-#include "main.h"
+#include <threads.h>
+#include <stdlib.h>
+#include "librace.h"
+#include "model-assert.h"
-extern unsigned pid;
-extern unsigned iterations;
-extern unsigned initial_nodes;
-extern private_t private;
-extern shared_mem_t* smp;
+#include "my_queue.h"
-void init_private()
-{
- private.node = 2 + initial_nodes + pid;
- private.value = 1 + initial_nodes + (pid * iterations);
+#define relaxed memory_order_relaxed
+#define release memory_order_release
+#define acquire memory_order_acquire
-}
+#define MAX_FREELIST 4 /* Each thread can own up to MAX_FREELIST free nodes */
+#define INITIAL_FREE 2 /* Each thread starts with INITIAL_FREE free nodes */
-void init_memory()
-{
-}
+#define POISON_IDX 0x666
+
+static unsigned int (*free_lists)[MAX_FREELIST];
-static unsigned new_node()
+/* Search this thread's free list for a "new" node */
+static unsigned int new_node()
{
- return private.node;
+ int i;
+ int t = get_thread_num();
+ for (i = 0; i < MAX_FREELIST; i++) {
+ unsigned int node = load_32(&free_lists[t][i]);
+ if (node) {
+ store_32(&free_lists[t][i], 0);
+ return node;
+ }
+ }
+ /* free_list is empty? */
+ MODEL_ASSERT(0);
+ return 0;
}
-static void reclaim(unsigned node)
+/* Place this node index back on this thread's free list */
+static void reclaim(unsigned int node)
{
- private.node = node;
+ int i;
+ int t = get_thread_num();
+
+ /* Don't reclaim NULL node */
+ MODEL_ASSERT(node);
+
+ for (i = 0; i < MAX_FREELIST; i++) {
+ /* Should never race with our own thread here */
+ unsigned int idx = load_32(&free_lists[t][i]);
+
+ /* Found empty spot in free list */
+ if (idx == 0) {
+ store_32(&free_lists[t][i], node);
+ return;
+ }
+ }
+ /* free list is full? */
+ MODEL_ASSERT(0);
}
-void init_queue()
+void init_queue(queue_t *q, int num_threads)
{
- unsigned i;
+ int i, j;
- /* initialize queue */
- smp->head.sep.ptr = 1;
- smp->head.sep.count = 0;
- smp->tail.sep.ptr = 1;
- smp->tail.sep.count = 0;
- smp->nodes[1].next.sep.ptr = NULL;
- smp->nodes[1].next.sep.count = 0;
-
- /* initialize avail list */
- for (i=2; i<MAX_NODES; i++) {
- smp->nodes[i].next.sep.ptr = i+1;
- smp->nodes[i].next.sep.count = 0;
- }
- smp->nodes[MAX_NODES].next.sep.ptr = NULL;
- smp->nodes[MAX_NODES].next.sep.count = 0;
-
- /* initialize queue contents */
- if (initial_nodes > 0) {
- for (i=2; i<initial_nodes+2; i++) {
- smp->nodes[i].value = i;
- smp->nodes[i-1].next.sep.ptr = i;
- smp->nodes[i].next.sep.ptr = NULL;
+ /* Initialize each thread's free list with INITIAL_FREE pointers */
+ /* The actual nodes are initialized with poison indexes */
+ free_lists = malloc(num_threads * sizeof(*free_lists));
+ for (i = 0; i < num_threads; i++) {
+ for (j = 0; j < INITIAL_FREE; j++) {
+ free_lists[i][j] = 2 + i * MAX_FREELIST + j;
+ atomic_init(&q->nodes[free_lists[i][j]].next, MAKE_POINTER(POISON_IDX, 0));
}
- smp->head.sep.ptr = 1;
- smp->tail.sep.ptr = 1 + initial_nodes;
}
+
+ /* initialize queue */
+ atomic_init(&q->head, MAKE_POINTER(1, 0));
+ atomic_init(&q->tail, MAKE_POINTER(1, 0));
+ atomic_init(&q->nodes[1].next, MAKE_POINTER(0, 0));
}
-void enqueue(unsigned val)
+void enqueue(queue_t *q, unsigned int val)
{
- unsigned success;
- unsigned node;
- pointer_t tail;
- pointer_t next;
+ int success = 0;
+ unsigned int node;
+ pointer tail;
+ pointer next;
+ pointer tmp;
node = new_node();
- smp->nodes[node].value = val;
- smp->nodes[node].next.sep.ptr = NULL;
-
- for (success = FALSE; success == FALSE; ) {
- tail.con = smp->tail.con;
- next.con = smp->nodes[tail.sep.ptr].next.con;
- if (tail.con == smp->tail.con) {
- if (next.sep.ptr == NULL) {
- success = cas(&smp->nodes[tail.sep.ptr].next,
- next.con,
- MAKE_LONG(node, next.sep.count+1));
+ store_32(&q->nodes[node].value, val);
+ tmp = atomic_load_explicit(&q->nodes[node].next, relaxed);
+ set_ptr(&tmp, 0); // NULL
+ atomic_store_explicit(&q->nodes[node].next, tmp, relaxed);
+
+ while (!success) {
+ tail = atomic_load_explicit(&q->tail, acquire);
+ next = atomic_load_explicit(&q->nodes[get_ptr(tail)].next, acquire);
+ if (tail == atomic_load_explicit(&q->tail, relaxed)) {
+
+ /* Check for uninitialized 'next' */
+ MODEL_ASSERT(get_ptr(next) != POISON_IDX);
+
+ if (get_ptr(next) == 0) { // == NULL
+ pointer value = MAKE_POINTER(node, get_count(next) + 1);
+ success = atomic_compare_exchange_strong_explicit(&q->nodes[get_ptr(tail)].next,
+ &next, value, release, release);
}
- if (success == FALSE) {
- cas(&smp->tail,
- tail.con,
- MAKE_LONG(smp->nodes[tail.sep.ptr].next.sep.ptr,
- tail.sep.count+1));
+ if (!success) {
+ unsigned int ptr = get_ptr(atomic_load_explicit(&q->nodes[get_ptr(tail)].next, acquire));
+ pointer value = MAKE_POINTER(ptr,
+ get_count(tail) + 1);
+ atomic_compare_exchange_strong_explicit(&q->tail,
+ &tail, value,
+ release, release);
thrd_yield();
}
}
}
- cas(&smp->tail,
- tail.con,
- MAKE_LONG(node, tail.sep.count+1));
+ atomic_compare_exchange_strong_explicit(&q->tail,
+ &tail,
+ MAKE_POINTER(node, get_count(tail) + 1),
+ release, release);
}
-unsigned dequeue()
+unsigned int dequeue(queue_t *q)
{
- unsigned value;
- unsigned success;
- pointer_t head;
- pointer_t tail;
- pointer_t next;
-
- for (success = FALSE; success == FALSE; ) {
- head.con = smp->head.con;
- tail.con = smp->tail.con;
- next.con = smp->nodes[head.sep.ptr].next.con;
- if (smp->head.con == head.con) {
- if (head.sep.ptr == tail.sep.ptr) {
- if (next.sep.ptr == NULL) {
- return NULL;
+ unsigned int value;
+ int success = 0;
+ pointer head;
+ pointer tail;
+ pointer next;
+
+ while (!success) {
+ head = atomic_load_explicit(&q->head, acquire);
+ tail = atomic_load_explicit(&q->tail, relaxed);
+ next = atomic_load_explicit(&q->nodes[get_ptr(head)].next, acquire);
+ if (atomic_load_explicit(&q->head, relaxed) == head) {
+ if (get_ptr(head) == get_ptr(tail)) {
+
+ /* Check for uninitialized 'next' */
+ MODEL_ASSERT(get_ptr(next) != POISON_IDX);
+
+ if (get_ptr(next) == 0) { // NULL
+ return 0; // NULL
}
- cas(&smp->tail,
- tail.con,
- MAKE_LONG(next.sep.ptr, tail.sep.count+1));
+ atomic_compare_exchange_strong_explicit(&q->tail,
+ &tail,
+ MAKE_POINTER(get_ptr(next), get_count(tail) + 1),
+ release, release);
thrd_yield();
} else {
- value = smp->nodes[next.sep.ptr].value;
- success = cas(&smp->head,
- head.con,
- MAKE_LONG(next.sep.ptr, head.sep.count+1));
- if (success == FALSE) {
+ value = load_32(&q->nodes[get_ptr(next)].value);
+ success = atomic_compare_exchange_strong_explicit(&q->head,
+ &head,
+ MAKE_POINTER(get_ptr(next), get_count(head) + 1),
+ release, release);
+ if (!success)
thrd_yield();
- }
}
}
}
- reclaim(head.sep.ptr);
+ reclaim(get_ptr(head));
return value;
}