static void InitializeCachePool()
{
int cpus = CPU_GetCount();
/* Round up to the next power of two. */
if (cpus <= 1)
cpus = 1;
else if (cpus <= 2)
cpus = 2;
else if (cpus <= 4)
cpus = 4;
else if (cpus <= 8)
cpus = 8;
#ifdef PAL_64BIT
else if (cpus > 16)
cpus = 32;
#endif
else
cpus = 16;
/* Everyone needs to mask off higher bits in case CPU_GetCurrent ever
* returns a number higher than it did now, or in case there are more than
* 32 CPUs. */
s_cpuMask = cpus - 1;
s_masterMask = ((ptrdiff_t)1 << cpus) - 1;
PAL_Atexit(ShutdownCachePool);
}
/**
* Run fn on each CPU in cs, passing the CPU number and the opaque
* argument, then join with all threads. If cs is NULL, fn is run on
* all affinitized CPU's (all CPU's by default). If results is
* non-NULL, it will be pointed to an array of *nResults results. The
* caller must free this array.
*/
void
CPU_RunOnSet(int (*fn)(int, void*), void *opaque, CPU_Set_t *cs,
int **results, int *nResults)
{
int r;
CPU_Set_t *aff = CPU_GetAffinity();
if (!cs)
cs = aff;
int count = CPU_GetCount(cs);
int cpus = cpuSetMax(cs);
struct runOnCPU roc[cpus];
pthread_t threads[cpus];
pthread_barrier_t barrier;
r = pthread_barrier_init(&barrier, NULL, count + 1);
if (r != 0) {
errno = r;
epanic("failed to create pthread barrier");
}
if (results)
*results = malloc(count * sizeof(*results));
if (nResults)
*nResults = count;
int i = 0;
for (int cpu = 0; cpu < cpus; ++cpu) {
if (numa_bitmask_isbitset(cs, cpu)) {
roc[cpu].fn = fn;
roc[cpu].cpu = cpu;
roc[cpu].i = i++;
roc[cpu].results = results ? *results : NULL;
roc[cpu].barrier = &barrier;
// Make sure the stack and such are allocated
// on this core.
CPU_Bind(cpu);
r = pthread_create(&threads[cpu], NULL, runOnCPUThread,
&roc[cpu]);
if (r != 0) {
errno = r;
epanic("failed to spawn thread on CPU %d", cpu);
}
}
}
CPU_SetAffinity(aff);
pthread_barrier_wait(&barrier);
for (int cpu = 0; cpu < cpus; ++cpu) {
if (numa_bitmask_isbitset(cs, cpu)) {
r = pthread_join(threads[cpu], NULL);
if (r != 0) {
errno = r;
epanic("failed to join thread on CPU %d", cpu);
}
}
}
pthread_barrier_destroy(&barrier);
CPU_FreeSet(aff);
}
