* Clock functions
*/
-#include <unistd.h>
#include <math.h>
-#include <sys/time.h>
-#include <time.h>
#include "fio.h"
-#include "smalloc.h"
-
-#include "hash.h"
#include "os/os.h"
#if defined(ARCH_HAVE_CPU_CLOCK)
#ifndef ARCH_CPU_CLOCK_CYCLES_PER_USEC
-static unsigned long cycles_per_usec;
+static unsigned long long cycles_per_msec;
static unsigned long long cycles_start;
static unsigned long long clock_mult;
static unsigned long long max_cycles_mask;
static unsigned int cycles_wrap;
#endif
#endif
-int tsc_reliable = 0;
+bool tsc_reliable = false;
struct tv_valid {
- uint64_t last_cycles;
- int last_tv_valid;
int warned;
};
#ifdef ARCH_HAVE_CPU_CLOCK
#endif /* FIO_DEBUG_TIME */
-#ifdef CONFIG_CLOCK_GETTIME
-static int fill_clock_gettime(struct timespec *ts)
+/*
+ * Queries the value of the monotonic clock if a monotonic clock is available
+ * or the wall clock time if no monotonic clock is available. Returns 0 if
+ * querying the clock succeeded or -1 if querying the clock failed.
+ */
+int fio_get_mono_time(struct timespec *ts)
{
-#if defined(CONFIG_CLOCK_MONOTONIC_RAW)
- return clock_gettime(CLOCK_MONOTONIC_RAW, ts);
-#elif defined(CONFIG_CLOCK_MONOTONIC)
- return clock_gettime(CLOCK_MONOTONIC, ts);
+ int ret;
+
+#ifdef CONFIG_CLOCK_GETTIME
+#if defined(CONFIG_CLOCK_MONOTONIC)
+ ret = clock_gettime(CLOCK_MONOTONIC, ts);
#else
- return clock_gettime(CLOCK_REALTIME, ts);
+ ret = clock_gettime(CLOCK_REALTIME, ts);
#endif
-}
+#else
+ struct timeval tv;
+
+ ret = gettimeofday(&tv, NULL);
+ if (ret == 0) {
+ ts->tv_sec = tv.tv_sec;
+ ts->tv_nsec = tv.tv_usec * 1000;
+ }
#endif
+ assert(ret <= 0);
+ return ret;
+}
static void __fio_gettime(struct timespec *tp)
{
#endif
#ifdef CONFIG_CLOCK_GETTIME
case CS_CGETTIME: {
- if (fill_clock_gettime(tp) < 0) {
- log_err("fio: clock_gettime fails\n");
+ if (fio_get_mono_time(tp) < 0) {
+ log_err("fio: fio_get_mono_time() fails\n");
assert(0);
}
break;
nsecs = multiples * nsecs_for_max_cycles;
nsecs += ((t & max_cycles_mask) * clock_mult) >> clock_shift;
#endif
- tv->last_cycles = t;
- tv->last_tv_valid = 1;
-
tp->tv_sec = nsecs / 1000000000ULL;
tp->tv_nsec = nsecs % 1000000000ULL;
break;
}
#if defined(ARCH_HAVE_CPU_CLOCK) && !defined(ARCH_CPU_CLOCK_CYCLES_PER_USEC)
-static unsigned long get_cycles_per_usec(void)
+static unsigned long get_cycles_per_msec(void)
{
struct timespec s, e;
uint64_t c_s, c_e;
- enum fio_cs old_cs = fio_clock_source;
uint64_t elapsed;
-#ifdef CONFIG_CLOCK_GETTIME
- fio_clock_source = CS_CGETTIME;
-#else
- fio_clock_source = CS_GTOD;
-#endif
- __fio_gettime(&s);
+ fio_get_mono_time(&s);
c_s = get_cpu_clock();
do {
- __fio_gettime(&e);
+ fio_get_mono_time(&e);
+ c_e = get_cpu_clock();
- elapsed = utime_since(&s, &e);
- if (elapsed >= 1280) {
- c_e = get_cpu_clock();
+ elapsed = ntime_since(&s, &e);
+ if (elapsed >= 1280000)
break;
- }
} while (1);
- fio_clock_source = old_cs;
- return (c_e - c_s) / elapsed;
+ return (c_e - c_s) * 1000000 / elapsed;
}
#define NR_TIME_ITERS 50
int i, samples, sft = 0;
unsigned long long tmp, max_ticks, max_mult;
- cycles[0] = get_cycles_per_usec();
+ cycles[0] = get_cycles_per_msec();
S = delta = mean = 0.0;
for (i = 0; i < NR_TIME_ITERS; i++) {
- cycles[i] = get_cycles_per_usec();
+ cycles[i] = get_cycles_per_msec();
delta = cycles[i] - mean;
if (delta) {
mean += delta / (i + 1.0);
dprint(FD_TIME, "cycles[%d]=%llu\n", i, (unsigned long long) cycles[i]);
avg /= samples;
- cycles_per_usec = avg;
- dprint(FD_TIME, "avg: %llu\n", (unsigned long long) avg);
- dprint(FD_TIME, "min=%llu, max=%llu, mean=%f, S=%f\n",
+ cycles_per_msec = avg;
+ dprint(FD_TIME, "min=%llu, max=%llu, mean=%f, S=%f, N=%d\n",
(unsigned long long) minc,
- (unsigned long long) maxc, mean, S);
+ (unsigned long long) maxc, mean, S, NR_TIME_ITERS);
+ dprint(FD_TIME, "trimmed mean=%llu, N=%d\n", (unsigned long long) avg, samples);
- max_ticks = MAX_CLOCK_SEC * cycles_per_usec * 1000000ULL;
- max_mult = ULLONG_MAX / max_ticks;
- dprint(FD_TIME, "\n\nmax_ticks=%llu, __builtin_clzll=%d, max_mult=%llu\n",
- max_ticks, __builtin_clzll(max_ticks), max_mult);
+ max_ticks = MAX_CLOCK_SEC * cycles_per_msec * 1000ULL;
+ max_mult = ULLONG_MAX / max_ticks;
+ dprint(FD_TIME, "\n\nmax_ticks=%llu, __builtin_clzll=%d, "
+ "max_mult=%llu\n", max_ticks,
+ __builtin_clzll(max_ticks), max_mult);
/*
* Find the largest shift count that will produce
* a multiplier that does not exceed max_mult
*/
- tmp = max_mult * cycles_per_usec / 1000;
+ tmp = max_mult * cycles_per_msec / 1000000;
while (tmp > 1) {
tmp >>= 1;
sft++;
dprint(FD_TIME, "tmp=%llu, sft=%u\n", tmp, sft);
}
- clock_shift = sft;
- clock_mult = (1ULL << sft) * 1000 / cycles_per_usec;
- dprint(FD_TIME, "clock_shift=%u, clock_mult=%llu\n", clock_shift, clock_mult);
+ clock_shift = sft;
+ clock_mult = (1ULL << sft) * 1000000 / cycles_per_msec;
+ dprint(FD_TIME, "clock_shift=%u, clock_mult=%llu\n", clock_shift,
+ clock_mult);
- // Find the greatest power of 2 clock ticks that is less than the ticks in MAX_CLOCK_SEC_2STAGE
+ /*
+ * Find the greatest power of 2 clock ticks that is less than the
+ * ticks in MAX_CLOCK_SEC_2STAGE
+ */
max_cycles_shift = max_cycles_mask = 0;
- tmp = MAX_CLOCK_SEC * 1000000ULL * cycles_per_usec;
- dprint(FD_TIME, "tmp=%llu, max_cycles_shift=%u\n", tmp, max_cycles_shift);
+ tmp = MAX_CLOCK_SEC * 1000ULL * cycles_per_msec;
+ dprint(FD_TIME, "tmp=%llu, max_cycles_shift=%u\n", tmp,
+ max_cycles_shift);
while (tmp > 1) {
tmp >>= 1;
max_cycles_shift++;
dprint(FD_TIME, "tmp=%llu, max_cycles_shift=%u\n", tmp, max_cycles_shift);
}
- // if use use (1ULL << max_cycles_shift) * 1000 / cycles_per_usec here we will
- // have a discontinuity every (1ULL << max_cycles_shift) cycles
- nsecs_for_max_cycles = ((1ULL << max_cycles_shift) * clock_mult) >> clock_shift;
+ /*
+ * if use use (1ULL << max_cycles_shift) * 1000 / cycles_per_msec
+ * here we will have a discontinuity every
+ * (1ULL << max_cycles_shift) cycles
+ */
+ nsecs_for_max_cycles = ((1ULL << max_cycles_shift) * clock_mult)
+ >> clock_shift;
- // Use a bitmask to calculate ticks % (1ULL << max_cycles_shift)
+ /* Use a bitmask to calculate ticks % (1ULL << max_cycles_shift) */
for (tmp = 0; tmp < max_cycles_shift; tmp++)
max_cycles_mask |= 1ULL << tmp;
- dprint(FD_TIME, "max_cycles_shift=%u, 2^max_cycles_shift=%llu, nsecs_for_max_cycles=%llu, max_cycles_mask=%016llx\n",
- max_cycles_shift, (1ULL << max_cycles_shift),
- nsecs_for_max_cycles, max_cycles_mask);
+ dprint(FD_TIME, "max_cycles_shift=%u, 2^max_cycles_shift=%llu, "
+ "nsecs_for_max_cycles=%llu, "
+ "max_cycles_mask=%016llx\n",
+ max_cycles_shift, (1ULL << max_cycles_shift),
+ nsecs_for_max_cycles, max_cycles_mask);
cycles_start = get_cpu_clock();
dprint(FD_TIME, "cycles_start=%llu\n", cycles_start);
}
#endif // ARCH_HAVE_CPU_CLOCK
-#ifndef CONFIG_TLS_THREAD
-void fio_local_clock_init(int is_thread)
+#if defined(ARCH_HAVE_CPU_CLOCK) && !defined(CONFIG_TLS_THREAD)
+void fio_local_clock_init(void)
{
struct tv_valid *t;
free(data);
}
#else
-void fio_local_clock_init(int is_thread)
+void fio_local_clock_init(void)
{
}
#endif
if (fio_clock_source == fio_clock_source_inited)
return;
-#ifndef CONFIG_TLS_THREAD
+#if defined(ARCH_HAVE_CPU_CLOCK) && !defined(CONFIG_TLS_THREAD)
if (pthread_key_create(&tv_tls_key, kill_tv_tls_key))
log_err("fio: can't create TLS key\n");
#endif
fio_clock_source_inited = fio_clock_source;
if (calibrate_cpu_clock())
- tsc_reliable = 0;
+ tsc_reliable = false;
/*
* If the arch sets tsc_reliable != 0, then it must be good enough
fio_clock_source = CS_CPUCLOCK;
} else if (fio_clock_source == CS_CPUCLOCK)
log_info("fio: clocksource=cpu may not be reliable\n");
+ dprint(FD_TIME, "gettime: clocksource=%d\n", (int) fio_clock_source);
}
uint64_t ntime_since(const struct timespec *s, const struct timespec *e)
return nsec + (sec * 1000000000LL);
}
+uint64_t ntime_since_now(const struct timespec *s)
+{
+ struct timespec now;
+
+ fio_gettime(&now, NULL);
+ return ntime_since(s, &now);
+}
+
uint64_t utime_since(const struct timespec *s, const struct timespec *e)
{
int64_t sec, usec;
return mtime_since(s, &t);
}
-uint64_t mtime_since(const struct timespec *s, const struct timespec *e)
+/*
+ * Returns *e - *s in milliseconds as a signed integer. Note: rounding is
+ * asymmetric. If the difference yields +1 ns then 0 is returned. If the
+ * difference yields -1 ns then -1 is returned.
+ */
+int64_t rel_time_since(const struct timespec *s, const struct timespec *e)
{
- int64_t sec, usec;
+ int64_t sec, nsec;
sec = e->tv_sec - s->tv_sec;
- usec = (e->tv_nsec - s->tv_nsec) / 1000;
- if (sec > 0 && usec < 0) {
+ nsec = e->tv_nsec - s->tv_nsec;
+ if (nsec < 0) {
sec--;
- usec += 1000000;
+ nsec += 1000ULL * 1000 * 1000;
}
+ assert(0 <= nsec && nsec < 1000ULL * 1000 * 1000);
- if (sec < 0 || (sec == 0 && usec < 0))
- return 0;
+ return sec * 1000 + nsec / (1000 * 1000);
+}
- sec *= 1000;
- usec /= 1000;
- return sec + usec;
+/*
+ * Returns *e - *s in milliseconds as an unsigned integer. Returns 0 if
+ * *e < *s.
+ */
+uint64_t mtime_since(const struct timespec *s, const struct timespec *e)
+{
+ return max(rel_time_since(s, e), (int64_t)0);
}
uint64_t time_since_now(const struct timespec *s)
}
#if defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK) && \
- defined(CONFIG_SFAA)
+ defined(CONFIG_SYNC_SYNC) && defined(CONFIG_CMP_SWAP)
#define CLOCK_ENTRIES_DEBUG 100000
-#define CLOCK_ENTRIES_TEST 10000
+#define CLOCK_ENTRIES_TEST 1000
struct clock_entry {
uint32_t seq;
pthread_t thread;
int cpu;
int debug;
- pthread_mutex_t lock;
- pthread_mutex_t started;
+ struct fio_sem lock;
unsigned long nr_entries;
uint32_t *seq;
struct clock_entry *entries;
};
-static inline uint32_t atomic32_inc_return(uint32_t *seq)
+static inline uint32_t atomic32_compare_and_swap(uint32_t *ptr, uint32_t old,
+ uint32_t new)
{
- return 1 + __sync_fetch_and_add(seq, 1);
+ return __sync_val_compare_and_swap(ptr, old, new);
}
static void *clock_thread_fn(void *data)
struct clock_thread *t = data;
struct clock_entry *c;
os_cpu_mask_t cpu_mask;
- uint32_t last_seq;
unsigned long long first;
int i;
goto err;
}
- pthread_mutex_lock(&t->lock);
- pthread_mutex_unlock(&t->started);
+ fio_sem_down(&t->lock);
first = get_cpu_clock();
- last_seq = 0;
c = &t->entries[0];
for (i = 0; i < t->nr_entries; i++, c++) {
uint32_t seq;
c->cpu = t->cpu;
do {
- seq = atomic32_inc_return(t->seq);
- if (seq < last_seq)
+ seq = *t->seq;
+ if (seq == UINT_MAX)
break;
+ __sync_synchronize();
tsc = get_cpu_clock();
- } while (seq != *t->seq);
+ } while (seq != atomic32_compare_and_swap(t->seq, seq, seq + 1));
+
+ if (seq == UINT_MAX)
+ break;
c->seq = seq;
c->tsc = tsc;
* The most common platform clock breakage is returning zero
* indefinitely. Check for that and return failure.
*/
- if (!t->entries[i - 1].tsc && !t->entries[0].tsc)
+ if (i > 1 && !t->entries[i - 1].tsc && !t->entries[0].tsc)
goto err;
fio_cpuset_exit(&cpu_mask);
int fio_monotonic_clocktest(int debug)
{
struct clock_thread *cthreads;
- unsigned int nr_cpus = cpus_online();
+ unsigned int seen_cpus, nr_cpus = cpus_online();
struct clock_entry *entries;
unsigned long nr_entries, tentries, failed = 0;
struct clock_entry *prev, *this;
uint32_t seq = 0;
unsigned int i;
+ os_cpu_mask_t mask;
+
+#ifdef FIO_HAVE_GET_THREAD_AFFINITY
+ fio_get_thread_affinity(mask);
+#else
+ memset(&mask, 0, sizeof(mask));
+ for (i = 0; i < nr_cpus; i++)
+ fio_cpu_set(&mask, i);
+#endif
if (debug) {
log_info("cs: reliable_tsc: %s\n", tsc_reliable ? "yes" : "no");
if (debug)
log_info("cs: Testing %u CPUs\n", nr_cpus);
+ seen_cpus = 0;
for (i = 0; i < nr_cpus; i++) {
struct clock_thread *t = &cthreads[i];
+ if (!fio_cpu_isset(&mask, i))
+ continue;
t->cpu = i;
t->debug = debug;
t->seq = &seq;
t->nr_entries = nr_entries;
- t->entries = &entries[i * nr_entries];
- pthread_mutex_init(&t->lock, NULL);
- pthread_mutex_init(&t->started, NULL);
- pthread_mutex_lock(&t->lock);
+ t->entries = &entries[seen_cpus * nr_entries];
+ __fio_sem_init(&t->lock, FIO_SEM_LOCKED);
if (pthread_create(&t->thread, NULL, clock_thread_fn, t)) {
failed++;
nr_cpus = i;
break;
}
+ seen_cpus++;
}
for (i = 0; i < nr_cpus; i++) {
struct clock_thread *t = &cthreads[i];
- pthread_mutex_lock(&t->started);
- }
-
- for (i = 0; i < nr_cpus; i++) {
- struct clock_thread *t = &cthreads[i];
-
- pthread_mutex_unlock(&t->lock);
+ if (!fio_cpu_isset(&mask, i))
+ continue;
+ fio_sem_up(&t->lock);
}
for (i = 0; i < nr_cpus; i++) {
struct clock_thread *t = &cthreads[i];
void *ret;
+ if (!fio_cpu_isset(&mask, i))
+ continue;
pthread_join(t->thread, &ret);
if (ret)
failed++;
+ __fio_sem_remove(&t->lock);
}
free(cthreads);
goto err;
}
+ tentries = nr_entries * seen_cpus;
qsort(entries, tentries, sizeof(struct clock_entry), clock_cmp);
/* silence silly gcc */