aboutsummaryrefslogtreecommitdiff
path: root/include/salticidae/queue.h
blob: 3df1529f351e08f79919736a63a47a6f11163a18 (plain) (blame)
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
#ifndef _SALTICIDAE_QUEUE_H
#define _SALTICIDAE_QUEUE_H

#include <atomic>
#include <vector>
#include <cassert>
#include <thread>

namespace salticidae {

static size_t const cacheline_size = 64;

class FreeList {
    public:
    struct Node {
        std::atomic<Node *> next;
        std::atomic<size_t> refcnt;
        Node(): next(nullptr), refcnt(1) {}
    };

    private:
    alignas(cacheline_size) std::atomic<Node *> top;

    public:
    FreeList(): top(nullptr) {}
    FreeList(const FreeList &) = delete;
    FreeList(FreeList &&) = delete;

    void release_ref(Node *u) {
        if (u->refcnt.fetch_sub(1, std::memory_order_relaxed) != 1) return;
        for (;;)
        {
            auto t = top.load(std::memory_order_relaxed);
            // repair the next pointer before CAS, otherwise u->next == nullptr
            // could lead to skipping elements
            u->next.store(t, std::memory_order_relaxed);
            // the replacement is ok even if ABA happens
            if (top.compare_exchange_weak(t, u, std::memory_order_release))
            {
                u->refcnt.store(1, std::memory_order_relaxed);
                break;
            }
        }
    }

    bool push(Node *u) {
        release_ref(u);
        return true;
    }

    bool pop(Node *&r) {
        bool loop = true;
        while (loop)
        {
            auto u = top.load(std::memory_order_acquire);
            /* the list is now empty */
            if (u == nullptr) return false;
            auto t = u->refcnt.load(std::memory_order_relaxed);
            /* let's wait for another round if u is a ghost (already popped) */
            if (!t) continue;
            /* otherwise t > 0, so with CAS, the invariant that zero refcnt can
             * never be increased is guaranteed */
            if (u->refcnt.compare_exchange_weak(t, t + 1, std::memory_order_relaxed))
            {
                /* here, nobody is able to change v->next (because v->next is
                 * only changed when pushed) even when ABA happens */
                auto v = u;
                auto nv = u->next.load(std::memory_order_relaxed);
                if (top.compare_exchange_weak(v, nv, std::memory_order_relaxed))
                {
                    /* manage to pop the head */
                    r = u;
                    loop = false;
                    /* do not need to try cas_push here because the current
                     * thread is the only one who can push u back */
                }
                /* release the refcnt and execute the delayed push call if
                 * necessary */
                release_ref(u);
            }
        }
        return true;
    }
};

const size_t MPMCQ_SIZE = 4096;

template<typename T>
class MPMCQueue {
    protected:
    struct Block: public FreeList::Node {
        alignas(cacheline_size) std::atomic<uint32_t> head;
        alignas(cacheline_size) std::atomic<uint32_t> tail;
        T elem[MPMCQ_SIZE];
        std::atomic<bool> avail[MPMCQ_SIZE];
        std::atomic<Block *> next;
    };

    FreeList blks;

    alignas(cacheline_size) std::atomic<Block *> head;
    alignas(cacheline_size) std::atomic<Block *> tail;

    template<typename U>
    bool _enqueue(U &&e, bool unbounded = true) {
        for (;;)
        {
            auto t = tail.load(std::memory_order_relaxed);
            auto tcnt = t->refcnt.load(std::memory_order_relaxed);
            if (!tcnt) continue;
            if (!t->refcnt.compare_exchange_weak(tcnt, tcnt + 1, std::memory_order_relaxed))
                continue;
            auto tt = t->tail.load(std::memory_order_relaxed);
            if (tt >= MPMCQ_SIZE)
            {
                if (t->next.load(std::memory_order_relaxed) == nullptr)
                {
                    FreeList::Node * _nblk;
                    if (!blks.pop(_nblk))
                    {
                        if (unbounded) _nblk = new Block();
                        else {
                            blks.release_ref(t);
                            return false;
                        }
                    }
                    auto nblk = static_cast<Block *>(_nblk);
                    nblk->head.store(0, std::memory_order_relaxed);
                    nblk->tail.store(0, std::memory_order_relaxed);
                    nblk->next.store(nullptr, std::memory_order_relaxed);
                    Block *tnext = nullptr;
                    if (!t->next.compare_exchange_weak(tnext, nblk, std::memory_order_acq_rel))
                        blks.push(_nblk);
                    else
                        tail.store(nblk, std::memory_order_relaxed);
                }
                blks.release_ref(t);
                continue;
            }
            auto tt2 = tt;
            if (t->tail.compare_exchange_weak(tt2, tt2 + 1, std::memory_order_relaxed))
            {
                new (&(t->elem[tt])) T(std::forward<U>(e));
                t->avail[tt].store(true, std::memory_order_release);
                blks.release_ref(t);
                break;
            }
            blks.release_ref(t);
        }
        return true;
    }

    public:
    MPMCQueue(const MPMCQueue &) = delete;
    MPMCQueue(MPMCQueue &&) = delete;

    MPMCQueue(): head(new Block()), tail(head.load()) {
        auto h = head.load();
        h->head = h->tail = 0;
        memset(h->avail, 0, sizeof(h->avail));
        h->next = nullptr;
    }

    ~MPMCQueue() {
        for (FreeList::Node *ptr; blks.pop(ptr); ) delete ptr;
        for (Block *ptr = head.load(), *nptr; ptr; ptr = nptr)
        {
            nptr = ptr->next;
            delete ptr;
        }
    }

    void set_capacity(size_t capacity = 0) {
        capacity = std::max(capacity / MPMCQ_SIZE, (size_t)1);
        while (capacity--) blks.push(new Block());
    }

    template<typename U>
    bool enqueue(U &&e, bool unbounded = true) {
        return _enqueue(e, unbounded);
    }

    template<typename U>
    bool try_enqueue(U &&e) {
        return _enqueue(e, false);
    }

    bool try_dequeue(T &e) {
        for (;;)
        {
            auto h = this->head.load(std::memory_order_relaxed);
            auto hcnt = h->refcnt.load(std::memory_order_relaxed);
            if (!hcnt) continue;
            if (!h->refcnt.compare_exchange_weak(hcnt, hcnt + 1, std::memory_order_relaxed))
                continue;

            auto hh = h->head.load(std::memory_order_relaxed);
            auto tt = h->tail.load(std::memory_order_relaxed);
            if (hh >= tt)
            {
                if (tt < MPMCQ_SIZE) { blks.release_ref(h); return false; }
                auto hnext = h->next.load(std::memory_order_relaxed);
                if (hnext == nullptr) { blks.release_ref(h); return false; }
                if (this->head.compare_exchange_weak(h, hnext, std::memory_order_relaxed))
                    this->blks.push(h);
                blks.release_ref(h);
                continue;
            }
            while (!h->avail[hh].load(std::memory_order_acquire))
                std::this_thread::yield();
            auto hh2 = hh;
            if (h->head.compare_exchange_weak(hh2, hh2 + 1, std::memory_order_relaxed))
            {
                e = std::move(h->elem[hh]);
                blks.release_ref(h);
                break;
            }
            blks.release_ref(h);
        }
        return true;
    }
};

template<typename T>
struct MPSCQueue: public MPMCQueue<T> {
    using MPMCQueue<T>::MPMCQueue;
    /* the same thread is calling the following functions */

    bool try_dequeue(T &e) {
        for (;;)
        {
            auto h = this->head.load(std::memory_order_relaxed);
            auto hh = h->head.load(std::memory_order_relaxed);
            auto tt = h->tail.load(std::memory_order_relaxed);
            if