Source file src/runtime/preempt.go

     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  	"internal/abi"
    57  	"internal/goarch"
    58  	"internal/stringslite"
    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 && asyncM.preemptGen.Load() == 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 := asyncM2.preemptGen.Load()
   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  // It saves GP registers (anything that might contain a pointer) to the G stack.
   296  // Hence, when stack scanning encounters an asyncPreempt frame, it scans that
   297  // frame and its parent frame conservatively.
   298  //
   299  // On some platforms, it saves large additional scalar-only register state such
   300  // as vector registers to an "extended register state" on the P.
   301  //
   302  // asyncPreempt is implemented in assembly.
   303  func asyncPreempt()
   304  
   305  // asyncPreempt2 is the Go continuation of asyncPreempt.
   306  //
   307  // It must be deeply nosplit because there's untyped data on the stack from
   308  // asyncPreempt.
   309  //
   310  // It must not have any write barriers because we need to limit the amount of
   311  // stack it uses.
   312  //
   313  //go:nosplit
   314  //go:nowritebarrierrec
   315  func asyncPreempt2() {
   316  	// We can't grow the stack with untyped data from asyncPreempt, so switch to
   317  	// the system stack right away.
   318  	mcall(func(gp *g) {
   319  		gp.asyncSafePoint = true
   320  
   321  		// Move the extended register state from the P to the G. We do this now that
   322  		// we're on the system stack to avoid stack splits.
   323  		xRegSave(gp)
   324  
   325  		if gp.preemptStop {
   326  			preemptPark(gp)
   327  		} else {
   328  			gopreempt_m(gp)
   329  		}
   330  		// The above functions never return.
   331  	})
   332  
   333  	// Do not grow the stack below here!
   334  
   335  	gp := getg()
   336  
   337  	// Put the extended register state back on the M so resumption can find it.
   338  	// We can't do this in asyncPreemptM because the park calls never return.
   339  	xRegRestore(gp)
   340  
   341  	gp.asyncSafePoint = false
   342  }
   343  
   344  // asyncPreemptStack is the bytes of stack space required to inject an
   345  // asyncPreempt call.
   346  var asyncPreemptStack = ^uintptr(0)
   347  
   348  func init() {
   349  	f := findfunc(abi.FuncPCABI0(asyncPreempt))
   350  	total := funcMaxSPDelta(f)
   351  	f = findfunc(abi.FuncPCABIInternal(asyncPreempt2))
   352  	total += funcMaxSPDelta(f)
   353  	f = findfunc(abi.FuncPCABIInternal(xRegRestore))
   354  	total += funcMaxSPDelta(f)
   355  	// Add some overhead for return PCs, etc.
   356  	asyncPreemptStack = uintptr(total) + 8*goarch.PtrSize
   357  	if asyncPreemptStack > stackNosplit {
   358  		// We need more than the nosplit limit. This isn't unsafe, but it may
   359  		// limit asynchronous preemption. Consider moving state into xRegState.
   360  		print("runtime: asyncPreemptStack=", asyncPreemptStack, "\n")
   361  		throw("async stack too large")
   362  	}
   363  }
   364  
   365  // wantAsyncPreempt returns whether an asynchronous preemption is
   366  // queued for gp.
   367  func wantAsyncPreempt(gp *g) bool {
   368  	// Check both the G and the P.
   369  	return (gp.preempt || gp.m.p != 0 && gp.m.p.ptr().preempt) && readgstatus(gp)&^_Gscan == _Grunning
   370  }
   371  
   372  // isAsyncSafePoint reports whether gp at instruction PC is an
   373  // asynchronous safe point. This indicates that:
   374  //
   375  // 1. It's safe to suspend gp and conservatively scan its stack and
   376  // registers. There are no potentially hidden pointer values and it's
   377  // not in the middle of an atomic sequence like a write barrier.
   378  //
   379  // 2. gp has enough stack space to inject the asyncPreempt call.
   380  //
   381  // 3. It's generally safe to interact with the runtime, even if we're
   382  // in a signal handler stopped here. For example, there are no runtime
   383  // locks held, so acquiring a runtime lock won't self-deadlock.
   384  //
   385  // In some cases the PC is safe for asynchronous preemption but it
   386  // also needs to adjust the resumption PC. The new PC is returned in
   387  // the second result.
   388  func isAsyncSafePoint(gp *g, pc, sp, lr uintptr) (bool, uintptr) {
   389  	mp := gp.m
   390  
   391  	// Only user Gs can have safe-points. We check this first
   392  	// because it's extremely common that we'll catch mp in the
   393  	// scheduler processing this G preemption.
   394  	if mp.curg != gp {
   395  		return false, 0
   396  	}
   397  
   398  	// Check M state.
   399  	if mp.p == 0 || !canPreemptM(mp) {
   400  		return false, 0
   401  	}
   402  
   403  	// Check stack space.
   404  	if sp < gp.stack.lo || sp-gp.stack.lo < asyncPreemptStack {
   405  		return false, 0
   406  	}
   407  
   408  	// Check if PC is an unsafe-point.
   409  	f := findfunc(pc)
   410  	if !f.valid() {
   411  		// Not Go code.
   412  		return false, 0
   413  	}
   414  	if (GOARCH == "mips" || GOARCH == "mipsle" || GOARCH == "mips64" || GOARCH == "mips64le") && lr == pc+8 && funcspdelta(f, pc) == 0 {
   415  		// We probably stopped at a half-executed CALL instruction,
   416  		// where the LR is updated but the PC has not. If we preempt
   417  		// here we'll see a seemingly self-recursive call, which is in
   418  		// fact not.
   419  		// This is normally ok, as we use the return address saved on
   420  		// stack for unwinding, not the LR value. But if this is a
   421  		// call to morestack, we haven't created the frame, and we'll
   422  		// use the LR for unwinding, which will be bad.
   423  		return false, 0
   424  	}
   425  	up, startpc := pcdatavalue2(f, abi.PCDATA_UnsafePoint, pc)
   426  	if up == abi.UnsafePointUnsafe {
   427  		// Unsafe-point marked by compiler. This includes
   428  		// atomic sequences (e.g., write barrier) and nosplit
   429  		// functions (except at calls).
   430  		return false, 0
   431  	}
   432  	if fd := funcdata(f, abi.FUNCDATA_LocalsPointerMaps); fd == nil || f.flag&abi.FuncFlagAsm != 0 {
   433  		// This is assembly code. Don't assume it's well-formed.
   434  		// TODO: Empirically we still need the fd == nil check. Why?
   435  		//
   436  		// TODO: Are there cases that are safe but don't have a
   437  		// locals pointer map, like empty frame functions?
   438  		// It might be possible to preempt any assembly functions
   439  		// except the ones that have funcFlag_SPWRITE set in f.flag.
   440  		return false, 0
   441  	}
   442  	// Check the inner-most name
   443  	u, uf := newInlineUnwinder(f, pc)
   444  	name := u.srcFunc(uf).name()
   445  	if stringslite.HasPrefix(name, "runtime.") ||
   446  		stringslite.HasPrefix(name, "internal/runtime/") ||
   447  		stringslite.HasPrefix(name, "reflect.") {
   448  		// For now we never async preempt the runtime or
   449  		// anything closely tied to the runtime. Known issues
   450  		// include: various points in the scheduler ("don't
   451  		// preempt between here and here"), much of the defer
   452  		// implementation (untyped info on stack), bulk write
   453  		// barriers (write barrier check), atomic functions in
   454  		// internal/runtime/atomic, reflect.{makeFuncStub,methodValueCall}.
   455  		//
   456  		// Note that this is a subset of the runtimePkgs in pkgspecial.go
   457  		// and these checks are theoretically redundant because the compiler
   458  		// marks "all points" in runtime functions as unsafe for async preemption.
   459  		// But for some reason, we can't eliminate these checks until https://go.dev/issue/72031
   460  		// is resolved.
   461  		//
   462  		// TODO(austin): We should improve this, or opt things
   463  		// in incrementally.
   464  		return false, 0
   465  	}
   466  	switch up {
   467  	case abi.UnsafePointRestart1, abi.UnsafePointRestart2:
   468  		// Restartable instruction sequence. Back off PC to
   469  		// the start PC.
   470  		if startpc == 0 || startpc > pc || pc-startpc > 20 {
   471  			throw("bad restart PC")
   472  		}
   473  		return true, startpc
   474  	case abi.UnsafePointRestartAtEntry:
   475  		// Restart from the function entry at resumption.
   476  		return true, f.entry()
   477  	}
   478  	return true, pc
   479  }
   480  

View as plain text