-
Notifications
You must be signed in to change notification settings - Fork 8
/
atomic_reference.hpp
309 lines (267 loc) · 6.9 KB
/
atomic_reference.hpp
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
#pragma once
#include <cassert>
#include <atomic>
#include "asm.hpp"
/**
* A std::shared_ptr<T>-like abstraction for reference counting,
* except we guarantee atomicity. While C++11 provides atomic_*
* specializations of functions which operate on shared_ptr objects,
* they are not quite desirable yet:
*
* A) As of gcc 4.7, an atomic implementation is not yet provided in libstdc++
*
* B) There's no builtin support in shared_ptrs for marked ptrs. Working around
* it with custom deleters is a bit messy. Easier to build an abstraction
* which understands marked pointers
*
* C) A custom implementation is more efficient memory-wise, because we can force
* ref counted pointers to implement a common interface (and save having to
* allocated a separate control block)
*/
class atomic_ref_counted {
protected:
// construction does NOT increment reference count
atomic_ref_counted() : count_(0) {}
~atomic_ref_counted()
{
assert(count_.load() == 0);
}
public:
inline void
inc()
{
count_++;
}
// returns true if last decrement
inline bool
dec()
{
assert(count_.load() > 0);
return --count_ == 0;
}
private:
std::atomic<uint32_t> count_;
};
class nop_ref_counted {
public:
inline void inc() {}
inline bool dec() { return false; }
};
namespace private_ {
template <typename T>
struct ptr_ops_mixin {
typedef intptr_t opaque_t;
static inline opaque_t
Mark(opaque_t p)
{
return p | 0x1;
}
static inline bool
IsMarked(opaque_t p)
{
return p & 0x1;
}
static inline T *
Ptr(opaque_t p)
{
return (T *) (p & ~0x1);
}
static inline opaque_t
BuildOpaque(T *ptr, opaque_t op)
{
return opaque_t(ptr) | (op & 0x1);
}
};
}
class nop_lock {
public:
inline void lock() {}
inline void unlock() {}
inline bool try_lock() { return true; }
};
// T must inherit atomic_ref_counted (or implement the same interface)
// this class also supports one-time marking of ptrs.
//
// Doesn't support custom deleter
template <typename T, typename LockImpl = spinlock>
class atomic_ref_ptr : public private_::ptr_ops_mixin<T> {
typedef typename private_::ptr_ops_mixin<T>::opaque_t opaque_t;
typedef LockImpl lock_type;
typedef std::lock_guard<lock_type> lock_guard;
public:
// nullptr constructor
atomic_ref_ptr() : ptr_(opaque_t(nullptr)), mutex_() {}
~atomic_ref_ptr() {
T *ptr = get();
if (ptr && ptr->dec())
delete ptr;
}
// constructors don't accept a marked ptr
explicit atomic_ref_ptr(T *ptr)
: ptr_(opaque_t(ptr)), mutex_()
{
if (ptr)
ptr->inc();
}
template <typename U>
explicit atomic_ref_ptr(U *ptr)
: ptr_(opaque_t(static_cast<T *>(ptr))), mutex_()
{
if (ptr)
ptr->inc();
}
// Copy construction/assignment
//
// NOTE: Copy assignments don't propagate the marks
//
// NOTE: Assigning to a reference preserves its current mark
atomic_ref_ptr(const atomic_ref_ptr &other)
: ptr_(opaque_t(nullptr)), mutex_()
{
assignFrom(other);
}
template <typename U, typename V>
atomic_ref_ptr(const atomic_ref_ptr<U, V> &other)
: ptr_(opaque_t(nullptr)), mutex_()
{
assignFrom(other);
}
atomic_ref_ptr &
operator=(const atomic_ref_ptr &other)
{
assignFrom(other);
return *this;
}
template <typename U, typename V>
atomic_ref_ptr &
operator=(const atomic_ref_ptr<U, V> &other)
{
assignFrom(other);
return *this;
}
explicit inline
operator bool() const
{
return get();
}
T &
operator*() const
{
return *get();
}
T *
operator->() const
{
return get();
}
template <typename U, typename V>
inline bool
operator==(const atomic_ref_ptr<U, V> &other) const
{
return get() == other.get();
}
template <typename U, typename V>
inline bool
operator!=(const atomic_ref_ptr<U, V> &other) const
{
return !operator==(other);
}
inline T *
get() const
{
return this->Ptr(get_raw());
}
inline bool
get_mark() const
{
return this->IsMarked(get_raw());
}
// returns when this ptr is marked- returns
// true if the caller was the one responsible for the marking
inline bool
mark()
{
retry:
opaque_t this_opaque = get_raw();
if (this->IsMarked(this_opaque))
return false;
opaque_t new_opaque = this->Mark(this_opaque);
if (!ptr_.compare_exchange_strong(this_opaque, new_opaque)) {
nop_pause();
goto retry;
}
assert(get_mark());
return true;
}
// desired_value is stable by default because it is pass by value, so we
// don't need to lock it
inline bool
compare_exchange_strong(
const atomic_ref_ptr &expected_value,
atomic_ref_ptr desired_value)
{
std::lock(mutex_, expected_value.mutex_);
lock_guard l0(mutex_, std::adopt_lock);
lock_guard l1(expected_value.mutex_, std::adopt_lock);
opaque_t expected_opaque = expected_value.ptr_.load(); // assume stable
opaque_t desired_opaque = desired_value.ptr_.load();
if (!ptr_.compare_exchange_strong(expected_opaque, desired_opaque))
return false;
T *expected_ptr = this->Ptr(expected_opaque);
T *desired_ptr = this->Ptr(desired_opaque);
if (expected_ptr == desired_ptr)
// self-exchange
return true;
if (desired_ptr)
desired_ptr->inc();
if (expected_ptr && expected_ptr->dec())
delete expected_ptr;
return true;
}
private:
template <typename U, typename V>
void
assignFrom(const atomic_ref_ptr<U, V> &other)
{
retry:
std::lock(mutex_, other.mutex_);
lock_guard l0(mutex_, std::adopt_lock);
typename atomic_ref_ptr<U, V>::lock_guard l1(other.mutex_, std::adopt_lock);
opaque_t this_opaque = get_raw();
T *this_ptr = this->Ptr(this_opaque);
T *that_ptr = other.get();
if (this_ptr == that_ptr) {
// self-assignment
return;
}
opaque_t new_opaque = this->BuildOpaque(that_ptr, this_opaque);
// could have a concurrent marker
if (!ptr_.compare_exchange_strong(this_opaque, new_opaque)) {
nop_pause();
goto retry;
}
if (that_ptr)
that_ptr->inc();
if (this_ptr && this_ptr->dec())
delete this_ptr;
}
inline opaque_t
get_raw() const
{
return ptr_.load();
}
std::atomic<opaque_t> ptr_;
// this spinlock guards Ptr(ptr_) from changing (marks can change w/o grabing
// mutex)
//
// Why do we need a mutex for ref counting? This is because we assume
// that the source of a copy assignment (ie v in p = v) is un-stable,
// that is, it can experience concurrent modification during the assignment.
// Note that without this assumption, this pointer is of limited use.
//
// Given this assumption, each assignment has a potental race condition!
// That is, it is not possible to do a load() from the source followed by
// an increment of the reference count *atomically* w/o a lock. Thus, we
// need a lock to allow us to atomically load and increment.
mutable lock_type mutex_;
};