1 // Copyright 2014 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 package runtime 6 7 import ( 8 "internal/goarch" 9 "internal/runtime/atomic" 10 "unsafe" 11 ) 12 13 // Solaris runtime-integrated network poller. 14 // 15 // Solaris uses event ports for scalable network I/O. Event 16 // ports are level-triggered, unlike epoll and kqueue which 17 // can be configured in both level-triggered and edge-triggered 18 // mode. Level triggering means we have to keep track of a few things 19 // ourselves. After we receive an event for a file descriptor, 20 // it's our responsibility to ask again to be notified for future 21 // events for that descriptor. When doing this we must keep track of 22 // what kind of events the goroutines are currently interested in, 23 // for example a fd may be open both for reading and writing. 24 // 25 // A description of the high level operation of this code 26 // follows. Networking code will get a file descriptor by some means 27 // and will register it with the netpolling mechanism by a code path 28 // that eventually calls runtime·netpollopen. runtime·netpollopen 29 // calls port_associate with an empty event set. That means that we 30 // will not receive any events at this point. The association needs 31 // to be done at this early point because we need to process the I/O 32 // readiness notification at some point in the future. If I/O becomes 33 // ready when nobody is listening, when we finally care about it, 34 // nobody will tell us anymore. 35 // 36 // Beside calling runtime·netpollopen, the networking code paths 37 // will call runtime·netpollarm each time goroutines are interested 38 // in doing network I/O. Because now we know what kind of I/O we 39 // are interested in (reading/writing), we can call port_associate 40 // passing the correct type of event set (POLLIN/POLLOUT). As we made 41 // sure to have already associated the file descriptor with the port, 42 // when we now call port_associate, we will unblock the main poller 43 // loop (in runtime·netpoll) right away if the socket is actually 44 // ready for I/O. 45 // 46 // The main poller loop runs in its own thread waiting for events 47 // using port_getn. When an event happens, it will tell the scheduler 48 // about it using runtime·netpollready. Besides doing this, it must 49 // also re-associate the events that were not part of this current 50 // notification with the file descriptor. Failing to do this would 51 // mean each notification will prevent concurrent code using the 52 // same file descriptor in parallel. 53 // 54 // The logic dealing with re-associations is encapsulated in 55 // runtime·netpollupdate. This function takes care to associate the 56 // descriptor only with the subset of events that were previously 57 // part of the association, except the one that just happened. We 58 // can't re-associate with that right away, because event ports 59 // are level triggered so it would cause a busy loop. Instead, that 60 // association is effected only by the runtime·netpollarm code path, 61 // when Go code actually asks for I/O. 62 // 63 // The open and arming mechanisms are serialized using the lock 64 // inside PollDesc. This is required because the netpoll loop runs 65 // asynchronously in respect to other Go code and by the time we get 66 // to call port_associate to update the association in the loop, the 67 // file descriptor might have been closed and reopened already. The 68 // lock allows runtime·netpollupdate to be called synchronously from 69 // the loop thread while preventing other threads operating to the 70 // same PollDesc, so once we unblock in the main loop, until we loop 71 // again we know for sure we are always talking about the same file 72 // descriptor and can safely access the data we want (the event set). 73 74 //go:cgo_import_dynamic libc_port_create port_create "libc.so" 75 //go:cgo_import_dynamic libc_port_associate port_associate "libc.so" 76 //go:cgo_import_dynamic libc_port_dissociate port_dissociate "libc.so" 77 //go:cgo_import_dynamic libc_port_getn port_getn "libc.so" 78 //go:cgo_import_dynamic libc_port_alert port_alert "libc.so" 79 80 //go:linkname libc_port_create libc_port_create 81 //go:linkname libc_port_associate libc_port_associate 82 //go:linkname libc_port_dissociate libc_port_dissociate 83 //go:linkname libc_port_getn libc_port_getn 84 //go:linkname libc_port_alert libc_port_alert 85 86 var ( 87 libc_port_create, 88 libc_port_associate, 89 libc_port_dissociate, 90 libc_port_getn, 91 libc_port_alert libcFunc 92 netpollWakeSig atomic.Uint32 // used to avoid duplicate calls of netpollBreak 93 ) 94 95 func errno() int32 { 96 return *getg().m.perrno 97 } 98 99 func port_create() int32 { 100 return int32(sysvicall0(&libc_port_create)) 101 } 102 103 func port_associate(port, source int32, object uintptr, events uint32, user uintptr) int32 { 104 return int32(sysvicall5(&libc_port_associate, uintptr(port), uintptr(source), object, uintptr(events), user)) 105 } 106 107 func port_dissociate(port, source int32, object uintptr) int32 { 108 return int32(sysvicall3(&libc_port_dissociate, uintptr(port), uintptr(source), object)) 109 } 110 111 func port_getn(port int32, evs *portevent, max uint32, nget *uint32, timeout *timespec) int32 { 112 return int32(sysvicall5(&libc_port_getn, uintptr(port), uintptr(unsafe.Pointer(evs)), uintptr(max), uintptr(unsafe.Pointer(nget)), uintptr(unsafe.Pointer(timeout)))) 113 } 114 115 func port_alert(port int32, flags, events uint32, user uintptr) int32 { 116 return int32(sysvicall4(&libc_port_alert, uintptr(port), uintptr(flags), uintptr(events), user)) 117 } 118 119 var portfd int32 = -1 120 121 func netpollinit() { 122 portfd = port_create() 123 if portfd >= 0 { 124 closeonexec(portfd) 125 return 126 } 127 128 print("runtime: port_create failed (errno=", errno(), ")\n") 129 throw("runtime: netpollinit failed") 130 } 131 132 func netpollIsPollDescriptor(fd uintptr) bool { 133 return fd == uintptr(portfd) 134 } 135 136 func netpollopen(fd uintptr, pd *pollDesc) int32 { 137 lock(&pd.lock) 138 // We don't register for any specific type of events yet, that's 139 // netpollarm's job. We merely ensure we call port_associate before 140 // asynchronous connect/accept completes, so when we actually want 141 // to do any I/O, the call to port_associate (from netpollarm, 142 // with the interested event set) will unblock port_getn right away 143 // because of the I/O readiness notification. 144 pd.user = 0 145 tp := taggedPointerPack(unsafe.Pointer(pd), pd.fdseq.Load()) 146 // Note that this won't work on a 32-bit system, 147 // as taggedPointer is always 64-bits but uintptr will be 32 bits. 148 // Fortunately we only support Solaris on amd64. 149 if goarch.PtrSize != 8 { 150 throw("runtime: netpollopen: unsupported pointer size") 151 } 152 r := port_associate(portfd, _PORT_SOURCE_FD, fd, 0, uintptr(tp)) 153 unlock(&pd.lock) 154 return r 155 } 156 157 func netpollclose(fd uintptr) int32 { 158 return port_dissociate(portfd, _PORT_SOURCE_FD, fd) 159 } 160 161 // Updates the association with a new set of interested events. After 162 // this call, port_getn will return one and only one event for that 163 // particular descriptor, so this function needs to be called again. 164 func netpollupdate(pd *pollDesc, set, clear uint32) { 165 if pd.info().closing() { 166 return 167 } 168 169 old := pd.user 170 events := (old & ^clear) | set 171 if old == events { 172 return 173 } 174 175 tp := taggedPointerPack(unsafe.Pointer(pd), pd.fdseq.Load()) 176 if events != 0 && port_associate(portfd, _PORT_SOURCE_FD, pd.fd, events, uintptr(tp)) != 0 { 177 print("runtime: port_associate failed (errno=", errno(), ")\n") 178 throw("runtime: netpollupdate failed") 179 } 180 pd.user = events 181 } 182 183 // subscribe the fd to the port such that port_getn will return one event. 184 func netpollarm(pd *pollDesc, mode int) { 185 lock(&pd.lock) 186 switch mode { 187 case 'r': 188 netpollupdate(pd, _POLLIN, 0) 189 case 'w': 190 netpollupdate(pd, _POLLOUT, 0) 191 default: 192 throw("runtime: bad mode") 193 } 194 unlock(&pd.lock) 195 } 196 197 // netpollBreak interrupts a port_getn wait. 198 func netpollBreak() { 199 // Failing to cas indicates there is an in-flight wakeup, so we're done here. 200 if !netpollWakeSig.CompareAndSwap(0, 1) { 201 return 202 } 203 204 // Use port_alert to put portfd into alert mode. 205 // This will wake up all threads sleeping in port_getn on portfd, 206 // and cause their calls to port_getn to return immediately. 207 // Further, until portfd is taken out of alert mode, 208 // all calls to port_getn will return immediately. 209 if port_alert(portfd, _PORT_ALERT_UPDATE, _POLLHUP, uintptr(unsafe.Pointer(&portfd))) < 0 { 210 if e := errno(); e != _EBUSY { 211 println("runtime: port_alert failed with", e) 212 throw("runtime: netpoll: port_alert failed") 213 } 214 } 215 } 216 217 // netpoll checks for ready network connections. 218 // Returns list of goroutines that become runnable. 219 // delay < 0: blocks indefinitely 220 // delay == 0: does not block, just polls 221 // delay > 0: block for up to that many nanoseconds 222 func netpoll(delay int64) (gList, int32) { 223 if portfd == -1 { 224 return gList{}, 0 225 } 226 227 var wait *timespec 228 var ts timespec 229 if delay < 0 { 230 wait = nil 231 } else if delay == 0 { 232 wait = &ts 233 } else { 234 ts.setNsec(delay) 235 if ts.tv_sec > 1e6 { 236 // An arbitrary cap on how long to wait for a timer. 237 // 1e6 s == ~11.5 days. 238 ts.tv_sec = 1e6 239 } 240 wait = &ts 241 } 242 243 var events [128]portevent 244 retry: 245 var n uint32 = 1 246 r := port_getn(portfd, &events[0], uint32(len(events)), &n, wait) 247 e := errno() 248 if r < 0 && e == _ETIME && n > 0 { 249 // As per port_getn(3C), an ETIME failure does not preclude the 250 // delivery of some number of events. Treat a timeout failure 251 // with delivered events as a success. 252 r = 0 253 } 254 if r < 0 { 255 if e != _EINTR && e != _ETIME { 256 print("runtime: port_getn on fd ", portfd, " failed (errno=", e, ")\n") 257 throw("runtime: netpoll failed") 258 } 259 // If a timed sleep was interrupted and there are no events, 260 // just return to recalculate how long we should sleep now. 261 if delay > 0 { 262 return gList{}, 0 263 } 264 goto retry 265 } 266 267 var toRun gList 268 delta := int32(0) 269 for i := 0; i < int(n); i++ { 270 ev := &events[i] 271 272 if ev.portev_source == _PORT_SOURCE_ALERT { 273 if ev.portev_events != _POLLHUP || unsafe.Pointer(ev.portev_user) != unsafe.Pointer(&portfd) { 274 throw("runtime: netpoll: bad port_alert wakeup") 275 } 276 if delay != 0 { 277 // Now that a blocking call to netpoll 278 // has seen the alert, take portfd 279 // back out of alert mode. 280 // See the comment in netpollBreak. 281 if port_alert(portfd, 0, 0, 0) < 0 { 282 e := errno() 283 println("runtime: port_alert failed with", e) 284 throw("runtime: netpoll: port_alert failed") 285 } 286 netpollWakeSig.Store(0) 287 } 288 continue 289 } 290 291 if ev.portev_events == 0 { 292 continue 293 } 294 295 tp := taggedPointer(uintptr(unsafe.Pointer(ev.portev_user))) 296 pd := (*pollDesc)(tp.pointer()) 297 if pd.fdseq.Load() != tp.tag() { 298 continue 299 } 300 301 var mode, clear int32 302 if (ev.portev_events & (_POLLIN | _POLLHUP | _POLLERR)) != 0 { 303 mode += 'r' 304 clear |= _POLLIN 305 } 306 if (ev.portev_events & (_POLLOUT | _POLLHUP | _POLLERR)) != 0 { 307 mode += 'w' 308 clear |= _POLLOUT 309 } 310 // To effect edge-triggered events, we need to be sure to 311 // update our association with whatever events were not 312 // set with the event. For example if we are registered 313 // for POLLIN|POLLOUT, and we get POLLIN, besides waking 314 // the goroutine interested in POLLIN we have to not forget 315 // about the one interested in POLLOUT. 316 if clear != 0 { 317 lock(&pd.lock) 318 netpollupdate(pd, 0, uint32(clear)) 319 unlock(&pd.lock) 320 } 321 322 if mode != 0 { 323 // TODO(mikio): Consider implementing event 324 // scanning error reporting once we are sure 325 // about the event port on SmartOS. 326 // 327 // See golang.org/x/issue/30840. 328 delta += netpollready(&toRun, pd, mode) 329 } 330 } 331 332 return toRun, delta 333 } 334