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 linux nacl netbsd openbsd solaris windows 6 7 package runtime 8 9 import "unsafe" 10 11 // Integrated network poller (platform-independent part). 12 // A particular implementation (epoll/kqueue) must define the following functions: 13 // func netpollinit() // to initialize the poller 14 // func netpollopen(fd uintptr, pd *pollDesc) int32 // to arm edge-triggered notifications 15 // and associate fd with pd. 16 // An implementation must call the following function to denote that the pd is ready. 17 // func netpollready(gpp **g, pd *pollDesc, mode int32) 18 19 // pollDesc contains 2 binary semaphores, rg and wg, to park reader and writer 20 // goroutines respectively. The semaphore can be in the following states: 21 // pdReady - io readiness notification is pending; 22 // a goroutine consumes the notification by changing the state to nil. 23 // pdWait - a goroutine prepares to park on the semaphore, but not yet parked; 24 // the goroutine commits to park by changing the state to G pointer, 25 // or, alternatively, concurrent io notification changes the state to READY, 26 // or, alternatively, concurrent timeout/close changes the state to nil. 27 // G pointer - the goroutine is blocked on the semaphore; 28 // io notification or timeout/close changes the state to READY or nil respectively 29 // and unparks the goroutine. 30 // nil - nothing of the above. 31 const ( 32 pdReady uintptr = 1 33 pdWait uintptr = 2 34 ) 35 36 const pollBlockSize = 4 * 1024 37 38 // Network poller descriptor. 39 type pollDesc struct { 40 link *pollDesc // in pollcache, protected by pollcache.lock 41 42 // The lock protects pollOpen, pollSetDeadline, pollUnblock and deadlineimpl operations. 43 // This fully covers seq, rt and wt variables. fd is constant throughout the PollDesc lifetime. 44 // pollReset, pollWait, pollWaitCanceled and runtimenetpollready (IO readiness notification) 45 // proceed w/o taking the lock. So closing, rg, rd, wg and wd are manipulated 46 // in a lock-free way by all operations. 47 // NOTE(dvyukov): the following code uses uintptr to store *g (rg/wg), 48 // that will blow up when GC starts moving objects. 49 lock mutex // protects the following fields 50 fd uintptr 51 closing bool 52 seq uintptr // protects from stale timers and ready notifications 53 rg uintptr // pdReady, pdWait, G waiting for read or nil 54 rt timer // read deadline timer (set if rt.f != nil) 55 rd int64 // read deadline 56 wg uintptr // pdReady, pdWait, G waiting for write or nil 57 wt timer // write deadline timer 58 wd int64 // write deadline 59 user uint32 // user settable cookie 60 } 61 62 type pollCache struct { 63 lock mutex 64 first *pollDesc 65 // PollDesc objects must be type-stable, 66 // because we can get ready notification from epoll/kqueue 67 // after the descriptor is closed/reused. 68 // Stale notifications are detected using seq variable, 69 // seq is incremented when deadlines are changed or descriptor is reused. 70 } 71 72 var ( 73 netpollInited uint32 74 pollcache pollCache 75 ) 76 77 //go:linkname net_runtime_pollServerInit net.runtime_pollServerInit 78 func net_runtime_pollServerInit() { 79 netpollinit() 80 atomicstore(&netpollInited, 1) 81 } 82 83 func netpollinited() bool { 84 return atomicload(&netpollInited) != 0 85 } 86 87 //go:linkname net_runtime_pollOpen net.runtime_pollOpen 88 func net_runtime_pollOpen(fd uintptr) (*pollDesc, int) { 89 pd := pollcache.alloc() 90 lock(&pd.lock) 91 if pd.wg != 0 && pd.wg != pdReady { 92 throw("netpollOpen: blocked write on free descriptor") 93 } 94 if pd.rg != 0 && pd.rg != pdReady { 95 throw("netpollOpen: blocked read on free descriptor") 96 } 97 pd.fd = fd 98 pd.closing = false 99 pd.seq++ 100 pd.rg = 0 101 pd.rd = 0 102 pd.wg = 0 103 pd.wd = 0 104 unlock(&pd.lock) 105 106 var errno int32 107 errno = netpollopen(fd, pd) 108 return pd, int(errno) 109 } 110 111 //go:linkname net_runtime_pollClose net.runtime_pollClose 112 func net_runtime_pollClose(pd *pollDesc) { 113 if !pd.closing { 114 throw("netpollClose: close w/o unblock") 115 } 116 if pd.wg != 0 && pd.wg != pdReady { 117 throw("netpollClose: blocked write on closing descriptor") 118 } 119 if pd.rg != 0 && pd.rg != pdReady { 120 throw("netpollClose: blocked read on closing descriptor") 121 } 122 netpollclose(uintptr(pd.fd)) 123 pollcache.free(pd) 124 } 125 126 func (c *pollCache) free(pd *pollDesc) { 127 lock(&c.lock) 128 pd.link = c.first 129 c.first = pd 130 unlock(&c.lock) 131 } 132 133 //go:linkname net_runtime_pollReset net.runtime_pollReset 134 func net_runtime_pollReset(pd *pollDesc, mode int) int { 135 err := netpollcheckerr(pd, int32(mode)) 136 if err != 0 { 137 return err 138 } 139 if mode == 'r' { 140 pd.rg = 0 141 } else if mode == 'w' { 142 pd.wg = 0 143 } 144 return 0 145 } 146 147 //go:linkname net_runtime_pollWait net.runtime_pollWait 148 func net_runtime_pollWait(pd *pollDesc, mode int) int { 149 err := netpollcheckerr(pd, int32(mode)) 150 if err != 0 { 151 return err 152 } 153 // As for now only Solaris uses level-triggered IO. 154 if GOOS == "solaris" { 155 netpollarm(pd, mode) 156 } 157 for !netpollblock(pd, int32(mode), false) { 158 err = netpollcheckerr(pd, int32(mode)) 159 if err != 0 { 160 return err 161 } 162 // Can happen if timeout has fired and unblocked us, 163 // but before we had a chance to run, timeout has been reset. 164 // Pretend it has not happened and retry. 165 } 166 return 0 167 } 168 169 //go:linkname net_runtime_pollWaitCanceled net.runtime_pollWaitCanceled 170 func net_runtime_pollWaitCanceled(pd *pollDesc, mode int) { 171 // This function is used only on windows after a failed attempt to cancel 172 // a pending async IO operation. Wait for ioready, ignore closing or timeouts. 173 for !netpollblock(pd, int32(mode), true) { 174 } 175 } 176 177 //go:linkname net_runtime_pollSetDeadline net.runtime_pollSetDeadline 178 func net_runtime_pollSetDeadline(pd *pollDesc, d int64, mode int) { 179 lock(&pd.lock) 180 if pd.closing { 181 unlock(&pd.lock) 182 return 183 } 184 pd.seq++ // invalidate current timers 185 // Reset current timers. 186 if pd.rt.f != nil { 187 deltimer(&pd.rt) 188 pd.rt.f = nil 189 } 190 if pd.wt.f != nil { 191 deltimer(&pd.wt) 192 pd.wt.f = nil 193 } 194 // Setup new timers. 195 if d != 0 && d <= nanotime() { 196 d = -1 197 } 198 if mode == 'r' || mode == 'r'+'w' { 199 pd.rd = d 200 } 201 if mode == 'w' || mode == 'r'+'w' { 202 pd.wd = d 203 } 204 if pd.rd > 0 && pd.rd == pd.wd { 205 pd.rt.f = netpollDeadline 206 pd.rt.when = pd.rd 207 // Copy current seq into the timer arg. 208 // Timer func will check the seq against current descriptor seq, 209 // if they differ the descriptor was reused or timers were reset. 210 pd.rt.arg = pd 211 pd.rt.seq = pd.seq 212 addtimer(&pd.rt) 213 } else { 214 if pd.rd > 0 { 215 pd.rt.f = netpollReadDeadline 216 pd.rt.when = pd.rd 217 pd.rt.arg = pd 218 pd.rt.seq = pd.seq 219 addtimer(&pd.rt) 220 } 221 if pd.wd > 0 { 222 pd.wt.f = netpollWriteDeadline 223 pd.wt.when = pd.wd 224 pd.wt.arg = pd 225 pd.wt.seq = pd.seq 226 addtimer(&pd.wt) 227 } 228 } 229 // If we set the new deadline in the past, unblock currently pending IO if any. 230 var rg, wg *g 231 atomicstorep(unsafe.Pointer(&wg), nil) // full memory barrier between stores to rd/wd and load of rg/wg in netpollunblock 232 if pd.rd < 0 { 233 rg = netpollunblock(pd, 'r', false) 234 } 235 if pd.wd < 0 { 236 wg = netpollunblock(pd, 'w', false) 237 } 238 unlock(&pd.lock) 239 if rg != nil { 240 goready(rg, 3) 241 } 242 if wg != nil { 243 goready(wg, 3) 244 } 245 } 246 247 //go:linkname net_runtime_pollUnblock net.runtime_pollUnblock 248 func net_runtime_pollUnblock(pd *pollDesc) { 249 lock(&pd.lock) 250 if pd.closing { 251 throw("netpollUnblock: already closing") 252 } 253 pd.closing = true 254 pd.seq++ 255 var rg, wg *g 256 atomicstorep(unsafe.Pointer(&rg), nil) // full memory barrier between store to closing and read of rg/wg in netpollunblock 257 rg = netpollunblock(pd, 'r', false) 258 wg = netpollunblock(pd, 'w', false) 259 if pd.rt.f != nil { 260 deltimer(&pd.rt) 261 pd.rt.f = nil 262 } 263 if pd.wt.f != nil { 264 deltimer(&pd.wt) 265 pd.wt.f = nil 266 } 267 unlock(&pd.lock) 268 if rg != nil { 269 goready(rg, 3) 270 } 271 if wg != nil { 272 goready(wg, 3) 273 } 274 } 275 276 // make pd ready, newly runnable goroutines (if any) are returned in rg/wg 277 // May run during STW, so write barriers are not allowed. 278 //go:nowritebarrier 279 func netpollready(gpp *guintptr, pd *pollDesc, mode int32) { 280 var rg, wg guintptr 281 if mode == 'r' || mode == 'r'+'w' { 282 rg.set(netpollunblock(pd, 'r', true)) 283 } 284 if mode == 'w' || mode == 'r'+'w' { 285 wg.set(netpollunblock(pd, 'w', true)) 286 } 287 if rg != 0 { 288 rg.ptr().schedlink = *gpp 289 *gpp = rg 290 } 291 if wg != 0 { 292 wg.ptr().schedlink = *gpp 293 *gpp = wg 294 } 295 } 296 297 func netpollcheckerr(pd *pollDesc, mode int32) int { 298 if pd.closing { 299 return 1 // errClosing 300 } 301 if (mode == 'r' && pd.rd < 0) || (mode == 'w' && pd.wd < 0) { 302 return 2 // errTimeout 303 } 304 return 0 305 } 306 307 func netpollblockcommit(gp *g, gpp unsafe.Pointer) bool { 308 return casuintptr((*uintptr)(gpp), pdWait, uintptr(unsafe.Pointer(gp))) 309 } 310 311 // returns true if IO is ready, or false if timedout or closed 312 // waitio - wait only for completed IO, ignore errors 313 func netpollblock(pd *pollDesc, mode int32, waitio bool) bool { 314 gpp := &pd.rg 315 if mode == 'w' { 316 gpp = &pd.wg 317 } 318 319 // set the gpp semaphore to WAIT 320 for { 321 old := *gpp 322 if old == pdReady { 323 *gpp = 0 324 return true 325 } 326 if old != 0 { 327 throw("netpollblock: double wait") 328 } 329 if casuintptr(gpp, 0, pdWait) { 330 break 331 } 332 } 333 334 // need to recheck error states after setting gpp to WAIT 335 // this is necessary because runtime_pollUnblock/runtime_pollSetDeadline/deadlineimpl 336 // do the opposite: store to closing/rd/wd, membarrier, load of rg/wg 337 if waitio || netpollcheckerr(pd, mode) == 0 { 338 gopark(netpollblockcommit, unsafe.Pointer(gpp), "IO wait", traceEvGoBlockNet, 5) 339 } 340 // be careful to not lose concurrent READY notification 341 old := xchguintptr(gpp, 0) 342 if old > pdWait { 343 throw("netpollblock: corrupted state") 344 } 345 return old == pdReady 346 } 347 348 func netpollunblock(pd *pollDesc, mode int32, ioready bool) *g { 349 gpp := &pd.rg 350 if mode == 'w' { 351 gpp = &pd.wg 352 } 353 354 for { 355 old := *gpp 356 if old == pdReady { 357 return nil 358 } 359 if old == 0 && !ioready { 360 // Only set READY for ioready. runtime_pollWait 361 // will check for timeout/cancel before waiting. 362 return nil 363 } 364 var new uintptr 365 if ioready { 366 new = pdReady 367 } 368 if casuintptr(gpp, old, new) { 369 if old == pdReady || old == pdWait { 370 old = 0 371 } 372 return (*g)(unsafe.Pointer(old)) 373 } 374 } 375 } 376 377 func netpolldeadlineimpl(pd *pollDesc, seq uintptr, read, write bool) { 378 lock(&pd.lock) 379 // Seq arg is seq when the timer was set. 380 // If it's stale, ignore the timer event. 381 if seq != pd.seq { 382 // The descriptor was reused or timers were reset. 383 unlock(&pd.lock) 384 return 385 } 386 var rg *g 387 if read { 388 if pd.rd <= 0 || pd.rt.f == nil { 389 throw("netpolldeadlineimpl: inconsistent read deadline") 390 } 391 pd.rd = -1 392 atomicstorep(unsafe.Pointer(&pd.rt.f), nil) // full memory barrier between store to rd and load of rg in netpollunblock 393 rg = netpollunblock(pd, 'r', false) 394 } 395 var wg *g 396 if write { 397 if pd.wd <= 0 || pd.wt.f == nil && !read { 398 throw("netpolldeadlineimpl: inconsistent write deadline") 399 } 400 pd.wd = -1 401 atomicstorep(unsafe.Pointer(&pd.wt.f), nil) // full memory barrier between store to wd and load of wg in netpollunblock 402 wg = netpollunblock(pd, 'w', false) 403 } 404 unlock(&pd.lock) 405 if rg != nil { 406 goready(rg, 0) 407 } 408 if wg != nil { 409 goready(wg, 0) 410 } 411 } 412 413 func netpollDeadline(arg interface{}, seq uintptr) { 414 netpolldeadlineimpl(arg.(*pollDesc), seq, true, true) 415 } 416 417 func netpollReadDeadline(arg interface{}, seq uintptr) { 418 netpolldeadlineimpl(arg.(*pollDesc), seq, true, false) 419 } 420 421 func netpollWriteDeadline(arg interface{}, seq uintptr) { 422 netpolldeadlineimpl(arg.(*pollDesc), seq, false, true) 423 } 424 425 func (c *pollCache) alloc() *pollDesc { 426 lock(&c.lock) 427 if c.first == nil { 428 const pdSize = unsafe.Sizeof(pollDesc{}) 429 n := pollBlockSize / pdSize 430 if n == 0 { 431 n = 1 432 } 433 // Must be in non-GC memory because can be referenced 434 // only from epoll/kqueue internals. 435 mem := persistentalloc(n*pdSize, 0, &memstats.other_sys) 436 for i := uintptr(0); i < n; i++ { 437 pd := (*pollDesc)(add(mem, i*pdSize)) 438 pd.link = c.first 439 c.first = pd 440 } 441 } 442 pd := c.first 443 c.first = pd.link 444 unlock(&c.lock) 445 return pd 446 } 447