[fio.git] / engines / libaio.c

/*
 * libaio engine
 *
 * IO engine using the Linux native aio interface.
 *
 */
#include <stdlib.h>
#include <unistd.h>
#include <errno.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"

/* Should be defined in newest aio_abi.h */
#ifndef IOCB_FLAG_IOPRIO
#define IOCB_FLAG_IOPRIO    (1 << 1)
#endif

/* Hack for libaio < 0.3.111 */
#ifndef CONFIG_LIBAIO_RW_FLAGS
#define aio_rw_flags __pad2
#endif

static int fio_libaio_commit(struct thread_data *td);
static int fio_libaio_init(struct thread_data *td);

struct libaio_data {
	io_context_t aio_ctx;
	struct io_event *aio_events;
	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
	 * that we know if the ring is full or empty, when
	 * 'head' == 'tail'. 'entries' is the ring size, and
	 * 'is_pow2' is just an optimization to use AND instead of
	 * modulus to get the remainder on ring increment.
	 */
	int is_pow2;
	unsigned int entries;
	unsigned int queued;
	unsigned int head;
	unsigned int tail;
};

struct libaio_options {
	void *pad;
	unsigned int userspace_reap;
	unsigned int cmdprio_percentage;
	unsigned int nowait;
};

static struct fio_option options[] = {
	{
		.name	= "userspace_reap",
		.lname	= "Libaio userspace reaping",
		.type	= FIO_OPT_STR_SET,
		.off1	= offsetof(struct libaio_options, userspace_reap),
		.help	= "Use alternative user-space reap implementation",
		.category = FIO_OPT_C_ENGINE,
		.group	= FIO_OPT_G_LIBAIO,
	},
#ifdef FIO_HAVE_IOPRIO_CLASS
	{
		.name	= "cmdprio_percentage",
		.lname	= "high priority percentage",
		.type	= FIO_OPT_INT,
		.off1	= offsetof(struct libaio_options, cmdprio_percentage),
		.minval	= 1,
		.maxval	= 100,
		.help	= "Send high priority I/O this percentage of the time",
		.category = FIO_OPT_C_ENGINE,
		.group	= FIO_OPT_G_LIBAIO,
	},
#else
	{
		.name	= "cmdprio_percentage",
		.lname	= "high priority percentage",
		.type	= FIO_OPT_UNSUPPORTED,
		.help	= "Your platform does not support I/O priority classes",
	},
#endif
	{
		.name	= "nowait",
		.lname	= "RWF_NOWAIT",
		.type	= FIO_OPT_BOOL,
		.off1	= offsetof(struct libaio_options, nowait),
		.help	= "Set RWF_NOWAIT for reads/writes",
		.category = FIO_OPT_C_ENGINE,
		.group	= FIO_OPT_G_LIBAIO,
	},
	{
		.name	= NULL,
	},
};

static inline void ring_inc(struct libaio_data *ld, unsigned int *val,
			    unsigned int add)
{
	if (ld->is_pow2)
		*val = (*val + add) & (ld->entries - 1);
	else
		*val = (*val + add) % ld->entries;
}

static int fio_libaio_prep(struct thread_data *td, struct io_u *io_u)
{
	struct libaio_options *o = td->eo;
	struct fio_file *f = io_u->file;
	struct iocb *iocb = &io_u->iocb;

	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->nowait)
			iocb->aio_rw_flags |= RWF_NOWAIT;
	} 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->nowait)
			iocb->aio_rw_flags |= RWF_NOWAIT;
	} else if (ddir_sync(io_u->ddir))
		io_prep_fsync(iocb, f->fd);

	return 0;
}

static void fio_libaio_prio_prep(struct thread_data *td, struct io_u *io_u)
{
	struct libaio_options *o = td->eo;
	if (rand_between(&td->prio_state, 0, 99) < o->cmdprio_percentage) {
		io_u->iocb.aio_reqprio = IOPRIO_CLASS_RT << IOPRIO_CLASS_SHIFT;
		io_u->iocb.u.c.flags |= IOCB_FLAG_IOPRIO;
		io_u->flags |= IO_U_F_PRIORITY;
	}
	return;
}

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 = container_of(ev->obj, struct io_u, iocb);

	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;
}

struct aio_ring {
	unsigned id;		 /** kernel internal index number */
	unsigned nr;		 /** number of io_events */
	unsigned head;
	unsigned tail;

	unsigned magic;
	unsigned compat_features;
	unsigned incompat_features;
	unsigned header_length;	/** size of aio_ring */

	struct io_event events[0];
};

#define AIO_RING_MAGIC	0xa10a10a1

static int user_io_getevents(io_context_t aio_ctx, unsigned int max,
			     struct io_event *events)
{
	long i = 0;
	unsigned head;
	struct aio_ring *ring = (struct aio_ring*) aio_ctx;

	while (i < max) {
		head = ring->head;

		if (head == ring->tail) {
			/* There are no more completions */
			break;
		} else {
			/* There is another completion to reap */
			events[i] = ring->events[head];
			atomic_store_release(&ring->head,
					     (head + 1) % ring->nr);
			i++;
		}
	}

	return i;
}

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_options *o = td->eo;
	unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
	struct timespec __lt, *lt = NULL;
	int r, events = 0;

	if (t) {
		__lt = *t;
		lt = &__lt;
	}

	do {
		if (o->userspace_reap == 1
		    && actual_min == 0
		    && ((struct aio_ring *)(ld->aio_ctx))->magic
				== AIO_RING_MAGIC) {
			r = user_io_getevents(ld->aio_ctx, max,
				ld->aio_events + events);
		} else {
			r = io_getevents(ld->aio_ctx, actual_min,
				max, ld->aio_events + events, lt);
		}
		if (r > 0)
			events += r;
		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 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_options *o = td->eo;

	fio_ro_check(td, io_u);

	if (ld->queued == 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->queued)
			return FIO_Q_BUSY;

		do_io_u_sync(td, io_u);
		return FIO_Q_COMPLETED;
	}

	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;
	}

	if (o->cmdprio_percentage)
		fio_libaio_prio_prep(td, io_u);

	ld->iocbs[ld->head] = &io_u->iocb;
	ld->io_us[ld->head] = io_u;
	ring_inc(ld, &ld->head, 1);
	ld->queued++;
	return FIO_Q_QUEUED;
}

static void fio_libaio_queued(struct thread_data *td, struct io_u **io_us,
			      unsigned int nr)
{
	struct timespec 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;
	struct timespec ts;
	int ret, wait_start = 0;

	if (!ld->queued)
		return 0;

	do {
		long nr = ld->queued;

		nr = min((unsigned int) nr, ld->entries - ld->tail);
		io_us = ld->io_us + ld->tail;
		iocbs = ld->iocbs + ld->tail;

		ret = io_submit(ld->aio_ctx, nr, iocbs);
		if (ret > 0) {
			fio_libaio_queued(td, io_us, ret);
			io_u_mark_submit(td, ret);

			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) {
			/*
			 * If we get EAGAIN, we should break out without
			 * error and let the upper layer reap some
			 * events for us. If we have no queued IO, we
			 * must loop here. If we loop for more than 30s,
			 * just error out, something must be buggy in the
			 * IO path.
			 */
			if (ld->queued) {
				ret = 0;
				break;
			}
			if (!wait_start) {
				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);

	return ret;
}

static int fio_libaio_cancel(struct thread_data *td, struct io_u *io_u)
{
	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;

	if (ld) {
		/*
		 * 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);
		free(ld);
	}
}

static int fio_libaio_post_init(struct thread_data *td)
{
	struct libaio_data *ld = td->io_ops_data;
	int err;

	err = io_queue_init(td->o.iodepth, &ld->aio_ctx);
	if (err) {
		td_verror(td, -err, "io_queue_init");
		return 1;
	}

	return 0;
}

static int fio_libaio_init(struct thread_data *td)
{
	struct libaio_data *ld;
	struct thread_options *to = &td->o;
	struct libaio_options *o = td->eo;

	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;
	/*
	 * Check for option conflicts
	 */
	if ((fio_option_is_set(to, ioprio) || fio_option_is_set(to, ioprio_class)) &&
			o->cmdprio_percentage != 0) {
		log_err("%s: cmdprio_percentage option and mutually exclusive "
				"prio or prioclass option is set, exiting\n", to->name);
		td_verror(td, EINVAL, "fio_libaio_init");
		return 1;
	}
	return 0;
}

FIO_STATIC struct ioengine_ops ioengine = {
	.name			= "libaio",
	.version		= FIO_IOOPS_VERSION,
	.flags			= FIO_ASYNCIO_SYNC_TRIM,
	.init			= fio_libaio_init,
	.post_init		= fio_libaio_post_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,
	.options		= options,
	.option_struct_size	= sizeof(struct libaio_options),
};

static void fio_init fio_libaio_register(void)
{
	register_ioengine(&ioengine);
}

static void fio_exit fio_libaio_unregister(void)
{
	unregister_ioengine(&ioengine);
}
Commit	Line	Data
	1	/*
	2	* libaio engine
	3	*
	4	* IO engine using the Linux native aio interface.
	5	*
	6	*/
	7	#include <stdlib.h>
	8	#include <unistd.h>
	9	#include <errno.h>
	10	#include <libaio.h>
	11	#include <sys/time.h>
	12	#include <sys/resource.h>
	13
	14	#include "../fio.h"
	15	#include "../lib/pow2.h"
	16	#include "../optgroup.h"
	17	#include "../lib/memalign.h"
	18
	19	/* Should be defined in newest aio_abi.h */
	20	#ifndef IOCB_FLAG_IOPRIO
	21	#define IOCB_FLAG_IOPRIO (1 << 1)
	22	#endif
	23
	24	/* Hack for libaio < 0.3.111 */
	25	#ifndef CONFIG_LIBAIO_RW_FLAGS
	26	#define aio_rw_flags __pad2
	27	#endif
	28
	29	static int fio_libaio_commit(struct thread_data *td);
	30	static int fio_libaio_init(struct thread_data *td);
	31
	32	struct libaio_data {
	33	io_context_t aio_ctx;
	34	struct io_event *aio_events;
	35	struct iocb **iocbs;
	36	struct io_u **io_us;
	37
	38	struct io_u **io_u_index;
	39
	40	/*
	41	* Basic ring buffer. 'head' is incremented in _queue(), and
	42	* 'tail' is incremented in _commit(). We keep 'queued' so
	43	* that we know if the ring is full or empty, when
	44	* 'head' == 'tail'. 'entries' is the ring size, and
	45	* 'is_pow2' is just an optimization to use AND instead of
	46	* modulus to get the remainder on ring increment.
	47	*/
	48	int is_pow2;
	49	unsigned int entries;
	50	unsigned int queued;
	51	unsigned int head;
	52	unsigned int tail;
	53	};
	54
	55	struct libaio_options {
	56	void *pad;
	57	unsigned int userspace_reap;
	58	unsigned int cmdprio_percentage;
	59	unsigned int nowait;
	60	};
	61
	62	static struct fio_option options[] = {
	63	{
	64	.name = "userspace_reap",
	65	.lname = "Libaio userspace reaping",
	66	.type = FIO_OPT_STR_SET,
	67	.off1 = offsetof(struct libaio_options, userspace_reap),
	68	.help = "Use alternative user-space reap implementation",
	69	.category = FIO_OPT_C_ENGINE,
	70	.group = FIO_OPT_G_LIBAIO,
	71	},
	72	#ifdef FIO_HAVE_IOPRIO_CLASS
	73	{
	74	.name = "cmdprio_percentage",
	75	.lname = "high priority percentage",
	76	.type = FIO_OPT_INT,
	77	.off1 = offsetof(struct libaio_options, cmdprio_percentage),
	78	.minval = 1,
	79	.maxval = 100,
	80	.help = "Send high priority I/O this percentage of the time",
	81	.category = FIO_OPT_C_ENGINE,
	82	.group = FIO_OPT_G_LIBAIO,
	83	},
	84	#else
	85	{
	86	.name = "cmdprio_percentage",
	87	.lname = "high priority percentage",
	88	.type = FIO_OPT_UNSUPPORTED,
	89	.help = "Your platform does not support I/O priority classes",
	90	},
	91	#endif
	92	{
	93	.name = "nowait",
	94	.lname = "RWF_NOWAIT",
	95	.type = FIO_OPT_BOOL,
	96	.off1 = offsetof(struct libaio_options, nowait),
	97	.help = "Set RWF_NOWAIT for reads/writes",
	98	.category = FIO_OPT_C_ENGINE,
	99	.group = FIO_OPT_G_LIBAIO,
	100	},
	101	{
	102	.name = NULL,
	103	},
	104	};
	105
	106	static inline void ring_inc(struct libaio_data ld, unsigned int val,
	107	unsigned int add)
	108	{
	109	if (ld->is_pow2)
	110	val = (val + add) & (ld->entries - 1);
	111	else
	112	val = (val + add) % ld->entries;
	113	}
	114
	115	static int fio_libaio_prep(struct thread_data td, struct io_u io_u)
	116	{
	117	struct libaio_options *o = td->eo;
	118	struct fio_file *f = io_u->file;
	119	struct iocb *iocb = &io_u->iocb;
	120
	121	if (io_u->ddir == DDIR_READ) {
	122	io_prep_pread(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
	123	if (o->nowait)
	124	iocb->aio_rw_flags \|= RWF_NOWAIT;
	125	} else if (io_u->ddir == DDIR_WRITE) {
	126	io_prep_pwrite(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
	127	if (o->nowait)
	128	iocb->aio_rw_flags \|= RWF_NOWAIT;
	129	} else if (ddir_sync(io_u->ddir))
	130	io_prep_fsync(iocb, f->fd);
	131
	132	return 0;
	133	}
	134
	135	static void fio_libaio_prio_prep(struct thread_data td, struct io_u io_u)
	136	{
	137	struct libaio_options *o = td->eo;
	138	if (rand_between(&td->prio_state, 0, 99) < o->cmdprio_percentage) {
	139	io_u->iocb.aio_reqprio = IOPRIO_CLASS_RT << IOPRIO_CLASS_SHIFT;
	140	io_u->iocb.u.c.flags \|= IOCB_FLAG_IOPRIO;
	141	io_u->flags \|= IO_U_F_PRIORITY;
	142	}
	143	return;
	144	}
	145
	146	static struct io_u fio_libaio_event(struct thread_data td, int event)
	147	{
	148	struct libaio_data *ld = td->io_ops_data;
	149	struct io_event *ev;
	150	struct io_u *io_u;
	151
	152	ev = ld->aio_events + event;
	153	io_u = container_of(ev->obj, struct io_u, iocb);
	154
	155	if (ev->res != io_u->xfer_buflen) {
	156	if (ev->res > io_u->xfer_buflen)
	157	io_u->error = -ev->res;
	158	else
	159	io_u->resid = io_u->xfer_buflen - ev->res;
	160	} else
	161	io_u->error = 0;
	162
	163	return io_u;
	164	}
	165
	166	struct aio_ring {
	167	unsigned id; /** kernel internal index number */
	168	unsigned nr; /** number of io_events */
	169	unsigned head;
	170	unsigned tail;
	171
	172	unsigned magic;
	173	unsigned compat_features;
	174	unsigned incompat_features;
	175	unsigned header_length; /** size of aio_ring */
	176
	177	struct io_event events[0];
	178	};
	179
	180	#define AIO_RING_MAGIC 0xa10a10a1
	181
	182	static int user_io_getevents(io_context_t aio_ctx, unsigned int max,
	183	struct io_event *events)
	184	{
	185	long i = 0;
	186	unsigned head;
	187	struct aio_ring ring = (struct aio_ring) aio_ctx;
	188
	189	while (i < max) {
	190	head = ring->head;
	191
	192	if (head == ring->tail) {
	193	/* There are no more completions */
	194	break;
	195	} else {
	196	/* There is another completion to reap */
	197	events[i] = ring->events[head];
	198	atomic_store_release(&ring->head,
	199	(head + 1) % ring->nr);
	200	i++;
	201	}
	202	}
	203
	204	return i;
	205	}
	206
	207	static int fio_libaio_getevents(struct thread_data *td, unsigned int min,
	208	unsigned int max, const struct timespec *t)
	209	{
	210	struct libaio_data *ld = td->io_ops_data;
	211	struct libaio_options *o = td->eo;
	212	unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
	213	struct timespec __lt, *lt = NULL;
	214	int r, events = 0;
	215
	216	if (t) {
	217	__lt = *t;
	218	lt = &__lt;
	219	}
	220
	221	do {
	222	if (o->userspace_reap == 1
	223	&& actual_min == 0
	224	&& ((struct aio_ring *)(ld->aio_ctx))->magic
	225	== AIO_RING_MAGIC) {
	226	r = user_io_getevents(ld->aio_ctx, max,
	227	ld->aio_events + events);
	228	} else {
	229	r = io_getevents(ld->aio_ctx, actual_min,
	230	max, ld->aio_events + events, lt);
	231	}
	232	if (r > 0)
	233	events += r;
	234	else if ((min && r == 0) \|\| r == -EAGAIN) {
	235	fio_libaio_commit(td);
	236	if (actual_min)
	237	usleep(10);
	238	} else if (r != -EINTR)
	239	break;
	240	} while (events < min);
	241
	242	return r < 0 ? r : events;
	243	}
	244
	245	static enum fio_q_status fio_libaio_queue(struct thread_data *td,
	246	struct io_u *io_u)
	247	{
	248	struct libaio_data *ld = td->io_ops_data;
	249	struct libaio_options *o = td->eo;
	250
	251	fio_ro_check(td, io_u);
	252
	253	if (ld->queued == td->o.iodepth)
	254	return FIO_Q_BUSY;
	255
	256	/*
	257	* fsync is tricky, since it can fail and we need to do it
	258	* serialized with other io. the reason is that linux doesn't
	259	* support aio fsync yet. So return busy for the case where we
	260	* have pending io, to let fio complete those first.
	261	*/
	262	if (ddir_sync(io_u->ddir)) {
	263	if (ld->queued)
	264	return FIO_Q_BUSY;
	265
	266	do_io_u_sync(td, io_u);
	267	return FIO_Q_COMPLETED;
	268	}
	269
	270	if (io_u->ddir == DDIR_TRIM) {
	271	if (ld->queued)
	272	return FIO_Q_BUSY;
	273
	274	do_io_u_trim(td, io_u);
	275	io_u_mark_submit(td, 1);
	276	io_u_mark_complete(td, 1);
	277	return FIO_Q_COMPLETED;
	278	}
	279
	280	if (o->cmdprio_percentage)
	281	fio_libaio_prio_prep(td, io_u);
	282
	283	ld->iocbs[ld->head] = &io_u->iocb;
	284	ld->io_us[ld->head] = io_u;
	285	ring_inc(ld, &ld->head, 1);
	286	ld->queued++;
	287	return FIO_Q_QUEUED;
	288	}
	289
	290	static void fio_libaio_queued(struct thread_data td, struct io_u *io_us,
	291	unsigned int nr)
	292	{
	293	struct timespec now;
	294	unsigned int i;
	295
	296	if (!fio_fill_issue_time(td))
	297	return;
	298
	299	fio_gettime(&now, NULL);
	300
	301	for (i = 0; i < nr; i++) {
	302	struct io_u *io_u = io_us[i];
	303
	304	memcpy(&io_u->issue_time, &now, sizeof(now));
	305	io_u_queued(td, io_u);
	306	}
	307	}
	308
	309	static int fio_libaio_commit(struct thread_data *td)
	310	{
	311	struct libaio_data *ld = td->io_ops_data;
	312	struct iocb **iocbs;
	313	struct io_u **io_us;
	314	struct timespec ts;
	315	int ret, wait_start = 0;
	316
	317	if (!ld->queued)
	318	return 0;
	319
	320	do {
	321	long nr = ld->queued;
	322
	323	nr = min((unsigned int) nr, ld->entries - ld->tail);
	324	io_us = ld->io_us + ld->tail;
	325	iocbs = ld->iocbs + ld->tail;
	326
	327	ret = io_submit(ld->aio_ctx, nr, iocbs);
	328	if (ret > 0) {
	329	fio_libaio_queued(td, io_us, ret);
	330	io_u_mark_submit(td, ret);
	331
	332	ld->queued -= ret;
	333	ring_inc(ld, &ld->tail, ret);
	334	ret = 0;
	335	wait_start = 0;
	336	} else if (ret == -EINTR \|\| !ret) {
	337	if (!ret)
	338	io_u_mark_submit(td, ret);
	339	wait_start = 0;
	340	continue;
	341	} else if (ret == -EAGAIN) {
	342	/*
	343	* If we get EAGAIN, we should break out without
	344	* error and let the upper layer reap some
	345	* events for us. If we have no queued IO, we
	346	* must loop here. If we loop for more than 30s,
	347	* just error out, something must be buggy in the
	348	* IO path.
	349	*/
	350	if (ld->queued) {
	351	ret = 0;
	352	break;
	353	}
	354	if (!wait_start) {
	355	fio_gettime(&ts, NULL);
	356	wait_start = 1;
	357	} else if (mtime_since_now(&ts) > 30000) {
	358	log_err("fio: aio appears to be stalled, giving up\n");
	359	break;
	360	}
	361	usleep(1);
	362	continue;
	363	} else if (ret == -ENOMEM) {
	364	/*
	365	* If we get -ENOMEM, reap events if we can. If
	366	* we cannot, treat it as a fatal event since there's
	367	* nothing we can do about it.
	368	*/
	369	if (ld->queued)
	370	ret = 0;
	371	break;
	372	} else
	373	break;
	374	} while (ld->queued);
	375
	376	return ret;
	377	}
	378
	379	static int fio_libaio_cancel(struct thread_data td, struct io_u io_u)
	380	{
	381	struct libaio_data *ld = td->io_ops_data;
	382
	383	return io_cancel(ld->aio_ctx, &io_u->iocb, ld->aio_events);
	384	}
	385
	386	static void fio_libaio_cleanup(struct thread_data *td)
	387	{
	388	struct libaio_data *ld = td->io_ops_data;
	389
	390	if (ld) {
	391	/*
	392	* Work-around to avoid huge RCU stalls at exit time. If we
	393	* don't do this here, then it'll be torn down by exit_aio().
	394	* But for that case we can parallellize the freeing, thus
	395	* speeding it up a lot.
	396	*/
	397	if (!(td->flags & TD_F_CHILD))
	398	io_destroy(ld->aio_ctx);
	399	free(ld->aio_events);
	400	free(ld->iocbs);
	401	free(ld->io_us);
	402	free(ld);
	403	}
	404	}
	405
	406	static int fio_libaio_post_init(struct thread_data *td)
	407	{
	408	struct libaio_data *ld = td->io_ops_data;
	409	int err;
	410
	411	err = io_queue_init(td->o.iodepth, &ld->aio_ctx);
	412	if (err) {
	413	td_verror(td, -err, "io_queue_init");
	414	return 1;
	415	}
	416
	417	return 0;
	418	}
	419
	420	static int fio_libaio_init(struct thread_data *td)
	421	{
	422	struct libaio_data *ld;
	423	struct thread_options *to = &td->o;
	424	struct libaio_options *o = td->eo;
	425
	426	ld = calloc(1, sizeof(*ld));
	427
	428	ld->entries = td->o.iodepth;
	429	ld->is_pow2 = is_power_of_2(ld->entries);
	430	ld->aio_events = calloc(ld->entries, sizeof(struct io_event));
	431	ld->iocbs = calloc(ld->entries, sizeof(struct iocb *));
	432	ld->io_us = calloc(ld->entries, sizeof(struct io_u *));
	433
	434	td->io_ops_data = ld;
	435	/*
	436	* Check for option conflicts
	437	*/
	438	if ((fio_option_is_set(to, ioprio) \|\| fio_option_is_set(to, ioprio_class)) &&
	439	o->cmdprio_percentage != 0) {
	440	log_err("%s: cmdprio_percentage option and mutually exclusive "
	441	"prio or prioclass option is set, exiting\n", to->name);
	442	td_verror(td, EINVAL, "fio_libaio_init");
	443	return 1;
	444	}
	445	return 0;
	446	}
	447
	448	FIO_STATIC struct ioengine_ops ioengine = {
	449	.name = "libaio",
	450	.version = FIO_IOOPS_VERSION,
	451	.flags = FIO_ASYNCIO_SYNC_TRIM,
	452	.init = fio_libaio_init,
	453	.post_init = fio_libaio_post_init,
	454	.prep = fio_libaio_prep,
	455	.queue = fio_libaio_queue,
	456	.commit = fio_libaio_commit,
	457	.cancel = fio_libaio_cancel,
	458	.getevents = fio_libaio_getevents,
	459	.event = fio_libaio_event,
	460	.cleanup = fio_libaio_cleanup,
	461	.open_file = generic_open_file,
	462	.close_file = generic_close_file,
	463	.get_file_size = generic_get_file_size,
	464	.options = options,
	465	.option_struct_size = sizeof(struct libaio_options),
	466	};
	467
	468	static void fio_init fio_libaio_register(void)
	469	{
	470	register_ioengine(&ioengine);
	471	}
	472
	473	static void fio_exit fio_libaio_unregister(void)
	474	{
	475	unregister_ioengine(&ioengine);
	476	}