coro.go

     1  // Copyright 2023 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 "unsafe"
     8  
     9  // A coro represents extra concurrency without extra parallelism,
    10  // as would be needed for a coroutine implementation.
    11  // The coro does not represent a specific coroutine, only the ability
    12  // to do coroutine-style control transfers.
    13  // It can be thought of as like a special channel that always has
    14  // a goroutine blocked on it. If another goroutine calls coroswitch(c),
    15  // the caller becomes the goroutine blocked in c, and the goroutine
    16  // formerly blocked in c starts running.
    17  // These switches continue until a call to coroexit(c),
    18  // which ends the use of the coro by releasing the blocked
    19  // goroutine in c and exiting the current goroutine.
    20  //
    21  // Coros are heap allocated and garbage collected, so that user code
    22  // can hold a pointer to a coro without causing potential dangling
    23  // pointer errors.
    24  type coro struct {
    25  	gp guintptr
    26  	f  func(*coro)
    27  
    28  	// State for validating thread-lock interactions.
    29  	mp        *m
    30  	lockedExt uint32 // mp's external LockOSThread counter at coro creation time.
    31  	lockedInt uint32 // mp's internal lockOSThread counter at coro creation time.
    32  }
    33  
    34  //go:linkname newcoro
    35  
    36  // newcoro creates a new coro containing a
    37  // goroutine blocked waiting to run f
    38  // and returns that coro.
    39  func newcoro(f func(*coro)) *coro {
    40  	c := new(coro)
    41  	c.f = f
    42  	pc := getcallerpc()
    43  	gp := getg()
    44  	systemstack(func() {
    45  		mp := gp.m
    46  		start := corostart
    47  		startfv := *(**funcval)(unsafe.Pointer(&start))
    48  		gp = newproc1(startfv, gp, pc, true, waitReasonCoroutine)
    49  
    50  		// Scribble down locked thread state if needed and/or donate
    51  		// thread-lock state to the new goroutine.
    52  		if mp.lockedExt+mp.lockedInt != 0 {
    53  			c.mp = mp
    54  			c.lockedExt = mp.lockedExt
    55  			c.lockedInt = mp.lockedInt
    56  		}
    57  	})
    58  	gp.coroarg = c
    59  	c.gp.set(gp)
    60  	return c
    61  }
    62  
    63  // corostart is the entry func for a new coroutine.
    64  // It runs the coroutine user function f passed to corostart
    65  // and then calls coroexit to remove the extra concurrency.
    66  func corostart() {
    67  	gp := getg()
    68  	c := gp.coroarg
    69  	gp.coroarg = nil
    70  
    71  	defer coroexit(c)
    72  	c.f(c)
    73  }
    74  
    75  // coroexit is like coroswitch but closes the coro
    76  // and exits the current goroutine
    77  func coroexit(c *coro) {
    78  	gp := getg()
    79  	gp.coroarg = c
    80  	gp.coroexit = true
    81  	mcall(coroswitch_m)
    82  }
    83  
    84  //go:linkname coroswitch
    85  
    86  // coroswitch switches to the goroutine blocked on c
    87  // and then blocks the current goroutine on c.
    88  func coroswitch(c *coro) {
    89  	gp := getg()
    90  	gp.coroarg = c
    91  	mcall(coroswitch_m)
    92  }
    93  
    94  // coroswitch_m is the implementation of coroswitch
    95  // that runs on the m stack.
    96  //
    97  // Note: Coroutine switches are expected to happen at
    98  // an order of magnitude (or more) higher frequency
    99  // than regular goroutine switches, so this path is heavily
   100  // optimized to remove unnecessary work.
   101  // The fast path here is three CAS: the one at the top on gp.atomicstatus,
   102  // the one in the middle to choose the next g,
   103  // and the one at the bottom on gnext.atomicstatus.
   104  // It is important not to add more atomic operations or other
   105  // expensive operations to the fast path.
   106  func coroswitch_m(gp *g) {
   107  	c := gp.coroarg
   108  	gp.coroarg = nil
   109  	exit := gp.coroexit
   110  	gp.coroexit = false
   111  	mp := gp.m
   112  
   113  	// Track and validate thread-lock interactions.
   114  	//
   115  	// The rules with thread-lock interactions are simple. When a coro goroutine is switched to,
   116  	// the same thread must be used, and the locked state must match with the thread-lock state of
   117  	// the goroutine which called newcoro. Thread-lock state consists of the thread and the number
   118  	// of internal (cgo callback, etc.) and external (LockOSThread) thread locks.
   119  	locked := gp.lockedm != 0
   120  	if c.mp != nil || locked {
   121  		if mp != c.mp || mp.lockedInt != c.lockedInt || mp.lockedExt != c.lockedExt {
   122  			print("coro: got thread ", unsafe.Pointer(mp), ", want ", unsafe.Pointer(c.mp), "\n")
   123  			print("coro: got lock internal ", mp.lockedInt, ", want ", c.lockedInt, "\n")
   124  			print("coro: got lock external ", mp.lockedExt, ", want ", c.lockedExt, "\n")
   125  			throw("coro: OS thread locking must match locking at coroutine creation")
   126  		}
   127  	}
   128  
   129  	// Acquire tracer for writing for the duration of this call.
   130  	//
   131  	// There's a lot of state manipulation performed with shortcuts
   132  	// but we need to make sure the tracer can only observe the
   133  	// start and end states to maintain a coherent model and avoid
   134  	// emitting an event for every single transition.
   135  	trace := traceAcquire()
   136  
   137  	if locked {
   138  		// Detach the goroutine from the thread; we'll attach to the goroutine we're
   139  		// switching to before returning.
   140  		gp.lockedm.set(nil)
   141  	}
   142  
   143  	if exit {
   144  		// The M might have a non-zero OS thread lock count when we get here, gdestroy
   145  		// will avoid destroying the M if the G isn't explicitly locked to it via lockedm,
   146  		// which we cleared above. It's fine to gdestroy here also, even when locked to
   147  		// the thread, because we'll be switching back to another goroutine anyway, which
   148  		// will take back its thread-lock state before returning.
   149  		gdestroy(gp)
   150  		gp = nil
   151  	} else {
   152  		// If we can CAS ourselves directly from running to waiting, so do,
   153  		// keeping the control transfer as lightweight as possible.
   154  		gp.waitreason = waitReasonCoroutine
   155  		if !gp.atomicstatus.CompareAndSwap(_Grunning, _Gwaiting) {
   156  			// The CAS failed: use casgstatus, which will take care of
   157  			// coordinating with the garbage collector about the state change.
   158  			casgstatus(gp, _Grunning, _Gwaiting)
   159  		}
   160  
   161  		// Clear gp.m.
   162  		setMNoWB(&gp.m, nil)
   163  	}
   164  
   165  	// The goroutine stored in c is the one to run next.
   166  	// Swap it with ourselves.
   167  	var gnext *g
   168  	for {
   169  		// Note: this is a racy load, but it will eventually
   170  		// get the right value, and if it gets the wrong value,
   171  		// the c.gp.cas will fail, so no harm done other than
   172  		// a wasted loop iteration.
   173  		// The cas will also sync c.gp's
   174  		// memory enough that the next iteration of the racy load
   175  		// should see the correct value.
   176  		// We are avoiding the atomic load to keep this path
   177  		// as lightweight as absolutely possible.
   178  		// (The atomic load is free on x86 but not free elsewhere.)
   179  		next := c.gp
   180  		if next.ptr() == nil {
   181  			throw("coroswitch on exited coro")
   182  		}
   183  		var self guintptr
   184  		self.set(gp)
   185  		if c.gp.cas(next, self) {
   186  			gnext = next.ptr()
   187  			break
   188  		}
   189  	}
   190  
   191  	// Check if we're switching to ourselves. This case is able to break our
   192  	// thread-lock invariants and an unbuffered channel implementation of
   193  	// coroswitch would deadlock. It's clear that this case should just not
   194  	// work.
   195  	if gnext == gp {
   196  		throw("coroswitch of a goroutine to itself")
   197  	}
   198  
   199  	// Emit the trace event after getting gnext but before changing curg.
   200  	// GoSwitch expects that the current G is running and that we haven't
   201  	// switched yet for correct status emission.
   202  	if trace.ok() {
   203  		trace.GoSwitch(gnext, exit)
   204  	}
   205  
   206  	// Start running next, without heavy scheduling machinery.
   207  	// Set mp.curg and gnext.m and then update scheduling state
   208  	// directly if possible.
   209  	setGNoWB(&mp.curg, gnext)
   210  	setMNoWB(&gnext.m, mp)
   211  
   212  	// Synchronize with any out-standing goroutine profile. We're about to start
   213  	// executing, and an invariant of the profiler is that we tryRecordGoroutineProfile
   214  	// whenever a goroutine is about to start running.
   215  	//
   216  	// N.B. We must do this before transitioning to _Grunning but after installing gnext
   217  	// in curg, so that we have a valid curg for allocation (tryRecordGoroutineProfile
   218  	// may allocate).
   219  	if goroutineProfile.active {
   220  		tryRecordGoroutineProfile(gnext, nil, osyield)
   221  	}
   222  
   223  	if !gnext.atomicstatus.CompareAndSwap(_Gwaiting, _Grunning) {
   224  		// The CAS failed: use casgstatus, which will take care of
   225  		// coordinating with the garbage collector about the state change.
   226  		casgstatus(gnext, _Gwaiting, _Grunnable)
   227  		casgstatus(gnext, _Grunnable, _Grunning)
   228  	}
   229  
   230  	// Donate locked state.
   231  	if locked {
   232  		mp.lockedg.set(gnext)
   233  		gnext.lockedm.set(mp)
   234  	}
   235  
   236  	// Release the trace locker. We've completed all the necessary transitions..
   237  	if trace.ok() {
   238  		traceRelease(trace)
   239  	}
   240  
   241  	// Switch to gnext. Does not return.
   242  	gogo(&gnext.sched)
   243  }
   244
View as plain text