netpoll_kqueue.go

     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  //go:build darwin || dragonfly || freebsd || netbsd || openbsd
     6  
     7  package runtime
     8  
     9  // Integrated network poller (kqueue-based implementation).
    10  
    11  import (
    12  	"internal/goarch"
    13  	"internal/runtime/atomic"
    14  	"unsafe"
    15  )
    16  
    17  var (
    18  	kq             int32         = -1
    19  	netpollWakeSig atomic.Uint32 // used to avoid duplicate calls of netpollBreak
    20  )
    21  
    22  func netpollinit() {
    23  	kq = kqueue()
    24  	if kq < 0 {
    25  		println("runtime: kqueue failed with", -kq)
    26  		throw("runtime: netpollinit failed")
    27  	}
    28  	closeonexec(kq)
    29  	addWakeupEvent(kq)
    30  }
    31  
    32  func netpollopen(fd uintptr, pd *pollDesc) int32 {
    33  	// Arm both EVFILT_READ and EVFILT_WRITE in edge-triggered mode (EV_CLEAR)
    34  	// for the whole fd lifetime. The notifications are automatically unregistered
    35  	// when fd is closed.
    36  	var ev [2]keventt
    37  	*(*uintptr)(unsafe.Pointer(&ev[0].ident)) = fd
    38  	ev[0].filter = _EVFILT_READ
    39  	ev[0].flags = _EV_ADD | _EV_CLEAR
    40  	ev[0].fflags = 0
    41  	ev[0].data = 0
    42  
    43  	if goarch.PtrSize == 4 {
    44  		// We only have a pointer-sized field to store into,
    45  		// so on a 32-bit system we get no sequence protection.
    46  		// TODO(iant): If we notice any problems we could at least
    47  		// steal the low-order 2 bits for a tiny sequence number.
    48  		ev[0].udata = (*byte)(unsafe.Pointer(pd))
    49  	} else {
    50  		tp := taggedPointerPack(unsafe.Pointer(pd), pd.fdseq.Load())
    51  		ev[0].udata = (*byte)(unsafe.Pointer(uintptr(tp)))
    52  	}
    53  	ev[1] = ev[0]
    54  	ev[1].filter = _EVFILT_WRITE
    55  	n := kevent(kq, &ev[0], 2, nil, 0, nil)
    56  	if n < 0 {
    57  		return -n
    58  	}
    59  	return 0
    60  }
    61  
    62  func netpollclose(fd uintptr) int32 {
    63  	// Don't need to unregister because calling close()
    64  	// on fd will remove any kevents that reference the descriptor.
    65  	return 0
    66  }
    67  
    68  func netpollarm(pd *pollDesc, mode int) {
    69  	throw("runtime: unused")
    70  }
    71  
    72  // netpollBreak interrupts a kevent.
    73  func netpollBreak() {
    74  	// Failing to cas indicates there is an in-flight wakeup, so we're done here.
    75  	if !netpollWakeSig.CompareAndSwap(0, 1) {
    76  		return
    77  	}
    78  
    79  	wakeNetpoll(kq)
    80  }
    81  
    82  // netpoll checks for ready network connections.
    83  // Returns list of goroutines that become runnable.
    84  // delay < 0: blocks indefinitely
    85  // delay == 0: does not block, just polls
    86  // delay > 0: block for up to that many nanoseconds
    87  func netpoll(delay int64) (gList, int32) {
    88  	if kq == -1 {
    89  		return gList{}, 0
    90  	}
    91  	var tp *timespec
    92  	var ts timespec
    93  	if delay < 0 {
    94  		tp = nil
    95  	} else if delay == 0 {
    96  		tp = &ts
    97  	} else {
    98  		ts.setNsec(delay)
    99  		if ts.tv_sec > 1e6 {
   100  			// Darwin returns EINVAL if the sleep time is too long.
   101  			ts.tv_sec = 1e6
   102  		}
   103  		tp = &ts
   104  	}
   105  	var events [64]keventt
   106  retry:
   107  	n := kevent(kq, nil, 0, &events[0], int32(len(events)), tp)
   108  	if n < 0 {
   109  		// Ignore the ETIMEDOUT error for now, but try to dive deep and
   110  		// figure out what really happened with n == ETIMEOUT,
   111  		// see https://go.dev/issue/59679 for details.
   112  		if n != -_EINTR && n != -_ETIMEDOUT {
   113  			println("runtime: kevent on fd", kq, "failed with", -n)
   114  			throw("runtime: netpoll failed")
   115  		}
   116  		// If a timed sleep was interrupted, just return to
   117  		// recalculate how long we should sleep now.
   118  		if delay > 0 {
   119  			return gList{}, 0
   120  		}
   121  		goto retry
   122  	}
   123  	var toRun gList
   124  	delta := int32(0)
   125  	for i := 0; i < int(n); i++ {
   126  		ev := &events[i]
   127  
   128  		if isWakeup(ev) {
   129  			if delay != 0 {
   130  				// netpollBreak could be picked up by a nonblocking poll.
   131  				// Only call drainWakeupEvent and reset the netpollWakeSig if blocking.
   132  				drainWakeupEvent(kq)
   133  				netpollWakeSig.Store(0)
   134  			}
   135  			continue
   136  		}
   137  
   138  		var mode int32
   139  		switch ev.filter {
   140  		case _EVFILT_READ:
   141  			mode += 'r'
   142  
   143  			// On some systems when the read end of a pipe
   144  			// is closed the write end will not get a
   145  			// _EVFILT_WRITE event, but will get a
   146  			// _EVFILT_READ event with EV_EOF set.
   147  			// Note that setting 'w' here just means that we
   148  			// will wake up a goroutine waiting to write;
   149  			// that goroutine will try the write again,
   150  			// and the appropriate thing will happen based
   151  			// on what that write returns (success, EPIPE, EAGAIN).
   152  			if ev.flags&_EV_EOF != 0 {
   153  				mode += 'w'
   154  			}
   155  		case _EVFILT_WRITE:
   156  			mode += 'w'
   157  		}
   158  		if mode != 0 {
   159  			var pd *pollDesc
   160  			var tag uintptr
   161  			if goarch.PtrSize == 4 {
   162  				// No sequence protection on 32-bit systems.
   163  				// See netpollopen for details.
   164  				pd = (*pollDesc)(unsafe.Pointer(ev.udata))
   165  				tag = 0
   166  			} else {
   167  				tp := taggedPointer(uintptr(unsafe.Pointer(ev.udata)))
   168  				pd = (*pollDesc)(tp.pointer())
   169  				tag = tp.tag()
   170  				if pd.fdseq.Load() != tag {
   171  					continue
   172  				}
   173  			}
   174  			pd.setEventErr(ev.flags == _EV_ERROR, tag)
   175  			delta += netpollready(&toRun, pd, mode)
   176  		}
   177  	}
   178  	return toRun, delta
   179  }
   180
View as plain text