ms-queue/queue_wildcard.c

   1 #include <threads.h>
   2 #include <stdlib.h>
   3 #include "librace.h"
   4 #include "model-assert.h"
   5
   6 #include "queue.h"
   7 #include "wildcard.h"
   8
   9 #define MAX_FREELIST 4 /* Each thread can own up to MAX_FREELIST free nodes */
  10 #define INITIAL_FREE 2 /* Each thread starts with INITIAL_FREE free nodes */
  11
  12 #define POISON_IDX 0x666
  13
  14 static unsigned int (*free_lists)[MAX_FREELIST];
  15
  16 /* Search this thread's free list for a "new" node */
  17 static unsigned int new_node()
  18 {
  19         int i;
  20         int t = get_thread_num();
  21         for (i = 0; i < MAX_FREELIST; i++) {
  22                 //unsigned int node = load_32(&free_lists[t][i]);
  23                 unsigned int node = free_lists[t][i];
  24                 if (node) {
  25                         //store_32(&free_lists[t][i], 0);
  26                         free_lists[t][i] = 0;
  27                         return node;
  28                 }
  29         }
  30         /* free_list is empty? */
  31         MODEL_ASSERT(0);
  32         return 0;
  33 }
  34
  35 /* Place this node index back on this thread's free list */
  36 static void reclaim(unsigned int node)
  37 {
  38         int i;
  39         int t = get_thread_num();
  40
  41         /* Don't reclaim NULL node */
  42         //MODEL_ASSERT(node);
  43
  44         for (i = 0; i < MAX_FREELIST; i++) {
  45                 /* Should never race with our own thread here */
  46                 //unsigned int idx = load_32(&free_lists[t][i]);
  47                 unsigned int idx = free_lists[t][i];
  48
  49                 /* Found empty spot in free list */
  50                 if (idx == 0) {
  51                         //store_32(&free_lists[t][i], node);
  52                         free_lists[t][i] = node;
  53                         return;
  54                 }
  55         }
  56         /* free list is full? */
  57         MODEL_ASSERT(0);
  58 }
  59
  60 void init_queue(queue_t *q, int num_threads)
  61 {
  62         int i, j;
  63
  64         /* Initialize each thread's free list with INITIAL_FREE pointers */
  65         /* The actual nodes are initialized with poison indexes */
  66         free_lists = malloc(num_threads * sizeof(*free_lists));
  67         for (i = 0; i < num_threads; i++) {
  68                 for (j = 0; j < INITIAL_FREE; j++) {
  69                         free_lists[i][j] = 2 + i * MAX_FREELIST + j;
  70                         atomic_init(&q->nodes[free_lists[i][j]].next, MAKE_POINTER(POISON_IDX, 0));
  71                 }
  72         }
  73
  74         /* initialize queue */
  75         atomic_init(&q->head, MAKE_POINTER(1, 0));
  76         atomic_init(&q->tail, MAKE_POINTER(1, 0));
  77         atomic_init(&q->nodes[1].next, MAKE_POINTER(0, 0));
  78 }
  79
  80 void enqueue(queue_t *q, unsigned int val)
  81 {
  82         int success = 0;
  83         unsigned int node;
  84         pointer tail;
  85         pointer next;
  86         pointer tmp;
  87
  88         node = new_node();
  89         //store_32(&q->nodes[node].value, val);
  90         q->nodes[node].value = val;
  91         tmp = atomic_load_explicit(&q->nodes[node].next, wildcard(1)); // relaxed
  92         set_ptr(&tmp, 0); // NULL
  93         atomic_store_explicit(&q->nodes[node].next, tmp, wildcard(2)); // relaxed
  94
  95         while (!success) {
  96                 tail = atomic_load_explicit(&q->tail, wildcard(3)); // acquire
  97                 // FIXME: SCFence makes this relaxed
  98                 next = atomic_load_explicit(&q->nodes[get_ptr(tail)].next, wildcard(4)); //acquire
  99                 if (tail == atomic_load_explicit(&q->tail, wildcard(5))) { // relaxed
 100
 101                         /* Check for uninitialized 'next' */
 102                         //MODEL_ASSERT(get_ptr(next) != POISON_IDX);
 103
 104                         if (get_ptr(next) == 0) { // == NULL
 105                                 pointer value = MAKE_POINTER(node, get_count(next) + 1);
 106
 107                                 // ***********************************
 108                                 // Inference analysis results have two choices here, it either
 109                                 // makes wildcard(6) acq_rel and wildcard(8) relaxed or
 110                                 // wildcard(6) release and wildcard(8) acquire. The
 111                                 // synchronization here is for enqueue() to dequeue(), and
 112                                 // actually either synchronization options will work!!!
 113                                 success = atomic_compare_exchange_strong_explicit(&q->nodes[get_ptr(tail)].next,
 114                                                 &next, value, wildcard(6), wildcard(7)); // release & relaxed
 115                         }
 116                         if (!success) {
 117                                 unsigned int ptr =
 118                                         get_ptr(atomic_load_explicit(&q->nodes[get_ptr(tail)].next, wildcard(8))); // acquire
 119                                 pointer value = MAKE_POINTER(ptr,
 120                                                 get_count(tail) + 1);
 121                                 atomic_compare_exchange_strong_explicit(&q->tail,
 122                                                 &tail, value,
 123                                                 wildcard(9), wildcard(10)); // release & relaxed
 124                                 thrd_yield();
 125                         }
 126                 }
 127         }
 128         atomic_compare_exchange_strong_explicit(&q->tail,
 129                         &tail,
 130                         MAKE_POINTER(node, get_count(tail) + 1),
 131                         wildcard(11), wildcard(12)); // release & relaxed
 132 }
 133
 134 bool dequeue(queue_t *q, unsigned int *retVal)
 135 {
 136         unsigned int value;
 137         int success = 0;
 138         pointer head;
 139         pointer tail;
 140         pointer next;
 141
 142         while (!success) {
 143                 head = atomic_load_explicit(&q->head, wildcard(13)); // acquire
 144                 // SCFence makes this acquire, and we actually need an acquire here!!!
 145                 tail = atomic_load_explicit(&q->tail, wildcard(14)); // relaxed
 146                 next = atomic_load_explicit(&q->nodes[get_ptr(head)].next, wildcard(15)); // acquire
 147                 if (atomic_load_explicit(&q->head, wildcard(16)) == head) { // relaxed
 148                         if (get_ptr(head) == get_ptr(tail)) {
 149
 150                                 /* Check for uninitialized 'next' */
 151                                 //MODEL_ASSERT(get_ptr(next) != POISON_IDX);
 152
 153                                 if (get_ptr(next) == 0) { // NULL
 154                                         return false; // NULL
 155                                 }
 156                                 atomic_compare_exchange_strong_explicit(&q->tail,
 157                                                 &tail,
 158                                                 MAKE_POINTER(get_ptr(next), get_count(tail) + 1),
 159                                                 wildcard(17), wildcard(18)); // release & relaxed
 160                                 thrd_yield();
 161                         } else {
 162                                 //value = load_32(&q->nodes[get_ptr(next)].value);
 163                                 value = q->nodes[get_ptr(next)].value;
 164                                 success = atomic_compare_exchange_strong_explicit(&q->head,
 165                                                 &head, MAKE_POINTER(get_ptr(next), get_count(head) + 1),
 166                                                 wildcard(19), wildcard(20)); // release & relaxed
 167                                 if (!success)
 168                                         thrd_yield();
 169                         }
 170                 }
 171         }
 172         reclaim(get_ptr(head));
 173         *retVal = value;
 174         return true;
 175 }