/*
- * native linux aio io engine
+ * libaio engine
+ *
+ * IO engine using the Linux native aio interface.
*
*/
#include <stdio.h>
#include <assert.h>
#include "../fio.h"
-#include "../os.h"
#ifdef FIO_HAVE_LIBAIO
struct libaio_data {
io_context_t aio_ctx;
struct io_event *aio_events;
+ struct iocb **iocbs;
+ struct io_u **io_us;
+ int iocbs_nr;
};
static int fio_libaio_prep(struct thread_data fio_unused *td, struct io_u *io_u)
io_prep_pread(&io_u->iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
else if (io_u->ddir == DDIR_WRITE)
io_prep_pwrite(&io_u->iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
- else if (io_u->ddir == DDIR_SYNC)
+ else if (ddir_sync(io_u->ddir))
io_prep_fsync(&io_u->iocb, f->fd);
else
return 1;
static struct io_u *fio_libaio_event(struct thread_data *td, int event)
{
struct libaio_data *ld = td->io_ops->data;
+ struct io_event *ev;
+ struct io_u *io_u;
+
+ ev = ld->aio_events + event;
+ io_u = ev_to_iou(ev);
- return ev_to_iou(ld->aio_events + event);
+ if (ev->res != io_u->xfer_buflen) {
+ if (ev->res > io_u->xfer_buflen)
+ io_u->error = -ev->res;
+ else
+ io_u->resid = io_u->xfer_buflen - ev->res;
+ } else
+ io_u->error = 0;
+
+ return io_u;
}
-static int fio_libaio_getevents(struct thread_data *td, int min, int max,
- struct timespec *t)
+static int fio_libaio_getevents(struct thread_data *td, unsigned int min,
+ unsigned int max, struct timespec *t)
{
struct libaio_data *ld = td->io_ops->data;
- long r;
+ int r;
do {
r = io_getevents(ld->aio_ctx, min, max, ld->aio_events, t);
- if (r >= min)
+ if (r >= (int) min)
break;
else if (r == -EAGAIN) {
usleep(100);
break;
} while (1);
- if (r < 0)
- r = -r;
-
- return (int) r;
+ return r;
}
static int fio_libaio_queue(struct thread_data *td, struct io_u *io_u)
{
struct libaio_data *ld = td->io_ops->data;
- struct iocb *iocb = &io_u->iocb;
- long ret;
+ fio_ro_check(td, io_u);
+
+ if (ld->iocbs_nr == (int) td->o.iodepth)
+ return FIO_Q_BUSY;
+
+ /*
+ * fsync is tricky, since it can fail and we need to do it
+ * serialized with other io. the reason is that linux doesn't
+ * support aio fsync yet. So return busy for the case where we
+ * have pending io, to let fio complete those first.
+ */
+ if (ddir_sync(io_u->ddir)) {
+ if (ld->iocbs_nr)
+ return FIO_Q_BUSY;
+
+ do_io_u_sync(td, io_u);
+ return FIO_Q_COMPLETED;
+ }
+
+ ld->iocbs[ld->iocbs_nr] = &io_u->iocb;
+ ld->io_us[ld->iocbs_nr] = io_u;
+ ld->iocbs_nr++;
+ return FIO_Q_QUEUED;
+}
+
+static void fio_libaio_queued(struct thread_data *td, struct io_u **io_us,
+ unsigned int nr)
+{
+ struct timeval now;
+ unsigned int i;
+
+ if (!fio_fill_issue_time(td))
+ return;
+
+ fio_gettime(&now, NULL);
+
+ for (i = 0; i < nr; i++) {
+ struct io_u *io_u = io_us[i];
+
+ memcpy(&io_u->issue_time, &now, sizeof(now));
+ io_u_queued(td, io_u);
+ }
+}
+
+static int fio_libaio_commit(struct thread_data *td)
+{
+ struct libaio_data *ld = td->io_ops->data;
+ struct iocb **iocbs;
+ struct io_u **io_us;
+ int ret;
+
+ if (!ld->iocbs_nr)
+ return 0;
+
+ io_us = ld->io_us;
+ iocbs = ld->iocbs;
do {
- ret = io_submit(ld->aio_ctx, 1, &iocb);
- if (ret == 1)
- return 0;
- else if (ret == -EAGAIN || !ret)
- usleep(100);
- else if (ret == -EINTR)
+ ret = io_submit(ld->aio_ctx, ld->iocbs_nr, iocbs);
+ if (ret > 0) {
+ fio_libaio_queued(td, io_us, ret);
+ io_u_mark_submit(td, ret);
+ ld->iocbs_nr -= ret;
+ io_us += ret;
+ iocbs += ret;
+ ret = 0;
+ } else if (!ret || ret == -EAGAIN || ret == -EINTR) {
+ if (!ret)
+ io_u_mark_submit(td, ret);
continue;
- else
+ } else
break;
- } while (1);
-
- if (ret <= 0) {
- io_u->resid = io_u->xfer_buflen;
- io_u->error = -ret;
- return 1;
- }
+ } while (ld->iocbs_nr);
- return 0;
+ return ret;
}
static int fio_libaio_cancel(struct thread_data *td, struct io_u *io_u)
if (ld) {
io_destroy(ld->aio_ctx);
- if (ld->aio_events)
- free(ld->aio_events);
-
+ free(ld->aio_events);
+ free(ld->iocbs);
+ free(ld->io_us);
free(ld);
- td->io_ops->data = NULL;
}
}
static int fio_libaio_init(struct thread_data *td)
{
struct libaio_data *ld = malloc(sizeof(*ld));
+ int err;
memset(ld, 0, sizeof(*ld));
- if (io_queue_init(td->iodepth, &ld->aio_ctx)) {
- td_verror(td, errno);
+
+ err = io_queue_init(td->o.iodepth, &ld->aio_ctx);
+ if (err) {
+ td_verror(td, -err, "io_queue_init");
+ log_err("fio: check /proc/sys/fs/aio-max-nr\n");
free(ld);
return 1;
}
- ld->aio_events = malloc(td->iodepth * sizeof(struct io_event));
- memset(ld->aio_events, 0, td->iodepth * sizeof(struct io_event));
+ ld->aio_events = malloc(td->o.iodepth * sizeof(struct io_event));
+ memset(ld->aio_events, 0, td->o.iodepth * sizeof(struct io_event));
+ ld->iocbs = malloc(td->o.iodepth * sizeof(struct iocb *));
+ memset(ld->iocbs, 0, sizeof(struct iocb *));
+ ld->io_us = malloc(td->o.iodepth * sizeof(struct io_u *));
+ memset(ld->io_us, 0, td->o.iodepth * sizeof(struct io_u *));
+ ld->iocbs_nr = 0;
+
td->io_ops->data = ld;
return 0;
}
.init = fio_libaio_init,
.prep = fio_libaio_prep,
.queue = fio_libaio_queue,
+ .commit = fio_libaio_commit,
.cancel = fio_libaio_cancel,
.getevents = fio_libaio_getevents,
.event = fio_libaio_event,
.cleanup = fio_libaio_cleanup,
+ .open_file = generic_open_file,
+ .close_file = generic_close_file,
+ .get_file_size = generic_get_file_size,
};
#else /* FIO_HAVE_LIBAIO */