#include "arch.h"
+/* Make this big enough to include the POWER5+ L3 cacheline size of 256B */
+#define CACHE_LINE_SIZE 4096
+
+/* hardcoded number of CPUs */
+#define NR_CPUS 16384
+
#if defined(_syscall0)
_syscall0(pid_t, gettid)
#elif defined(__NR_gettid)
struct per_thread_lock {
pthread_mutex_t lock;
-} __attribute__((aligned(128))); /* cache-line aligned */
+} __attribute__((aligned(CACHE_LINE_SIZE))); /* cache-line aligned */
static struct per_thread_lock *per_thread_lock;
static volatile int test_go, test_stop;
-static int wdelay;
+static unsigned long wdelay;
static volatile struct test_array test_array = { 8 };
cpu_relax();
}
+static int verbose_mode;
+
+#define printf_verbose(fmt, args...) \
+ do { \
+ if (verbose_mode) \
+ printf(fmt, args); \
+ } while (0)
+
+static unsigned int cpu_affinities[NR_CPUS];
+static unsigned int next_aff = 0;
+static int use_affinity = 0;
+
+pthread_mutex_t affinity_mutex = PTHREAD_MUTEX_INITIALIZER;
+
+static void set_affinity(void)
+{
+ cpu_set_t mask;
+ int cpu;
+ int ret;
+
+ if (!use_affinity)
+ return;
+
+ ret = pthread_mutex_lock(&affinity_mutex);
+ if (ret) {
+ perror("Error in pthread mutex lock");
+ exit(-1);
+ }
+ cpu = cpu_affinities[next_aff++];
+ ret = pthread_mutex_unlock(&affinity_mutex);
+ if (ret) {
+ perror("Error in pthread mutex unlock");
+ exit(-1);
+ }
+ CPU_ZERO(&mask);
+ CPU_SET(cpu, &mask);
+ sched_setaffinity(0, sizeof(mask), &mask);
+}
+
/*
* returns 0 if test should end.
*/
static unsigned long long __thread nr_writes;
static unsigned long long __thread nr_reads;
-static unsigned long long __attribute__((aligned(128))) *tot_nr_writes;
-static unsigned long long __attribute__((aligned(128))) *tot_nr_reads;
+static
+unsigned long long __attribute__((aligned(CACHE_LINE_SIZE))) *tot_nr_writes;
+static
+unsigned long long __attribute__((aligned(CACHE_LINE_SIZE))) *tot_nr_reads;
static unsigned int nr_readers;
static unsigned int nr_writers;
{
unsigned long tidx = (unsigned long)data;
- printf("thread_begin %s, thread id : %lx, tid %lu\n",
+ printf_verbose("thread_begin %s, thread id : %lx, tid %lu\n",
"reader", pthread_self(), (unsigned long)gettid());
+ set_affinity();
+
while (!test_go)
{
}
}
tot_nr_reads[tidx] = nr_reads;
- printf("thread_end %s, thread id : %lx, tid %lu\n",
+ printf_verbose("thread_end %s, thread id : %lx, tid %lu\n",
"reader", pthread_self(), (unsigned long)gettid());
return ((void*)1);
unsigned long wtidx = (unsigned long)data;
long tidx;
- printf("thread_begin %s, thread id : %lx, tid %lu\n",
+ printf_verbose("thread_begin %s, thread id : %lx, tid %lu\n",
"writer", pthread_self(), (unsigned long)gettid());
+ set_affinity();
+
while (!test_go)
{
}
}
test_array.a = 0;
test_array.a = 8;
- for (tidx = nr_readers - 1; tidx >= 0; tidx--) {
+ for (tidx = (long)nr_readers - 1; tidx >= 0; tidx--) {
pthread_mutex_unlock(&per_thread_lock[tidx].lock);
}
nr_writes++;
if (unlikely(!test_duration_write()))
break;
if (unlikely(wdelay))
- usleep(wdelay);
+ loop_sleep(wdelay);
}
- printf("thread_end %s, thread id : %lx, tid %lu\n",
+ printf_verbose("thread_end %s, thread id : %lx, tid %lu\n",
"writer", pthread_self(), (unsigned long)gettid());
tot_nr_writes[wtidx] = nr_writes;
return ((void*)2);
#endif
printf(" [-d delay] (writer period (us))");
printf(" [-c duration] (reader C.S. duration (in loops))");
+ printf(" [-v] (verbose output)");
printf(" [-a cpu#] [-a cpu#]... (affinity)");
printf("\n");
}
-cpu_set_t affinity;
-
int main(int argc, char **argv)
{
int err;
unsigned long long *count_reader, *count_writer;
unsigned long long tot_reads = 0, tot_writes = 0;
int i, a;
- int use_affinity = 0;
if (argc < 4) {
show_usage(argc, argv);
return -1;
}
- CPU_ZERO(&affinity);
-
for (i = 4; i < argc; i++) {
if (argv[i][0] != '-')
continue;
return -1;
}
a = atoi(argv[++i]);
- CPU_SET(a, &affinity);
+ cpu_affinities[next_aff++] = a;
use_affinity = 1;
- printf("Adding CPU %d affinity\n", a);
+ printf_verbose("Adding CPU %d affinity\n", a);
break;
case 'c':
if (argc < i + 2) {
show_usage(argc, argv);
return -1;
}
- rduration = atoi(argv[++i]);
+ rduration = atol(argv[++i]);
break;
case 'd':
if (argc < i + 2) {
show_usage(argc, argv);
return -1;
}
- wdelay = atoi(argv[++i]);
+ wdelay = atol(argv[++i]);
+ break;
+ case 'v':
+ verbose_mode = 1;
break;
}
}
- printf("running test for %lu seconds, %u readers, %u writers.\n",
+ printf_verbose("running test for %lu seconds, %u readers, %u writers.\n",
duration, nr_readers, nr_writers);
- printf("Writer delay : %u us.\n", wdelay);
- printf("thread %-6s, thread id : %lx, tid %lu\n",
+ printf_verbose("Writer delay : %lu loops.\n", wdelay);
+ printf_verbose("Reader duration : %lu loops.\n", rduration);
+ printf_verbose("thread %-6s, thread id : %lx, tid %lu\n",
"main", pthread_self(), (unsigned long)gettid());
- if (use_affinity
- && sched_setaffinity(0, sizeof(affinity), &affinity) < 0) {
- perror("sched_setaffinity");
- exit(-1);
- }
-
tid_reader = malloc(sizeof(*tid_reader) * nr_readers);
tid_writer = malloc(sizeof(*tid_writer) * nr_writers);
count_reader = malloc(sizeof(*count_reader) * nr_readers);
tot_nr_writes = malloc(sizeof(*tot_nr_writes) * nr_writers);
per_thread_lock = malloc(sizeof(*per_thread_lock) * nr_readers);
+ next_aff = 0;
+
for (i = 0; i < nr_readers; i++) {
err = pthread_create(&tid_reader[i], NULL, thr_reader,
(void *)(long)i);
exit(1);
tot_writes += tot_nr_writes[i];
}
-
- printf("total number of reads : %llu, writes %llu\n", tot_reads,
+
+ printf_verbose("total number of reads : %llu, writes %llu\n", tot_reads,
tot_writes);
+ printf("SUMMARY %-25s testdur %4lu nr_readers %3u rdur %6lu "
+ "nr_writers %3u "
+ "wdelay %6lu nr_reads %12llu nr_writes %12llu nr_ops %12llu\n",
+ argv[0], duration, nr_readers, rduration,
+ nr_writers, wdelay, tot_reads, tot_writes,
+ tot_reads + tot_writes);
+
free(tid_reader);
free(tid_writer);
free(count_reader);