static int fio_syncio_queue(struct thread_data *td, struct io_u *io_u)
{
struct fio_file *f = io_u->file;
int ret;
fio_ro_check(td, io_u);
if (io_u->ddir == DDIR_READ)
ret = read(f->fd, io_u->xfer_buf, io_u->xfer_buflen);
else if (io_u->ddir == DDIR_WRITE)
ret = write(f->fd, io_u->xfer_buf, io_u->xfer_buflen);
else if (io_u->ddir == DDIR_TRIM) {
do_io_u_trim(td, io_u);
return FIO_Q_COMPLETED;
} else
ret = do_io_u_sync(td, io_u);
return fio_io_end(td, io_u, ret);
}
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;
}
static enum fio_q_status fio_ioring_queue(struct thread_data *td,
struct io_u *io_u)
{
struct ioring_data *ld = td->io_ops_data;
struct io_sq_ring *ring = &ld->sq_ring;
unsigned tail, next_tail;
fio_ro_check(td, io_u);
if (ld->queued == ld->iodepth)
return FIO_Q_BUSY;
if (io_u->ddir == DDIR_TRIM) {
if (ld->queued)
return FIO_Q_BUSY;
do_io_u_trim(td, io_u);
io_u_mark_submit(td, 1);
io_u_mark_complete(td, 1);
return FIO_Q_COMPLETED;
}
tail = *ring->tail;
next_tail = tail + 1;
read_barrier();
if (next_tail == *ring->head)
return FIO_Q_BUSY;
/* ensure sqe stores are ordered with tail update */
write_barrier();
ring->array[tail & ld->sq_ring_mask] = io_u->index;
*ring->tail = next_tail;
write_barrier();
ld->queued++;
return FIO_Q_QUEUED;
}
static int fio_pvsyncio_queue(struct thread_data *td, struct io_u *io_u)
{
struct syncio_data *sd = td->io_ops->data;
struct iovec *iov = &sd->iovecs[0];
struct fio_file *f = io_u->file;
int ret;
fio_ro_check(td, io_u);
iov->iov_base = io_u->xfer_buf;
iov->iov_len = io_u->xfer_buflen;
if (io_u->ddir == DDIR_READ)
ret = preadv(f->fd, iov, 1, io_u->offset);
else if (io_u->ddir == DDIR_WRITE)
ret = pwritev(f->fd, iov, 1, io_u->offset);
else if (io_u->ddir == DDIR_TRIM) {
do_io_u_trim(td, io_u);
return FIO_Q_COMPLETED;
} else
ret = do_io_u_sync(td, io_u);
return fio_io_end(td, io_u, ret);
}
/*
* For splice writing, we can vmsplice our data buffer directly into a
* pipe and then splice that to a file.
*/
static int fio_splice_write(struct thread_data *td, struct io_u *io_u)
{
struct spliceio_data *sd = td->io_ops->data;
struct iovec iov = {
.iov_base = io_u->xfer_buf,
.iov_len = io_u->xfer_buflen,
};
struct pollfd pfd = { .fd = sd->pipe[1], .events = POLLOUT, };
struct fio_file *f = io_u->file;
off_t off = io_u->offset;
int ret, ret2;
while (iov.iov_len) {
if (poll(&pfd, 1, -1) < 0)
return errno;
ret = vmsplice(sd->pipe[1], &iov, 1, SPLICE_F_NONBLOCK);
if (ret < 0)
return -errno;
iov.iov_len -= ret;
iov.iov_base += ret;
while (ret) {
ret2 = splice(sd->pipe[0], NULL, f->fd, &off, ret, 0);
if (ret2 < 0)
return -errno;
ret -= ret2;
}
}
return io_u->xfer_buflen;
}
static int fio_spliceio_queue(struct thread_data *td, struct io_u *io_u)
{
struct spliceio_data *sd = td->io_ops->data;
int ret = 0;
fio_ro_check(td, io_u);
if (io_u->ddir == DDIR_READ) {
if (sd->vmsplice_to_user) {
ret = fio_splice_read(td, io_u);
/*
* This kernel doesn't support vmsplice to user
* space. Reset the vmsplice_to_user flag, so that
* we retry below and don't hit this path again.
*/
if (ret == -EBADF)
sd->vmsplice_to_user = 0;
}
if (!sd->vmsplice_to_user)
ret = fio_splice_read_old(td, io_u);
} else if (io_u->ddir == DDIR_WRITE)
ret = fio_splice_write(td, io_u);
else if (io_u->ddir == DDIR_TRIM)
ret = do_io_u_trim(td, io_u);
else
ret = do_io_u_sync(td, io_u);
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) {
td_verror(td, io_u->error, "xfer");
if (io_u->error == EINVAL)
log_err("fio: looks like splice doesn't work on this"
" file system\n");
}
return FIO_Q_COMPLETED;
}
static void fio_spliceio_cleanup(struct thread_data *td)
{
struct spliceio_data *sd = td->io_ops->data;
if (sd) {
close(sd->pipe[0]);
close(sd->pipe[1]);
free(sd);
}
}
static int fio_spliceio_init(struct thread_data *td)
{
struct spliceio_data *sd = malloc(sizeof(*sd));
if (pipe(sd->pipe) < 0) {
td_verror(td, errno, "pipe");
free(sd);
return 1;
}
/*
* Assume this work, we'll reset this if it doesn't
*/
sd->vmsplice_to_user = 1;
/*
* Works with "real" vmsplice to user, eg mapping pages directly.
* Reset if we fail.
*/
sd->vmsplice_to_user_map = 1;
/*
* And if vmsplice_to_user works, we definitely need aligned
* buffers. Just set ->odirect to force that.
*/
if (td_read(td))
td->o.mem_align = 1;
td->io_ops->data = sd;
return 0;
}
static struct ioengine_ops ioengine = {
.name = "splice",
.version = FIO_IOOPS_VERSION,
.init = fio_spliceio_init,
.queue = fio_spliceio_queue,
.cleanup = fio_spliceio_cleanup,
.open_file = generic_open_file,
.close_file = generic_close_file,
.get_file_size = generic_get_file_size,
.flags = FIO_SYNCIO | FIO_PIPEIO,
};
static void fio_init fio_spliceio_register(void)
{
register_ioengine(&ioengine);
}
static void fio_exit fio_spliceio_unregister(void)
{
unregister_ioengine(&ioengine);
}
