/* this is the meat of the state machine. There is a list of
* active operations structs, and as each one finishes the required
* io it is moved to a list of finished operations. Once they have
* all finished whatever stage they were in, they are given the chance
* to restart and pick a different stage (read/write/random read etc)
*
* various timings are printed in between the stages, along with
* thread synchronization if there are more than one threads.
*/
int worker(struct thread_info *t)
{
struct io_oper *oper;
char *this_stage = NULL;
struct timeval stage_time;
int status = 0;
int iteration = 0;
int cnt;
aio_setup(&t->io_ctx, 512);
restart:
if (num_threads > 1) {
pthread_mutex_lock(&stage_mutex);
threads_starting++;
if (threads_starting == num_threads) {
threads_ending = 0;
gettimeofday(&global_stage_start_time, NULL);
pthread_cond_broadcast(&stage_cond);
}
while (threads_starting != num_threads)
pthread_cond_wait(&stage_cond, &stage_mutex);
pthread_mutex_unlock(&stage_mutex);
}
if (t->active_opers) {
this_stage = stage_name(t->active_opers->rw);
gettimeofday(&stage_time, NULL);
t->stage_mb_trans = 0;
}
cnt = 0;
/* first we send everything through aio */
while(t->active_opers && (cnt < iterations || iterations == RUN_FOREVER)) {
if (stonewall && threads_ending) {
oper = t->active_opers;
oper->stonewalled = 1;
oper_list_del(oper, &t->active_opers);
oper_list_add(oper, &t->finished_opers);
} else {
run_active_list(t, io_iter, max_io_submit);
}
cnt++;
}
if (latency_stats)
print_latency(t);
if (completion_latency_stats)
print_completion_latency(t);
/* then we wait for all the operations to finish */
oper = t->finished_opers;
do {
if (!oper)
break;
io_oper_wait(t, oper);
oper = oper->next;
} while(oper != t->finished_opers);
/* then we do an fsync to get the timing for any future operations
* right, and check to see if any of these need to get restarted
*/
oper = t->finished_opers;
while(oper) {
if (fsync_stages)
fsync(oper->fd);
t->stage_mb_trans += oper_mb_trans(oper);
if (restart_oper(oper)) {
oper_list_del(oper, &t->finished_opers);
oper_list_add(oper, &t->active_opers);
oper = t->finished_opers;
continue;
}
oper = oper->next;
if (oper == t->finished_opers)
break;
}
if (t->stage_mb_trans && t->num_files > 0) {
double seconds = time_since_now(&stage_time);
fprintf(stderr, "thread %llu %s totals (%.2f MB/s) %.2f MB in %.2fs\n",
(unsigned long long)(t - global_thread_info), this_stage,
t->stage_mb_trans/seconds, t->stage_mb_trans, seconds);
}
if (num_threads > 1) {
pthread_mutex_lock(&stage_mutex);
threads_ending++;
if (threads_ending == num_threads) {
threads_starting = 0;
pthread_cond_broadcast(&stage_cond);
global_thread_throughput(t, this_stage);
}
while(threads_ending != num_threads)
pthread_cond_wait(&stage_cond, &stage_mutex);
pthread_mutex_unlock(&stage_mutex);
}
/* someone got restarted, go back to the beginning */
if (t->active_opers && (cnt < iterations || iterations == RUN_FOREVER)) {
iteration++;
goto restart;
}
/* finally, free all the ram */
while(t->finished_opers) {
oper = t->finished_opers;
oper_list_del(oper, &t->finished_opers);
status = finish_oper(t, oper);
}
if (t->num_global_pending) {
fprintf(stderr, "global num pending is %d\n", t->num_global_pending);
}
io_queue_release(t->io_ctx);
return status;
}
int main (int argc, char **argv)
{
ssize_t ret_size;
struct stat st;
int ret, flags;
int part_request;
long long this_time;
double part_min, part_max, time_min, time_max;
double time_sum, time_sum2, time_mdev, time_avg;
double part_sum, part_sum2, part_mdev, part_avg;
long long time_now, time_next, period_deadline;
setvbuf(stdout, NULL, _IOLBF, 0);
parse_options(argc, argv);
interval_ts.tv_sec = interval / 1000000;
interval_ts.tv_nsec = (interval % 1000000) * 1000;
if (!size)
size = default_size;
if (size <= 0)
errx(1, "request size must be greather than zero");
#ifdef MAX_RW_COUNT
if (size > MAX_RW_COUNT)
warnx("this platform supports requests %u bytes at most",
MAX_RW_COUNT);
#endif
if (wsize)
temp_wsize = wsize;
else if (size > temp_wsize)
temp_wsize = size;
flags = O_RDONLY;
#if !defined(HAVE_POSIX_FADVICE) && !defined(HAVE_NOCACHE_IO)
# if defined(HAVE_DIRECT_IO)
direct |= !cached;
# else
if (!cached && !write_test) {
warnx("non-cached read I/O not supported by this platform");
warnx("you can use write I/O to get reliable results");
cached = 1;
}
# endif
#endif
if (write_test) {
flags = O_RDWR;
make_request = do_pwrite;
}
if (async)
aio_setup();
if (direct)
#ifdef HAVE_DIRECT_IO
flags |= O_DIRECT;
#else
errx(1, "direct I/O not supported by this platform");
#endif
#ifdef __MINGW32__
flags |= O_BINARY;
#endif
if (stat(path, &st))
err(2, "stat \"%s\" failed", path);
if (!S_ISDIR(st.st_mode) && write_test && write_test < 3)
errx(2, "think twice, then use -WWW to shred this target");
if (S_ISDIR(st.st_mode) || S_ISREG(st.st_mode)) {
if (S_ISDIR(st.st_mode))
st.st_size = offset + temp_wsize;
parse_device(st.st_dev);
} else if (S_ISBLK(st.st_mode) || S_ISCHR(st.st_mode)) {
fd = open(path, flags);
if (fd < 0)
err(2, "failed to open \"%s\"", path);
if (get_device_size(fd, &st)) {
if (!S_ISCHR(st.st_mode))
err(2, "block get size ioctl failed");
st.st_size = offset + temp_wsize;
fstype = "character";
device = "device";
} else {
device_size = st.st_size;
fstype = "block";
device = "device ";
}
if (!cached && write_test && fdatasync(fd)) {
warnx("fdatasync not supported by \"%s\", "
"enable cached requests", path);
cached = 1;
}
} else {
errx(2, "unsupported destination: \"%s\"", path);
}
if (wsize > st.st_size || offset > st.st_size - wsize)
errx(2, "target is too small for this");
if (!wsize)
wsize = st.st_size - offset;
if (size > wsize)
errx(2, "request size is too big for this target");
ret = posix_memalign(&buf, 0x1000, size);
if (ret)
errx(2, "buffer allocation failed");
random_memory(buf, size);
if (S_ISDIR(st.st_mode)) {
fd = create_temp(path, "ioping.tmp");
if (fd < 0)
err(2, "failed to create temporary file at \"%s\"", path);
if (keep_file) {
if (fstat(fd, &st))
err(2, "fstat at \"%s\" failed", path);
if (st.st_size >= offset + wsize)
#ifndef __MINGW32__
if (st.st_blocks >= (st.st_size + 511) / 512)
#endif
goto skip_preparation;
}
for (woffset = 0 ; woffset < wsize ; woffset += ret_size) {
ret_size = size;
if (woffset + ret_size > wsize)
ret_size = wsize - woffset;
if (woffset)
random_memory(buf, ret_size);
ret_size = pwrite(fd, buf, ret_size, offset + woffset);
if (ret_size <= 0)
err(2, "preparation write failed");
}
skip_preparation:
if (fsync(fd))
err(2, "fsync failed");
} else if (S_ISREG(st.st_mode)) {
fd = open(path, flags);
if (fd < 0)
err(2, "failed to open \"%s\"", path);
}
if (!cached) {
#ifdef HAVE_POSIX_FADVICE
ret = posix_fadvise(fd, offset, wsize, POSIX_FADV_RANDOM);
if (ret)
err(2, "fadvise failed");
#endif
#ifdef HAVE_NOCACHE_IO
ret = fcntl(fd, F_NOCACHE, 1);
if (ret)
err(2, "fcntl nocache failed");
#endif
}
srandom(now());
if (deadline)
deadline += now();
set_signal();
request = 0;
woffset = 0;
part_request = 0;
part_min = time_min = LLONG_MAX;
part_max = time_max = LLONG_MIN;
part_sum = time_sum = 0;
part_sum2 = time_sum2 = 0;
time_now = now();
period_deadline = time_now + period_time;
while (!exiting) {
request++;
part_request++;
if (randomize)
woffset = random() % (wsize / size) * size;
#ifdef HAVE_POSIX_FADVICE
if (!cached) {
ret = posix_fadvise(fd, offset + woffset, size,
POSIX_FADV_DONTNEED);
if (ret)
err(3, "fadvise failed");
}
#endif
if (write_test)
shake_memory(buf, size);
this_time = now();
ret_size = make_request(fd, buf, size, offset + woffset);
if (ret_size < 0) {
if (errno != EINTR)
err(3, "request failed");
} else if (ret_size < size)
warnx("request returned less than expected: %zu", ret_size);
else if (ret_size > size)
errx(3, "request returned more than expected: %zu", ret_size);
time_now = now();
this_time = time_now - this_time;
time_next = time_now + interval;
part_sum += this_time;
part_sum2 += this_time * this_time;
if (this_time < part_min)
part_min = this_time;
if (this_time > part_max)
part_max = this_time;
if (!quiet) {
print_size(ret_size);
printf(" %s %s (%s %s", write_test ? "to" : "from",
path, fstype, device);
if (device_size)
print_size(device_size);
printf("): request=%d time=", request);
print_time(this_time);
printf("\n");
}
if ((period_request && (part_request >= period_request)) ||
(period_time && (time_next >= period_deadline))) {
part_avg = part_sum / part_request;
part_mdev = sqrt(part_sum2 / part_request - part_avg * part_avg);
printf("%d %.0f %.0f %.0f %.0f %.0f %.0f %.0f\n",
part_request, part_sum,
1000000. * part_request / part_sum,
1000000. * part_request * size / part_sum,
part_min, part_avg,
part_max, part_mdev);
time_sum += part_sum;
time_sum2 += part_sum2;
if (part_min < time_min)
time_min = part_min;
if (part_max > time_max)
time_max = part_max;
part_min = LLONG_MAX;
part_max = LLONG_MIN;
part_sum = part_sum2 = 0;
part_request = 0;
period_deadline = time_now + period_time;
}
if (!randomize) {
woffset += size;
if (woffset + size > wsize)
woffset = 0;
}
if (exiting)
break;
if (stop_at_request && request >= stop_at_request)
break;
if (deadline && time_next >= deadline)
break;
if (interval)
nanosleep(&interval_ts, NULL);
}
time_sum += part_sum;
time_sum2 += part_sum2;
if (part_min < time_min)
time_min = part_min;
if (part_max > time_max)
time_max = part_max;
time_avg = time_sum / request;
time_mdev = sqrt(time_sum2 / request - time_avg * time_avg);
if (batch_mode) {
printf("%d %.0f %.0f %.0f %.0f %.0f %.0f %.0f\n",
request, time_sum,
1000000. * request / time_sum,
1000000. * request * size / time_sum,
time_min, time_avg,
time_max, time_mdev);
} else if (!quiet || (!period_time && !period_request)) {
printf("\n--- %s (%s %s", path, fstype, device);
if (device_size)
print_size(device_size);
printf(") ioping statistics ---\n");
print_int(request);
printf(" requests completed in ");
print_time(time_sum);
printf(", ");
print_size((long long)request * size);
printf(" %s, ", write_test ? "written" : "read");
print_int(1000000. * request / time_sum);
printf(" iops, ");
print_size(1000000. * request * size / time_sum);
printf("/s\n");
printf("min/avg/max/mdev = ");
print_time(time_min);
printf(" / ");
print_time(time_avg);
printf(" / ");
print_time(time_max);
printf(" / ");
print_time(time_mdev);
printf("\n");
}
return 0;
}