static int fill_io_u(struct thread_data *td, struct io_u *io_u)
{
if (td->io_ops->flags & FIO_NOIO)
goto out;
set_rw_ddir(td, io_u);
/*
* fsync() or fdatasync() or trim etc, we are done
*/
if (!ddir_rw(io_u->ddir))
goto out;
/*
* See if it's time to switch to a new zone
*/
if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
td->zone_bytes = 0;
io_u->file->file_offset += td->o.zone_range + td->o.zone_skip;
io_u->file->last_pos = io_u->file->file_offset;
td->io_skip_bytes += td->o.zone_skip;
}
/*
* No log, let the seq/rand engine retrieve the next buflen and
* position.
*/
if (get_next_offset(td, io_u)) {
dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
return 1;
}
io_u->buflen = get_next_buflen(td, io_u);
if (!io_u->buflen) {
dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
return 1;
}
if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
dprint(FD_IO, "io_u %p, offset too large\n", io_u);
dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset,
io_u->buflen, io_u->file->real_file_size);
return 1;
}
/*
* mark entry before potentially trimming io_u
*/
if (td_random(td) && file_randommap(td, io_u->file))
mark_random_map(td, io_u);
/*
* If using a write iolog, store this entry.
*/
out:
dprint_io_u(io_u, "fill_io_u");
td->zone_bytes += io_u->buflen;
log_io_u(td, io_u);
return 0;
}
/*
* Must be serialized by caller.
*/
int workqueue_enqueue(struct workqueue *wq, struct io_u *io_u)
{
struct submit_worker *sw;
sw = get_submit_worker(wq);
if (sw) {
const enum fio_ddir ddir = acct_ddir(io_u);
struct thread_data *parent = wq->td;
if (ddir_rw(ddir)) {
parent->io_issues[ddir]++;
parent->io_issue_bytes[ddir] += io_u->xfer_buflen;
parent->rate_io_issue_bytes[ddir] += io_u->xfer_buflen;
}
pthread_mutex_lock(&sw->lock);
flist_add_tail(&io_u->verify_list, &sw->work_list);
sw->seq = ++wq->work_seq;
sw->flags &= ~SW_F_IDLE;
pthread_mutex_unlock(&sw->lock);
pthread_cond_signal(&sw->cond);
return FIO_Q_QUEUED;
}
return FIO_Q_BUSY;
}
static int get_next_block(struct thread_data *td, struct io_u *io_u,
enum fio_ddir ddir, int rw_seq, unsigned long long *b)
{
struct fio_file *f = io_u->file;
int ret;
assert(ddir_rw(ddir));
if (rw_seq) {
if (td_random(td))
ret = get_next_rand_block(td, f, ddir, b);
else
ret = get_next_seq_block(td, f, ddir, b);
} else {
io_u->flags |= IO_U_F_BUSY_OK;
if (td->o.rw_seq == RW_SEQ_SEQ) {
ret = get_next_seq_block(td, f, ddir, b);
if (ret)
ret = get_next_rand_block(td, f, ddir, b);
} else if (td->o.rw_seq == RW_SEQ_IDENT) {
if (f->last_start != -1ULL)
*b = (f->last_start - f->file_offset)
/ td->o.min_bs[ddir];
else
*b = 0;
ret = 0;
} else {
log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
ret = 1;
}
}
return ret;
}
/*
* For random io, generate a random new block and see if it's used. Repeat
* until we find a free one. For sequential io, just return the end of
* the last io issued.
*/
static int __get_next_offset(struct thread_data *td, struct io_u *io_u,
unsigned int *is_random)
{
struct fio_file *f = io_u->file;
enum fio_ddir ddir = io_u->ddir;
int rw_seq_hit = 0;
assert(ddir_rw(ddir));
if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
rw_seq_hit = 1;
td->ddir_seq_nr = td->o.ddir_seq_nr;
}
if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
return 1;
if (io_u->offset >= f->io_size) {
dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
(unsigned long long) io_u->offset,
(unsigned long long) f->io_size);
return 1;
}
io_u->offset += f->file_offset;
if (io_u->offset >= f->real_file_size) {
dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
(unsigned long long) io_u->offset,
(unsigned long long) f->real_file_size);
return 1;
}
return 0;
}
static uint64_t last_block(struct thread_data *td, struct fio_file *f,
enum fio_ddir ddir)
{
uint64_t max_blocks;
uint64_t max_size;
assert(ddir_rw(ddir));
/*
* Hmm, should we make sure that ->io_size <= ->real_file_size?
*/
max_size = f->io_size;
if (max_size > f->real_file_size)
max_size = f->real_file_size;
if (td->o.zone_range)
max_size = td->o.zone_range;
if (td->o.min_bs[ddir] > td->o.ba[ddir])
max_size -= td->o.min_bs[ddir] - td->o.ba[ddir];
max_blocks = max_size / (uint64_t) td->o.ba[ddir];
if (!max_blocks)
return 0;
return max_blocks;
}
static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
{
const int ddir = io_u->ddir;
unsigned int uninitialized_var(buflen);
unsigned int minbs, maxbs;
unsigned long r, rand_max;
assert(ddir_rw(ddir));
minbs = td->o.min_bs[ddir];
maxbs = td->o.max_bs[ddir];
if (minbs == maxbs)
return minbs;
/*
* If we can't satisfy the min block size from here, then fail
*/
if (!io_u_fits(td, io_u, minbs))
return 0;
if (td->o.use_os_rand)
rand_max = OS_RAND_MAX;
else
rand_max = FRAND_MAX;
do {
if (td->o.use_os_rand)
r = os_random_long(&td->bsrange_state);
else
r = __rand(&td->__bsrange_state);
if (!td->o.bssplit_nr[ddir]) {
buflen = 1 + (unsigned int) ((double) maxbs *
(r / (rand_max + 1.0)));
if (buflen < minbs)
buflen = minbs;
} else {
long perc = 0;
unsigned int i;
for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
struct bssplit *bsp = &td->o.bssplit[ddir][i];
buflen = bsp->bs;
perc += bsp->perc;
if ((r <= ((rand_max / 100L) * perc)) &&
io_u_fits(td, io_u, buflen))
break;
}
}
if (!td->o.bs_unaligned && is_power_of_2(minbs))
buflen = (buflen + minbs - 1) & ~(minbs - 1);
} while (!io_u_fits(td, io_u, buflen));
return buflen;
}
static int get_next_seq_block(struct thread_data *td, struct fio_file *f,
enum fio_ddir ddir, unsigned long long *b)
{
assert(ddir_rw(ddir));
if (f->last_pos < f->real_file_size) {
*b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir];
return 0;
}
return 1;
}
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 get_next_block(struct thread_data *td, struct io_u *io_u,
enum fio_ddir ddir, int rw_seq)
{
struct fio_file *f = io_u->file;
unsigned long long b, offset;
int ret;
assert(ddir_rw(ddir));
b = offset = -1ULL;
if (rw_seq) {
if (td_random(td))
ret = get_next_rand_block(td, f, ddir, &b);
else
ret = get_next_seq_offset(td, f, ddir, &offset);
} else {
io_u->flags |= IO_U_F_BUSY_OK;
if (td->o.rw_seq == RW_SEQ_SEQ) {
ret = get_next_seq_offset(td, f, ddir, &offset);
if (ret)
ret = get_next_rand_block(td, f, ddir, &b);
} else if (td->o.rw_seq == RW_SEQ_IDENT) {
if (f->last_start != -1ULL)
offset = f->last_start - f->file_offset;
else
offset = 0;
ret = 0;
} else {
log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
ret = 1;
}
}
if (!ret) {
if (offset != -1ULL)
io_u->offset = offset;
else if (b != -1ULL)
io_u->offset = b * td->o.ba[ddir];
else {
log_err("fio: bug in offset generation\n");
ret = 1;
}
}
return ret;
}
static int fio_syncio_prep(struct thread_data *td, struct io_u *io_u)
{
struct fio_file *f = io_u->file;
if (!ddir_rw(io_u->ddir))
return 0;
if (LAST_POS(f) != -1ULL && LAST_POS(f) == io_u->offset)
return 0;
if (lseek(f->fd, io_u->offset, SEEK_SET) == -1) {
td_verror(td, errno, "lseek");
return 1;
}
return 0;
}
static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
{
const int ddir = io_u->ddir;
unsigned int uninitialized_var(buflen);
unsigned int minbs, maxbs;
long r;
assert(ddir_rw(ddir));
minbs = td->o.min_bs[ddir];
maxbs = td->o.max_bs[ddir];
if (minbs == maxbs)
buflen = minbs;
else {
r = os_random_long(&td->bsrange_state);
if (!td->o.bssplit_nr[ddir]) {
buflen = 1 + (unsigned int) ((double) maxbs *
(r / (OS_RAND_MAX + 1.0)));
if (buflen < minbs)
buflen = minbs;
} else {
long perc = 0;
unsigned int i;
for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
struct bssplit *bsp = &td->o.bssplit[ddir][i];
buflen = bsp->bs;
perc += bsp->perc;
if (r <= ((OS_RAND_MAX / 100L) * perc))
break;
}
}
if (!td->o.bs_unaligned && is_power_of_2(minbs))
buflen = (buflen + minbs - 1) & ~(minbs - 1);
}
if (io_u->offset + buflen > io_u->file->real_file_size) {
dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen,
minbs, ddir);
buflen = minbs;
}
return buflen;
}
static int __setup_rate(struct thread_data *td, enum fio_ddir ddir)
{
unsigned int bs = td->o.min_bs[ddir];
assert(ddir_rw(ddir));
if (td->o.rate[ddir])
td->rate_bps[ddir] = td->o.rate[ddir];
else
td->rate_bps[ddir] = td->o.rate_iops[ddir] * bs;
if (!td->rate_bps[ddir]) {
log_err("rate lower than supported\n");
return -1;
}
td->rate_pending_usleep[ddir] = 0;
return 0;
}
void requeue_io_u(struct thread_data *td, struct io_u **io_u)
{
struct io_u *__io_u = *io_u;
dprint(FD_IO, "requeue %p\n", __io_u);
td_io_u_lock(td);
__io_u->flags |= IO_U_F_FREE;
if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(__io_u->ddir))
td->io_issues[__io_u->ddir]--;
__io_u->flags &= ~IO_U_F_FLIGHT;
if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
td->cur_depth--;
flist_del(&__io_u->list);
flist_add_tail(&__io_u->list, &td->io_u_requeues);
td_io_u_unlock(td);
*io_u = NULL;
}
static int fio_gf_prep(struct thread_data *td, struct io_u *io_u)
{
struct fio_file *f = io_u->file;
struct gf_data *g = td->io_ops_data;
dprint(FD_FILE, "fio prep\n");
if (!ddir_rw(io_u->ddir))
return 0;
if (LAST_POS(f) != -1ULL && LAST_POS(f) == io_u->offset)
return 0;
if (glfs_lseek(g->fd, io_u->offset, SEEK_SET) < 0) {
td_verror(td, errno, "lseek");
return 1;
}
return 0;
}
static int fio_gf_queue(struct thread_data *td, struct io_u *io_u)
{
struct gf_data *g = td->io_ops_data;
int ret = 0;
dprint(FD_FILE, "fio queue len %lu\n", io_u->xfer_buflen);
fio_ro_check(td, io_u);
if (io_u->ddir == DDIR_READ)
ret = glfs_read(g->fd, io_u->xfer_buf, io_u->xfer_buflen, 0);
else if (io_u->ddir == DDIR_WRITE)
ret = glfs_write(g->fd, io_u->xfer_buf, io_u->xfer_buflen, 0);
else if (io_u->ddir == DDIR_SYNC)
ret = glfs_fsync(g->fd);
else if (io_u->ddir == DDIR_DATASYNC)
ret = glfs_fdatasync(g->fd);
else {
log_err("unsupported operation.\n");
return -EINVAL;
}
dprint(FD_FILE, "fio len %lu ret %d\n", io_u->xfer_buflen, ret);
if (io_u->file && ret >= 0 && ddir_rw(io_u->ddir))
LAST_POS(io_u->file) = io_u->offset + ret;
if (ret != (int)io_u->xfer_buflen) {
if (ret >= 0) {
io_u->resid = io_u->xfer_buflen - ret;
io_u->error = 0;
return FIO_Q_COMPLETED;
} else
io_u->error = errno;
}
if (io_u->error) {
log_err("IO failed.\n");
td_verror(td, io_u->error, "xfer");
}
return FIO_Q_COMPLETED;
}
static enum fio_q_status fio_mmapio_queue(struct thread_data *td,
struct io_u *io_u)
{
struct fio_file *f = io_u->file;
struct fio_mmap_data *fmd = FILE_ENG_DATA(f);
fio_ro_check(td, io_u);
if (io_u->ddir == DDIR_READ)
memcpy(io_u->xfer_buf, io_u->mmap_data, io_u->xfer_buflen);
else if (io_u->ddir == DDIR_WRITE)
memcpy(io_u->mmap_data, io_u->xfer_buf, io_u->xfer_buflen);
else if (ddir_sync(io_u->ddir)) {
if (msync(fmd->mmap_ptr, fmd->mmap_sz, MS_SYNC)) {
io_u->error = errno;
td_verror(td, io_u->error, "msync");
}
} else if (io_u->ddir == DDIR_TRIM) {
int ret = do_io_u_trim(td, io_u);
if (!ret)
td_verror(td, io_u->error, "trim");
}
/*
* not really direct, but should drop the pages from the cache
*/
if (td->o.odirect && ddir_rw(io_u->ddir)) {
if (msync(io_u->mmap_data, io_u->xfer_buflen, MS_SYNC) < 0) {
io_u->error = errno;
td_verror(td, io_u->error, "msync");
}
if (posix_madvise(io_u->mmap_data, io_u->xfer_buflen, POSIX_MADV_DONTNEED) < 0) {
io_u->error = errno;
td_verror(td, io_u->error, "madvise");
}
}
return FIO_Q_COMPLETED;
}
void requeue_io_u(struct thread_data *td, struct io_u **io_u)
{
struct io_u *__io_u = *io_u;
enum fio_ddir ddir = acct_ddir(__io_u);
dprint(FD_IO, "requeue %p\n", __io_u);
td_io_u_lock(td);
__io_u->flags |= IO_U_F_FREE;
if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
td->io_issues[ddir]--;
__io_u->flags &= ~IO_U_F_FLIGHT;
if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
td->cur_depth--;
io_u_rpush(&td->io_u_requeues, __io_u);
td_io_u_unlock(td);
*io_u = NULL;
}
static unsigned long long last_block(struct thread_data *td, struct fio_file *f,
enum fio_ddir ddir)
{
unsigned long long max_blocks;
unsigned long long max_size;
assert(ddir_rw(ddir));
/*
* Hmm, should we make sure that ->io_size <= ->real_file_size?
*/
max_size = f->io_size;
if (max_size > f->real_file_size)
max_size = f->real_file_size;
max_blocks = max_size / (unsigned long long) td->o.ba[ddir];
if (!max_blocks)
return 0;
return max_blocks;
}
static int fio_io_end(struct thread_data *td, struct io_u *io_u, int ret)
{
if (io_u->file && ret >= 0 && ddir_rw(io_u->ddir))
LAST_POS(io_u->file) = io_u->offset + ret;
if (ret != (int) io_u->xfer_buflen) {
if (ret >= 0) {
io_u->resid = io_u->xfer_buflen - ret;
io_u->error = 0;
return FIO_Q_COMPLETED;
} else
io_u->error = errno;
}
if (io_u->error) {
io_u_log_error(td, io_u);
td_verror(td, io_u->error, "xfer");
}
return FIO_Q_COMPLETED;
}
static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
enum fio_ddir ddir, uint64_t *offset)
{
assert(ddir_rw(ddir));
if (f->last_pos >= f->io_size + get_start_offset(td) && td->o.time_based)
f->last_pos = f->last_pos - f->io_size;
if (f->last_pos < f->real_file_size) {
uint64_t pos;
if (f->last_pos == f->file_offset && td->o.ddir_seq_add < 0)
f->last_pos = f->real_file_size;
pos = f->last_pos - f->file_offset;
if (pos)
pos += td->o.ddir_seq_add;
*offset = pos;
return 0;
}
return 1;
}
static void io_completed(struct thread_data *td, struct io_u *io_u,
struct io_completion_data *icd)
{
struct fio_file *f;
dprint_io_u(io_u, "io complete");
td_io_u_lock(td);
assert(io_u->flags & IO_U_F_FLIGHT);
io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
/*
* Mark IO ok to verify
*/
if (io_u->ipo) {
io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
write_barrier();
}
td_io_u_unlock(td);
if (ddir_sync(io_u->ddir)) {
td->last_was_sync = 1;
f = io_u->file;
if (f) {
f->first_write = -1ULL;
f->last_write = -1ULL;
}
return;
}
td->last_was_sync = 0;
td->last_ddir = io_u->ddir;
if (!io_u->error && ddir_rw(io_u->ddir)) {
unsigned int bytes = io_u->buflen - io_u->resid;
const enum fio_ddir idx = io_u->ddir;
const enum fio_ddir odx = io_u->ddir ^ 1;
int ret;
td->io_blocks[idx]++;
td->this_io_blocks[idx]++;
td->io_bytes[idx] += bytes;
if (!(io_u->flags & IO_U_F_VER_LIST))
td->this_io_bytes[idx] += bytes;
if (idx == DDIR_WRITE) {
f = io_u->file;
if (f) {
if (f->first_write == -1ULL ||
io_u->offset < f->first_write)
f->first_write = io_u->offset;
if (f->last_write == -1ULL ||
((io_u->offset + bytes) > f->last_write))
f->last_write = io_u->offset + bytes;
}
}
if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
td->runstate == TD_VERIFYING)) {
account_io_completion(td, io_u, icd, idx, bytes);
if (__should_check_rate(td, idx)) {
td->rate_pending_usleep[idx] =
(usec_for_io(td, idx) -
utime_since_now(&td->start));
}
if (idx != DDIR_TRIM && __should_check_rate(td, odx))
td->rate_pending_usleep[odx] =
(usec_for_io(td, odx) -
utime_since_now(&td->start));
}
icd->bytes_done[idx] += bytes;
if (io_u->end_io) {
ret = io_u->end_io(td, io_u);
if (ret && !icd->error)
icd->error = ret;
}
} else if (io_u->error) {
icd->error = io_u->error;
io_u_log_error(td, io_u);
}
if (icd->error) {
enum error_type_bit eb = td_error_type(io_u->ddir, icd->error);
if (!td_non_fatal_error(td, eb, icd->error))
return;
/*
* If there is a non_fatal error, then add to the error count
* and clear all the errors.
*/
update_error_count(td, icd->error);
td_clear_error(td);
icd->error = 0;
io_u->error = 0;
}
}
/*
* Return an io_u to be processed. Gets a buflen and offset, sets direction,
* etc. The returned io_u is fully ready to be prepped and submitted.
*/
struct io_u *get_io_u(struct thread_data *td)
{
struct fio_file *f;
struct io_u *io_u;
int do_scramble = 0;
long ret = 0;
io_u = __get_io_u(td);
if (!io_u) {
dprint(FD_IO, "__get_io_u failed\n");
return NULL;
}
if (check_get_verify(td, io_u))
goto out;
if (check_get_trim(td, io_u))
goto out;
/*
* from a requeue, io_u already setup
*/
if (io_u->file)
goto out;
/*
* If using an iolog, grab next piece if any available.
*/
if (td->flags & TD_F_READ_IOLOG) {
if (read_iolog_get(td, io_u))
goto err_put;
} else if (set_io_u_file(td, io_u)) {
ret = -EBUSY;
dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
goto err_put;
}
f = io_u->file;
if (!f) {
dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
goto err_put;
}
assert(fio_file_open(f));
if (ddir_rw(io_u->ddir)) {
if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
goto err_put;
}
f->last_start = io_u->offset;
f->last_pos = io_u->offset + io_u->buflen;
if (io_u->ddir == DDIR_WRITE) {
if (td->flags & TD_F_REFILL_BUFFERS) {
io_u_fill_buffer(td, io_u,
io_u->xfer_buflen, io_u->xfer_buflen);
} else if (td->flags & TD_F_SCRAMBLE_BUFFERS)
do_scramble = 1;
if (td->flags & TD_F_VER_NONE) {
populate_verify_io_u(td, io_u);
do_scramble = 0;
}
} else if (io_u->ddir == DDIR_READ) {
/*
* Reset the buf_filled parameters so next time if the
* buffer is used for writes it is refilled.
*/
io_u->buf_filled_len = 0;
}
}
/*
* Set io data pointers.
*/
io_u->xfer_buf = io_u->buf;
io_u->xfer_buflen = io_u->buflen;
out:
assert(io_u->file);
if (!td_io_prep(td, io_u)) {
if (!td->o.disable_slat)
fio_gettime(&io_u->start_time, NULL);
if (do_scramble)
small_content_scramble(io_u);
return io_u;
}
err_put:
dprint(FD_IO, "get_io_u failed\n");
put_io_u(td, io_u);
return ERR_PTR(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;
}
static void io_completed(struct thread_data *td, struct io_u *io_u,
struct io_completion_data *icd)
{
struct fio_file *f;
dprint_io_u(io_u, "io complete");
td_io_u_lock(td);
assert(io_u->flags & IO_U_F_FLIGHT);
io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
td_io_u_unlock(td);
if (ddir_sync(io_u->ddir)) {
td->last_was_sync = 1;
f = io_u->file;
if (f) {
f->first_write = -1ULL;
f->last_write = -1ULL;
}
return;
}
td->last_was_sync = 0;
td->last_ddir = io_u->ddir;
if (!io_u->error && ddir_rw(io_u->ddir)) {
unsigned int bytes = io_u->buflen - io_u->resid;
const enum fio_ddir idx = io_u->ddir;
const enum fio_ddir odx = io_u->ddir ^ 1;
int ret;
td->io_blocks[idx]++;
td->this_io_blocks[idx]++;
td->io_bytes[idx] += bytes;
if (!(io_u->flags & IO_U_F_VER_LIST))
td->this_io_bytes[idx] += bytes;
if (idx == DDIR_WRITE) {
f = io_u->file;
if (f) {
if (f->first_write == -1ULL ||
io_u->offset < f->first_write)
f->first_write = io_u->offset;
if (f->last_write == -1ULL ||
((io_u->offset + bytes) > f->last_write))
f->last_write = io_u->offset + bytes;
}
}
if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
td->runstate == TD_VERIFYING)) {
account_io_completion(td, io_u, icd, idx, bytes);
if (__should_check_rate(td, idx)) {
td->rate_pending_usleep[idx] =
(usec_for_io(td, idx) -
utime_since_now(&td->start));
}
if (__should_check_latency(td, idx)) {
unsigned long lusec = utime_since(
&io_u->issue_time, &icd->time);
/* Linear increase and logarithmic decrease */
if (lusec > td->o.shed_latency[idx]) {
if (td->shed_count[idx] < MAX_SHED_COUNT ) {
td->shed_count[idx] += (1<<SHED_FRAC_BITS);
}
}
else if (td->shed_count[idx]) {
td->shed_count[idx] -= get_used_bits(td->shed_count[idx]);
}
if (td->shed_count[idx]) {
lusec = (lusec * td->shed_count[idx]) >> SHED_FRAC_BITS;
if (lusec > td->rate_pending_usleep[idx]) {
td->rate_pending_usleep[idx] = lusec;
}
}
}
