[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"

#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 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 hipri;
};

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,
	},
	{
		.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 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 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(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 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];
			read_barrier();
			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;

	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;
	}

	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_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;
	int err = 0;

	err = fio_libaio_queue_init(ld, td->o.iodepth, o->hipri);
	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;

	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;
	return 0;
}

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
2866c82d	1	/*
da751ca9 JA	2	* libaio engine
	3	*
	4	* IO engine using the Linux native aio interface.
2866c82d JA	5	*
2866c82d JA	6	*/
2866c82d JA	7	#include <stdlib.h>
	8	#include <unistd.h>
	9	#include <errno.h>
67bf9823	10	#include <libaio.h>
46961748 JA	11	#include <sys/time.h>
46961748 JA	12	#include <sys/resource.h>
5f350952 JA	13
5f350952 JA	14	#include "../fio.h"
0f38bbef	15	#include "../lib/pow2.h"
d220c761	16	#include "../optgroup.h"
90a8c9b2	17	#include "../lib/memalign.h"
2866c82d	18
ee636f3f JA	19	#ifndef IOCB_FLAG_HIPRI
	20	#define IOCB_FLAG_HIPRI (1 << 2)
	21	#endif
2da0f5d5	22
2da0f5d5	23	#ifndef IOCTX_FLAG_IOPOLL
702906e9	24	#define IOCTX_FLAG_IOPOLL (1 << 0)
2041bd34	25	#endif
a391d73d	26
0fc2e103 JA	27	static int fio_libaio_commit(struct thread_data *td);
0fc2e103 JA	28
2866c82d JA	29	struct libaio_data {
	30	io_context_t aio_ctx;
	31	struct io_event *aio_events;
755200a3	32	struct iocb **iocbs;
7e77dd02	33	struct io_u **io_us;
acd45e02	34
3c3168e9 JA	35	struct io_u **io_u_index;
3c3168e9 JA	36
acd45e02 JA	37	/*
	38	* Basic ring buffer. 'head' is incremented in _queue(), and
	39	* 'tail' is incremented in _commit(). We keep 'queued' so
	40	* that we know if the ring is full or empty, when
	41	* 'head' == 'tail'. 'entries' is the ring size, and
	42	* 'is_pow2' is just an optimization to use AND instead of
	43	* modulus to get the remainder on ring increment.
	44	*/
	45	int is_pow2;
	46	unsigned int entries;
	47	unsigned int queued;
	48	unsigned int head;
	49	unsigned int tail;
2866c82d JA	50	};
2866c82d JA	51
de890a1e	52	struct libaio_options {
a1f871c7	53	void *pad;
de890a1e	54	unsigned int userspace_reap;
a391d73d	55	unsigned int hipri;
de890a1e SL	56	};
	57
	58	static struct fio_option options[] = {
	59	{
	60	.name = "userspace_reap",
e8b0e958	61	.lname = "Libaio userspace reaping",
de890a1e SL	62	.type = FIO_OPT_STR_SET,
	63	.off1 = offsetof(struct libaio_options, userspace_reap),
	64	.help = "Use alternative user-space reap implementation",
e90a0adf	65	.category = FIO_OPT_C_ENGINE,
5a3cd5f3	66	.group = FIO_OPT_G_LIBAIO,
de890a1e	67	},
a391d73d JA	68	{
a391d73d JA	69	.name = "hipri",
ee636f3f	70	.lname = "High Priority",
a391d73d JA	71	.type = FIO_OPT_STR_SET,
a391d73d JA	72	.off1 = offsetof(struct libaio_options, hipri),
6acc46d8	73	.help = "Use polled IO completions",
a391d73d JA	74	.category = FIO_OPT_C_ENGINE,
	75	.group = FIO_OPT_G_LIBAIO,
	76	},
de890a1e SL	77	{
	78	.name = NULL,
	79	},
	80	};
	81
acd45e02 JA	82	static inline void ring_inc(struct libaio_data ld, unsigned int val,
	83	unsigned int add)
	84	{
	85	if (ld->is_pow2)
	86	val = (val + add) & (ld->entries - 1);
	87	else
	88	val = (val + add) % ld->entries;
	89	}
	90
7a16dd02	91	static int fio_libaio_prep(struct thread_data fio_unused td, struct io_u io_u)
2866c82d	92	{
53cdc686	93	struct fio_file *f = io_u->file;
a391d73d	94	struct libaio_options *o = td->eo;
3c3168e9 JA	95	struct iocb *iocb;
3c3168e9 JA	96
702906e9	97	iocb = &io_u->iocb;
53cdc686	98
a391d73d	99	if (io_u->ddir == DDIR_READ) {
702906e9 JA	100	io_prep_pread(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
	101	if (o->hipri)
	102	iocb->u.c.flags \|= IOCB_FLAG_HIPRI;
a391d73d	103	} else if (io_u->ddir == DDIR_WRITE) {
702906e9 JA	104	io_prep_pwrite(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
	105	if (o->hipri)
	106	iocb->u.c.flags \|= IOCB_FLAG_HIPRI;
a391d73d	107	} else if (ddir_sync(io_u->ddir))
3c3168e9	108	io_prep_fsync(iocb, f->fd);
2866c82d JA	109
	110	return 0;
	111	}
	112
	113	static struct io_u fio_libaio_event(struct thread_data td, int event)
	114	{
565e784d	115	struct libaio_data *ld = td->io_ops_data;
f423479d JA	116	struct io_event *ev;
f423479d JA	117	struct io_u *io_u;
2866c82d	118
f423479d	119	ev = ld->aio_events + event;
702906e9	120	io_u = container_of(ev->obj, struct io_u, iocb);
f423479d JA	121
	122	if (ev->res != io_u->xfer_buflen) {
	123	if (ev->res > io_u->xfer_buflen)
	124	io_u->error = -ev->res;
	125	else
	126	io_u->resid = io_u->xfer_buflen - ev->res;
	127	} else
	128	io_u->error = 0;
	129
	130	return io_u;
2866c82d JA	131	}
2866c82d JA	132
675012f0 DE	133	struct aio_ring {
	134	unsigned id; /** kernel internal index number */
	135	unsigned nr; /** number of io_events */
	136	unsigned head;
	137	unsigned tail;
c44b1ff5	138
675012f0 DE	139	unsigned magic;
	140	unsigned compat_features;
	141	unsigned incompat_features;
	142	unsigned header_length; /** size of aio_ring */
	143
	144	struct io_event events[0];
	145	};
	146
	147	#define AIO_RING_MAGIC 0xa10a10a1
	148
	149	static int user_io_getevents(io_context_t aio_ctx, unsigned int max,
c44b1ff5	150	struct io_event *events)
675012f0 DE	151	{
	152	long i = 0;
	153	unsigned head;
c44b1ff5	154	struct aio_ring ring = (struct aio_ring) aio_ctx;
675012f0 DE	155
	156	while (i < max) {
	157	head = ring->head;
	158
	159	if (head == ring->tail) {
	160	/* There are no more completions */
	161	break;
	162	} else {
	163	/* There is another completion to reap */
	164	events[i] = ring->events[head];
	165	read_barrier();
c44b1ff5	166	ring->head = (head + 1) % ring->nr;
675012f0 DE	167	i++;
	168	}
	169	}
	170
	171	return i;
	172	}
	173
e7d2e616	174	static int fio_libaio_getevents(struct thread_data *td, unsigned int min,
1f440ece	175	unsigned int max, const struct timespec *t)
2866c82d	176	{
565e784d	177	struct libaio_data *ld = td->io_ops_data;
de890a1e	178	struct libaio_options *o = td->eo;
82407585	179	unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
1f440ece	180	struct timespec __lt, *lt = NULL;
0b7fdba7	181	int r, events = 0;
2866c82d	182
1f440ece JA	183	if (t) {
	184	__lt = *t;
	185	lt = &__lt;
	186	}
	187
2866c82d	188	do {
de890a1e	189	if (o->userspace_reap == 1
675012f0 DE	190	&& actual_min == 0
	191	&& ((struct aio_ring *)(ld->aio_ctx))->magic
	192	== AIO_RING_MAGIC) {
	193	r = user_io_getevents(ld->aio_ctx, max,
	194	ld->aio_events + events);
	195	} else {
	196	r = io_getevents(ld->aio_ctx, actual_min,
1f440ece	197	max, ld->aio_events + events, lt);
675012f0	198	}
3441a52d	199	if (r > 0)
0b7fdba7	200	events += r;
3441a52d	201	else if ((min && r == 0) \|\| r == -EAGAIN) {
0fc2e103	202	fio_libaio_commit(td);
6347e43d JA	203	if (actual_min)
6347e43d JA	204	usleep(10);
0fc2e103	205	} else if (r != -EINTR)
a31dc2dd	206	break;
0b7fdba7 DE	207	} while (events < min);
	208
	209	return r < 0 ? r : events;
2866c82d JA	210	}
2866c82d JA	211
2e4ef4fb JA	212	static enum fio_q_status fio_libaio_queue(struct thread_data *td,
2e4ef4fb JA	213	struct io_u *io_u)
2866c82d	214	{
565e784d	215	struct libaio_data *ld = td->io_ops_data;
2866c82d	216
7101d9c2 JA	217	fio_ro_check(td, io_u);
7101d9c2 JA	218
acd45e02	219	if (ld->queued == td->o.iodepth)
755200a3 JA	220	return FIO_Q_BUSY;
	221
	222	/*
	223	* fsync is tricky, since it can fail and we need to do it
	224	* serialized with other io. the reason is that linux doesn't
	225	* support aio fsync yet. So return busy for the case where we
	226	* have pending io, to let fio complete those first.
	227	*/
f011531e	228	if (ddir_sync(io_u->ddir)) {
acd45e02	229	if (ld->queued)
755200a3	230	return FIO_Q_BUSY;
5f9099ea	231
f011531e	232	do_io_u_sync(td, io_u);
755200a3 JA	233	return FIO_Q_COMPLETED;
	234	}
	235
a5f3027c	236	if (io_u->ddir == DDIR_TRIM) {
acd45e02	237	if (ld->queued)
a5f3027c JA	238	return FIO_Q_BUSY;
	239
	240	do_io_u_trim(td, io_u);
c0681c9d VF	241	io_u_mark_submit(td, 1);
c0681c9d VF	242	io_u_mark_complete(td, 1);
a5f3027c JA	243	return FIO_Q_COMPLETED;
	244	}
	245
702906e9	246	ld->iocbs[ld->head] = &io_u->iocb;
acd45e02 JA	247	ld->io_us[ld->head] = io_u;
	248	ring_inc(ld, &ld->head, 1);
	249	ld->queued++;
755200a3 JA	250	return FIO_Q_QUEUED;
	251	}
	252
7e77dd02 JA	253	static void fio_libaio_queued(struct thread_data td, struct io_u *io_us,
	254	unsigned int nr)
	255	{
8b6a404c	256	struct timespec now;
7e77dd02 JA	257	unsigned int i;
7e77dd02 JA	258
12d9d841 JA	259	if (!fio_fill_issue_time(td))
	260	return;
	261
7e77dd02 JA	262	fio_gettime(&now, NULL);
	263
	264	for (i = 0; i < nr; i++) {
	265	struct io_u *io_u = io_us[i];
	266
	267	memcpy(&io_u->issue_time, &now, sizeof(now));
	268	io_u_queued(td, io_u);
	269	}
	270	}
	271
755200a3 JA	272	static int fio_libaio_commit(struct thread_data *td)
755200a3 JA	273	{
565e784d	274	struct libaio_data *ld = td->io_ops_data;
755200a3	275	struct iocb **iocbs;
7e77dd02	276	struct io_u **io_us;
8b6a404c	277	struct timespec ts;
a120ca7f	278	int ret, wait_start = 0;
755200a3	279
acd45e02	280	if (!ld->queued)
755200a3 JA	281	return 0;
755200a3 JA	282
2866c82d	283	do {
acd45e02 JA	284	long nr = ld->queued;
	285
	286	nr = min((unsigned int) nr, ld->entries - ld->tail);
	287	io_us = ld->io_us + ld->tail;
	288	iocbs = ld->iocbs + ld->tail;
	289
	290	ret = io_submit(ld->aio_ctx, nr, iocbs);
5e00c2c4	291	if (ret > 0) {
7e77dd02	292	fio_libaio_queued(td, io_us, ret);
838bc709	293	io_u_mark_submit(td, ret);
acd45e02 JA	294
	295	ld->queued -= ret;
	296	ring_inc(ld, &ld->tail, ret);
5e00c2c4	297	ret = 0;
e3b4e568	298	wait_start = 0;
acd45e02	299	} else if (ret == -EINTR \|\| !ret) {
838bc709 JA	300	if (!ret)
838bc709 JA	301	io_u_mark_submit(td, ret);
e3b4e568	302	wait_start = 0;
2866c82d	303	continue;
acd45e02 JA	304	} else if (ret == -EAGAIN) {
	305	/*
	306	* If we get EAGAIN, we should break out without
	307	* error and let the upper layer reap some
a120ca7f JA	308	* events for us. If we have no queued IO, we
	309	* must loop here. If we loop for more than 30s,
	310	* just error out, something must be buggy in the
	311	* IO path.
acd45e02	312	*/
a120ca7f JA	313	if (ld->queued) {
	314	ret = 0;
	315	break;
	316	}
	317	if (!wait_start) {
8b6a404c	318	fio_gettime(&ts, NULL);
d36b072d	319	wait_start = 1;
8b6a404c	320	} else if (mtime_since_now(&ts) > 30000) {
a120ca7f JA	321	log_err("fio: aio appears to be stalled, giving up\n");
	322	break;
	323	}
	324	usleep(1);
	325	continue;
a31dc2dd JA	326	} else if (ret == -ENOMEM) {
	327	/*
	328	* If we get -ENOMEM, reap events if we can. If
	329	* we cannot, treat it as a fatal event since there's
	330	* nothing we can do about it.
	331	*/
	332	if (ld->queued)
	333	ret = 0;
	334	break;
838bc709	335	} else
2866c82d	336	break;
2c3a4ae9	337	} while (ld->queued);
2866c82d	338
36167d82	339	return ret;
2866c82d JA	340	}
	341
	342	static int fio_libaio_cancel(struct thread_data td, struct io_u io_u)
	343	{
565e784d	344	struct libaio_data *ld = td->io_ops_data;
2866c82d JA	345
	346	return io_cancel(ld->aio_ctx, &io_u->iocb, ld->aio_events);
	347	}
	348
	349	static void fio_libaio_cleanup(struct thread_data *td)
	350	{
565e784d	351	struct libaio_data *ld = td->io_ops_data;
2866c82d JA	352
2866c82d JA	353	if (ld) {
f24c2649 JA	354	/*
	355	* Work-around to avoid huge RCU stalls at exit time. If we
	356	* don't do this here, then it'll be torn down by exit_aio().
	357	* But for that case we can parallellize the freeing, thus
	358	* speeding it up a lot.
	359	*/
	360	if (!(td->flags & TD_F_CHILD))
	361	io_destroy(ld->aio_ctx);
7e77dd02 JA	362	free(ld->aio_events);
	363	free(ld->iocbs);
	364	free(ld->io_us);
2866c82d	365	free(ld);
2866c82d JA	366	}
	367	}
	368
e65ef959	369	static int fio_libaio_old_queue_init(struct libaio_data *ld, unsigned int depth,
702906e9	370	bool hipri)
ebec344d	371	{
ebec344d JA	372	if (hipri) {
	373	log_err("fio: polled aio not available on your platform\n");
	374	return 1;
	375	}
3c3168e9	376
ebec344d	377	return io_queue_init(depth, &ld->aio_ctx);
e65ef959 JA	378	}
	379
	380	static int fio_libaio_queue_init(struct libaio_data *ld, unsigned int depth,
702906e9	381	bool hipri)
e65ef959 JA	382	{
	383	#ifdef __NR_sys_io_setup2
	384	int ret, flags = 0;
	385
	386	if (hipri)
	387	flags \|= IOCTX_FLAG_IOPOLL;
e65ef959	388
702906e9 JA	389	ret = syscall(__NR_sys_io_setup2, depth, flags, NULL, NULL,
702906e9 JA	390	&ld->aio_ctx);
e65ef959 JA	391	if (!ret)
e65ef959 JA	392	return 0;
0fcbc009	393	/* fall through to old syscall */
ebec344d	394	#endif
702906e9	395	return fio_libaio_old_queue_init(ld, depth, hipri);
2041bd34 JA	396	}
	397
	398	static int fio_libaio_post_init(struct thread_data *td)
	399	{
	400	struct libaio_data *ld = td->io_ops_data;
	401	struct libaio_options *o = td->eo;
2041bd34 JA	402	int err = 0;
2041bd34 JA	403
702906e9	404	err = fio_libaio_queue_init(ld, td->o.iodepth, o->hipri);
2041bd34 JA	405	if (err) {
	406	td_verror(td, -err, "io_queue_init");
	407	return 1;
	408	}
	409
	410	return 0;
ebec344d JA	411	}
ebec344d JA	412
2866c82d JA	413	static int fio_libaio_init(struct thread_data *td)
2866c82d JA	414	{
acd45e02	415	struct libaio_data *ld;
2866c82d	416
acd45e02	417	ld = calloc(1, sizeof(*ld));
c1db2dce	418
acd45e02 JA	419	ld->entries = td->o.iodepth;
	420	ld->is_pow2 = is_power_of_2(ld->entries);
	421	ld->aio_events = calloc(ld->entries, sizeof(struct io_event));
	422	ld->iocbs = calloc(ld->entries, sizeof(struct iocb *));
	423	ld->io_us = calloc(ld->entries, sizeof(struct io_u *));
755200a3	424
565e784d	425	td->io_ops_data = ld;
2866c82d JA	426	return 0;
	427	}
	428
5f350952	429	static struct ioengine_ops ioengine = {
de890a1e SL	430	.name = "libaio",
de890a1e SL	431	.version = FIO_IOOPS_VERSION,
04ba61df	432	.flags = FIO_ASYNCIO_SYNC_TRIM,
de890a1e	433	.init = fio_libaio_init,
2041bd34	434	.post_init = fio_libaio_post_init,
de890a1e SL	435	.prep = fio_libaio_prep,
	436	.queue = fio_libaio_queue,
	437	.commit = fio_libaio_commit,
	438	.cancel = fio_libaio_cancel,
	439	.getevents = fio_libaio_getevents,
	440	.event = fio_libaio_event,
	441	.cleanup = fio_libaio_cleanup,
	442	.open_file = generic_open_file,
	443	.close_file = generic_close_file,
	444	.get_file_size = generic_get_file_size,
	445	.options = options,
	446	.option_struct_size = sizeof(struct libaio_options),
2866c82d	447	};
34cfcdaf	448
5f350952 JA	449	static void fio_init fio_libaio_register(void)
	450	{
	451	register_ioengine(&ioengine);
	452	}
	453
	454	static void fio_exit fio_libaio_unregister(void)
	455	{
	456	unregister_ioengine(&ioengine);
	457	}