1 // Copyright 2019 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 // Goroutine preemption 6 // 7 // A goroutine can be preempted at any safe-point. Currently, there 8 // are a few categories of safe-points: 9 // 10 // 1. A blocked safe-point occurs for the duration that a goroutine is 11 // descheduled, blocked on synchronization, or in a system call. 12 // 13 // 2. Synchronous safe-points occur when a running goroutine checks 14 // for a preemption request. 15 // 16 // 3. Asynchronous safe-points occur at any instruction in user code 17 // where the goroutine can be safely paused and a conservative 18 // stack and register scan can find stack roots. The runtime can 19 // stop a goroutine at an async safe-point using a signal. 20 // 21 // At both blocked and synchronous safe-points, a goroutine's CPU 22 // state is minimal and the garbage collector has complete information 23 // about its entire stack. This makes it possible to deschedule a 24 // goroutine with minimal space, and to precisely scan a goroutine's 25 // stack. 26 // 27 // Synchronous safe-points are implemented by overloading the stack 28 // bound check in function prologues. To preempt a goroutine at the 29 // next synchronous safe-point, the runtime poisons the goroutine's 30 // stack bound to a value that will cause the next stack bound check 31 // to fail and enter the stack growth implementation, which will 32 // detect that it was actually a preemption and redirect to preemption 33 // handling. 34 // 35 // Preemption at asynchronous safe-points is implemented by suspending 36 // the thread using an OS mechanism (e.g., signals) and inspecting its 37 // state to determine if the goroutine was at an asynchronous 38 // safe-point. Since the thread suspension itself is generally 39 // asynchronous, it also checks if the running goroutine wants to be 40 // preempted, since this could have changed. If all conditions are 41 // satisfied, it adjusts the signal context to make it look like the 42 // signaled thread just called asyncPreempt and resumes the thread. 43 // asyncPreempt spills all registers and enters the scheduler. 44 // 45 // (An alternative would be to preempt in the signal handler itself. 46 // This would let the OS save and restore the register state and the 47 // runtime would only need to know how to extract potentially 48 // pointer-containing registers from the signal context. However, this 49 // would consume an M for every preempted G, and the scheduler itself 50 // is not designed to run from a signal handler, as it tends to 51 // allocate memory and start threads in the preemption path.) 52 53 package runtime 54 55 import ( 56 "runtime/internal/atomic" 57 "runtime/internal/sys" 58 "unsafe" 59 ) 60 61 type suspendGState struct { 62 g *g 63 64 // dead indicates the goroutine was not suspended because it 65 // is dead. This goroutine could be reused after the dead 66 // state was observed, so the caller must not assume that it 67 // remains dead. 68 dead bool 69 70 // stopped indicates that this suspendG transitioned the G to 71 // _Gwaiting via g.preemptStop and thus is responsible for 72 // readying it when done. 73 stopped bool 74 } 75 76 // suspendG suspends goroutine gp at a safe-point and returns the 77 // state of the suspended goroutine. The caller gets read access to 78 // the goroutine until it calls resumeG. 79 // 80 // It is safe for multiple callers to attempt to suspend the same 81 // goroutine at the same time. The goroutine may execute between 82 // subsequent successful suspend operations. The current 83 // implementation grants exclusive access to the goroutine, and hence 84 // multiple callers will serialize. However, the intent is to grant 85 // shared read access, so please don't depend on exclusive access. 86 // 87 // This must be called from the system stack and the user goroutine on 88 // the current M (if any) must be in a preemptible state. This 89 // prevents deadlocks where two goroutines attempt to suspend each 90 // other and both are in non-preemptible states. There are other ways 91 // to resolve this deadlock, but this seems simplest. 92 // 93 // TODO(austin): What if we instead required this to be called from a 94 // user goroutine? Then we could deschedule the goroutine while 95 // waiting instead of blocking the thread. If two goroutines tried to 96 // suspend each other, one of them would win and the other wouldn't 97 // complete the suspend until it was resumed. We would have to be 98 // careful that they couldn't actually queue up suspend for each other 99 // and then both be suspended. This would also avoid the need for a 100 // kernel context switch in the synchronous case because we could just 101 // directly schedule the waiter. The context switch is unavoidable in 102 // the signal case. 103 // 104 //go:systemstack 105 func suspendG(gp *g) suspendGState { 106 if mp := getg().m; mp.curg != nil && readgstatus(mp.curg) == _Grunning { 107 // Since we're on the system stack of this M, the user 108 // G is stuck at an unsafe point. If another goroutine 109 // were to try to preempt m.curg, it could deadlock. 110 throw("suspendG from non-preemptible goroutine") 111 } 112 113 // See https://golang.org/cl/21503 for justification of the yield delay. 114 const yieldDelay = 10 * 1000 115 var nextYield int64 116 117 // Drive the goroutine to a preemption point. 118 stopped := false 119 var asyncM *m 120 var asyncGen uint32 121 var nextPreemptM int64 122 for i := 0; ; i++ { 123 switch s := readgstatus(gp); s { 124 default: 125 if s&_Gscan != 0 { 126 // Someone else is suspending it. Wait 127 // for them to finish. 128 // 129 // TODO: It would be nicer if we could 130 // coalesce suspends. 131 break 132 } 133 134 dumpgstatus(gp) 135 throw("invalid g status") 136 137 case _Gdead: 138 // Nothing to suspend. 139 // 140 // preemptStop may need to be cleared, but 141 // doing that here could race with goroutine 142 // reuse. Instead, goexit0 clears it. 143 return suspendGState{dead: true} 144 145 case _Gcopystack: 146 // The stack is being copied. We need to wait 147 // until this is done. 148 149 case _Gpreempted: 150 // We (or someone else) suspended the G. Claim 151 // ownership of it by transitioning it to 152 // _Gwaiting. 153 if !casGFromPreempted(gp, _Gpreempted, _Gwaiting) { 154 break 155 } 156 157 // We stopped the G, so we have to ready it later. 158 stopped = true 159 160 s = _Gwaiting 161 fallthrough 162 163 case _Grunnable, _Gsyscall, _Gwaiting: 164 // Claim goroutine by setting scan bit. 165 // This may race with execution or readying of gp. 166 // The scan bit keeps it from transition state. 167 if !castogscanstatus(gp, s, s|_Gscan) { 168 break 169 } 170 171 // Clear the preemption request. It's safe to 172 // reset the stack guard because we hold the 173 // _Gscan bit and thus own the stack. 174 gp.preemptStop = false 175 gp.preempt = false 176 gp.stackguard0 = gp.stack.lo + _StackGuard 177 178 // The goroutine was already at a safe-point 179 // and we've now locked that in. 180 // 181 // TODO: It would be much better if we didn't 182 // leave it in _Gscan, but instead gently 183 // prevented its scheduling until resumption. 184 // Maybe we only use this to bump a suspended 185 // count and the scheduler skips suspended 186 // goroutines? That wouldn't be enough for 187 // {_Gsyscall,_Gwaiting} -> _Grunning. Maybe 188 // for all those transitions we need to check 189 // suspended and deschedule? 190 return suspendGState{g: gp, stopped: stopped} 191 192 case _Grunning: 193 // Optimization: if there is already a pending preemption request 194 // (from the previous loop iteration), don't bother with the atomics. 195 if gp.preemptStop && gp.preempt && gp.stackguard0 == stackPreempt && asyncM == gp.m && atomic.Load(&asyncM.preemptGen) == asyncGen { 196 break 197 } 198 199 // Temporarily block state transitions. 200 if !castogscanstatus(gp, _Grunning, _Gscanrunning) { 201 break 202 } 203 204 // Request synchronous preemption. 205 gp.preemptStop = true 206 gp.preempt = true 207 gp.stackguard0 = stackPreempt 208 209 // Prepare for asynchronous preemption. 210 asyncM2 := gp.m 211 asyncGen2 := atomic.Load(&asyncM2.preemptGen) 212 needAsync := asyncM != asyncM2 || asyncGen != asyncGen2 213 asyncM = asyncM2 214 asyncGen = asyncGen2 215 216 casfrom_Gscanstatus(gp, _Gscanrunning, _Grunning) 217 218 // Send asynchronous preemption. We do this 219 // after CASing the G back to _Grunning 220 // because preemptM may be synchronous and we 221 // don't want to catch the G just spinning on 222 // its status. 223 if preemptMSupported && debug.asyncpreemptoff == 0 && needAsync { 224 // Rate limit preemptM calls. This is 225 // particularly important on Windows 226 // where preemptM is actually 227 // synchronous and the spin loop here 228 // can lead to live-lock. 229 now := nanotime() 230 if now >= nextPreemptM { 231 nextPreemptM = now + yieldDelay/2 232 preemptM(asyncM) 233 } 234 } 235 } 236 237 // TODO: Don't busy wait. This loop should really only 238 // be a simple read/decide/CAS loop that only fails if 239 // there's an active race. Once the CAS succeeds, we 240 // should queue up the preemption (which will require 241 // it to be reliable in the _Grunning case, not 242 // best-effort) and then sleep until we're notified 243 // that the goroutine is suspended. 244 if i == 0 { 245 nextYield = nanotime() + yieldDelay 246 } 247 if nanotime() < nextYield { 248 procyield(10) 249 } else { 250 osyield() 251 nextYield = nanotime() + yieldDelay/2 252 } 253 } 254 } 255 256 // resumeG undoes the effects of suspendG, allowing the suspended 257 // goroutine to continue from its current safe-point. 258 func resumeG(state suspendGState) { 259 if state.dead { 260 // We didn't actually stop anything. 261 return 262 } 263 264 gp := state.g 265 switch s := readgstatus(gp); s { 266 default: 267 dumpgstatus(gp) 268 throw("unexpected g status") 269 270 case _Grunnable | _Gscan, 271 _Gwaiting | _Gscan, 272 _Gsyscall | _Gscan: 273 casfrom_Gscanstatus(gp, s, s&^_Gscan) 274 } 275 276 if state.stopped { 277 // We stopped it, so we need to re-schedule it. 278 ready(gp, 0, true) 279 } 280 } 281 282 // canPreemptM reports whether mp is in a state that is safe to preempt. 283 // 284 // It is nosplit because it has nosplit callers. 285 // 286 //go:nosplit 287 func canPreemptM(mp *m) bool { 288 return mp.locks == 0 && mp.mallocing == 0 && mp.preemptoff == "" && mp.p.ptr().status == _Prunning 289 } 290 291 //go:generate go run mkpreempt.go 292 293 // asyncPreempt saves all user registers and calls asyncPreempt2. 294 // 295 // When stack scanning encounters an asyncPreempt frame, it scans that 296 // frame and its parent frame conservatively. 297 // 298 // asyncPreempt is implemented in assembly. 299 func asyncPreempt() 300 301 //go:nosplit 302 func asyncPreempt2() { 303 gp := getg() 304 gp.asyncSafePoint = true 305 if gp.preemptStop { 306 mcall(preemptPark) 307 } else { 308 mcall(gopreempt_m) 309 } 310 gp.asyncSafePoint = false 311 } 312 313 // asyncPreemptStack is the bytes of stack space required to inject an 314 // asyncPreempt call. 315 var asyncPreemptStack = ^uintptr(0) 316 317 func init() { 318 f := findfunc(funcPC(asyncPreempt)) 319 total := funcMaxSPDelta(f) 320 f = findfunc(funcPC(asyncPreempt2)) 321 total += funcMaxSPDelta(f) 322 // Add some overhead for return PCs, etc. 323 asyncPreemptStack = uintptr(total) + 8*sys.PtrSize 324 if asyncPreemptStack > _StackLimit { 325 // We need more than the nosplit limit. This isn't 326 // unsafe, but it may limit asynchronous preemption. 327 // 328 // This may be a problem if we start using more 329 // registers. In that case, we should store registers 330 // in a context object. If we pre-allocate one per P, 331 // asyncPreempt can spill just a few registers to the 332 // stack, then grab its context object and spill into 333 // it. When it enters the runtime, it would allocate a 334 // new context for the P. 335 print("runtime: asyncPreemptStack=", asyncPreemptStack, "\n") 336 throw("async stack too large") 337 } 338 } 339 340 // wantAsyncPreempt returns whether an asynchronous preemption is 341 // queued for gp. 342 func wantAsyncPreempt(gp *g) bool { 343 // Check both the G and the P. 344 return (gp.preempt || gp.m.p != 0 && gp.m.p.ptr().preempt) && readgstatus(gp)&^_Gscan == _Grunning 345 } 346 347 // isAsyncSafePoint reports whether gp at instruction PC is an 348 // asynchronous safe point. This indicates that: 349 // 350 // 1. It's safe to suspend gp and conservatively scan its stack and 351 // registers. There are no potentially hidden pointer values and it's 352 // not in the middle of an atomic sequence like a write barrier. 353 // 354 // 2. gp has enough stack space to inject the asyncPreempt call. 355 // 356 // 3. It's generally safe to interact with the runtime, even if we're 357 // in a signal handler stopped here. For example, there are no runtime 358 // locks held, so acquiring a runtime lock won't self-deadlock. 359 // 360 // In some cases the PC is safe for asynchronous preemption but it 361 // also needs to adjust the resumption PC. The new PC is returned in 362 // the second result. 363 func isAsyncSafePoint(gp *g, pc, sp, lr uintptr) (bool, uintptr) { 364 mp := gp.m 365 366 // Only user Gs can have safe-points. We check this first 367 // because it's extremely common that we'll catch mp in the 368 // scheduler processing this G preemption. 369 if mp.curg != gp { 370 return false, 0 371 } 372 373 // Check M state. 374 if mp.p == 0 || !canPreemptM(mp) { 375 return false, 0 376 } 377 378 // Check stack space. 379 if sp < gp.stack.lo || sp-gp.stack.lo < asyncPreemptStack { 380 return false, 0 381 } 382 383 // Check if PC is an unsafe-point. 384 f := findfunc(pc) 385 if !f.valid() { 386 // Not Go code. 387 return false, 0 388 } 389 if (GOARCH == "mips" || GOARCH == "mipsle" || GOARCH == "mips64" || GOARCH == "mips64le") && lr == pc+8 && funcspdelta(f, pc, nil) == 0 { 390 // We probably stopped at a half-executed CALL instruction, 391 // where the LR is updated but the PC has not. If we preempt 392 // here we'll see a seemingly self-recursive call, which is in 393 // fact not. 394 // This is normally ok, as we use the return address saved on 395 // stack for unwinding, not the LR value. But if this is a 396 // call to morestack, we haven't created the frame, and we'll 397 // use the LR for unwinding, which will be bad. 398 return false, 0 399 } 400 up, startpc := pcdatavalue2(f, _PCDATA_UnsafePoint, pc) 401 if up != _PCDATA_UnsafePointSafe { 402 // Unsafe-point marked by compiler. This includes 403 // atomic sequences (e.g., write barrier) and nosplit 404 // functions (except at calls). 405 return false, 0 406 } 407 if fd := funcdata(f, _FUNCDATA_LocalsPointerMaps); fd == nil || fd == unsafe.Pointer(&no_pointers_stackmap) { 408 // This is assembly code. Don't assume it's 409 // well-formed. We identify assembly code by 410 // checking that it has either no stack map, or 411 // no_pointers_stackmap, which is the stack map 412 // for ones marked as NO_LOCAL_POINTERS. 413 // 414 // TODO: Are there cases that are safe but don't have a 415 // locals pointer map, like empty frame functions? 416 // It might be possible to preempt any assembly functions 417 // except the ones that have funcFlag_SPWRITE set in f.flag. 418 return false, 0 419 } 420 name := funcname(f) 421 if inldata := funcdata(f, _FUNCDATA_InlTree); inldata != nil { 422 inltree := (*[1 << 20]inlinedCall)(inldata) 423 ix := pcdatavalue(f, _PCDATA_InlTreeIndex, pc, nil) 424 if ix >= 0 { 425 name = funcnameFromNameoff(f, inltree[ix].func_) 426 } 427 } 428 if hasPrefix(name, "runtime.") || 429 hasPrefix(name, "runtime/internal/") || 430 hasPrefix(name, "reflect.") { 431 // For now we never async preempt the runtime or 432 // anything closely tied to the runtime. Known issues 433 // include: various points in the scheduler ("don't 434 // preempt between here and here"), much of the defer 435 // implementation (untyped info on stack), bulk write 436 // barriers (write barrier check), 437 // reflect.{makeFuncStub,methodValueCall}. 438 // 439 // TODO(austin): We should improve this, or opt things 440 // in incrementally. 441 return false, 0 442 } 443 switch up { 444 case _PCDATA_Restart1, _PCDATA_Restart2: 445 // Restartable instruction sequence. Back off PC to 446 // the start PC. 447 if startpc == 0 || startpc > pc || pc-startpc > 20 { 448 throw("bad restart PC") 449 } 450 return true, startpc 451 case _PCDATA_RestartAtEntry: 452 // Restart from the function entry at resumption. 453 return true, f.entry 454 } 455 return true, pc 456 } 457 458 var no_pointers_stackmap uint64 // defined in assembly, for NO_LOCAL_POINTERS macro 459