* IO engine using the Linux native aio interface.
*
*/
-#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <errno.h>
-#include <assert.h>
#include <libaio.h>
+#include <sys/time.h>
+#include <sys/resource.h>
#include "../fio.h"
+#include "../lib/pow2.h"
+#include "../optgroup.h"
+#include "../lib/memalign.h"
+
+#ifndef IOCB_FLAG_HIPRI
+#define IOCB_FLAG_HIPRI (1 << 2)
+#endif
+
+#ifndef IOCTX_FLAG_IOPOLL
+#define IOCTX_FLAG_IOPOLL (1 << 0)
+#endif
+
+static int fio_libaio_commit(struct thread_data *td);
struct libaio_data {
io_context_t aio_ctx;
struct iocb **iocbs;
struct io_u **io_us;
+ struct io_u **io_u_index;
+
/*
* Basic ring buffer. 'head' is incremented in _queue(), and
* 'tail' is incremented in _commit(). We keep 'queued' so
};
struct libaio_options {
- struct thread_data *td;
+ void *pad;
unsigned int userspace_reap;
+ unsigned int hipri;
};
static struct fio_option options[] = {
.category = FIO_OPT_C_ENGINE,
.group = FIO_OPT_G_LIBAIO,
},
+ {
+ .name = "hipri",
+ .lname = "High Priority",
+ .type = FIO_OPT_STR_SET,
+ .off1 = offsetof(struct libaio_options, hipri),
+ .help = "Use polled IO completions",
+ .category = FIO_OPT_C_ENGINE,
+ .group = FIO_OPT_G_LIBAIO,
+ },
{
.name = NULL,
},
static int fio_libaio_prep(struct thread_data fio_unused *td, struct io_u *io_u)
{
struct fio_file *f = io_u->file;
+ struct libaio_options *o = td->eo;
+ struct iocb *iocb;
+
+ iocb = &io_u->iocb;
- if (io_u->ddir == DDIR_READ)
- 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 (ddir_sync(io_u->ddir))
- io_prep_fsync(&io_u->iocb, f->fd);
+ if (io_u->ddir == DDIR_READ) {
+ io_prep_pread(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
+ if (o->hipri)
+ iocb->u.c.flags |= IOCB_FLAG_HIPRI;
+ } else if (io_u->ddir == DDIR_WRITE) {
+ io_prep_pwrite(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
+ if (o->hipri)
+ iocb->u.c.flags |= IOCB_FLAG_HIPRI;
+ } else if (ddir_sync(io_u->ddir))
+ io_prep_fsync(iocb, f->fd);
return 0;
}
static struct io_u *fio_libaio_event(struct thread_data *td, int event)
{
- struct libaio_data *ld = td->io_ops->data;
+ struct libaio_data *ld = td->io_ops_data;
struct io_event *ev;
struct io_u *io_u;
static int fio_libaio_getevents(struct thread_data *td, unsigned int min,
unsigned int max, const struct timespec *t)
{
- struct libaio_data *ld = td->io_ops->data;
+ struct libaio_data *ld = td->io_ops_data;
struct libaio_options *o = td->eo;
- unsigned actual_min = td->o.iodepth_batch_complete == 0 ? 0 : min;
+ unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
struct timespec __lt, *lt = NULL;
int r, events = 0;
r = io_getevents(ld->aio_ctx, actual_min,
max, ld->aio_events + events, lt);
}
- if (r >= 0)
+ if (r > 0)
events += r;
- else if (r == -EAGAIN)
- usleep(100);
+ else if ((min && r == 0) || r == -EAGAIN) {
+ fio_libaio_commit(td);
+ if (actual_min)
+ usleep(10);
+ } else if (r != -EINTR)
+ break;
} while (events < min);
return r < 0 ? r : events;
}
-static int fio_libaio_queue(struct thread_data *td, struct io_u *io_u)
+static enum fio_q_status fio_libaio_queue(struct thread_data *td,
+ struct io_u *io_u)
{
- struct libaio_data *ld = td->io_ops->data;
+ struct libaio_data *ld = td->io_ops_data;
fio_ro_check(td, io_u);
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;
}
static void fio_libaio_queued(struct thread_data *td, struct io_u **io_us,
unsigned int nr)
{
- struct timeval now;
+ struct timespec now;
unsigned int i;
if (!fio_fill_issue_time(td))
static int fio_libaio_commit(struct thread_data *td)
{
- struct libaio_data *ld = td->io_ops->data;
+ struct libaio_data *ld = td->io_ops_data;
struct iocb **iocbs;
struct io_u **io_us;
- struct timeval tv;
+ struct timespec ts;
int ret, wait_start = 0;
if (!ld->queued)
ld->queued -= ret;
ring_inc(ld, &ld->tail, ret);
ret = 0;
+ wait_start = 0;
} else if (ret == -EINTR || !ret) {
if (!ret)
io_u_mark_submit(td, ret);
+ wait_start = 0;
continue;
} else if (ret == -EAGAIN) {
/*
break;
}
if (!wait_start) {
- fio_gettime(&tv, NULL);
- wait_start = 0;
- } else if (mtime_since_now(&tv) > 30000) {
+ fio_gettime(&ts, NULL);
+ wait_start = 1;
+ } else if (mtime_since_now(&ts) > 30000) {
log_err("fio: aio appears to be stalled, giving up\n");
break;
}
usleep(1);
continue;
+ } else if (ret == -ENOMEM) {
+ /*
+ * If we get -ENOMEM, reap events if we can. If
+ * we cannot, treat it as a fatal event since there's
+ * nothing we can do about it.
+ */
+ if (ld->queued)
+ ret = 0;
+ break;
} else
break;
} while (ld->queued);
static int fio_libaio_cancel(struct thread_data *td, struct io_u *io_u)
{
- struct libaio_data *ld = td->io_ops->data;
+ struct libaio_data *ld = td->io_ops_data;
return io_cancel(ld->aio_ctx, &io_u->iocb, ld->aio_events);
}
static void fio_libaio_cleanup(struct thread_data *td)
{
- struct libaio_data *ld = td->io_ops->data;
+ struct libaio_data *ld = td->io_ops_data;
if (ld) {
- io_destroy(ld->aio_ctx);
+ /*
+ * Work-around to avoid huge RCU stalls at exit time. If we
+ * don't do this here, then it'll be torn down by exit_aio().
+ * But for that case we can parallellize the freeing, thus
+ * speeding it up a lot.
+ */
+ if (!(td->flags & TD_F_CHILD))
+ io_destroy(ld->aio_ctx);
free(ld->aio_events);
free(ld->iocbs);
free(ld->io_us);
}
}
-static int fio_libaio_init(struct thread_data *td)
+static int fio_libaio_old_queue_init(struct libaio_data *ld, unsigned int depth,
+ bool hipri)
+{
+ if (hipri) {
+ log_err("fio: polled aio not available on your platform\n");
+ return 1;
+ }
+
+ return io_queue_init(depth, &ld->aio_ctx);
+}
+
+static int fio_libaio_queue_init(struct libaio_data *ld, unsigned int depth,
+ bool hipri)
{
+#ifdef __NR_sys_io_setup2
+ int ret, flags = 0;
+
+ if (hipri)
+ flags |= IOCTX_FLAG_IOPOLL;
+
+ ret = syscall(__NR_sys_io_setup2, depth, flags, NULL, NULL,
+ &ld->aio_ctx);
+ if (!ret)
+ return 0;
+ /* fall through to old syscall */
+#endif
+ return fio_libaio_old_queue_init(ld, depth, hipri);
+}
+
+static int fio_libaio_post_init(struct thread_data *td)
+{
+ struct libaio_data *ld = td->io_ops_data;
struct libaio_options *o = td->eo;
- struct libaio_data *ld;
int err = 0;
- ld = calloc(1, sizeof(*ld));
-
- /*
- * First try passing in 0 for queue depth, since we don't
- * care about the user ring. If that fails, the kernel is too old
- * and we need the right depth.
- */
- if (!o->userspace_reap)
- err = io_queue_init(INT_MAX, &ld->aio_ctx);
- if (o->userspace_reap || err == -EINVAL)
- err = io_queue_init(td->o.iodepth, &ld->aio_ctx);
+ err = fio_libaio_queue_init(ld, td->o.iodepth, o->hipri);
if (err) {
td_verror(td, -err, "io_queue_init");
- log_err("fio: check /proc/sys/fs/aio-max-nr\n");
- free(ld);
return 1;
}
+ return 0;
+}
+
+static int fio_libaio_init(struct thread_data *td)
+{
+ struct libaio_data *ld;
+
+ ld = calloc(1, sizeof(*ld));
+
ld->entries = td->o.iodepth;
ld->is_pow2 = is_power_of_2(ld->entries);
ld->aio_events = calloc(ld->entries, sizeof(struct io_event));
ld->iocbs = calloc(ld->entries, sizeof(struct iocb *));
ld->io_us = calloc(ld->entries, sizeof(struct io_u *));
- td->io_ops->data = ld;
+ td->io_ops_data = ld;
return 0;
}
.name = "libaio",
.version = FIO_IOOPS_VERSION,
.init = fio_libaio_init,
+ .post_init = fio_libaio_post_init,
.prep = fio_libaio_prep,
.queue = fio_libaio_queue,
.commit = fio_libaio_commit,