intern.go

     1  // Copyright 2020 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 intern lets you make smaller comparable values by boxing
     6  // a larger comparable value (such as a 16 byte string header) down
     7  // into a globally unique 8 byte pointer.
     8  //
     9  // The globally unique pointers are garbage collected with weak
    10  // references and finalizers. This package hides that.
    11  package intern
    12  
    13  import (
    14  	"internal/godebug"
    15  	"runtime"
    16  	"sync"
    17  	"unsafe"
    18  )
    19  
    20  // A Value pointer is the handle to an underlying comparable value.
    21  // See func Get for how Value pointers may be used.
    22  type Value struct {
    23  	_      [0]func() // prevent people from accidentally using value type as comparable
    24  	cmpVal any
    25  	// resurrected is guarded by mu (for all instances of Value).
    26  	// It is set true whenever v is synthesized from a uintptr.
    27  	resurrected bool
    28  }
    29  
    30  // Get returns the comparable value passed to the Get func
    31  // that returned v.
    32  func (v *Value) Get() any { return v.cmpVal }
    33  
    34  // key is a key in our global value map.
    35  // It contains type-specialized fields to avoid allocations
    36  // when converting common types to empty interfaces.
    37  type key struct {
    38  	s      string
    39  	cmpVal any
    40  	// isString reports whether key contains a string.
    41  	// Without it, the zero value of key is ambiguous.
    42  	isString bool
    43  }
    44  
    45  // keyFor returns a key to use with cmpVal.
    46  func keyFor(cmpVal any) key {
    47  	if s, ok := cmpVal.(string); ok {
    48  		return key{s: s, isString: true}
    49  	}
    50  	return key{cmpVal: cmpVal}
    51  }
    52  
    53  // Value returns a *Value built from k.
    54  func (k key) Value() *Value {
    55  	if k.isString {
    56  		return &Value{cmpVal: k.s}
    57  	}
    58  	return &Value{cmpVal: k.cmpVal}
    59  }
    60  
    61  var (
    62  	// mu guards valMap, a weakref map of *Value by underlying value.
    63  	// It also guards the resurrected field of all *Values.
    64  	mu      sync.Mutex
    65  	valMap  = map[key]uintptr{} // to uintptr(*Value)
    66  	valSafe = safeMap()         // non-nil in safe+leaky mode
    67  )
    68  
    69  // safeMap returns a non-nil map if we're in safe-but-leaky mode,
    70  // as controlled by GODEBUG=intern=leaky
    71  func safeMap() map[key]*Value {
    72  	if godebug.Get("intern") == "leaky" {
    73  		return map[key]*Value{}
    74  	}
    75  	return nil
    76  }
    77  
    78  // Get returns a pointer representing the comparable value cmpVal.
    79  //
    80  // The returned pointer will be the same for Get(v) and Get(v2)
    81  // if and only if v == v2, and can be used as a map key.
    82  func Get(cmpVal any) *Value {
    83  	return get(keyFor(cmpVal))
    84  }
    85  
    86  // GetByString is identical to Get, except that it is specialized for strings.
    87  // This avoids an allocation from putting a string into an interface{}
    88  // to pass as an argument to Get.
    89  func GetByString(s string) *Value {
    90  	return get(key{s: s, isString: true})
    91  }
    92  
    93  // We play unsafe games that violate Go's rules (and assume a non-moving
    94  // collector). So we quiet Go here.
    95  // See the comment below Get for more implementation details.
    96  //go:nocheckptr
    97  func get(k key) *Value {
    98  	mu.Lock()
    99  	defer mu.Unlock()
   100  
   101  	var v *Value
   102  	if valSafe != nil {
   103  		v = valSafe[k]
   104  	} else if addr, ok := valMap[k]; ok {
   105  		v = (*Value)(unsafe.Pointer(addr))
   106  		v.resurrected = true
   107  	}
   108  	if v != nil {
   109  		return v
   110  	}
   111  	v = k.Value()
   112  	if valSafe != nil {
   113  		valSafe[k] = v
   114  	} else {
   115  		// SetFinalizer before uintptr conversion (theoretical concern;
   116  		// see https://github.com/go4org/intern/issues/13)
   117  		runtime.SetFinalizer(v, finalize)
   118  		valMap[k] = uintptr(unsafe.Pointer(v))
   119  	}
   120  	return v
   121  }
   122  
   123  func finalize(v *Value) {
   124  	mu.Lock()
   125  	defer mu.Unlock()
   126  	if v.resurrected {
   127  		// We lost the race. Somebody resurrected it while we
   128  		// were about to finalize it. Try again next round.
   129  		v.resurrected = false
   130  		runtime.SetFinalizer(v, finalize)
   131  		return
   132  	}
   133  	delete(valMap, keyFor(v.cmpVal))
   134  }
   135  
   136  // Interning is simple if you don't require that unused values be
   137  // garbage collectable. But we do require that; we don't want to be
   138  // DOS vector. We do this by using a uintptr to hide the pointer from
   139  // the garbage collector, and using a finalizer to eliminate the
   140  // pointer when no other code is using it.
   141  //
   142  // The obvious implementation of this is to use a
   143  // map[interface{}]uintptr-of-*interface{}, and set up a finalizer to
   144  // delete from the map. Unfortunately, this is racy. Because pointers
   145  // are being created in violation of Go's unsafety rules, it's
   146  // possible to create a pointer to a value concurrently with the GC
   147  // concluding that the value can be collected. There are other races
   148  // that break the equality invariant as well, but the use-after-free
   149  // will cause a runtime crash.
   150  //
   151  // To make this work, the finalizer needs to know that no references
   152  // have been unsafely created since the finalizer was set up. To do
   153  // this, values carry a "resurrected" sentinel, which gets set
   154  // whenever a pointer is unsafely created. If the finalizer encounters
   155  // the sentinel, it clears the sentinel and delays collection for one
   156  // additional GC cycle, by re-installing itself as finalizer. This
   157  // ensures that the unsafely created pointer is visible to the GC, and
   158  // will correctly prevent collection.
   159  //
   160  // This technique does mean that interned values that get reused take
   161  // at least 3 GC cycles to fully collect (1 to clear the sentinel, 1
   162  // to clean up the unsafe map, 1 to be actually deleted).
   163  //
   164  // @ianlancetaylor commented in
   165  // https://github.com/golang/go/issues/41303#issuecomment-717401656
   166  // that it is possible to implement weak references in terms of
   167  // finalizers without unsafe. Unfortunately, the approach he outlined
   168  // does not work here, for two reasons. First, there is no way to
   169  // construct a strong pointer out of a weak pointer; our map stores
   170  // weak pointers, but we must return strong pointers to callers.
   171  // Second, and more fundamentally, we must return not just _a_ strong
   172  // pointer to callers, but _the same_ strong pointer to callers. In
   173  // order to return _the same_ strong pointer to callers, we must track
   174  // it, which is exactly what we cannot do with strong pointers.
   175  //
   176  // See https://github.com/inetaf/netaddr/issues/53 for more
   177  // discussion, and https://github.com/go4org/intern/issues/2 for an
   178  // illustration of the subtleties at play.
   179
View as plain text