static void disk_util_timer_arm(void);
static void print_thread_status(void);
+extern unsigned long long mlock_size;
+
/*
* thread life cycle
*/
enum {
TD_NOT_CREATED = 0,
TD_CREATED,
+ TD_INITIALIZED,
TD_RUNNING,
TD_VERIFYING,
TD_EXITED,
static sem_t startup_sem;
#define TERMINATE_ALL (-1)
+#define JOB_START_TIMEOUT (5 * 1000)
static void terminate_threads(int group_id)
{
if (td->cur_depth)
cleanup_pending_aio(td);
- if (should_fsync(td) && td->fsync_blocks)
+ if (should_fsync(td) && td->end_fsync)
sync_td(td);
}
return fio_posixaio_init(td);
else if (td->io_engine == FIO_SGIO)
return fio_sgio_init(td);
+ else if (td->io_engine == FIO_SPLICEIO)
+ return fio_spliceio_init(td);
else {
fprintf(stderr, "bad io_engine %d\n", td->io_engine);
return 1;
return;
/*
- * for md/dm, there's no queue dir. we already have the right place
+ * If there's a ../queue/ directory there, we are inside a partition.
+ * Check if that is the case and jump back. For loop/md/dm etc we
+ * are already in the right spot.
*/
- sprintf(tmp, "%s/stat", foo);
- if (stat(tmp, &st)) {
- /*
- * if this is inside a partition dir, jump back to parent
- */
- sprintf(tmp, "%s/queue", foo);
+ sprintf(tmp, "%s/../queue", foo);
+ if (!stat(tmp, &st)) {
+ p = dirname(foo);
+ sprintf(tmp, "%s/queue", p);
if (stat(tmp, &st)) {
- p = dirname(foo);
- sprintf(tmp, "%s/queue", p);
- if (stat(tmp, &st)) {
- fprintf(stderr, "unknown sysfs layout\n");
- return;
- }
- sprintf(foo, "%s", p);
+ fprintf(stderr, "unknown sysfs layout\n");
+ return;
}
+ sprintf(foo, "%s", p);
}
disk_util_add(dev, foo);
}
}
+ if (init_random_state(td))
+ goto err;
+
+ td_set_runstate(td, TD_INITIALIZED);
sem_post(&startup_sem);
sem_wait(&td->mutex);
if (!td->create_serialize && setup_file(td))
goto err;
- if (init_random_state(td))
- goto err;
-
gettimeofday(&td->epoch, NULL);
while (td->loops--) {
case TD_CREATED:
c = 'C';
break;
+ case TD_INITIALIZED:
+ c = 'I';
+ break;
case TD_NOT_CREATED:
c = 'P';
break;
if (td->timeout && eta_sec > (td->timeout - elapsed))
eta_sec = td->timeout - elapsed;
- } else if (td->runstate == TD_NOT_CREATED || td->runstate == TD_CREATED) {
+ } else if (td->runstate == TD_NOT_CREATED || td->runstate == TD_CREATED
+ || td->runstate == TD_INITIALIZED) {
int t_eta = 0, r_eta = 0;
/*
}
}
+static void fio_unpin_memory(void *pinned)
+{
+ if (pinned) {
+ if (munlock(pinned, mlock_size) < 0)
+ perror("munlock");
+ munmap(pinned, mlock_size);
+ }
+}
+
+static void *fio_pin_memory(void)
+{
+ long pagesize, pages;
+ void *ptr;
+
+ if (!mlock_size)
+ return NULL;
+
+ /*
+ * Don't allow mlock of more than real_mem-128MB
+ */
+ pagesize = sysconf(_SC_PAGESIZE);
+ pages = sysconf(_SC_PHYS_PAGES);
+ if (pages != -1 && pagesize != -1) {
+ unsigned long long real_mem = pages * pagesize;
+
+ if ((mlock_size + 128 * 1024 * 1024) > real_mem) {
+ mlock_size = real_mem - 128 * 1024 * 1024;
+ printf("fio: limiting mlocked memory to %lluMiB\n",
+ mlock_size >> 20);
+ }
+ }
+
+ ptr = mmap(NULL, mlock_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0);
+ if (!ptr) {
+ perror("malloc locked mem");
+ return NULL;
+ }
+ if (mlock(ptr, mlock_size) < 0) {
+ munmap(ptr, mlock_size);
+ perror("mlock");
+ return NULL;
+ }
+
+ return ptr;
+}
+
static void run_threads(void)
{
struct thread_data *td;
unsigned long spent;
int i, todo, nr_running, m_rate, t_rate, nr_started;
+ void *mlocked_mem;
+
+ mlocked_mem = fio_pin_memory();
printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
fflush(stdout);
gettimeofday(&genesis, NULL);
while (todo) {
+ struct thread_data *map[MAX_JOBS];
+ struct timeval this_start;
+ int this_jobs = 0, left;
+
/*
* create threads (TD_NOT_CREATED -> TD_CREATED)
*/
if (td->stonewall && (nr_started || nr_running))
break;
+ /*
+ * Set state to created. Thread will transition
+ * to TD_INITIALIZED when it's done setting up.
+ */
td_set_runstate(td, TD_CREATED);
+ map[this_jobs++] = td;
sem_init(&startup_sem, 0, 1);
- todo--;
nr_started++;
if (td->use_thread) {
}
/*
- * start created threads (TD_CREATED -> TD_RUNNING)
+ * Wait for the started threads to transition to
+ * TD_INITIALIZED.
+ */
+ printf("fio: Waiting for threads to initialize...\n");
+ gettimeofday(&this_start, NULL);
+ left = this_jobs;
+ while (left) {
+ if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
+ break;
+
+ usleep(100000);
+
+ for (i = 0; i < this_jobs; i++) {
+ td = map[i];
+ if (!td)
+ continue;
+ if (td->runstate == TD_INITIALIZED ||
+ td->runstate >= TD_EXITED) {
+ map[i] = NULL;
+ left--;
+ continue;
+ }
+ }
+ }
+
+ if (left) {
+ fprintf(stderr, "fio: %d jobs failed to start\n", left);
+ for (i = 0; i < this_jobs; i++) {
+ td = map[i];
+ if (!td)
+ continue;
+ kill(td->pid, SIGTERM);
+ }
+ break;
+ }
+
+ /*
+ * start created threads (TD_INITIALIZED -> TD_RUNNING)
*/
+ printf("fio: Go for launch\n");
for (i = 0; i < thread_number; i++) {
td = &threads[i];
- if (td->runstate != TD_CREATED)
+ if (td->runstate != TD_INITIALIZED)
continue;
td_set_runstate(td, TD_RUNNING);
nr_started--;
m_rate += td->ratemin;
t_rate += td->rate;
+ todo--;
sem_post(&td->mutex);
}
}
update_io_ticks();
+ fio_unpin_memory(mlocked_mem);
}
static void show_group_stats(struct group_run_stats *rs, int id)