6494  
6495  		WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
6496  		WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6497  		tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
6498  
6499  		/* RFC1323: The window in SYN & SYN/ACK segments is
6500  		 * never scaled.
6501  		 */
6502  		tp->snd_wnd    = ntohs(th->window);
6503  		tp->snd_wl1    = TCP_SKB_CB(skb)->seq;
6504  		tp->max_window = tp->snd_wnd;
6505  
6506  		tcp_ecn_rcv_syn(tp, th);
6507  
6508  		tcp_mtup_init(sk);
6509  		tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
6510  		tcp_initialize_rcv_mss(sk);
6511  
6512  		tcp_send_synack(sk);
6513  #if 0
6514  		/* Note, we could accept data and URG from this segment.
6515  		 * There are no obstacles to make this (except that we must
6516  		 * either change tcp_recvmsg() to prevent it from returning data
6517  		 * before 3WHS completes per RFC793, or employ TCP Fast Open).
6518  		 *
6519  		 * However, if we ignore data in ACKless segments sometimes,
6520  		 * we have no reasons to accept it sometimes.
6521  		 * Also, seems the code doing it in step6 of tcp_rcv_state_process
6522  		 * is not flawless. So, discard packet for sanity.
6523  		 * Uncomment this return to process the data.
6524  		 */
6525  		return -1;
6526  #else
6527  		goto consume;
6528  #endif
6529  	}
6530  	/* "fifth, if neither of the SYN or RST bits is set then
6531  	 * drop the segment and return."
6532  	 */
6533  
6534  discard_and_undo:
6535  	tcp_clear_options(&tp->rx_opt);
6536  	tp->rx_opt.mss_clamp = saved_clamp;
6537  	tcp_drop_reason(sk, skb, reason);
6538  	return 0;
6539  
6540  reset_and_undo:
6541  	tcp_clear_options(&tp->rx_opt);
6542  	tp->rx_opt.mss_clamp = saved_clamp;
6543  	return 1;
6544  }
6545  
tcp_rcv_synrecv_state_fastopen(struct sock * sk)6546  static void tcp_rcv_synrecv_state_fastopen(struct sock *sk)
6547  {
6548  	struct tcp_sock *tp = tcp_sk(sk);
6549  	struct request_sock *req;
6550  
6551  	/* If we are still handling the SYNACK RTO, see if timestamp ECR allows
6552  	 * undo. If peer SACKs triggered fast recovery, we can't undo here.
6553  	 */
6554  	if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss && !tp->packets_out)
6555  		tcp_try_undo_recovery(sk);
6556  
6557  	tcp_update_rto_time(tp);
6558  	inet_csk(sk)->icsk_retransmits = 0;
6559  	/* In tcp_fastopen_synack_timer() on the first SYNACK RTO we set
6560  	 * retrans_stamp but don't enter CA_Loss, so in case that happened we
6561  	 * need to zero retrans_stamp here to prevent spurious
6562  	 * retransmits_timed_out(). However, if the ACK of our SYNACK caused us
6563  	 * to enter CA_Recovery then we need to leave retrans_stamp as it was
6564  	 * set entering CA_Recovery, for correct retransmits_timed_out() and
6565  	 * undo behavior.
6566  	 */
6567  	tcp_retrans_stamp_cleanup(sk);
6568  
6569  	/* Once we leave TCP_SYN_RECV or TCP_FIN_WAIT_1,
6570  	 * we no longer need req so release it.
6571  	 */
6572  	req = rcu_dereference_protected(tp->fastopen_rsk,
6573  					lockdep_sock_is_held(sk));
6574  	reqsk_fastopen_remove(sk, req, false);
6575  
6576  	/* Re-arm the timer because data may have been sent out.
6577  	 * This is similar to the regular data transmission case
6578  	 * when new data has just been ack'ed.
6579  	 *
6580  	 * (TFO) - we could try to be more aggressive and
6581  	 * retransmitting any data sooner based on when they
6582  	 * are sent out.
6583  	 */
6584  	tcp_rearm_rto(sk);
6585  }
6586  
6587  /*
6588   *	This function implements the receiving procedure of RFC 793 for
6589   *	all states except ESTABLISHED and TIME_WAIT.
6590   *	It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
6591   *	address independent.
6592   */
6593  
tcp_rcv_state_process(struct sock * sk,struct sk_buff * skb)6594  int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
6595  {
6596  	struct tcp_sock *tp = tcp_sk(sk);
6597  	struct inet_connection_sock *icsk = inet_csk(sk);
6598  	const struct tcphdr *th = tcp_hdr(skb);
6599  	struct request_sock *req;
6600  	int queued = 0;
6601  	bool acceptable;
6602  	SKB_DR(reason);
6603  
6604  	switch (sk->sk_state) {
6605  	case TCP_CLOSE:
6606  		SKB_DR_SET(reason, TCP_CLOSE);
6607  		goto discard;
6608  
6609  	case TCP_LISTEN:
6610  		if (th->ack)
6611  			return 1;
6612  
6613  		if (th->rst) {
6614  			SKB_DR_SET(reason, TCP_RESET);
6615  			goto discard;
6616  		}
6617  		if (th->syn) {
6618  			if (th->fin) {
6619  				SKB_DR_SET(reason, TCP_FLAGS);
6620  				goto discard;
6621  			}
6622  			/* It is possible that we process SYN packets from backlog,
6623  			 * so we need to make sure to disable BH and RCU right there.
6624  			 */
6625  			rcu_read_lock();
6626  			local_bh_disable();
6627  			acceptable = icsk->icsk_af_ops->conn_request(sk, skb) >= 0;
6628  			local_bh_enable();
6629  			rcu_read_unlock();
6630  
6631  			if (!acceptable)
6632  				return 1;
6633  			consume_skb(skb);
6634  			return 0;
6635  		}
6636  		SKB_DR_SET(reason, TCP_FLAGS);
6637  		goto discard;
6638  
6639  	case TCP_SYN_SENT:
6640  		tp->rx_opt.saw_tstamp = 0;
6641  		tcp_mstamp_refresh(tp);
6642  		queued = tcp_rcv_synsent_state_process(sk, skb, th);
6643  		if (queued >= 0)
6644  			return queued;
6645  
6646  		/* Do step6 onward by hand. */
6647  		tcp_urg(sk, skb, th);
6648  		__kfree_skb(skb);
6649  		tcp_data_snd_check(sk);
6650  		return 0;
6651  	}
6652  
6653  	tcp_mstamp_refresh(tp);
6654  	tp->rx_opt.saw_tstamp = 0;
6655  	req = rcu_dereference_protected(tp->fastopen_rsk,
6656  					lockdep_sock_is_held(sk));
6657  	if (req) {
6658  		bool req_stolen;
6659  
6660  		WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
6661  		    sk->sk_state != TCP_FIN_WAIT1);
6662  
6663  		if (!tcp_check_req(sk, skb, req, true, &req_stolen)) {
6664  			SKB_DR_SET(reason, TCP_FASTOPEN);
6665  			goto discard;
6666  		}
6667  	}
6668  
6669  	if (!th->ack && !th->rst && !th->syn) {
6670  		SKB_DR_SET(reason, TCP_FLAGS);
6671  		goto discard;
6672  	}
6673  	if (!tcp_validate_incoming(sk, skb, th, 0))
6674  		return 0;
6675  
6676  	/* step 5: check the ACK field */
6677  	acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
6678  				      FLAG_UPDATE_TS_RECENT |
6679  				      FLAG_NO_CHALLENGE_ACK) > 0;
6680  
6681  	if (!acceptable) {
6682  		if (sk->sk_state == TCP_SYN_RECV)
6683  			return 1;	/* send one RST */
6684  		tcp_send_challenge_ack(sk);
6685  		SKB_DR_SET(reason, TCP_OLD_ACK);
6686  		goto discard;
6687  	}
6688  	switch (sk->sk_state) {
6689  	case TCP_SYN_RECV:
6690  		tp->delivered++; /* SYN-ACK delivery isn't tracked in tcp_ack */
6691  		if (!tp->srtt_us)
6692  			tcp_synack_rtt_meas(sk, req);
6693  
6694  		if (req) {
6695  			tcp_rcv_synrecv_state_fastopen(sk);
6696  		} else {
6697  			tcp_try_undo_spurious_syn(sk);
6698  			tp->retrans_stamp = 0;
6699  			tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB,
6700  					  skb);
6701  			WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6702  		}
6703  		smp_mb();
6704  		tcp_set_state(sk, TCP_ESTABLISHED);
6705  		sk->sk_state_change(sk);
6706  
6707  		/* Note, that this wakeup is only for marginal crossed SYN case.
6708  		 * Passively open sockets are not waked up, because
6709  		 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
6710  		 */
6711  		if (sk->sk_socket)
6712  			sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6713  
6714  		tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
6715  		tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
6716  		tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6717  
6718  		if (tp->rx_opt.tstamp_ok)
6719  			tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6720  
6721  		if (!inet_csk(sk)->icsk_ca_ops->cong_control)
6722  			tcp_update_pacing_rate(sk);
6723  
6724  		/* Prevent spurious tcp_cwnd_restart() on first data packet */
6725  		tp->lsndtime = tcp_jiffies32;
6726  
6727  		tcp_initialize_rcv_mss(sk);
6728  		tcp_fast_path_on(tp);
6729  		if (sk->sk_shutdown & SEND_SHUTDOWN)
6730  			tcp_shutdown(sk, SEND_SHUTDOWN);
6731  		break;
6732  
6733  	case TCP_FIN_WAIT1: {
6734  		int tmo;
6735  
6736  		if (req)
6737  			tcp_rcv_synrecv_state_fastopen(sk);
6738  
6739  		if (tp->snd_una != tp->write_seq)
6740  			break;
6741  
6742  		tcp_set_state(sk, TCP_FIN_WAIT2);
6743  		WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | SEND_SHUTDOWN);
6744  
6745  		sk_dst_confirm(sk);
6746  
6747  		if (!sock_flag(sk, SOCK_DEAD)) {
6748  			/* Wake up lingering close() */
6749  			sk->sk_state_change(sk);
6750  			break;
6751  		}
6752  
6753  		if (READ_ONCE(tp->linger2) < 0) {
6754  			tcp_done(sk);
6755  			NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6756  			return 1;
6757  		}
6758  		if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6759  		    after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6760  			/* Receive out of order FIN after close() */
6761  			if (tp->syn_fastopen && th->fin)
6762  				tcp_fastopen_active_disable(sk);
6763  			tcp_done(sk);
6764  			NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6765  			return 1;
6766  		}
6767  
6768  		tmo = tcp_fin_time(sk);
6769  		if (tmo > TCP_TIMEWAIT_LEN) {
6770  			inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
6771  		} else if (th->fin || sock_owned_by_user(sk)) {
6772  			/* Bad case. We could lose such FIN otherwise.
6773  			 * It is not a big problem, but it looks confusing
6774  			 * and not so rare event. We still can lose it now,
6775  			 * if it spins in bh_lock_sock(), but it is really
6776  			 * marginal case.
6777  			 */
6778  			inet_csk_reset_keepalive_timer(sk, tmo);
6779  		} else {
6780  			tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6781  			goto consume;
6782  		}
6783  		break;
6784  	}
6785  
6786  	case TCP_CLOSING:
6787  		if (tp->snd_una == tp->write_seq) {
6788  			tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6789  			goto consume;
6790  		}
6791  		break;
6792  
6793  	case TCP_LAST_ACK:
6794  		if (tp->snd_una == tp->write_seq) {
6795  			tcp_update_metrics(sk);
6796  			tcp_done(sk);
6797  			goto consume;
6798  		}
6799  		break;
6800  	}
6801  
6802  	/* step 6: check the URG bit */
6803  	tcp_urg(sk, skb, th);
6804  
6805  	/* step 7: process the segment text */
6806  	switch (sk->sk_state) {
6807  	case TCP_CLOSE_WAIT:
6808  	case TCP_CLOSING:
6809  	case TCP_LAST_ACK:
6810  		if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
6811  			/* If a subflow has been reset, the packet should not
6812  			 * continue to be processed, drop the packet.
6813  			 */
6814  			if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb))
6815  				goto discard;
6816  			break;
6817  		}
6818  		fallthrough;
6819  	case TCP_FIN_WAIT1:
6820  	case TCP_FIN_WAIT2:
6821  		/* RFC 793 says to queue data in these states,
6822  		 * RFC 1122 says we MUST send a reset.
6823  		 * BSD 4.4 also does reset.
6824  		 */
6825  		if (sk->sk_shutdown & RCV_SHUTDOWN) {
6826  			if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6827  			    after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6828  				NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6829  				tcp_reset(sk, skb);
6830  				return 1;
6831  			}
6832  		}
6833  		fallthrough;
6834  	case TCP_ESTABLISHED:
6835  		tcp_data_queue(sk, skb);
6836  		queued = 1;
6837  		break;
6838  	}
6839  
6840  	/* tcp_data could move socket to TIME-WAIT */
6841  	if (sk->sk_state != TCP_CLOSE) {
6842  		tcp_data_snd_check(sk);
6843  		tcp_ack_snd_check(sk);
6844  	}
6845  
6846  	if (!queued) {
6847  discard:
6848  		tcp_drop_reason(sk, skb, reason);
6849  	}
6850  	return 0;
6851  
6852  consume:
6853  	__kfree_skb(skb);
6854  	return 0;
6855  }
6856  EXPORT_SYMBOL(tcp_rcv_state_process);
6857  
pr_drop_req(struct request_sock * req,__u16 port,int family)6858  static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6859  {
6860  	struct inet_request_sock *ireq = inet_rsk(req);
6861  
6862  	if (family == AF_INET)
6863  		net_dbg_ratelimited("drop open request from %pI4/%u\n",
6864  				    &ireq->ir_rmt_addr, port);
6865  #if IS_ENABLED(CONFIG_IPV6)
6866  	else if (family == AF_INET6)
6867  		net_dbg_ratelimited("drop open request from %pI6/%u\n",
6868  				    &ireq->ir_v6_rmt_addr, port);
6869  #endif
6870  }
6871  
6872  /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6873   *
6874   * If we receive a SYN packet with these bits set, it means a
6875   * network is playing bad games with TOS bits. In order to
6876   * avoid possible false congestion notifications, we disable
6877   * TCP ECN negotiation.
6878   *
6879   * Exception: tcp_ca wants ECN. This is required for DCTCP
6880   * congestion control: Linux DCTCP asserts ECT on all packets,
6881   * including SYN, which is most optimal solution; however,
6882   * others, such as FreeBSD do not.
6883   *
6884   * Exception: At least one of the reserved bits of the TCP header (th->res1) is
6885   * set, indicating the use of a future TCP extension (such as AccECN). See
6886   * RFC8311 §4.3 which updates RFC3168 to allow the development of such
6887   * extensions.
6888   */
tcp_ecn_create_request(struct request_sock * req,const struct sk_buff * skb,const struct sock * listen_sk,const struct dst_entry * dst)6889  static void tcp_ecn_create_request(struct request_sock *req,
6890  				   const struct sk_buff *skb,
6891  				   const struct sock *listen_sk,
6892  				   const struct dst_entry *dst)
6893  {
6894  	const struct tcphdr *th = tcp_hdr(skb);
6895  	const struct net *net = sock_net(listen_sk);
6896  	bool th_ecn = th->ece && th->cwr;
6897  	bool ect, ecn_ok;
6898  	u32 ecn_ok_dst;
6899  
6900  	if (!th_ecn)
6901  		return;
6902  
6903  	ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
6904  	ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6905  	ecn_ok = READ_ONCE(net->ipv4.sysctl_tcp_ecn) || ecn_ok_dst;
6906  
6907  	if (((!ect || th->res1) && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
6908  	    (ecn_ok_dst & DST_FEATURE_ECN_CA) ||
6909  	    tcp_bpf_ca_needs_ecn((struct sock *)req))
6910  		inet_rsk(req)->ecn_ok = 1;
6911  }
6912  
tcp_openreq_init(struct request_sock * req,const struct tcp_options_received * rx_opt,struct sk_buff * skb,const struct sock * sk)6913  static void tcp_openreq_init(struct request_sock *req,
6914  			     const struct tcp_options_received *rx_opt,
6915  			     struct sk_buff *skb, const struct sock *sk)
6916  {
6917  	struct inet_request_sock *ireq = inet_rsk(req);
6918  
6919  	req->rsk_rcv_wnd = 0;		/* So that tcp_send_synack() knows! */
6920  	tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6921  	tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6922  	tcp_rsk(req)->snt_synack = 0;
6923  	tcp_rsk(req)->last_oow_ack_time = 0;
6924  	req->mss = rx_opt->mss_clamp;
6925  	req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6926  	ireq->tstamp_ok = rx_opt->tstamp_ok;
6927  	ireq->sack_ok = rx_opt->sack_ok;
6928  	ireq->snd_wscale = rx_opt->snd_wscale;
6929  	ireq->wscale_ok = rx_opt->wscale_ok;
6930  	ireq->acked = 0;
6931  	ireq->ecn_ok = 0;
6932  	ireq->ir_rmt_port = tcp_hdr(skb)->source;
6933  	ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6934  	ireq->ir_mark = inet_request_mark(sk, skb);
6935  #if IS_ENABLED(CONFIG_SMC)
6936  	ireq->smc_ok = rx_opt->smc_ok && !(tcp_sk(sk)->smc_hs_congested &&
6937  			tcp_sk(sk)->smc_hs_congested(sk));
6938  #endif
6939  }
6940  
inet_reqsk_alloc(const struct request_sock_ops * ops,struct sock * sk_listener,bool attach_listener)6941  struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6942  				      struct sock *sk_listener,
6943  				      bool attach_listener)
6944  {
6945  	struct request_sock *req = reqsk_alloc(ops, sk_listener,
6946  					       attach_listener);
6947  
6948  	if (req) {
6949  		struct inet_request_sock *ireq = inet_rsk(req);
6950  
6951  		ireq->ireq_opt = NULL;
6952  #if IS_ENABLED(CONFIG_IPV6)
6953  		ireq->pktopts = NULL;
6954  #endif
6955  		atomic64_set(&ireq->ir_cookie, 0);
6956  		ireq->ireq_state = TCP_NEW_SYN_RECV;
6957  		write_pnet(&ireq->ireq_net, sock_net(sk_listener));
6958  		ireq->ireq_family = sk_listener->sk_family;
6959  		req->timeout = TCP_TIMEOUT_INIT;
6960  	}
6961  
6962  	return req;
6963  }
6964  EXPORT_SYMBOL(inet_reqsk_alloc);
6965  
6966  /*
6967   * Return true if a syncookie should be sent
6968   */
tcp_syn_flood_action(const struct sock * sk,const char * proto)6969  static bool tcp_syn_flood_action(const struct sock *sk, const char *proto)
6970  {
6971  	struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
6972  	const char *msg = "Dropping request";
6973  	struct net *net = sock_net(sk);
6974  	bool want_cookie = false;
6975  	u8 syncookies;
6976  
6977  	syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies);
6978  
6979  #ifdef CONFIG_SYN_COOKIES
6980  	if (syncookies) {
6981  		msg = "Sending cookies";
6982  		want_cookie = true;
6983  		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
6984  	} else
6985  #endif
6986  		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
6987  
6988  	if (!READ_ONCE(queue->synflood_warned) && syncookies != 2 &&
6989  	    xchg(&queue->synflood_warned, 1) == 0) {
6990  		if (IS_ENABLED(CONFIG_IPV6) && sk->sk_family == AF_INET6) {
6991  			net_info_ratelimited("%s: Possible SYN flooding on port [%pI6c]:%u. %s.\n",
6992  					proto, inet6_rcv_saddr(sk),
6993  					sk->sk_num, msg);
6994  		} else {
6995  			net_info_ratelimited("%s: Possible SYN flooding on port %pI4:%u. %s.\n",
6996  					proto, &sk->sk_rcv_saddr,
6997  					sk->sk_num, msg);
6998  		}
6999  	}
7000  
7001  	return want_cookie;
7002  }
7003  
tcp_reqsk_record_syn(const struct sock * sk,struct request_sock * req,const struct sk_buff * skb)7004  static void tcp_reqsk_record_syn(const struct sock *sk,
7005  				 struct request_sock *req,
7006  				 const struct sk_buff *skb)
7007  {
7008  	if (tcp_sk(sk)->save_syn) {
7009  		u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
7010  		struct saved_syn *saved_syn;
7011  		u32 mac_hdrlen;
7012  		void *base;
7013  
7014  		if (tcp_sk(sk)->save_syn == 2) {  /* Save full header. */
7015  			base = skb_mac_header(skb);
7016  			mac_hdrlen = skb_mac_header_len(skb);
7017  			len += mac_hdrlen;
7018  		} else {
7019  			base = skb_network_header(skb);
7020  			mac_hdrlen = 0;
7021  		}
7022  
7023  		saved_syn = kmalloc(struct_size(saved_syn, data, len),
7024  				    GFP_ATOMIC);
7025  		if (saved_syn) {
7026  			saved_syn->mac_hdrlen = mac_hdrlen;
7027  			saved_syn->network_hdrlen = skb_network_header_len(skb);
7028  			saved_syn->tcp_hdrlen = tcp_hdrlen(skb);
7029  			memcpy(saved_syn->data, base, len);
7030  			req->saved_syn = saved_syn;
7031  		}
7032  	}
7033  }
7034  
7035  /* If a SYN cookie is required and supported, returns a clamped MSS value to be
7036   * used for SYN cookie generation.
7037   */
tcp_get_syncookie_mss(struct request_sock_ops * rsk_ops,const struct tcp_request_sock_ops * af_ops,struct sock * sk,struct tcphdr * th)7038  u16 tcp_get_syncookie_mss(struct request_sock_ops *rsk_ops,
7039  			  const struct tcp_request_sock_ops *af_ops,
7040  			  struct sock *sk, struct tcphdr *th)
7041  {
7042  	struct tcp_sock *tp = tcp_sk(sk);
7043  	u16 mss;
7044  
7045  	if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_syncookies) != 2 &&
7046  	    !inet_csk_reqsk_queue_is_full(sk))
7047  		return 0;
7048  
7049  	if (!tcp_syn_flood_action(sk, rsk_ops->slab_name))
7050  		return 0;
7051  
7052  	if (sk_acceptq_is_full(sk)) {
7053  		NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
7054  		return 0;
7055  	}
7056  
7057  	mss = tcp_parse_mss_option(th, tp->rx_opt.user_mss);
7058  	if (!mss)
7059  		mss = af_ops->mss_clamp;
7060  
7061  	return mss;
7062  }
7063  EXPORT_SYMBOL_GPL(tcp_get_syncookie_mss);
7064  
tcp_conn_request(struct request_sock_ops * rsk_ops,const struct tcp_request_sock_ops * af_ops,struct sock * sk,struct sk_buff * skb)7065  int tcp_conn_request(struct request_sock_ops *rsk_ops,
7066  		     const struct tcp_request_sock_ops *af_ops,
7067  		     struct sock *sk, struct sk_buff *skb)
7068  {
7069  	struct tcp_fastopen_cookie foc = { .len = -1 };
7070  	__u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
7071  	struct tcp_options_received tmp_opt;
7072  	struct tcp_sock *tp = tcp_sk(sk);
7073  	struct net *net = sock_net(sk);
7074  	struct sock *fastopen_sk = NULL;
7075  	struct request_sock *req;
7076  	bool want_cookie = false;
7077  	struct dst_entry *dst;
7078  	struct flowi fl;
7079  	u8 syncookies;
7080  
7081  	syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies);
7082  
7083  	/* TW buckets are converted to open requests without
7084  	 * limitations, they conserve resources and peer is
7085  	 * evidently real one.
7086  	 */
7087  	if ((syncookies == 2 || inet_csk_reqsk_queue_is_full(sk)) && !isn) {
7088  		want_cookie = tcp_syn_flood_action(sk, rsk_ops->slab_name);
7089  		if (!want_cookie)
7090  			goto drop;
7091  	}
7092  
7093  	if (sk_acceptq_is_full(sk)) {
7094  		NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
7095  		goto drop;
7096  	}
7097  
7098  	req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
7099  	if (!req)
7100  		goto drop;
7101  
7102  	req->syncookie = want_cookie;
7103  	tcp_rsk(req)->af_specific = af_ops;
7104  	tcp_rsk(req)->ts_off = 0;
7105  #if IS_ENABLED(CONFIG_MPTCP)
7106  	tcp_rsk(req)->is_mptcp = 0;
7107  #endif
7108  
7109  	tcp_clear_options(&tmp_opt);
7110  	tmp_opt.mss_clamp = af_ops->mss_clamp;
7111  	tmp_opt.user_mss  = tp->rx_opt.user_mss;
7112  	tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0,
7113  			  want_cookie ? NULL : &foc);
7114  
7115  	if (want_cookie && !tmp_opt.saw_tstamp)
7116  		tcp_clear_options(&tmp_opt);
7117  
7118  	if (IS_ENABLED(CONFIG_SMC) && want_cookie)
7119  		tmp_opt.smc_ok = 0;
7120  
7121  	tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
7122  	tcp_openreq_init(req, &tmp_opt, skb, sk);
7123  	inet_rsk(req)->no_srccheck = inet_test_bit(TRANSPARENT, sk);
7124  
7125  	/* Note: tcp_v6_init_req() might override ir_iif for link locals */
7126  	inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
7127  
7128  	dst = af_ops->route_req(sk, skb, &fl, req);
7129  	if (!dst)
7130  		goto drop_and_free;
7131  
7132  	if (tmp_opt.tstamp_ok)
7133  		tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb);
7134  
7135  	if (!want_cookie && !isn) {
7136  		int max_syn_backlog = READ_ONCE(net->ipv4.sysctl_max_syn_backlog);
7137  
7138  		/* Kill the following clause, if you dislike this way. */
7139  		if (!syncookies &&
7140  		    (max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
7141  		     (max_syn_backlog >> 2)) &&
7142  		    !tcp_peer_is_proven(req, dst)) {
7143  			/* Without syncookies last quarter of
7144  			 * backlog is filled with destinations,
7145  			 * proven to be alive.
7146  			 * It means that we continue to communicate
7147  			 * to destinations, already remembered
7148  			 * to the moment of synflood.
7149  			 */
7150  			pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
7151  				    rsk_ops->family);
7152  			goto drop_and_release;
7153  		}
7154  
7155  		isn = af_ops->init_seq(skb);
7156  	}
7157  
7158  	tcp_ecn_create_request(req, skb, sk, dst);
7159  
7160  	if (want_cookie) {
7161  		isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
7162  		if (!tmp_opt.tstamp_ok)
7163  			inet_rsk(req)->ecn_ok = 0;
7164  	}
7165  
7166  	tcp_rsk(req)->snt_isn = isn;
7167  	tcp_rsk(req)->txhash = net_tx_rndhash();
7168  	tcp_rsk(req)->syn_tos = TCP_SKB_CB(skb)->ip_dsfield;
7169  	tcp_openreq_init_rwin(req, sk, dst);
7170  	sk_rx_queue_set(req_to_sk(req), skb);
7171  	if (!want_cookie) {
7172  		tcp_reqsk_record_syn(sk, req, skb);
7173  		fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
7174  	}
7175  	if (fastopen_sk) {
7176  		af_ops->send_synack(fastopen_sk, dst, &fl, req,
7177  				    &foc, TCP_SYNACK_FASTOPEN, skb);
7178  		/* Add the child socket directly into the accept queue */
7179  		if (!inet_csk_reqsk_queue_add(sk, req, fastopen_sk)) {
7180  			reqsk_fastopen_remove(fastopen_sk, req, false);
7181  			bh_unlock_sock(fastopen_sk);
7182  			sock_put(fastopen_sk);
7183  			goto drop_and_free;
7184  		}
7185  		sk->sk_data_ready(sk);
7186  		bh_unlock_sock(fastopen_sk);
7187  		sock_put(fastopen_sk);
7188  	} else {
7189  		tcp_rsk(req)->tfo_listener = false;
7190  		if (!want_cookie) {
7191  			req->timeout = tcp_timeout_init((struct sock *)req);
7192  			if (unlikely(!inet_csk_reqsk_queue_hash_add(sk, req,
7193  								    req->timeout))) {
7194  				reqsk_free(req);
7195  				dst_release(dst);
7196  				return 0;
7197  			}
7198  
7199  		}
7200  		af_ops->send_synack(sk, dst, &fl, req, &foc,
7201  				    !want_cookie ? TCP_SYNACK_NORMAL :
7202  						   TCP_SYNACK_COOKIE,
7203  				    skb);
7204  		if (want_cookie) {
7205  			reqsk_free(req);
7206  			return 0;
7207  		}
7208  	}
7209  	reqsk_put(req);
7210  	return 0;
7211  
7212  drop_and_release:
7213  	dst_release(dst);
7214  drop_and_free:
7215  	__reqsk_free(req);
7216  drop:
7217  	tcp_listendrop(sk);
7218  	return 0;
7219  }
7220  EXPORT_SYMBOL(tcp_conn_request);
7221