#include <signal.h>
+#include <unistd.h>
+#ifdef CONFIG_HAVE_TIMERFD_CREATE
+#include <sys/timerfd.h>
+#endif
#ifdef CONFIG_VALGRIND_DEV
#include <valgrind/drd.h>
#else
#define DRD_IGNORE_VAR(x) do { } while (0)
#endif
+#ifdef WIN32
+#include "os/os-windows.h"
+#endif
+
#include "fio.h"
#include "smalloc.h"
#include "helper_thread.h"
#include "steadystate.h"
#include "pshared.h"
+static int sleep_accuracy_ms;
+static int timerfd = -1;
+
enum action {
A_EXIT = 1,
A_RESET = 2,
struct fio_sem *startup_sem;
} *helper_data;
+struct interval_timer {
+ const char *name;
+ struct timespec expires;
+ uint32_t interval_ms;
+ int (*func)(void);
+};
+
void helper_thread_destroy(void)
{
if (!helper_data)
pthread_join(helper_data->thread, NULL);
}
+/* Resets timers and returns the time in milliseconds until the next event. */
+static int reset_timers(struct interval_timer timer[], int num_timers,
+ struct timespec *now)
+{
+ uint32_t msec_to_next_event = INT_MAX;
+ int i;
+
+ for (i = 0; i < num_timers; ++i) {
+ timer[i].expires = *now;
+ timespec_add_msec(&timer[i].expires, timer[i].interval_ms);
+ msec_to_next_event = min_not_zero(msec_to_next_event,
+ timer[i].interval_ms);
+ }
+
+ return msec_to_next_event;
+}
+
/*
* Waits for an action from fd during at least timeout_ms. `fd` must be in
* non-blocking mode.
};
fd_set rfds, efds;
uint8_t action = 0;
+ uint64_t exp;
int res;
res = read_from_pipe(fd, &action, sizeof(action));
FD_SET(fd, &rfds);
FD_ZERO(&efds);
FD_SET(fd, &efds);
- res = select(fd + 1, &rfds, NULL, &efds, &timeout);
+#ifdef CONFIG_HAVE_TIMERFD_CREATE
+ {
+ /*
+ * If the timer frequency is 100 Hz, select() will round up
+ * `timeout` to the next multiple of 1 / 100 Hz = 10 ms. Hence
+ * use a high-resolution timer if possible to increase
+ * select() timeout accuracy.
+ */
+ struct itimerspec delta = {};
+
+ delta.it_value.tv_sec = timeout.tv_sec;
+ delta.it_value.tv_nsec = timeout.tv_usec * 1000;
+ res = timerfd_settime(timerfd, 0, &delta, NULL);
+ assert(res == 0);
+ FD_SET(timerfd, &rfds);
+ }
+#endif
+ res = select(max(fd, timerfd) + 1, &rfds, NULL, &efds,
+ timerfd >= 0 ? NULL : &timeout);
if (res < 0) {
log_err("fio: select() call in helper thread failed: %s",
strerror(errno));
}
if (FD_ISSET(fd, &rfds))
read_from_pipe(fd, &action, sizeof(action));
+ if (timerfd >= 0 && FD_ISSET(timerfd, &rfds)) {
+ res = read(timerfd, &exp, sizeof(exp));
+ assert(res == sizeof(exp));
+ }
return action;
}
-static unsigned int task_helper(struct timespec *last, struct timespec *now, unsigned int period, void do_task())
+/*
+ * Verify whether or not timer @it has expired. If timer @it has expired, call
+ * @it->func(). @now is the current time. @msec_to_next_event is an
+ * input/output parameter that represents the time until the next event.
+ */
+static int eval_timer(struct interval_timer *it, const struct timespec *now,
+ unsigned int *msec_to_next_event)
{
- unsigned int next, since;
-
- since = mtime_since(last, now);
- if (since >= period || period - since < 10) {
- do_task();
- timespec_add_msec(last, since);
- if (since > period)
- next = period - (since - period);
- else
- next = period;
- } else
- next = period - since;
-
- return next;
+ int64_t delta_ms;
+ bool expired;
+
+ /* interval == 0 means that the timer is disabled. */
+ if (it->interval_ms == 0)
+ return 0;
+
+ delta_ms = rel_time_since(now, &it->expires);
+ expired = delta_ms <= sleep_accuracy_ms;
+ if (expired) {
+ timespec_add_msec(&it->expires, it->interval_ms);
+ delta_ms = rel_time_since(now, &it->expires);
+ if (delta_ms < it->interval_ms - sleep_accuracy_ms ||
+ delta_ms > it->interval_ms + sleep_accuracy_ms) {
+ dprint(FD_HELPERTHREAD,
+ "%s: delta = %" PRIi64 " <> %u. Clock jump?\n",
+ it->name, delta_ms, it->interval_ms);
+ delta_ms = it->interval_ms;
+ it->expires = *now;
+ timespec_add_msec(&it->expires, it->interval_ms);
+ }
+ }
+ *msec_to_next_event = min((unsigned int)delta_ms, *msec_to_next_event);
+ return expired ? it->func() : 0;
}
static void *helper_thread_main(void *data)
{
struct helper_data *hd = data;
- unsigned int msec_to_next_event, next_log, next_si = status_interval;
- unsigned int next_ss = STEADYSTATE_MSEC;
- struct timespec ts, last_du, last_ss, last_si;
+ unsigned int msec_to_next_event, next_log;
+ struct interval_timer timer[] = {
+ {
+ .name = "disk_util",
+ .interval_ms = DISK_UTIL_MSEC,
+ .func = update_io_ticks,
+ },
+ {
+ .name = "status_interval",
+ .interval_ms = status_interval,
+ .func = __show_running_run_stats,
+ },
+ {
+ .name = "steadystate",
+ .interval_ms = steadystate_enabled ? STEADYSTATE_MSEC :
+ 0,
+ .func = steadystate_check,
+ }
+ };
+ struct timespec ts;
+ long clk_tck;
int ret = 0;
+ os_clk_tck(&clk_tck);
+
+ dprint(FD_HELPERTHREAD, "clk_tck = %ld\n", clk_tck);
+ assert(clk_tck > 0);
+ sleep_accuracy_ms = (1000 + clk_tck - 1) / clk_tck;
+
+#ifdef CONFIG_HAVE_TIMERFD_CREATE
+ timerfd = timerfd_create(CLOCK_MONOTONIC, TFD_NONBLOCK);
+ assert(timerfd >= 0);
+ sleep_accuracy_ms = 1;
+#endif
+
sk_out_assign(hd->sk_out);
/* Let another thread handle signals. */
block_signals();
fio_get_mono_time(&ts);
- memcpy(&last_du, &ts, sizeof(ts));
- memcpy(&last_ss, &ts, sizeof(ts));
- memcpy(&last_si, &ts, sizeof(ts));
+ msec_to_next_event = reset_timers(timer, FIO_ARRAY_SIZE(timer), &ts);
fio_sem_up(hd->startup_sem);
- msec_to_next_event = DISK_UTIL_MSEC;
while (!ret && !hd->exit) {
- uint64_t since_du;
uint8_t action;
+ int i;
action = wait_for_action(hd->pipe[0], msec_to_next_event);
if (action == A_EXIT)
fio_get_mono_time(&ts);
- if (action == A_RESET) {
- last_du = ts;
- last_ss = ts;
- }
+ msec_to_next_event = INT_MAX;
+
+ if (action == A_RESET)
+ msec_to_next_event = reset_timers(timer,
+ FIO_ARRAY_SIZE(timer), &ts);
- since_du = mtime_since(&last_du, &ts);
- if (since_du >= DISK_UTIL_MSEC || DISK_UTIL_MSEC - since_du < 10) {
- ret = update_io_ticks();
- timespec_add_msec(&last_du, DISK_UTIL_MSEC);
- msec_to_next_event = DISK_UTIL_MSEC;
- if (since_du >= DISK_UTIL_MSEC)
- msec_to_next_event -= (since_du - DISK_UTIL_MSEC);
- } else
- msec_to_next_event = DISK_UTIL_MSEC - since_du;
+ for (i = 0; i < FIO_ARRAY_SIZE(timer); ++i)
+ ret = eval_timer(&timer[i], &ts, &msec_to_next_event);
if (action == A_DO_STAT)
__show_running_run_stats();
- if (status_interval) {
- next_si = task_helper(&last_si, &ts, status_interval, __show_running_run_stats);
- msec_to_next_event = min(next_si, msec_to_next_event);
- }
-
next_log = calc_log_samples();
if (!next_log)
next_log = DISK_UTIL_MSEC;
- if (steadystate_enabled) {
- next_ss = task_helper(&last_ss, &ts, STEADYSTATE_MSEC, steadystate_check);
- msec_to_next_event = min(next_ss, msec_to_next_event);
- }
-
msec_to_next_event = min(next_log, msec_to_next_event);
- dprint(FD_HELPERTHREAD, "next_si: %u, next_ss: %u, next_log: %u, msec_to_next_event: %u\n",
- next_si, next_ss, next_log, msec_to_next_event);
+ dprint(FD_HELPERTHREAD,
+ "next_log: %u, msec_to_next_event: %u\n",
+ next_log, msec_to_next_event);
if (!is_backend)
print_thread_status();
}
+ if (timerfd >= 0) {
+ close(timerfd);
+ timerfd = -1;
+ }
+
fio_writeout_logs(false);
sk_out_drop();