static int client_rw(int out_fd, int file_fd, int offset)
{
int loops = client_loops;
struct timeval start;
struct stat sb;
char *buf;
unsigned long long size;
unsigned long msecs;
if (fstat(file_fd, &sb) < 0)
return error("fstat");
buf = malloc(splice_size);
gettimeofday(&start, NULL);
again:
if (lseek(file_fd, 0, SEEK_SET) < 0)
return error("lseek");
size = sb.st_size;
while (size) {
int this_len = min(size, (unsigned long long) splice_size);
int ret = read(file_fd, buf, this_len);
if (ret < 0)
return error("read");
size -= ret;
while (ret) {
int written = write(out_fd, buf, ret);
if (written < 0)
return error("write");
ret -= written;
}
}
loops--;
if ((mtime_since_now(&start) < max_client_run * 1000) && loops)
goto again;
free(buf);
size = sb.st_size >> 10;
size *= (client_loops - loops);
msecs = mtime_since_now(&start);
fprintf(stdout, "Client%d (rw): %Lu MiB/sec (%LuMiB in %lu msecs)\n", offset, size / (unsigned long long) msecs, size >> 10, msecs);
return 0;
}
static int client_mmap(int out_fd, int file_fd, int offset)
{
int loops = client_loops;
struct timeval start;
struct stat sb;
void *mmap_area, *buf;
unsigned long long size;
unsigned long msecs;
if (fstat(file_fd, &sb) < 0)
return error("fstat");
mmap_area = mmap(NULL, sb.st_size, PROT_READ, MAP_SHARED, file_fd, 0);
if (mmap_area == MAP_FAILED)
return error("mmap");
if (madvise(mmap_area, sb.st_size, MADV_WILLNEED) < 0)
return error("madvise");
gettimeofday(&start, NULL);
again:
buf = mmap_area;
size = sb.st_size;
while (size) {
int this_len = min(size, (unsigned long long) splice_size);
int ret = write(out_fd, buf, this_len);
if (ret < 0)
return error("write");
buf += ret;
size -= ret;
}
loops--;
if ((mtime_since_now(&start) < max_client_run * 1000) && loops)
goto again;
size = sb.st_size >> 10;
size *= (client_loops - loops);
msecs = mtime_since_now(&start);
fprintf(stdout, "Client%d (mmap): %Lu MiB/sec (%LuMiB in %lu msecs)\n", offset, size / (unsigned long long) msecs, size >> 10, msecs);
munmap(mmap_area, sb.st_size);
return 0;
}
static int client_splice_loop(int out_fd, int fd, int *pfd, int offset)
{
struct timeval start;
unsigned long long size;
unsigned long msecs;
struct stat sb;
int loops = client_loops;
loff_t off;
if (fstat(fd, &sb) < 0)
return error("fstat");
gettimeofday(&start, NULL);
again:
size = sb.st_size;
off = 0;
do {
int ret = ssplice(fd, &off, pfd[1], NULL, min(size, (unsigned long long) splice_size), 0);
if (ret <= 0)
return error("splice-in");
size -= ret;
while (ret > 0) {
int flags = size ? SPLICE_F_MORE : 0;
int written = ssplice(pfd[0], NULL, out_fd, NULL, ret, flags);
if (written <= 0)
return error("splice-out");
ret -= written;
}
} while (size);
loops--;
if ((mtime_since_now(&start) < max_client_run * 1000) && loops)
goto again;
size = sb.st_size >> 10;
size *= (client_loops - loops);
msecs = mtime_since_now(&start);
fprintf(stdout, "Client%d (splice): %Lu MiB/sec (%LuMiB in %lu msecs)\n", offset, size / (unsigned long long) msecs, size >> 10, msecs);
return 0;
}
/*
* Best effort calculation of the estimated pending runtime of a job.
*/
static int thread_eta(struct thread_data *td)
{
unsigned long long bytes_total, bytes_done;
unsigned long eta_sec = 0;
unsigned long elapsed;
uint64_t timeout;
elapsed = (mtime_since_now(&td->epoch) + 999) / 1000;
timeout = td->o.timeout / 1000000UL;
bytes_total = td->total_io_size;
if (td->o.fill_device && td->o.size == -1ULL) {
if (!td->fill_device_size || td->fill_device_size == -1ULL)
return 0;
bytes_total = td->fill_device_size;
}
if (td->o.zone_size && td->o.zone_skip && bytes_total) {
unsigned int nr_zones;
uint64_t zone_bytes;
zone_bytes = bytes_total + td->o.zone_size + td->o.zone_skip;
nr_zones = (zone_bytes - 1) / (td->o.zone_size + td->o.zone_skip);
bytes_total -= nr_zones * td->o.zone_skip;
}
/*
* if writing and verifying afterwards, bytes_total will be twice the
* size. In a mixed workload, verify phase will be the size of the
* first stage writes.
*/
if (td->o.do_verify && td->o.verify && td_write(td)) {
if (td_rw(td)) {
unsigned int perc = 50;
if (td->o.rwmix[DDIR_WRITE])
perc = td->o.rwmix[DDIR_WRITE];
bytes_total += (bytes_total * perc) / 100;
} else
bytes_total <<= 1;
}
if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING) {
double perc, perc_t;
bytes_done = ddir_rw_sum(td->io_bytes);
if (bytes_total) {
perc = (double) bytes_done / (double) bytes_total;
if (perc > 1.0)
perc = 1.0;
} else
perc = 0.0;
if (td->o.time_based) {
if (timeout) {
perc_t = (double) elapsed / (double) timeout;
if (perc_t < perc)
perc = perc_t;
} else {
/*
* Will never hit, we can't have time_based
* without a timeout set.
*/
perc = 0.0;
}
}
eta_sec = (unsigned long) (elapsed * (1.0 / perc)) - elapsed;
if (td->o.timeout &&
eta_sec > (timeout + done_secs - elapsed))
eta_sec = timeout + done_secs - elapsed;
} else if (td->runstate == TD_NOT_CREATED || td->runstate == TD_CREATED
|| td->runstate == TD_INITIALIZED
|| td->runstate == TD_SETTING_UP
|| td->runstate == TD_RAMP
|| td->runstate == TD_PRE_READING) {
int t_eta = 0, r_eta = 0;
unsigned long long rate_bytes;
/*
* We can only guess - assume it'll run the full timeout
* if given, otherwise assume it'll run at the specified rate.
*/
if (td->o.timeout) {
uint64_t timeout = td->o.timeout;
uint64_t start_delay = td->o.start_delay;
uint64_t ramp_time = td->o.ramp_time;
t_eta = timeout + start_delay + ramp_time;
t_eta /= 1000000ULL;
if (in_ramp_time(td)) {
unsigned long ramp_left;
ramp_left = mtime_since_now(&td->epoch);
ramp_left = (ramp_left + 999) / 1000;
if (ramp_left <= t_eta)
t_eta -= ramp_left;
}
}
rate_bytes = ddir_rw_sum(td->o.rate);
if (rate_bytes) {
r_eta = (bytes_total / 1024) / rate_bytes;
r_eta += (td->o.start_delay / 1000000ULL);
}
if (r_eta && t_eta)
eta_sec = min(r_eta, t_eta);
else if (r_eta)
eta_sec = r_eta;
else if (t_eta)
eta_sec = t_eta;
else
eta_sec = 0;
} else {
/*
* thread is already done or waiting for fsync
*/
eta_sec = 0;
}
return eta_sec;
}
/*
* Best effort calculation of the estimated pending runtime of a job.
*/
static int thread_eta(struct thread_data *td)
{
unsigned long long bytes_total, bytes_done;
unsigned long eta_sec = 0;
unsigned long elapsed;
elapsed = (mtime_since_now(&td->epoch) + 999) / 1000;
bytes_total = td->total_io_size;
/*
* if writing, bytes_total will be twice the size. If mixing,
* assume a 50/50 split and thus bytes_total will be 50% larger.
*/
if (td->o.do_verify && td->o.verify && td_write(td)) {
if (td_rw(td))
bytes_total = bytes_total * 3 / 2;
else
bytes_total <<= 1;
}
if (td->o.zone_size && td->o.zone_skip)
bytes_total /= (td->o.zone_skip / td->o.zone_size);
if (td->o.fill_device && td->o.size == -1ULL)
return 0;
if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING) {
double perc, perc_t;
bytes_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE];
perc = (double) bytes_done / (double) bytes_total;
if (perc > 1.0)
perc = 1.0;
if (td->o.time_based) {
perc_t = (double) elapsed / (double) td->o.timeout;
if (perc_t < perc)
perc = perc_t;
}
eta_sec = (unsigned long) (elapsed * (1.0 / perc)) - elapsed;
if (td->o.timeout &&
eta_sec > (td->o.timeout + done_secs - elapsed))
eta_sec = td->o.timeout + done_secs - elapsed;
} else if (td->runstate == TD_NOT_CREATED || td->runstate == TD_CREATED
|| td->runstate == TD_INITIALIZED
|| td->runstate == TD_RAMP
|| td->runstate == TD_PRE_READING) {
int t_eta = 0, r_eta = 0;
/*
* We can only guess - assume it'll run the full timeout
* if given, otherwise assume it'll run at the specified rate.
*/
if (td->o.timeout) {
t_eta = td->o.timeout + td->o.start_delay;
if (in_ramp_time(td)) {
unsigned long ramp_left;
ramp_left = mtime_since_now(&td->start);
ramp_left = (ramp_left + 999) / 1000;
if (ramp_left <= t_eta)
t_eta -= ramp_left;
}
}
if (td->o.rate[0] || td->o.rate[1]) {
r_eta = (bytes_total / 1024) / (td->o.rate[0] + td->o.rate[1]);
r_eta += td->o.start_delay;
}
if (r_eta && t_eta)
eta_sec = min(r_eta, t_eta);
else if (r_eta)
eta_sec = r_eta;
else if (t_eta)
eta_sec = t_eta;
else
eta_sec = 0;
} else {
/*
* thread is already done or waiting for fsync
*/
eta_sec = 0;
}
return eta_sec;
}
unsigned long time_since_now(struct timeval *s)
{
return mtime_since_now(s) / 1000;
}
unsigned long mtime_since_genesis(void)
{
return mtime_since_now(&genesis);
}
uint64_t mtime_since_genesis(void)
{
return mtime_since_now(&genesis);
}
