1// Copyright 2009 The Go Authors. All rights reserved.
2// Use of this source code is governed by a BSD-style
3// license that can be found in the LICENSE file.
4
5package net
6
7import (
8	"context"
9	"internal/itoa"
10	"io"
11	"net/netip"
12	"os"
13	"syscall"
14	"time"
15)
16
17// BUG(mikio): On JS and Windows, the File method of TCPConn and
18// TCPListener is not implemented.
19
20// TCPAddr represents the address of a TCP end point.
21type TCPAddr struct {
22	IP   IP
23	Port int
24	Zone string // IPv6 scoped addressing zone
25}
26
27// AddrPort returns the [TCPAddr] a as a [netip.AddrPort].
28//
29// If a.Port does not fit in a uint16, it's silently truncated.
30//
31// If a is nil, a zero value is returned.
32func (a *TCPAddr) AddrPort() netip.AddrPort {
33	if a == nil {
34		return netip.AddrPort{}
35	}
36	na, _ := netip.AddrFromSlice(a.IP)
37	na = na.WithZone(a.Zone)
38	return netip.AddrPortFrom(na, uint16(a.Port))
39}
40
41// Network returns the address's network name, "tcp".
42func (a *TCPAddr) Network() string { return "tcp" }
43
44func (a *TCPAddr) String() string {
45	if a == nil {
46		return "<nil>"
47	}
48	ip := ipEmptyString(a.IP)
49	if a.Zone != "" {
50		return JoinHostPort(ip+"%"+a.Zone, itoa.Itoa(a.Port))
51	}
52	return JoinHostPort(ip, itoa.Itoa(a.Port))
53}
54
55func (a *TCPAddr) isWildcard() bool {
56	if a == nil || a.IP == nil {
57		return true
58	}
59	return a.IP.IsUnspecified()
60}
61
62func (a *TCPAddr) opAddr() Addr {
63	if a == nil {
64		return nil
65	}
66	return a
67}
68
69// ResolveTCPAddr returns an address of TCP end point.
70//
71// The network must be a TCP network name.
72//
73// If the host in the address parameter is not a literal IP address or
74// the port is not a literal port number, ResolveTCPAddr resolves the
75// address to an address of TCP end point.
76// Otherwise, it parses the address as a pair of literal IP address
77// and port number.
78// The address parameter can use a host name, but this is not
79// recommended, because it will return at most one of the host name's
80// IP addresses.
81//
82// See func [Dial] for a description of the network and address
83// parameters.
84func ResolveTCPAddr(network, address string) (*TCPAddr, error) {
85	switch network {
86	case "tcp", "tcp4", "tcp6":
87	case "": // a hint wildcard for Go 1.0 undocumented behavior
88		network = "tcp"
89	default:
90		return nil, UnknownNetworkError(network)
91	}
92	addrs, err := DefaultResolver.internetAddrList(context.Background(), network, address)
93	if err != nil {
94		return nil, err
95	}
96	return addrs.forResolve(network, address).(*TCPAddr), nil
97}
98
99// TCPAddrFromAddrPort returns addr as a [TCPAddr]. If addr.IsValid() is false,
100// then the returned TCPAddr will contain a nil IP field, indicating an
101// address family-agnostic unspecified address.
102func TCPAddrFromAddrPort(addr netip.AddrPort) *TCPAddr {
103	return &TCPAddr{
104		IP:   addr.Addr().AsSlice(),
105		Zone: addr.Addr().Zone(),
106		Port: int(addr.Port()),
107	}
108}
109
110// TCPConn is an implementation of the [Conn] interface for TCP network
111// connections.
112type TCPConn struct {
113	conn
114}
115
116// KeepAliveConfig contains TCP keep-alive options.
117//
118// If the Idle, Interval, or Count fields are zero, a default value is chosen.
119// If a field is negative, the corresponding socket-level option will be left unchanged.
120//
121// Note that prior to Windows 10 version 1709, neither setting Idle and Interval
122// separately nor changing Count (which is usually 10) is supported.
123// Therefore, it's recommended to set both Idle and Interval to non-negative values
124// in conjunction with a -1 for Count on those old Windows if you intend to customize
125// the TCP keep-alive settings.
126// By contrast, if only one of Idle and Interval is set to a non-negative value,
127// the other will be set to the system default value, and ultimately,
128// set both Idle and Interval to negative values if you want to leave them unchanged.
129//
130// Note that Solaris and its derivatives do not support setting Interval to a non-negative value
131// and Count to a negative value, or vice-versa.
132type KeepAliveConfig struct {
133	// If Enable is true, keep-alive probes are enabled.
134	Enable bool
135
136	// Idle is the time that the connection must be idle before
137	// the first keep-alive probe is sent.
138	// If zero, a default value of 15 seconds is used.
139	Idle time.Duration
140
141	// Interval is the time between keep-alive probes.
142	// If zero, a default value of 15 seconds is used.
143	Interval time.Duration
144
145	// Count is the maximum number of keep-alive probes that
146	// can go unanswered before dropping a connection.
147	// If zero, a default value of 9 is used.
148	Count int
149}
150
151// SyscallConn returns a raw network connection.
152// This implements the [syscall.Conn] interface.
153func (c *TCPConn) SyscallConn() (syscall.RawConn, error) {
154	if !c.ok() {
155		return nil, syscall.EINVAL
156	}
157	return newRawConn(c.fd), nil
158}
159
160// ReadFrom implements the [io.ReaderFrom] ReadFrom method.
161func (c *TCPConn) ReadFrom(r io.Reader) (int64, error) {
162	if !c.ok() {
163		return 0, syscall.EINVAL
164	}
165	n, err := c.readFrom(r)
166	if err != nil && err != io.EOF {
167		err = &OpError{Op: "readfrom", Net: c.fd.net, Source: c.fd.laddr, Addr: c.fd.raddr, Err: err}
168	}
169	return n, err
170}
171
172// WriteTo implements the io.WriterTo WriteTo method.
173func (c *TCPConn) WriteTo(w io.Writer) (int64, error) {
174	if !c.ok() {
175		return 0, syscall.EINVAL
176	}
177	n, err := c.writeTo(w)
178	if err != nil && err != io.EOF {
179		err = &OpError{Op: "writeto", Net: c.fd.net, Source: c.fd.laddr, Addr: c.fd.raddr, Err: err}
180	}
181	return n, err
182}
183
184// CloseRead shuts down the reading side of the TCP connection.
185// Most callers should just use Close.
186func (c *TCPConn) CloseRead() error {
187	if !c.ok() {
188		return syscall.EINVAL
189	}
190	if err := c.fd.closeRead(); err != nil {
191		return &OpError{Op: "close", Net: c.fd.net, Source: c.fd.laddr, Addr: c.fd.raddr, Err: err}
192	}
193	return nil
194}
195
196// CloseWrite shuts down the writing side of the TCP connection.
197// Most callers should just use Close.
198func (c *TCPConn) CloseWrite() error {
199	if !c.ok() {
200		return syscall.EINVAL
201	}
202	if err := c.fd.closeWrite(); err != nil {
203		return &OpError{Op: "close", Net: c.fd.net, Source: c.fd.laddr, Addr: c.fd.raddr, Err: err}
204	}
205	return nil
206}
207
208// SetLinger sets the behavior of Close on a connection which still
209// has data waiting to be sent or to be acknowledged.
210//
211// If sec < 0 (the default), the operating system finishes sending the
212// data in the background.
213//
214// If sec == 0, the operating system discards any unsent or
215// unacknowledged data.
216//
217// If sec > 0, the data is sent in the background as with sec < 0.
218// On some operating systems including Linux, this may cause Close to block
219// until all data has been sent or discarded.
220// On some operating systems after sec seconds have elapsed any remaining
221// unsent data may be discarded.
222func (c *TCPConn) SetLinger(sec int) error {
223	if !c.ok() {
224		return syscall.EINVAL
225	}
226	if err := setLinger(c.fd, sec); err != nil {
227		return &OpError{Op: "set", Net: c.fd.net, Source: c.fd.laddr, Addr: c.fd.raddr, Err: err}
228	}
229	return nil
230}
231
232// SetKeepAlive sets whether the operating system should send
233// keep-alive messages on the connection.
234func (c *TCPConn) SetKeepAlive(keepalive bool) error {
235	if !c.ok() {
236		return syscall.EINVAL
237	}
238	if err := setKeepAlive(c.fd, keepalive); err != nil {
239		return &OpError{Op: "set", Net: c.fd.net, Source: c.fd.laddr, Addr: c.fd.raddr, Err: err}
240	}
241	return nil
242}
243
244// SetKeepAlivePeriod sets the duration the connection needs to
245// remain idle before TCP starts sending keepalive probes.
246//
247// Note that calling this method on Windows prior to Windows 10 version 1709
248// will reset the KeepAliveInterval to the default system value, which is normally 1 second.
249func (c *TCPConn) SetKeepAlivePeriod(d time.Duration) error {
250	if !c.ok() {
251		return syscall.EINVAL
252	}
253	if err := setKeepAliveIdle(c.fd, d); err != nil {
254		return &OpError{Op: "set", Net: c.fd.net, Source: c.fd.laddr, Addr: c.fd.raddr, Err: err}
255	}
256	return nil
257}
258
259// SetNoDelay controls whether the operating system should delay
260// packet transmission in hopes of sending fewer packets (Nagle's
261// algorithm).  The default is true (no delay), meaning that data is
262// sent as soon as possible after a Write.
263func (c *TCPConn) SetNoDelay(noDelay bool) error {
264	if !c.ok() {
265		return syscall.EINVAL
266	}
267	if err := setNoDelay(c.fd, noDelay); err != nil {
268		return &OpError{Op: "set", Net: c.fd.net, Source: c.fd.laddr, Addr: c.fd.raddr, Err: err}
269	}
270	return nil
271}
272
273// MultipathTCP reports whether the ongoing connection is using MPTCP.
274//
275// If Multipath TCP is not supported by the host, by the other peer or
276// intentionally / accidentally filtered out by a device in between, a
277// fallback to TCP will be done. This method does its best to check if
278// MPTCP is still being used or not.
279//
280// On Linux, more conditions are verified on kernels >= v5.16, improving
281// the results.
282func (c *TCPConn) MultipathTCP() (bool, error) {
283	if !c.ok() {
284		return false, syscall.EINVAL
285	}
286	return isUsingMultipathTCP(c.fd), nil
287}
288
289func newTCPConn(fd *netFD, keepAliveIdle time.Duration, keepAliveCfg KeepAliveConfig, preKeepAliveHook func(*netFD), keepAliveHook func(KeepAliveConfig)) *TCPConn {
290	setNoDelay(fd, true)
291	if !keepAliveCfg.Enable && keepAliveIdle >= 0 {
292		keepAliveCfg = KeepAliveConfig{
293			Enable: true,
294			Idle:   keepAliveIdle,
295		}
296	}
297	c := &TCPConn{conn{fd}}
298	if keepAliveCfg.Enable {
299		if preKeepAliveHook != nil {
300			preKeepAliveHook(fd)
301		}
302		c.SetKeepAliveConfig(keepAliveCfg)
303		if keepAliveHook != nil {
304			keepAliveHook(keepAliveCfg)
305		}
306	}
307	return c
308}
309
310// DialTCP acts like [Dial] for TCP networks.
311//
312// The network must be a TCP network name; see func Dial for details.
313//
314// If laddr is nil, a local address is automatically chosen.
315// If the IP field of raddr is nil or an unspecified IP address, the
316// local system is assumed.
317func DialTCP(network string, laddr, raddr *TCPAddr) (*TCPConn, error) {
318	switch network {
319	case "tcp", "tcp4", "tcp6":
320	default:
321		return nil, &OpError{Op: "dial", Net: network, Source: laddr.opAddr(), Addr: raddr.opAddr(), Err: UnknownNetworkError(network)}
322	}
323	if raddr == nil {
324		return nil, &OpError{Op: "dial", Net: network, Source: laddr.opAddr(), Addr: nil, Err: errMissingAddress}
325	}
326	sd := &sysDialer{network: network, address: raddr.String()}
327	c, err := sd.dialTCP(context.Background(), laddr, raddr)
328	if err != nil {
329		return nil, &OpError{Op: "dial", Net: network, Source: laddr.opAddr(), Addr: raddr.opAddr(), Err: err}
330	}
331	return c, nil
332}
333
334// TCPListener is a TCP network listener. Clients should typically
335// use variables of type [Listener] instead of assuming TCP.
336type TCPListener struct {
337	fd *netFD
338	lc ListenConfig
339}
340
341// SyscallConn returns a raw network connection.
342// This implements the [syscall.Conn] interface.
343//
344// The returned RawConn only supports calling Control. Read and
345// Write return an error.
346func (l *TCPListener) SyscallConn() (syscall.RawConn, error) {
347	if !l.ok() {
348		return nil, syscall.EINVAL
349	}
350	return newRawListener(l.fd), nil
351}
352
353// AcceptTCP accepts the next incoming call and returns the new
354// connection.
355func (l *TCPListener) AcceptTCP() (*TCPConn, error) {
356	if !l.ok() {
357		return nil, syscall.EINVAL
358	}
359	c, err := l.accept()
360	if err != nil {
361		return nil, &OpError{Op: "accept", Net: l.fd.net, Source: nil, Addr: l.fd.laddr, Err: err}
362	}
363	return c, nil
364}
365
366// Accept implements the Accept method in the [Listener] interface; it
367// waits for the next call and returns a generic [Conn].
368func (l *TCPListener) Accept() (Conn, error) {
369	if !l.ok() {
370		return nil, syscall.EINVAL
371	}
372	c, err := l.accept()
373	if err != nil {
374		return nil, &OpError{Op: "accept", Net: l.fd.net, Source: nil, Addr: l.fd.laddr, Err: err}
375	}
376	return c, nil
377}
378
379// Close stops listening on the TCP address.
380// Already Accepted connections are not closed.
381func (l *TCPListener) Close() error {
382	if !l.ok() {
383		return syscall.EINVAL
384	}
385	if err := l.close(); err != nil {
386		return &OpError{Op: "close", Net: l.fd.net, Source: nil, Addr: l.fd.laddr, Err: err}
387	}
388	return nil
389}
390
391// Addr returns the listener's network address, a [*TCPAddr].
392// The Addr returned is shared by all invocations of Addr, so
393// do not modify it.
394func (l *TCPListener) Addr() Addr { return l.fd.laddr }
395
396// SetDeadline sets the deadline associated with the listener.
397// A zero time value disables the deadline.
398func (l *TCPListener) SetDeadline(t time.Time) error {
399	if !l.ok() {
400		return syscall.EINVAL
401	}
402	return l.fd.SetDeadline(t)
403}
404
405// File returns a copy of the underlying [os.File].
406// It is the caller's responsibility to close f when finished.
407// Closing l does not affect f, and closing f does not affect l.
408//
409// The returned os.File's file descriptor is different from the
410// connection's. Attempting to change properties of the original
411// using this duplicate may or may not have the desired effect.
412func (l *TCPListener) File() (f *os.File, err error) {
413	if !l.ok() {
414		return nil, syscall.EINVAL
415	}
416	f, err = l.file()
417	if err != nil {
418		return nil, &OpError{Op: "file", Net: l.fd.net, Source: nil, Addr: l.fd.laddr, Err: err}
419	}
420	return
421}
422
423// ListenTCP acts like [Listen] for TCP networks.
424//
425// The network must be a TCP network name; see func Dial for details.
426//
427// If the IP field of laddr is nil or an unspecified IP address,
428// ListenTCP listens on all available unicast and anycast IP addresses
429// of the local system.
430// If the Port field of laddr is 0, a port number is automatically
431// chosen.
432func ListenTCP(network string, laddr *TCPAddr) (*TCPListener, error) {
433	switch network {
434	case "tcp", "tcp4", "tcp6":
435	default:
436		return nil, &OpError{Op: "listen", Net: network, Source: nil, Addr: laddr.opAddr(), Err: UnknownNetworkError(network)}
437	}
438	if laddr == nil {
439		laddr = &TCPAddr{}
440	}
441	sl := &sysListener{network: network, address: laddr.String()}
442	ln, err := sl.listenTCP(context.Background(), laddr)
443	if err != nil {
444		return nil, &OpError{Op: "listen", Net: network, Source: nil, Addr: laddr.opAddr(), Err: err}
445	}
446	return ln, nil
447}
448
449// roundDurationUp rounds d to the next multiple of to.
450func roundDurationUp(d time.Duration, to time.Duration) time.Duration {
451	return (d + to - 1) / to
452}
453