1// Copyright 2013 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
5// +build darwin dragonfly freebsd netbsd openbsd
6
7package runtime
8
9// Integrated network poller (kqueue-based implementation).
10
11import "unsafe"
12
13//extern kqueue
14func kqueue() int32
15
16//go:noescape
17//extern kevent
18func kevent(kq int32, ch *keventt, nch uintptr, ev *keventt, nev uintptr, ts *timespec) int32
19
20var (
21	kq int32 = -1
22
23	netpollBreakRd, netpollBreakWr uintptr // for netpollBreak
24)
25
26func netpollinit() {
27	kq = kqueue()
28	if kq < 0 {
29		println("netpollinit: kqueue failed with", errno())
30		throw("runtime: netpollinit failed")
31	}
32	closeonexec(kq)
33	r, w, errno := nonblockingPipe()
34	if errno != 0 {
35		println("runtime: pipe failed with", -errno)
36		throw("runtime: pipe failed")
37	}
38	ev := keventt{
39		filter: _EVFILT_READ,
40		flags:  _EV_ADD,
41	}
42	*(*uintptr)(unsafe.Pointer(&ev.ident)) = uintptr(r)
43	n := kevent(kq, &ev, 1, nil, 0, nil)
44	if n < 0 {
45		println("runtime: kevent failed with", -n)
46		throw("runtime: kevent failed")
47	}
48	netpollBreakRd = uintptr(r)
49	netpollBreakWr = uintptr(w)
50}
51
52func netpollIsPollDescriptor(fd uintptr) bool {
53	return fd == uintptr(kq) || fd == netpollBreakRd || fd == netpollBreakWr
54}
55
56func netpollopen(fd uintptr, pd *pollDesc) int32 {
57	// Arm both EVFILT_READ and EVFILT_WRITE in edge-triggered mode (EV_CLEAR)
58	// for the whole fd lifetime. The notifications are automatically unregistered
59	// when fd is closed.
60	var ev [2]keventt
61	*(*uintptr)(unsafe.Pointer(&ev[0].ident)) = fd
62	ev[0].filter = _EVFILT_READ
63	ev[0].flags = _EV_ADD | _EV_CLEAR
64	ev[0].fflags = 0
65	ev[0].data = 0
66	ev[0].udata = (*byte)(unsafe.Pointer(pd))
67	ev[1] = ev[0]
68	ev[1].filter = _EVFILT_WRITE
69	n := kevent(kq, &ev[0], 2, nil, 0, nil)
70	if n < 0 {
71		return int32(errno())
72	}
73	return 0
74}
75
76func netpollclose(fd uintptr) int32 {
77	// Don't need to unregister because calling close()
78	// on fd will remove any kevents that reference the descriptor.
79	return 0
80}
81
82func netpollarm(pd *pollDesc, mode int) {
83	throw("runtime: unused")
84}
85
86// netpollBreak interrupts an epollwait.
87func netpollBreak() {
88	for {
89		var b byte
90		n := write(netpollBreakWr, unsafe.Pointer(&b), 1)
91		if n == 1 || n == -_EAGAIN {
92			break
93		}
94		if n == -_EINTR {
95			continue
96		}
97		println("runtime: netpollBreak write failed with", -n)
98		throw("runtime: netpollBreak write failed")
99	}
100}
101
102// netpoll checks for ready network connections.
103// Returns list of goroutines that become runnable.
104// delay < 0: blocks indefinitely
105// delay == 0: does not block, just polls
106// delay > 0: block for up to that many nanoseconds
107func netpoll(delay int64) gList {
108	if kq == -1 {
109		return gList{}
110	}
111	var tp *timespec
112	var ts timespec
113	if delay < 0 {
114		tp = nil
115	} else if delay == 0 {
116		tp = &ts
117	} else {
118		ts.setNsec(delay)
119		if ts.tv_sec > 1e6 {
120			// Darwin returns EINVAL if the sleep time is too long.
121			ts.tv_sec = 1e6
122		}
123		tp = &ts
124	}
125	var events [64]keventt
126retry:
127	n := kevent(kq, nil, 0, &events[0], uintptr(len(events)), tp)
128	if n < 0 {
129		e := errno()
130		if e != _EINTR {
131			println("runtime: kevent on fd", kq, "failed with", e)
132			throw("runtime: netpoll failed")
133		}
134		// If a timed sleep was interrupted, just return to
135		// recalculate how long we should sleep now.
136		if delay > 0 {
137			return gList{}
138		}
139		goto retry
140	}
141	var toRun gList
142	for i := 0; i < int(n); i++ {
143		ev := &events[i]
144
145		if uintptr(ev.ident) == netpollBreakRd {
146			if ev.filter != _EVFILT_READ {
147				println("runtime: netpoll: break fd ready for", ev.filter)
148				throw("runtime: netpoll: break fd ready for something unexpected")
149			}
150			if delay != 0 {
151				// netpollBreak could be picked up by a
152				// nonblocking poll. Only read the byte
153				// if blocking.
154				var tmp [16]byte
155				read(int32(netpollBreakRd), noescape(unsafe.Pointer(&tmp[0])), int32(len(tmp)))
156			}
157			continue
158		}
159
160		var mode int32
161		switch ev.filter {
162		case _EVFILT_READ:
163			mode += 'r'
164
165			// On some systems when the read end of a pipe
166			// is closed the write end will not get a
167			// _EVFILT_WRITE event, but will get a
168			// _EVFILT_READ event with EV_EOF set.
169			// Note that setting 'w' here just means that we
170			// will wake up a goroutine waiting to write;
171			// that goroutine will try the write again,
172			// and the appropriate thing will happen based
173			// on what that write returns (success, EPIPE, EAGAIN).
174			if ev.flags&_EV_EOF != 0 {
175				mode += 'w'
176			}
177		case _EVFILT_WRITE:
178			mode += 'w'
179		}
180		if mode != 0 {
181			pd := (*pollDesc)(unsafe.Pointer(ev.udata))
182			pd.everr = false
183			if ev.flags == _EV_ERROR {
184				pd.everr = true
185			}
186			netpollready(&toRun, pd, mode)
187		}
188	}
189	return toRun
190}
191