static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
{
enum fio_ddir odir = ddir ^ 1;
struct timeval t;
long usec;
assert(ddir_rw(ddir));
if (td->rate_pending_usleep[ddir] <= 0)
return ddir;
/*
* We have too much pending sleep in this direction. See if we
* should switch.
*/
if (td_rw(td) && td->o.rwmix[odir]) {
/*
* Other direction does not have too much pending, switch
*/
if (td->rate_pending_usleep[odir] < 100000)
return odir;
/*
* Both directions have pending sleep. Sleep the minimum time
* and deduct from both.
*/
if (td->rate_pending_usleep[ddir] <=
td->rate_pending_usleep[odir]) {
usec = td->rate_pending_usleep[ddir];
} else {
usec = td->rate_pending_usleep[odir];
ddir = odir;
}
} else
usec = td->rate_pending_usleep[ddir];
io_u_quiesce(td);
fio_gettime(&t, NULL);
usec_sleep(td, usec);
usec = utime_since_now(&t);
td->rate_pending_usleep[ddir] -= usec;
odir = ddir ^ 1;
if (td_rw(td) && __should_check_rate(td, odir))
td->rate_pending_usleep[odir] -= usec;
if (ddir_trim(ddir))
return ddir;
return ddir;
}
static int verify_io_u_meta(struct verify_header *hdr, struct vcont *vc)
{
struct thread_data *td = vc->td;
struct vhdr_meta *vh = hdr_priv(hdr);
struct io_u *io_u = vc->io_u;
int ret = EILSEQ;
dprint(FD_VERIFY, "meta verify io_u %p, len %u\n", io_u, hdr->len);
if (vh->offset == io_u->offset + vc->hdr_num * td->o.verify_interval)
ret = 0;
if (td->o.verify_pattern_bytes)
ret |= verify_io_u_pattern(hdr, vc);
/*
* For read-only workloads, the program cannot be certain of the
* last numberio written to a block. Checking of numberio will be done
* only for workloads that write data.
* For verify_only, numberio will be checked in the last iteration when
* the correct state of numberio, that would have been written to each
* block in a previous run of fio, has been reached.
*/
if (td_write(td) || td_rw(td))
if (!td->o.verify_only || td->o.loops == 0)
if (vh->numberio != io_u->numberio)
ret = EILSEQ;
if (!ret)
return 0;
vc->name = "meta";
log_verify_failure(hdr, vc);
return ret;
}
/*
* Sets the status of the 'td' in the printed status map.
*/
static void check_str_update(struct thread_data *td)
{
char c = run_str[td->thread_number - 1];
switch (td->runstate) {
case TD_REAPED:
c = '_';
break;
case TD_EXITED:
c = 'E';
break;
case TD_RAMP:
c = '/';
break;
case TD_RUNNING:
if (td_rw(td)) {
if (td_random(td))
c = 'm';
else
c = 'M';
} else if (td_read(td)) {
if (td_random(td))
c = 'r';
else
c = 'R';
} else {
if (td_random(td))
c = 'w';
else
c = 'W';
}
break;
case TD_PRE_READING:
c = 'p';
break;
case TD_VERIFYING:
c = 'V';
break;
case TD_FSYNCING:
c = 'F';
break;
case TD_CREATED:
c = 'C';
break;
case TD_INITIALIZED:
c = 'I';
break;
case TD_NOT_CREATED:
c = 'P';
break;
default:
log_err("state %d\n", td->runstate);
}
run_str[td->thread_number - 1] = c;
}
/*
* Return the data direction for the next io_u. If the job is a
* mixed read/write workload, check the rwmix cycle and switch if
* necessary.
*/
static enum fio_ddir get_rw_ddir(struct thread_data *td)
{
enum fio_ddir ddir;
/*
* see if it's time to fsync
*/
if (td->o.fsync_blocks &&
!(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
td->io_issues[DDIR_WRITE] && should_fsync(td))
return DDIR_SYNC;
/*
* see if it's time to fdatasync
*/
if (td->o.fdatasync_blocks &&
!(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
td->io_issues[DDIR_WRITE] && should_fsync(td))
return DDIR_DATASYNC;
/*
* see if it's time to sync_file_range
*/
if (td->sync_file_range_nr &&
!(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
td->io_issues[DDIR_WRITE] && should_fsync(td))
return DDIR_SYNC_FILE_RANGE;
if (td_rw(td)) {
/*
* Check if it's time to seed a new data direction.
*/
if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
/*
* Put a top limit on how many bytes we do for
* one data direction, to avoid overflowing the
* ranges too much
*/
ddir = get_rand_ddir(td);
if (ddir != td->rwmix_ddir)
set_rwmix_bytes(td);
td->rwmix_ddir = ddir;
}
ddir = td->rwmix_ddir;
} else if (td_read(td))
ddir = DDIR_READ;
else if (td_write(td))
ddir = DDIR_WRITE;
else
ddir = DDIR_TRIM;
td->rwmix_ddir = rate_ddir(td, ddir);
return td->rwmix_ddir;
}
static int fio_mmapio_open(struct thread_data *td, struct fio_file *f)
{
int ret, flags;
ret = generic_open_file(td, f);
if (ret)
return ret;
/*
* for size checkup, don't mmap anything.
*/
if (!f->io_size)
return 0;
if (td_rw(td))
flags = PROT_READ | PROT_WRITE;
else if (td_write(td)) {
flags = PROT_WRITE;
if (td->o.verify != VERIFY_NONE)
flags |= PROT_READ;
} else
flags = PROT_READ;
f->mmap = mmap(NULL, f->io_size, flags, MAP_SHARED, f->fd, f->file_offset);
if (f->mmap == MAP_FAILED) {
f->mmap = NULL;
td_verror(td, errno, "mmap");
goto err;
}
if (file_invalidate_cache(td, f))
goto err;
if (!td_random(td)) {
if (madvise(f->mmap, f->io_size, MADV_SEQUENTIAL) < 0) {
td_verror(td, errno, "madvise");
goto err;
}
} else {
if (madvise(f->mmap, f->io_size, MADV_RANDOM) < 0) {
td_verror(td, errno, "madvise");
goto err;
}
}
return 0;
err:
td->io_ops->close_file(td, f);
return 1;
}
static int fio_mmap_file(struct thread_data *td, struct fio_file *f,
size_t length, off_t off)
{
struct fio_mmap_data *fmd = FILE_ENG_DATA(f);
int flags = 0;
if (td_rw(td) && !td->o.verify_only)
flags = PROT_READ | PROT_WRITE;
else if (td_write(td) && !td->o.verify_only) {
flags = PROT_WRITE;
if (td->o.verify != VERIFY_NONE)
flags |= PROT_READ;
} else
flags = PROT_READ;
fmd->mmap_ptr = mmap(NULL, length, flags, MAP_SHARED, f->fd, off);
if (fmd->mmap_ptr == MAP_FAILED) {
fmd->mmap_ptr = NULL;
td_verror(td, errno, "mmap");
goto err;
}
if (!fio_madvise_file(td, f, length))
goto err;
if (posix_madvise(fmd->mmap_ptr, length, POSIX_MADV_DONTNEED) < 0) {
td_verror(td, errno, "madvise");
goto err;
}
#ifdef FIO_MADV_FREE
if (f->filetype == FIO_TYPE_BLOCK)
(void) posix_madvise(fmd->mmap_ptr, fmd->mmap_sz, FIO_MADV_FREE);
#endif
err:
if (td->error && fmd->mmap_ptr)
munmap(fmd->mmap_ptr, length);
return td->error;
}
/*
* This is the mmap execution function
*/
static int fio_libpmem_file(struct thread_data *td, struct fio_file *f,
size_t length, off_t off)
{
struct fio_libpmem_data *fdd = FILE_ENG_DATA(f);
int flags = 0;
void *addr = NULL;
dprint(FD_IO, "DEBUG fio_libpmem_file\n");
if (td_rw(td))
flags = PROT_READ | PROT_WRITE;
else if (td_write(td)) {
flags = PROT_WRITE;
if (td->o.verify != VERIFY_NONE)
flags |= PROT_READ;
} else
flags = PROT_READ;
dprint(FD_IO, "f->file_name = %s td->o.verify = %d \n", f->file_name,
td->o.verify);
dprint(FD_IO, "length = %ld flags = %d f->fd = %d off = %ld \n",
length, flags, f->fd,off);
addr = util_map_hint(length, 0);
fdd->libpmem_ptr = mmap(addr, length, flags, MAP_SHARED, f->fd, off);
if (fdd->libpmem_ptr == MAP_FAILED) {
fdd->libpmem_ptr = NULL;
td_verror(td, errno, "mmap");
}
if (td->error && fdd->libpmem_ptr)
munmap(fdd->libpmem_ptr, length);
return td->error;
}
void rate_throttle(struct thread_data *td, unsigned long time_spent,
unsigned int bytes)
{
unsigned long usec_cycle;
unsigned int bs;
if (!td->o.rate && !td->o.rate_iops)
return;
if (td_rw(td))
bs = td->o.rw_min_bs;
else if (td_read(td))
bs = td->o.min_bs[DDIR_READ];
else
bs = td->o.min_bs[DDIR_WRITE];
usec_cycle = td->rate_usec_cycle * (bytes / bs);
if (time_spent < usec_cycle) {
unsigned long s = usec_cycle - time_spent;
td->rate_pending_usleep += s;
if (td->rate_pending_usleep >= 100000) {
struct timeval t;
fio_gettime(&t, NULL);
usec_sleep(td, td->rate_pending_usleep);
td->rate_pending_usleep -= utime_since_now(&t);
}
} else {
long overtime = time_spent - usec_cycle;
td->rate_pending_usleep -= overtime;
}
}
/*
* 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;
}
static int verify_header(struct io_u *io_u, struct thread_data *td,
struct verify_header *hdr, unsigned int hdr_num,
unsigned int hdr_len)
{
void *p = hdr;
uint32_t crc;
if (hdr->magic != FIO_HDR_MAGIC) {
log_err("verify: bad magic header %x, wanted %x",
hdr->magic, FIO_HDR_MAGIC);
goto err;
}
if (hdr->len != hdr_len) {
log_err("verify: bad header length %u, wanted %u",
hdr->len, hdr_len);
goto err;
}
if (hdr->rand_seed != io_u->rand_seed) {
log_err("verify: bad header rand_seed %"PRIu64
", wanted %"PRIu64,
hdr->rand_seed, io_u->rand_seed);
goto err;
}
if (hdr->offset != io_u->offset + hdr_num * td->o.verify_interval) {
log_err("verify: bad header offset %"PRIu64
", wanted %llu",
hdr->offset, io_u->offset);
goto err;
}
/*
* For read-only workloads, the program cannot be certain of the
* last numberio written to a block. Checking of numberio will be
* done only for workloads that write data. For verify_only,
* numberio will be checked in the last iteration when the correct
* state of numberio, that would have been written to each block
* in a previous run of fio, has been reached.
*/
if ((td_write(td) || td_rw(td)) && (td_min_bs(td) == td_max_bs(td)) &&
!td->o.time_based)
if (!td->o.verify_only || td->o.loops == 0)
if (hdr->numberio != io_u->numberio) {
log_err("verify: bad header numberio %"PRIu16
", wanted %"PRIu16,
hdr->numberio, io_u->numberio);
goto err;
}
crc = fio_crc32c(p, offsetof(struct verify_header, crc32));
if (crc != hdr->crc32) {
log_err("verify: bad header crc %x, calculated %x",
hdr->crc32, crc);
goto err;
}
return 0;
err:
log_err(" at file %s offset %llu, length %u\n",
io_u->file->file_name,
io_u->offset + hdr_num * hdr_len, hdr_len);
if (td->o.verify_dump)
dump_buf(p, hdr_len, io_u->offset + hdr_num * hdr_len,
"hdr_fail", io_u->file);
return EILSEQ;
}
static int fio_rdmaio_init(struct thread_data *td)
{
struct rdmaio_data *rd = td->io_ops->data;
struct rdmaio_options *o = td->eo;
unsigned int max_bs;
int ret, i;
if (td_rw(td)) {
log_err("fio: rdma connections must be read OR write\n");
return 1;
}
if (td_random(td)) {
log_err("fio: RDMA network IO can't be random\n");
return 1;
}
if (compat_options(td))
return 1;
if (!o->port) {
log_err("fio: no port has been specified which is required "
"for the rdma engine\n");
return 1;
}
if (check_set_rlimits(td))
return 1;
rd->rdma_protocol = o->verb;
rd->cq_event_num = 0;
rd->cm_channel = rdma_create_event_channel();
if (!rd->cm_channel) {
log_err("fio: rdma_create_event_channel fail\n");
return 1;
}
ret = rdma_create_id(rd->cm_channel, &rd->cm_id, rd, RDMA_PS_TCP);
if (ret) {
log_err("fio: rdma_create_id fail\n");
return 1;
}
if ((rd->rdma_protocol == FIO_RDMA_MEM_WRITE) ||
(rd->rdma_protocol == FIO_RDMA_MEM_READ)) {
rd->rmt_us =
malloc(FIO_RDMA_MAX_IO_DEPTH * sizeof(struct remote_u));
memset(rd->rmt_us, 0,
FIO_RDMA_MAX_IO_DEPTH * sizeof(struct remote_u));
rd->rmt_nr = 0;
}
rd->io_us_queued = malloc(td->o.iodepth * sizeof(struct io_u *));
memset(rd->io_us_queued, 0, td->o.iodepth * sizeof(struct io_u *));
rd->io_u_queued_nr = 0;
rd->io_us_flight = malloc(td->o.iodepth * sizeof(struct io_u *));
memset(rd->io_us_flight, 0, td->o.iodepth * sizeof(struct io_u *));
rd->io_u_flight_nr = 0;
rd->io_us_completed = malloc(td->o.iodepth * sizeof(struct io_u *));
memset(rd->io_us_completed, 0, td->o.iodepth * sizeof(struct io_u *));
rd->io_u_completed_nr = 0;
if (td_read(td)) { /* READ as the server */
rd->is_client = 0;
td->flags |= TD_F_NO_PROGRESS;
/* server rd->rdma_buf_len will be setup after got request */
ret = fio_rdmaio_setup_listen(td, o->port);
} else { /* WRITE as the client */
rd->is_client = 1;
ret = fio_rdmaio_setup_connect(td, td->o.filename, o->port);
}
max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
rd->send_buf.max_bs = htonl(max_bs);
/* register each io_u in the free list */
for (i = 0; i < td->io_u_freelist.nr; i++) {
struct io_u *io_u = td->io_u_freelist.io_us[i];
io_u->engine_data = malloc(sizeof(struct rdma_io_u_data));
memset(io_u->engine_data, 0, sizeof(struct rdma_io_u_data));
((struct rdma_io_u_data *)io_u->engine_data)->wr_id = i;
io_u->mr = ibv_reg_mr(rd->pd, io_u->buf, max_bs,
IBV_ACCESS_LOCAL_WRITE |
IBV_ACCESS_REMOTE_READ |
IBV_ACCESS_REMOTE_WRITE);
if (io_u->mr == NULL) {
log_err("fio: ibv_reg_mr io_u failed\n");
return 1;
}
rd->send_buf.rmt_us[i].buf =
htonll((uint64_t) (unsigned long)io_u->buf);
rd->send_buf.rmt_us[i].rkey = htonl(io_u->mr->rkey);
rd->send_buf.rmt_us[i].size = htonl(max_bs);
#if 0
log_info("fio: Send rkey %x addr %" PRIx64 " len %d to client\n", io_u->mr->rkey, io_u->buf, max_bs); */
#endif
}
rd->send_buf.nr = htonl(i);
return ret;
}
static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
{
enum fio_ddir odir = ddir ^ 1;
struct timeval t;
long usec;
assert(ddir_rw(ddir));
if (td->rate_pending_usleep[ddir] <= 0)
return ddir;
/*
* We have too much pending sleep in this direction. See if we
* should switch.
*/
if (td_rw(td)) {
/*
* Other direction does not have too much pending, switch
*/
if (td->rate_pending_usleep[odir] < 100000)
return odir;
/*
* Both directions have pending sleep. Sleep the minimum time
* and deduct from both.
*/
if (td->rate_pending_usleep[ddir] <=
td->rate_pending_usleep[odir]) {
usec = td->rate_pending_usleep[ddir];
} else {
usec = td->rate_pending_usleep[odir];
ddir = odir;
}
} else
usec = td->rate_pending_usleep[ddir];
/*
* We are going to sleep, ensure that we flush anything pending as
* not to skew our latency numbers.
*
* Changed to only monitor 'in flight' requests here instead of the
* td->cur_depth, b/c td->cur_depth does not accurately represent
* io's that have been actually submitted to an async engine,
* and cur_depth is meaningless for sync engines.
*/
if (td->io_u_in_flight) {
int fio_unused ret;
ret = io_u_queued_complete(td, td->io_u_in_flight, NULL);
}
fio_gettime(&t, NULL);
usec_sleep(td, usec);
usec = utime_since_now(&t);
td->rate_pending_usleep[ddir] -= usec;
odir = ddir ^ 1;
if (td_rw(td) && __should_check_rate(td, odir))
td->rate_pending_usleep[odir] -= usec;
if (ddir_trim(ddir))
return ddir;
return ddir;
}
int generic_open_file(struct thread_data *td, struct fio_file *f)
{
int is_std = 0;
int flags = 0;
int from_hash = 0;
dprint(FD_FILE, "fd open %s\n", f->file_name);
if (td_trim(td) && f->filetype != FIO_TYPE_BD) {
log_err("fio: trim only applies to block device\n");
return 1;
}
if (!strcmp(f->file_name, "-")) {
if (td_rw(td)) {
log_err("fio: can't read/write to stdin/out\n");
return 1;
}
is_std = 1;
/*
* move output logging to stderr, if we are writing to stdout
*/
if (td_write(td))
f_out = stderr;
}
if (td_trim(td))
goto skip_flags;
if (td->o.odirect)
flags |= OS_O_DIRECT;
if (td->o.sync_io)
flags |= O_SYNC;
if (td->o.create_on_open)
flags |= O_CREAT;
skip_flags:
if (f->filetype != FIO_TYPE_FILE)
flags |= FIO_O_NOATIME;
open_again:
if (td_write(td)) {
if (!read_only)
flags |= O_RDWR;
if (f->filetype == FIO_TYPE_FILE)
flags |= O_CREAT;
if (is_std)
f->fd = dup(STDOUT_FILENO);
else
from_hash = file_lookup_open(f, flags);
} else if (td_read(td)) {
if (f->filetype == FIO_TYPE_CHAR && !read_only)
flags |= O_RDWR;
else
flags |= O_RDONLY;
if (is_std)
f->fd = dup(STDIN_FILENO);
else
from_hash = file_lookup_open(f, flags);
} else { //td trim
flags |= O_RDWR;
from_hash = file_lookup_open(f, flags);
}
if (f->fd == -1) {
char buf[FIO_VERROR_SIZE];
int __e = errno;
if (__e == EPERM && (flags & FIO_O_NOATIME)) {
flags &= ~FIO_O_NOATIME;
goto open_again;
}
if (__e == EMFILE && file_close_shadow_fds(td))
goto open_again;
snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
if (__e == EINVAL && (flags & OS_O_DIRECT)) {
log_err("fio: looks like your file system does not " \
"support direct=1/buffered=0\n");
}
td_verror(td, __e, buf);
}
if (!from_hash && f->fd != -1) {
if (add_file_hash(f)) {
int fio_unused ret;
/*
* Stash away descriptor for later close. This is to
* work-around a "feature" on Linux, where a close of
* an fd that has been opened for write will trigger
* udev to call blkid to check partitions, fs id, etc.
* That polutes the device cache, which can slow down
* unbuffered accesses.
*/
if (f->shadow_fd == -1)
f->shadow_fd = f->fd;
else {
/*
* OK to ignore, we haven't done anything
* with it
*/
ret = generic_close_file(td, f);
}
goto open_again;
}
}
return 0;
}
int generic_open_file(struct thread_data *td, struct fio_file *f)
{
int is_std = 0;
int flags = 0;
int from_hash = 0;
dprint(FD_FILE, "fd open %s\n", f->file_name);
if (!strcmp(f->file_name, "-")) {
if (td_rw(td)) {
log_err("fio: can't read/write to stdin/out\n");
return 1;
}
is_std = 1;
/*
* move output logging to stderr, if we are writing to stdout
*/
if (td_write(td))
f_out = stderr;
}
if (td->o.odirect)
flags |= OS_O_DIRECT;
if (td->o.sync_io)
flags |= O_SYNC;
if (f->filetype != FIO_TYPE_FILE)
flags |= FIO_O_NOATIME;
if (td->o.create_on_open)
flags |= O_CREAT;
open_again:
if (td_write(td)) {
if (!read_only)
flags |= O_RDWR;
if (f->filetype == FIO_TYPE_FILE)
flags |= O_CREAT;
if (is_std)
f->fd = dup(STDOUT_FILENO);
else
from_hash = file_lookup_open(f, flags);
} else {
if (f->filetype == FIO_TYPE_CHAR && !read_only)
flags |= O_RDWR;
else
flags |= O_RDONLY;
if (is_std)
f->fd = dup(STDIN_FILENO);
else
from_hash = file_lookup_open(f, flags);
}
if (f->fd == -1) {
char buf[FIO_VERROR_SIZE];
int __e = errno;
if (__e == EPERM && (flags & FIO_O_NOATIME)) {
flags &= ~FIO_O_NOATIME;
goto open_again;
}
snprintf(buf, sizeof(buf) - 1, "open(%s)", f->file_name);
td_verror(td, __e, buf);
}
if (!from_hash && f->fd != -1) {
if (add_file_hash(f)) {
int fio_unused ret;
/*
* OK to ignore, we haven't done anything with it
*/
ret = generic_close_file(td, f);
goto open_again;
}
}
return 0;
}
/*
* Sets the status of the 'td' in the printed status map.
*/
static void check_str_update(struct thread_data *td)
{
char c = __run_str[td->thread_number - 1];
switch (td->runstate) {
case TD_REAPED:
if (td->error)
c = 'X';
else if (td->sig)
c = 'K';
else
c = '_';
break;
case TD_EXITED:
c = 'E';
break;
case TD_RAMP:
c = '/';
break;
case TD_RUNNING:
if (td_rw(td)) {
if (td_random(td)) {
if (td->o.rwmix[DDIR_READ] == 100)
c = 'r';
else if (td->o.rwmix[DDIR_WRITE] == 100)
c = 'w';
else
c = 'm';
} else {
if (td->o.rwmix[DDIR_READ] == 100)
c = 'R';
else if (td->o.rwmix[DDIR_WRITE] == 100)
c = 'W';
else
c = 'M';
}
} else if (td_read(td)) {
if (td_random(td))
c = 'r';
else
c = 'R';
} else if (td_write(td)) {
if (td_random(td))
c = 'w';
else
c = 'W';
} else {
if (td_random(td))
c = 'd';
else
c = 'D';
}
break;
case TD_PRE_READING:
c = 'p';
break;
case TD_VERIFYING:
c = 'V';
break;
case TD_FSYNCING:
c = 'F';
break;
case TD_FINISHING:
c = 'f';
break;
case TD_CREATED:
c = 'C';
break;
case TD_INITIALIZED:
case TD_SETTING_UP:
c = 'I';
break;
case TD_NOT_CREATED:
c = 'P';
break;
default:
log_err("state %d\n", td->runstate);
}
__run_str[td->thread_number - 1] = c;
update_condensed_str(__run_str, run_str);
}
/*
* Lazy way of fixing up options that depend on each other. We could also
* define option callback handlers, but this is easier.
*/
static int fixup_options(struct thread_data *td)
{
struct thread_options *o = &td->o;
int ret = 0;
#ifndef FIO_HAVE_PSHARED_MUTEX
if (!o->use_thread) {
log_info("fio: this platform does not support process shared"
" mutexes, forcing use of threads. Use the 'thread'"
" option to get rid of this warning.\n");
o->use_thread = 1;
ret = warnings_fatal;
}
#endif
if (o->write_iolog_file && o->read_iolog_file) {
log_err("fio: read iolog overrides write_iolog\n");
free(o->write_iolog_file);
o->write_iolog_file = NULL;
ret = warnings_fatal;
}
/*
* only really works with 1 file
*/
if (o->zone_size && o->open_files == 1)
o->zone_size = 0;
/*
* If zone_range isn't specified, backward compatibility dictates it
* should be made equal to zone_size.
*/
if (o->zone_size && !o->zone_range)
o->zone_range = o->zone_size;
/*
* Reads can do overwrites, we always need to pre-create the file
*/
if (td_read(td) || td_rw(td))
o->overwrite = 1;
if (!o->min_bs[DDIR_READ])
o->min_bs[DDIR_READ] = o->bs[DDIR_READ];
if (!o->max_bs[DDIR_READ])
o->max_bs[DDIR_READ] = o->bs[DDIR_READ];
if (!o->min_bs[DDIR_WRITE])
o->min_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
if (!o->max_bs[DDIR_WRITE])
o->max_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
o->rw_min_bs = min(o->min_bs[DDIR_READ], o->min_bs[DDIR_WRITE]);
/*
* For random IO, allow blockalign offset other than min_bs.
*/
if (!o->ba[DDIR_READ] || !td_random(td))
o->ba[DDIR_READ] = o->min_bs[DDIR_READ];
if (!o->ba[DDIR_WRITE] || !td_random(td))
o->ba[DDIR_WRITE] = o->min_bs[DDIR_WRITE];
if ((o->ba[DDIR_READ] != o->min_bs[DDIR_READ] ||
o->ba[DDIR_WRITE] != o->min_bs[DDIR_WRITE]) &&
!o->norandommap) {
log_err("fio: Any use of blockalign= turns off randommap\n");
o->norandommap = 1;
ret = warnings_fatal;
}
if (!o->file_size_high)
o->file_size_high = o->file_size_low;
if (o->norandommap && o->verify != VERIFY_NONE
&& !fixed_block_size(o)) {
log_err("fio: norandommap given for variable block sizes, "
"verify disabled\n");
o->verify = VERIFY_NONE;
ret = warnings_fatal;
}
if (o->bs_unaligned && (o->odirect || td->io_ops->flags & FIO_RAWIO))
log_err("fio: bs_unaligned may not work with raw io\n");
/*
* thinktime_spin must be less than thinktime
*/
if (o->thinktime_spin > o->thinktime)
o->thinktime_spin = o->thinktime;
/*
* The low water mark cannot be bigger than the iodepth
*/
if (o->iodepth_low > o->iodepth || !o->iodepth_low) {
/*
* syslet work around - if the workload is sequential,
* we want to let the queue drain all the way down to
* avoid seeking between async threads
*/
if (!strcmp(td->io_ops->name, "syslet-rw") && !td_random(td))
o->iodepth_low = 1;
else
o->iodepth_low = o->iodepth;
}
/*
* If batch number isn't set, default to the same as iodepth
*/
if (o->iodepth_batch > o->iodepth || !o->iodepth_batch)
o->iodepth_batch = o->iodepth;
if (o->nr_files > td->files_index)
o->nr_files = td->files_index;
if (o->open_files > o->nr_files || !o->open_files)
o->open_files = o->nr_files;
if (((o->rate[0] + o->rate[1]) && (o->rate_iops[0] + o->rate_iops[1]))||
((o->ratemin[0] + o->ratemin[1]) && (o->rate_iops_min[0] +
o->rate_iops_min[1]))) {
log_err("fio: rate and rate_iops are mutually exclusive\n");
ret = 1;
}
if ((o->rate[0] < o->ratemin[0]) || (o->rate[1] < o->ratemin[1]) ||
(o->rate_iops[0] < o->rate_iops_min[0]) ||
(o->rate_iops[1] < o->rate_iops_min[1])) {
log_err("fio: minimum rate exceeds rate\n");
ret = 1;
}
if (!o->timeout && o->time_based) {
log_err("fio: time_based requires a runtime/timeout setting\n");
o->time_based = 0;
ret = warnings_fatal;
}
if (o->fill_device && !o->size)
o->size = -1ULL;
if (o->verify != VERIFY_NONE) {
if (td_write(td) && o->do_verify && o->numjobs > 1) {
log_info("Multiple writers may overwrite blocks that "
"belong to other jobs. This can cause "
"verification failures.\n");
ret = warnings_fatal;
}
o->refill_buffers = 1;
if (o->max_bs[DDIR_WRITE] != o->min_bs[DDIR_WRITE] &&
!o->verify_interval)
o->verify_interval = o->min_bs[DDIR_WRITE];
}
if (o->pre_read) {
o->invalidate_cache = 0;
if (td->io_ops->flags & FIO_PIPEIO) {
log_info("fio: cannot pre-read files with an IO engine"
" that isn't seekable. Pre-read disabled.\n");
ret = warnings_fatal;
}
}
#ifndef FIO_HAVE_FDATASYNC
if (o->fdatasync_blocks) {
log_info("fio: this platform does not support fdatasync()"
" falling back to using fsync(). Use the 'fsync'"
" option instead of 'fdatasync' to get rid of"
" this warning\n");
o->fsync_blocks = o->fdatasync_blocks;
o->fdatasync_blocks = 0;
ret = warnings_fatal;
}
#endif
#ifdef WIN32
/*
* Windows doesn't support O_DIRECT or O_SYNC with the _open interface,
* so fail if we're passed those flags
*/
if ((td->io_ops->flags & FIO_SYNCIO) && (td->o.odirect || td->o.sync_io)) {
log_err("fio: Windows does not support direct or non-buffered io with"
" the synchronous ioengines. Use the 'windowsaio' ioengine"
" with 'direct=1' and 'iodepth=1' instead.\n");
ret = 1;
}
#endif
/*
* For fully compressible data, just zero them at init time.
* It's faster than repeatedly filling it.
*/
if (td->o.compress_percentage == 100) {
td->o.zero_buffers = 1;
td->o.compress_percentage = 0;
}
return ret;
}
/*
* Lazy way of fixing up options that depend on each other. We could also
* define option callback handlers, but this is easier.
*/
static int fixup_options(struct thread_data *td)
{
struct thread_options *o = &td->o;
int ret = 0;
#ifndef FIO_HAVE_PSHARED_MUTEX
if (!o->use_thread) {
log_info("fio: this platform does not support process shared"
" mutexes, forcing use of threads. Use the 'thread'"
" option to get rid of this warning.\n");
o->use_thread = 1;
ret = warnings_fatal;
}
#endif
if (o->write_iolog_file && o->read_iolog_file) {
log_err("fio: read iolog overrides write_iolog\n");
free(o->write_iolog_file);
o->write_iolog_file = NULL;
ret = warnings_fatal;
}
/*
* only really works with 1 file
*/
if (o->zone_size && o->open_files > 1)
o->zone_size = 0;
/*
* If zone_range isn't specified, backward compatibility dictates it
* should be made equal to zone_size.
*/
if (o->zone_size && !o->zone_range)
o->zone_range = o->zone_size;
/*
* Reads can do overwrites, we always need to pre-create the file
*/
if (td_read(td) || td_rw(td))
o->overwrite = 1;
if (!o->min_bs[DDIR_READ])
o->min_bs[DDIR_READ] = o->bs[DDIR_READ];
if (!o->max_bs[DDIR_READ])
o->max_bs[DDIR_READ] = o->bs[DDIR_READ];
if (!o->min_bs[DDIR_WRITE])
o->min_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
if (!o->max_bs[DDIR_WRITE])
o->max_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
if (!o->min_bs[DDIR_TRIM])
o->min_bs[DDIR_TRIM] = o->bs[DDIR_TRIM];
if (!o->max_bs[DDIR_TRIM])
o->max_bs[DDIR_TRIM] = o->bs[DDIR_TRIM];
o->rw_min_bs = min(o->min_bs[DDIR_READ], o->min_bs[DDIR_WRITE]);
o->rw_min_bs = min(o->min_bs[DDIR_TRIM], o->rw_min_bs);
/*
* For random IO, allow blockalign offset other than min_bs.
*/
if (!o->ba[DDIR_READ] || !td_random(td))
o->ba[DDIR_READ] = o->min_bs[DDIR_READ];
if (!o->ba[DDIR_WRITE] || !td_random(td))
o->ba[DDIR_WRITE] = o->min_bs[DDIR_WRITE];
if (!o->ba[DDIR_TRIM] || !td_random(td))
o->ba[DDIR_TRIM] = o->min_bs[DDIR_TRIM];
if ((o->ba[DDIR_READ] != o->min_bs[DDIR_READ] ||
o->ba[DDIR_WRITE] != o->min_bs[DDIR_WRITE] ||
o->ba[DDIR_TRIM] != o->min_bs[DDIR_TRIM]) &&
!o->norandommap) {
log_err("fio: Any use of blockalign= turns off randommap\n");
o->norandommap = 1;
ret = warnings_fatal;
}
if (!o->file_size_high)
o->file_size_high = o->file_size_low;
if (o->start_delay_high)
o->start_delay = get_rand_start_delay(td);
if (o->norandommap && o->verify != VERIFY_NONE
&& !fixed_block_size(o)) {
log_err("fio: norandommap given for variable block sizes, "
"verify disabled\n");
o->verify = VERIFY_NONE;
ret = warnings_fatal;
}
if (o->bs_unaligned && (o->odirect || td->io_ops->flags & FIO_RAWIO))
log_err("fio: bs_unaligned may not work with raw io\n");
/*
* thinktime_spin must be less than thinktime
*/
if (o->thinktime_spin > o->thinktime)
o->thinktime_spin = o->thinktime;
/*
* The low water mark cannot be bigger than the iodepth
*/
if (o->iodepth_low > o->iodepth || !o->iodepth_low)
o->iodepth_low = o->iodepth;
/*
* If batch number isn't set, default to the same as iodepth
*/
if (o->iodepth_batch > o->iodepth || !o->iodepth_batch)
o->iodepth_batch = o->iodepth;
if (o->nr_files > td->files_index)
o->nr_files = td->files_index;
if (o->open_files > o->nr_files || !o->open_files)
o->open_files = o->nr_files;
if (((o->rate[DDIR_READ] + o->rate[DDIR_WRITE] + o->rate[DDIR_TRIM]) &&
(o->rate_iops[DDIR_READ] + o->rate_iops[DDIR_WRITE] + o->rate_iops[DDIR_TRIM])) ||
((o->ratemin[DDIR_READ] + o->ratemin[DDIR_WRITE] + o->ratemin[DDIR_TRIM]) &&
(o->rate_iops_min[DDIR_READ] + o->rate_iops_min[DDIR_WRITE] + o->rate_iops_min[DDIR_TRIM]))) {
log_err("fio: rate and rate_iops are mutually exclusive\n");
ret = 1;
}
if ((o->rate[DDIR_READ] < o->ratemin[DDIR_READ]) ||
(o->rate[DDIR_WRITE] < o->ratemin[DDIR_WRITE]) ||
(o->rate[DDIR_TRIM] < o->ratemin[DDIR_TRIM]) ||
(o->rate_iops[DDIR_READ] < o->rate_iops_min[DDIR_READ]) ||
(o->rate_iops[DDIR_WRITE] < o->rate_iops_min[DDIR_WRITE]) ||
(o->rate_iops[DDIR_TRIM] < o->rate_iops_min[DDIR_TRIM])) {
log_err("fio: minimum rate exceeds rate\n");
ret = 1;
}
if (!o->timeout && o->time_based) {
log_err("fio: time_based requires a runtime/timeout setting\n");
o->time_based = 0;
ret = warnings_fatal;
}
if (o->fill_device && !o->size)
o->size = -1ULL;
if (o->verify != VERIFY_NONE) {
if (td_write(td) && o->do_verify && o->numjobs > 1) {
log_info("Multiple writers may overwrite blocks that "
"belong to other jobs. This can cause "
"verification failures.\n");
ret = warnings_fatal;
}
o->refill_buffers = 1;
if (o->max_bs[DDIR_WRITE] != o->min_bs[DDIR_WRITE] &&
!o->verify_interval)
o->verify_interval = o->min_bs[DDIR_WRITE];
/*
* Verify interval must be smaller or equal to the
* write size.
*/
if (o->verify_interval > o->min_bs[DDIR_WRITE])
o->verify_interval = o->min_bs[DDIR_WRITE];
else if (td_read(td) && o->verify_interval > o->min_bs[DDIR_READ])
o->verify_interval = o->min_bs[DDIR_READ];
}
if (o->pre_read) {
o->invalidate_cache = 0;
if (td->io_ops->flags & FIO_PIPEIO) {
log_info("fio: cannot pre-read files with an IO engine"
" that isn't seekable. Pre-read disabled.\n");
ret = warnings_fatal;
}
}
if (!o->unit_base) {
if (td->io_ops->flags & FIO_BIT_BASED)
o->unit_base = 1;
else
o->unit_base = 8;
}
#ifndef CONFIG_FDATASYNC
if (o->fdatasync_blocks) {
log_info("fio: this platform does not support fdatasync()"
" falling back to using fsync(). Use the 'fsync'"
" option instead of 'fdatasync' to get rid of"
" this warning\n");
o->fsync_blocks = o->fdatasync_blocks;
o->fdatasync_blocks = 0;
ret = warnings_fatal;
}
#endif
#ifdef WIN32
/*
* Windows doesn't support O_DIRECT or O_SYNC with the _open interface,
* so fail if we're passed those flags
*/
if ((td->io_ops->flags & FIO_SYNCIO) && (td->o.odirect || td->o.sync_io)) {
log_err("fio: Windows does not support direct or non-buffered io with"
" the synchronous ioengines. Use the 'windowsaio' ioengine"
" with 'direct=1' and 'iodepth=1' instead.\n");
ret = 1;
}
#endif
/*
* For fully compressible data, just zero them at init time.
* It's faster than repeatedly filling it.
*/
if (td->o.compress_percentage == 100) {
td->o.zero_buffers = 1;
td->o.compress_percentage = 0;
}
/*
* Using a non-uniform random distribution excludes usage of
* a random map
*/
if (td->o.random_distribution != FIO_RAND_DIST_RANDOM)
td->o.norandommap = 1;
/*
* If size is set but less than the min block size, complain
*/
if (o->size && o->size < td_min_bs(td)) {
log_err("fio: size too small, must be larger than the IO size: %llu\n", (unsigned long long) o->size);
ret = 1;
}
/*
* O_ATOMIC implies O_DIRECT
*/
if (td->o.oatomic)
td->o.odirect = 1;
/*
* If randseed is set, that overrides randrepeat
*/
if (td->o.rand_seed)
td->o.rand_repeatable = 0;
if ((td->io_ops->flags & FIO_NOEXTEND) && td->o.file_append) {
log_err("fio: can't append/extent with IO engine %s\n", td->io_ops->name);
ret = 1;
}
return ret;
}
