#include <core/system.h>
#include <ds/hashmap.h>
#include <ds/queue.h>
#include <ds/ringbuf.h>
#include <net/inet.h>
#include <sys/callout.h>

Data Structures
struct	tcp_hdr
	A TCP segment's header. More...
struct	tcp_pcb
	A connection (its control block, or PCB). More...

Macros
#define	TH_FIN 0x01
#define	TH_SYN 0x02
#define	TH_RST 0x04
#define	TH_PSH 0x08
#define	TH_ACK 0x10
#define	TH_URG 0x20
#define	SEQ_LT(a, b)
#define	SEQ_LEQ(a, b)
#define	SEQ_GT(a, b)
#define	SEQ_GEQ(a, b)
#define	TCP_BUFSIZE 65536
#define	TCP_MSS_DEFAULT 536

Enumerations
enum	tcp_state { TCP_CLOSED , TCP_LISTEN , TCP_SYN_SENT , TCP_SYN_RECEIVED , TCP_ESTABLISHED , TCP_FIN_WAIT_1 , TCP_FIN_WAIT_2 , TCP_CLOSE_WAIT , TCP_CLOSING , TCP_LAST_ACK , TCP_TIME_WAIT }

Functions
void	tcp_input (struct netif netif, struct netbuf nb, const struct ipv4_hdr *ih)
	Takes in a TCP segment that IPv4 received, and passes it to its connection or listener.
void	tcp_send_keepalive (struct tcp_pcb *pcb)
	Sends a keepalive probe: a segment one byte before what the peer has acknowledged (with no data), which the peer answers with an ACK.
void	tcp_output (struct tcp_pcb *pcb)
	Sends as much of the send buffer as the windows allow, in segments of at most mss bytes.
void	tcp_send_ack (struct tcp_pcb *pcb)
	Sends an ACK, now.
void	tcp_send_syn (struct tcp_pcb *pcb)
	Sends the SYN (or, in SYN_RECEIVED, the SYN and ACK), the first time or again, and starts the retransmission timer.
void	tcp_send_rst (struct tcp_pcb *pcb)
	Sends a reset on a connection.
void	tcp_reply_rst (struct netif netif, const struct ipv4_hdr ih, const struct tcp_hdr *th, size_t len)
	Answers a segment that is for no connection with a reset (RFC 9293 3.10.7.1).
void	tcp_rexmt_start (struct tcp_pcb *pcb)
	Starts the retransmission timer, if it isn't running.
void	tcp_rtt_update (struct tcp_pcb *pcb, uint64_t rtt)
	Updates the round trip estimates with a new measurement, and the retransmission timeout from them (RFC 6298).
uint16_t	tcp_rcv_window (struct tcp_pcb *pcb)
	Computes the window to advertise: the room in the receive buffer.
void	tcp_dack (struct tcp_pcb *pcb)
	Acknowledges a received segment: at once if it's the second one not acknowledged yet, or else after DACK_DELAY.
struct tcp_pcb *	tcp_pcb_new (void)
	Makes a connection, CLOSED, with its send and receive buffers, on the list of connections.
void	tcp_pcb_free (struct tcp_pcb *pcb)
	Frees a connection: its timers stop, it leaves the lists, and its buffers and waiting segments are freed.
struct tcp_pcb *	tcp_lookup (uint32_t laddr, uint16_t lport, uint32_t raddr, uint16_t rport)
	Finds a connection by both ends' addresses and ports.
struct tcp_pcb *	tcp_listener (uint32_t laddr, uint16_t lport)
	Finds the listener for a local address and port: one bound to that address, or else one bound to any address.
int	tcp_insert (struct tcp_pcb *pcb)
	Adds a connection to the hashmap of connections, by its addresses and ports, which it must have now.
uint32_t	tcp_iss (struct tcp_pcb *pcb)
	Chooses a connection's initial sequence number (RFC 6528): a clock that ticks every 4 us, plus a hash of the connection's addresses and ports with a secret, so that it can't be guessed.
void	tcp_fin_wait_2 (struct tcp_pcb *pcb)
	Handles our FIN being acknowledged (FIN_WAIT_2): if the socket is closed, the connection waits for the peer's FIN only for a while.
void	tcp_keepalive_reset (struct tcp_pcb *pcb)
	Restarts the keepalive timer, when the peer was heard from.
void	tcp_established (struct tcp_pcb *pcb)
	Moves a connection to ESTABLISHED: its congestion window starts, and a listener's connection is queued for accept().
void	tcp_drop (struct tcp_pcb *pcb, int err)
	Drops a connection that is lost, giving its socket an error.
void	tcp_closed (struct tcp_pcb *pcb)
	Moves a connection to CLOSED: it leaves the hashmap, its timers stop, and its waiters wake up.
void	tcp_time_wait (struct tcp_pcb *pcb)
	Moves a connection to TIME_WAIT, which it leaves after 2 MSL.
void	tcp_wakeup (struct tcp_pcb *pcb)
	Wakes up the threads waiting on a connection (in accept(), connect(), recv() or send()).

Variables
struct tcp_pcb *	tcp_pcbs

Macro Definition Documentation

◆ SEQ_GEQ

#define SEQ_GEQ	(		a,
			b )

Value:

((int32_t) ((a) - (b)) >= 0)

int32_t

signed int int32_t

Definition stddef.h:18

Definition at line 53 of file tcp.h.

Referenced by ack_input().

◆ SEQ_GT

#define SEQ_GT	(		a,
			b )

Value:

((int32_t) ((a) - (b)) > 0)

Definition at line 52 of file tcp.h.

Referenced by ack_input(), retransmit_first(), rexmt_timeout(), rtt_sample(), segment_input(), syn_sent_input(), tcp_output(), tcp_rcv_window(), and window_update().

◆ SEQ_LEQ

#define SEQ_LEQ	(		a,
			b )

Value:

((int32_t) ((a) - (b)) <= 0)

Definition at line 51 of file tcp.h.

Referenced by ack_input(), ooo_drain(), segment_input(), and syn_sent_input().

◆ SEQ_LT

#define SEQ_LT	(		a,
			b )

Value:

((int32_t) ((a) - (b)) < 0)

Definition at line 50 of file tcp.h.

Referenced by ack_input(), ooo_insert(), and segment_input().

◆ TCP_BUFSIZE

#define TCP_BUFSIZE 65536

Definition at line 56 of file tcp.h.

Referenced by socket_tcp_can_write(), socket_tcp_getsockopt(), socket_tcp_send(), tcp_pcb_new(), and tcp_rcv_window().

◆ TCP_MSS_DEFAULT

#define TCP_MSS_DEFAULT 536

Definition at line 59 of file tcp.h.

Referenced by our_mss(), set_mss(), and tcp_pcb_new().

◆ TH_ACK

#define TH_ACK 0x10

Definition at line 31 of file tcp.h.

Referenced by listen_input(), rexmt_timeout(), segment(), segment_input(), syn_sent_input(), tcp_output(), tcp_reply_rst(), tcp_send_ack(), tcp_send_keepalive(), tcp_send_rst(), and tcp_send_syn().

◆ TH_FIN

#define TH_FIN 0x01

Definition at line 27 of file tcp.h.

Referenced by acceptable(), ack_input(), segment_input(), tcp_output(), and tcp_reply_rst().

◆ TH_PSH

#define TH_PSH 0x08

Definition at line 30 of file tcp.h.

Referenced by tcp_output().

◆ TH_RST

#define TH_RST 0x04

Definition at line 29 of file tcp.h.

Referenced by listen_input(), segment_input(), syn_sent_input(), tcp_reply_rst(), and tcp_send_rst().

◆ TH_SYN

#define TH_SYN 0x02

Definition at line 28 of file tcp.h.

Referenced by acceptable(), listen_input(), segment(), segment_input(), syn_sent_input(), tcp_reply_rst(), and tcp_send_syn().

◆ TH_URG

#define TH_URG 0x20

Definition at line 32 of file tcp.h.

Enumeration Type Documentation

◆ tcp_state

enum tcp_state

Enumerator
TCP_CLOSED
TCP_LISTEN
TCP_SYN_SENT
TCP_SYN_RECEIVED
TCP_ESTABLISHED
TCP_FIN_WAIT_1
TCP_FIN_WAIT_2
TCP_CLOSE_WAIT
TCP_CLOSING
TCP_LAST_ACK
TCP_TIME_WAIT

Definition at line 35 of file tcp.h.

Function Documentation

◆ tcp_closed()

void tcp_closed ( struct tcp_pcb * pcb )

Moves a connection to CLOSED: it leaves the hashmap, its timers stop, and its waiters wake up.

It is freed if nothing holds it anymore; otherwise its socket keeps it.

Parameters

pcb	the connection

Definition at line 295 of file tcp_socket.c.

References callout_stop(), conns, tcp_pcb::dack, tcp_pcb::detached, tcp_pcb::hnode, NULL, tcp_pcb::parent, tcp_pcb::rexmt, tcp_pcb::state, TCP_CLOSED, tcp_closed(), tcp_pcb_free(), tcp_wakeup(), tcp_pcb::timer2, and unlink_parent().

Referenced by ack_input(), segment_input(), tcp_closed(), tcp_drop(), and timer2_closed().

◆ tcp_dack()

void tcp_dack ( struct tcp_pcb * pcb )

Acknowledges a received segment: at once if it's the second one not acknowledged yet, or else after DACK_DELAY.

Parameters

pcb	the connection

Definition at line 344 of file tcp_output.c.

References tcp_pcb::ack_pending, tcp_pcb::dack, DACK_DELAY, dack_timeout(), net_callout_set(), callout::pending, tcp_dack(), and tcp_send_ack().

Referenced by segment_input(), and tcp_dack().

◆ tcp_drop()

void tcp_drop	(	struct tcp_pcb *	pcb,
		int	err )

Drops a connection that is lost, giving its socket an error.

Parameters

pcb	the connection
err	the error for the socket (a positive errno), or 0

Definition at line 323 of file tcp_socket.c.

References tcp_pcb::error, tcp_closed(), and tcp_drop().

Referenced by keepalive_timeout(), rexmt_timeout(), segment_input(), socket_tcp_shutdown(), syn_sent_input(), and tcp_drop().

◆ tcp_established()

void tcp_established ( struct tcp_pcb * pcb )

Moves a connection to ESTABLISHED: its congestion window starts, and a listener's connection is queued for accept().

Parameters

pcb	the connection

Definition at line 335 of file tcp_socket.c.

References tcp_pcb::accept_head, tcp_pcb::accept_next, tcp_pcb::accept_tail, tcp_pcb::connected, tcp_pcb::cwnd, MAX, MIN, tcp_pcb::mss, tcp_pcb::parent, tcp_pcb::snd_max, tcp_pcb::snd_nxt, tcp_pcb::snd_una, tcp_pcb::state, TCP_ESTABLISHED, tcp_established(), tcp_keepalive_reset(), and tcp_wakeup().

Referenced by segment_input(), syn_sent_input(), and tcp_established().

◆ tcp_fin_wait_2()

void tcp_fin_wait_2 ( struct tcp_pcb * pcb )

Handles our FIN being acknowledged (FIN_WAIT_2): if the socket is closed, the connection waits for the peer's FIN only for a while.

Parameters

pcb	the connection

Definition at line 387 of file tcp_socket.c.

References tcp_pcb::detached, FIN_WAIT_2_TIME, net_callout_set(), tcp_fin_wait_2(), tcp_pcb::timer2, and timer2_closed().

Referenced by ack_input(), socket_tcp_shutdown(), and tcp_fin_wait_2().

◆ tcp_input()

void tcp_input	(	struct netif *	netif,
		struct netbuf *	nb,
		const struct ipv4_hdr *	ih )

Takes in a TCP segment that IPv4 received, and passes it to its connection or listener.

A segment with a bad checksum or header, or not sent to one of our addresses, is dropped.

Parameters

netif	the interface it came in on
nb	the segment, from its TCP header; it is freed or kept
ih	its IPv4 header

Definition at line 585 of file tcp_input.c.

References tcp_hdr::ack, netbuf::data, ipv4_hdr::dst, tcp_hdr::dst, tcp_hdr::flags, inet_checksum_pseudo(), IP_PROTO_TCP, ipv4_is_local(), netbuf::len, seg::len, listen_input(), netbuf_free(), ntohl, ntohs, tcp_hdr::off, options(), segment_input(), tcp_hdr::seq, ipv4_hdr::src, tcp_hdr::src, tcp_pcb::state, netif::stats, syn_sent_input(), tcp_input(), tcp_listener(), tcp_lookup(), tcp_reply_rst(), TCP_SYN_SENT, and tcp_hdr::win.

Referenced by ipv4_input(), and tcp_input().

◆ tcp_insert()

int tcp_insert ( struct tcp_pcb * pcb )

Adds a connection to the hashmap of connections, by its addresses and ports, which it must have now.

Parameters

pcb	the connection

Returns: 0, -EADDRINUSE if another connection has the same ones, or -ENOMEM

Definition at line 112 of file tcp_socket.c.

References conns, tcp_pcb::hnode, key_eq(), key_hash(), tcp_pcb::laddr, tcp_pcb::lport, tcp_pcb::raddr, tcp_pcb::rport, and tcp_insert().

Referenced by listen_input(), socket_tcp_connect(), and tcp_insert().

◆ tcp_iss()

uint32_t tcp_iss ( struct tcp_pcb * pcb )

Chooses a connection's initial sequence number (RFC 6528): a clock that ticks every 4 us, plus a hash of the connection's addresses and ports with a secret, so that it can't be guessed.

Parameters

pcb	the connection, with its addresses and ports

Returns: the initial sequence number

Definition at line 231 of file tcp_socket.c.

References arch_rtime_ns(), iss_secret, tcp_pcb::laddr, tcp_pcb::lport, memcpy(), tcp_pcb::raddr, random_get(), tcp_pcb::rport, and tcp_iss().

Referenced by listen_input(), socket_tcp_connect(), and tcp_iss().

◆ tcp_keepalive_reset()

void tcp_keepalive_reset ( struct tcp_pcb * pcb )

Restarts the keepalive timer, when the peer was heard from.

Parameters

pcb	the connection

Definition at line 417 of file tcp_socket.c.

References KEEP_IDLE, tcp_pcb::keep_probes, tcp_pcb::keepalive, keepalive_timeout(), net_callout_set(), tcp_pcb::state, TCP_ESTABLISHED, tcp_keepalive_reset(), and tcp_pcb::timer2.

Referenced by segment_input(), socket_tcp_setsockopt(), tcp_established(), and tcp_keepalive_reset().

◆ tcp_listener()

struct tcp_pcb * tcp_listener	(	uint32_t	laddr,
		uint16_t	lport )

Finds the listener for a local address and port: one bound to that address, or else one bound to any address.

Parameters

laddr	the local address
lport	the local port

Returns: the listener, or NULL

Definition at line 138 of file tcp_socket.c.

References tcp_pcb::laddr, tcp_pcb::lport, tcp_pcb::next, NULL, TCP_LISTEN, tcp_listener(), and tcp_pcbs.

Referenced by tcp_input(), and tcp_listener().

◆ tcp_lookup()

struct tcp_pcb * tcp_lookup	(	uint32_t	laddr,
		uint16_t	lport,
		uint32_t	raddr,
		uint16_t	rport )

Finds a connection by both ends' addresses and ports.

Parameters

laddr	the local address
lport	the local port
raddr	the remote address
rport	the remote port

Returns: the connection, or NULL

Definition at line 96 of file tcp_socket.c.

References conns, hashmap_node::entry, key_hash(), tcp_key::laddr, tcp_key::lport, NULL, tcp_key::raddr, tcp_key::rport, and tcp_lookup().

Referenced by tcp_input(), and tcp_lookup().

◆ tcp_output()

void tcp_output ( struct tcp_pcb * pcb )

Sends as much of the send buffer as the windows allow, in segments of at most mss bytes.

A smaller segment goes only if it has the last of the data and nothing is unacknowledged (Nagle's algorithm, unless TCP_NODELAY), and never just because the peer's window is small (the sender's silly window avoidance). The FIN goes with the last of the data. If the peer's window is zero, the persist timer probes it.

Parameters

pcb	the connection

Definition at line 249 of file tcp_output.c.

References arch_rtime_ns(), tcp_pcb::cwnd, tcp_pcb::fin_queued, tcp_hdr::flags, MIN, tcp_pcb::mss, net_callout_set(), tcp_pcb::nodelay, callout::pending, tcp_pcb::persist, tcp_pcb::rexmt, rexmt_timeout(), tcp_pcb::rto, tcp_pcb::rtt_seq, tcp_pcb::rtt_start, tcp_pcb::rtt_timing, segment(), sending(), tcp_hdr::seq, SEQ_GT, tcp_pcb::snd_max, tcp_pcb::snd_nxt, tcp_pcb::snd_una, tcp_pcb::snd_wnd, tcp_pcb::sndbuf, tcp_output(), tcp_rexmt_start(), TH_ACK, TH_FIN, and TH_PSH.

Referenced by ack_input(), retransmit_first(), rexmt_timeout(), segment_input(), send_fin(), socket_tcp_send(), socket_tcp_setsockopt(), and tcp_output().

◆ tcp_pcb_free()

void tcp_pcb_free ( struct tcp_pcb * pcb )

Frees a connection: its timers stop, it leaves the lists, and its buffers and waiting segments are freed.

Parameters

pcb	the connection

Definition at line 196 of file tcp_socket.c.

References callout_stop(), conns, tcp_pcb::dack, tcp_pcb::hnode, kfree(), netbuf_free(), netbuf::next, tcp_pcb::next, tcp_pcb::ooo, tcp_pcb::rcvbuf, tcp_pcb::rexmt, tcp_pcb::sndbuf, tcp_pcb_free(), tcp_pcbs, and tcp_pcb::timer2.

Referenced by listen_input(), socket_tcp_create(), socket_tcp_shutdown(), tcp_closed(), and tcp_pcb_free().

◆ tcp_pcb_new()

struct tcp_pcb * tcp_pcb_new ( void )

Makes a connection, CLOSED, with its send and receive buffers, on the list of connections.

Returns: the connection, or NULL if there was no memory

Definition at line 162 of file tcp_socket.c.

References kfree(), kmalloc(), M_ZERO, tcp_pcb::mss, tcp_pcb::next, NULL, tcp_pcb::rcvbuf, tcp_pcb::rto, RTO_INITIAL, tcp_pcb::sndbuf, tcp_pcb::ssthresh, tcp_pcb::state, TCP_BUFSIZE, TCP_CLOSED, TCP_MSS_DEFAULT, tcp_pcb_new(), and tcp_pcbs.

Referenced by listen_input(), socket_tcp_create(), and tcp_pcb_new().

◆ tcp_rcv_window()

uint16_t tcp_rcv_window ( struct tcp_pcb * pcb )

Computes the window to advertise: the room in the receive buffer.

Its right edge only moves forward by at least a segment (or half the buffer), so that the peer isn't offered small windows (RFC 9293 3.8.6.2.2, the receiver's silly window avoidance), and it never moves back.

Parameters