[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 "../fio.h"
#include "../lib/pow2.h"
#include "../optgroup.h"

#define IOCB_CMD_PREAD_POLL 9
#define IOCB_CMD_PWRITE_POLL 10

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;

	/*
	 * 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	= "RWF_HIPRI",
		.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;

	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);
		if (o->hipri)
			io_u->iocb.aio_lio_opcode = IOCB_CMD_PREAD_POLL;
	} 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);
		if (o->hipri)
			io_u->iocb.aio_lio_opcode = IOCB_CMD_PWRITE_POLL;
	} else if (ddir_sync(io_u->ddir))
		io_prep_fsync(&io_u->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_init(struct thread_data *td)
{
	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);
	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->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,
	.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,
	.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>
5f350952 JA	11
5f350952 JA	12	#include "../fio.h"
0f38bbef	13	#include "../lib/pow2.h"
d220c761	14	#include "../optgroup.h"
2866c82d	15
a391d73d JA	16	#define IOCB_CMD_PREAD_POLL 9
	17	#define IOCB_CMD_PWRITE_POLL 10
	18
0fc2e103 JA	19	static int fio_libaio_commit(struct thread_data *td);
0fc2e103 JA	20
2866c82d JA	21	struct libaio_data {
	22	io_context_t aio_ctx;
	23	struct io_event *aio_events;
755200a3	24	struct iocb **iocbs;
7e77dd02	25	struct io_u **io_us;
acd45e02 JA	26
	27	/*
	28	* Basic ring buffer. 'head' is incremented in _queue(), and
	29	* 'tail' is incremented in _commit(). We keep 'queued' so
	30	* that we know if the ring is full or empty, when
	31	* 'head' == 'tail'. 'entries' is the ring size, and
	32	* 'is_pow2' is just an optimization to use AND instead of
	33	* modulus to get the remainder on ring increment.
	34	*/
	35	int is_pow2;
	36	unsigned int entries;
	37	unsigned int queued;
	38	unsigned int head;
	39	unsigned int tail;
2866c82d JA	40	};
2866c82d JA	41
de890a1e	42	struct libaio_options {
a1f871c7	43	void *pad;
de890a1e	44	unsigned int userspace_reap;
a391d73d	45	unsigned int hipri;
de890a1e SL	46	};
	47
	48	static struct fio_option options[] = {
	49	{
	50	.name = "userspace_reap",
e8b0e958	51	.lname = "Libaio userspace reaping",
de890a1e SL	52	.type = FIO_OPT_STR_SET,
	53	.off1 = offsetof(struct libaio_options, userspace_reap),
	54	.help = "Use alternative user-space reap implementation",
e90a0adf	55	.category = FIO_OPT_C_ENGINE,
5a3cd5f3	56	.group = FIO_OPT_G_LIBAIO,
de890a1e	57	},
a391d73d JA	58	{
	59	.name = "hipri",
	60	.lname = "RWF_HIPRI",
	61	.type = FIO_OPT_STR_SET,
	62	.off1 = offsetof(struct libaio_options, hipri),
6acc46d8	63	.help = "Use polled IO completions",
a391d73d JA	64	.category = FIO_OPT_C_ENGINE,
	65	.group = FIO_OPT_G_LIBAIO,
	66	},
de890a1e SL	67	{
	68	.name = NULL,
	69	},
	70	};
	71
acd45e02 JA	72	static inline void ring_inc(struct libaio_data ld, unsigned int val,
	73	unsigned int add)
	74	{
	75	if (ld->is_pow2)
	76	val = (val + add) & (ld->entries - 1);
	77	else
	78	val = (val + add) % ld->entries;
	79	}
	80
7a16dd02	81	static int fio_libaio_prep(struct thread_data fio_unused td, struct io_u io_u)
2866c82d	82	{
53cdc686	83	struct fio_file *f = io_u->file;
a391d73d	84	struct libaio_options *o = td->eo;
53cdc686	85
a391d73d	86	if (io_u->ddir == DDIR_READ) {
cec6b55d	87	io_prep_pread(&io_u->iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
a391d73d JA	88	if (o->hipri)
	89	io_u->iocb.aio_lio_opcode = IOCB_CMD_PREAD_POLL;
	90	} else if (io_u->ddir == DDIR_WRITE) {
cec6b55d	91	io_prep_pwrite(&io_u->iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
a391d73d JA	92	if (o->hipri)
	93	io_u->iocb.aio_lio_opcode = IOCB_CMD_PWRITE_POLL;
	94	} else if (ddir_sync(io_u->ddir))
87dc1ab1	95	io_prep_fsync(&io_u->iocb, f->fd);
2866c82d JA	96
	97	return 0;
	98	}
	99
	100	static struct io_u fio_libaio_event(struct thread_data td, int event)
	101	{
565e784d	102	struct libaio_data *ld = td->io_ops_data;
f423479d JA	103	struct io_event *ev;
f423479d JA	104	struct io_u *io_u;
2866c82d	105
f423479d	106	ev = ld->aio_events + event;
0cc6ee59	107	io_u = container_of(ev->obj, struct io_u, iocb);
f423479d JA	108
	109	if (ev->res != io_u->xfer_buflen) {
	110	if (ev->res > io_u->xfer_buflen)
	111	io_u->error = -ev->res;
	112	else
	113	io_u->resid = io_u->xfer_buflen - ev->res;
	114	} else
	115	io_u->error = 0;
	116
	117	return io_u;
2866c82d JA	118	}
2866c82d JA	119
675012f0 DE	120	struct aio_ring {
	121	unsigned id; /** kernel internal index number */
	122	unsigned nr; /** number of io_events */
	123	unsigned head;
	124	unsigned tail;
c44b1ff5	125
675012f0 DE	126	unsigned magic;
	127	unsigned compat_features;
	128	unsigned incompat_features;
	129	unsigned header_length; /** size of aio_ring */
	130
	131	struct io_event events[0];
	132	};
	133
	134	#define AIO_RING_MAGIC 0xa10a10a1
	135
	136	static int user_io_getevents(io_context_t aio_ctx, unsigned int max,
c44b1ff5	137	struct io_event *events)
675012f0 DE	138	{
	139	long i = 0;
	140	unsigned head;
c44b1ff5	141	struct aio_ring ring = (struct aio_ring) aio_ctx;
675012f0 DE	142
	143	while (i < max) {
	144	head = ring->head;
	145
	146	if (head == ring->tail) {
	147	/* There are no more completions */
	148	break;
	149	} else {
	150	/* There is another completion to reap */
	151	events[i] = ring->events[head];
	152	read_barrier();
c44b1ff5	153	ring->head = (head + 1) % ring->nr;
675012f0 DE	154	i++;
	155	}
	156	}
	157
	158	return i;
	159	}
	160
e7d2e616	161	static int fio_libaio_getevents(struct thread_data *td, unsigned int min,
1f440ece	162	unsigned int max, const struct timespec *t)
2866c82d	163	{
565e784d	164	struct libaio_data *ld = td->io_ops_data;
de890a1e	165	struct libaio_options *o = td->eo;
82407585	166	unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
1f440ece	167	struct timespec __lt, *lt = NULL;
0b7fdba7	168	int r, events = 0;
2866c82d	169
1f440ece JA	170	if (t) {
	171	__lt = *t;
	172	lt = &__lt;
	173	}
	174
2866c82d	175	do {
de890a1e	176	if (o->userspace_reap == 1
675012f0 DE	177	&& actual_min == 0
	178	&& ((struct aio_ring *)(ld->aio_ctx))->magic
	179	== AIO_RING_MAGIC) {
	180	r = user_io_getevents(ld->aio_ctx, max,
	181	ld->aio_events + events);
	182	} else {
	183	r = io_getevents(ld->aio_ctx, actual_min,
1f440ece	184	max, ld->aio_events + events, lt);
675012f0	185	}
3441a52d	186	if (r > 0)
0b7fdba7	187	events += r;
3441a52d	188	else if ((min && r == 0) \|\| r == -EAGAIN) {
0fc2e103	189	fio_libaio_commit(td);
6347e43d JA	190	if (actual_min)
6347e43d JA	191	usleep(10);
0fc2e103	192	} else if (r != -EINTR)
a31dc2dd	193	break;
0b7fdba7 DE	194	} while (events < min);
	195
	196	return r < 0 ? r : events;
2866c82d JA	197	}
2866c82d JA	198
2e4ef4fb JA	199	static enum fio_q_status fio_libaio_queue(struct thread_data *td,
2e4ef4fb JA	200	struct io_u *io_u)
2866c82d	201	{
565e784d	202	struct libaio_data *ld = td->io_ops_data;
2866c82d	203
7101d9c2 JA	204	fio_ro_check(td, io_u);
7101d9c2 JA	205
acd45e02	206	if (ld->queued == td->o.iodepth)
755200a3 JA	207	return FIO_Q_BUSY;
	208
	209	/*
	210	* fsync is tricky, since it can fail and we need to do it
	211	* serialized with other io. the reason is that linux doesn't
	212	* support aio fsync yet. So return busy for the case where we
	213	* have pending io, to let fio complete those first.
	214	*/
f011531e	215	if (ddir_sync(io_u->ddir)) {
acd45e02	216	if (ld->queued)
755200a3	217	return FIO_Q_BUSY;
5f9099ea	218
f011531e	219	do_io_u_sync(td, io_u);
755200a3 JA	220	return FIO_Q_COMPLETED;
	221	}
	222
a5f3027c	223	if (io_u->ddir == DDIR_TRIM) {
acd45e02	224	if (ld->queued)
a5f3027c JA	225	return FIO_Q_BUSY;
	226
	227	do_io_u_trim(td, io_u);
c0681c9d VF	228	io_u_mark_submit(td, 1);
c0681c9d VF	229	io_u_mark_complete(td, 1);
a5f3027c JA	230	return FIO_Q_COMPLETED;
	231	}
	232
acd45e02 JA	233	ld->iocbs[ld->head] = &io_u->iocb;
	234	ld->io_us[ld->head] = io_u;
	235	ring_inc(ld, &ld->head, 1);
	236	ld->queued++;
755200a3 JA	237	return FIO_Q_QUEUED;
	238	}
	239
7e77dd02 JA	240	static void fio_libaio_queued(struct thread_data td, struct io_u *io_us,
	241	unsigned int nr)
	242	{
8b6a404c	243	struct timespec now;
7e77dd02 JA	244	unsigned int i;
7e77dd02 JA	245
12d9d841 JA	246	if (!fio_fill_issue_time(td))
	247	return;
	248
7e77dd02 JA	249	fio_gettime(&now, NULL);
	250
	251	for (i = 0; i < nr; i++) {
	252	struct io_u *io_u = io_us[i];
	253
	254	memcpy(&io_u->issue_time, &now, sizeof(now));
	255	io_u_queued(td, io_u);
	256	}
	257	}
	258
755200a3 JA	259	static int fio_libaio_commit(struct thread_data *td)
755200a3 JA	260	{
565e784d	261	struct libaio_data *ld = td->io_ops_data;
755200a3	262	struct iocb **iocbs;
7e77dd02	263	struct io_u **io_us;
8b6a404c	264	struct timespec ts;
a120ca7f	265	int ret, wait_start = 0;
755200a3	266
acd45e02	267	if (!ld->queued)
755200a3 JA	268	return 0;
755200a3 JA	269
2866c82d	270	do {
acd45e02 JA	271	long nr = ld->queued;
	272
	273	nr = min((unsigned int) nr, ld->entries - ld->tail);
	274	io_us = ld->io_us + ld->tail;
	275	iocbs = ld->iocbs + ld->tail;
	276
	277	ret = io_submit(ld->aio_ctx, nr, iocbs);
5e00c2c4	278	if (ret > 0) {
7e77dd02	279	fio_libaio_queued(td, io_us, ret);
838bc709	280	io_u_mark_submit(td, ret);
acd45e02 JA	281
	282	ld->queued -= ret;
	283	ring_inc(ld, &ld->tail, ret);
5e00c2c4	284	ret = 0;
e3b4e568	285	wait_start = 0;
acd45e02	286	} else if (ret == -EINTR \|\| !ret) {
838bc709 JA	287	if (!ret)
838bc709 JA	288	io_u_mark_submit(td, ret);
e3b4e568	289	wait_start = 0;
2866c82d	290	continue;
acd45e02 JA	291	} else if (ret == -EAGAIN) {
	292	/*
	293	* If we get EAGAIN, we should break out without
	294	* error and let the upper layer reap some
a120ca7f JA	295	* events for us. If we have no queued IO, we
	296	* must loop here. If we loop for more than 30s,
	297	* just error out, something must be buggy in the
	298	* IO path.
acd45e02	299	*/
a120ca7f JA	300	if (ld->queued) {
	301	ret = 0;
	302	break;
	303	}
	304	if (!wait_start) {
8b6a404c	305	fio_gettime(&ts, NULL);
d36b072d	306	wait_start = 1;
8b6a404c	307	} else if (mtime_since_now(&ts) > 30000) {
a120ca7f JA	308	log_err("fio: aio appears to be stalled, giving up\n");
	309	break;
	310	}
	311	usleep(1);
	312	continue;
a31dc2dd JA	313	} else if (ret == -ENOMEM) {
	314	/*
	315	* If we get -ENOMEM, reap events if we can. If
	316	* we cannot, treat it as a fatal event since there's
	317	* nothing we can do about it.
	318	*/
	319	if (ld->queued)
	320	ret = 0;
	321	break;
838bc709	322	} else
2866c82d	323	break;
2c3a4ae9	324	} while (ld->queued);
2866c82d	325
36167d82	326	return ret;
2866c82d JA	327	}
	328
	329	static int fio_libaio_cancel(struct thread_data td, struct io_u io_u)
	330	{
565e784d	331	struct libaio_data *ld = td->io_ops_data;
2866c82d JA	332
	333	return io_cancel(ld->aio_ctx, &io_u->iocb, ld->aio_events);
	334	}
	335
	336	static void fio_libaio_cleanup(struct thread_data *td)
	337	{
565e784d	338	struct libaio_data *ld = td->io_ops_data;
2866c82d JA	339
2866c82d JA	340	if (ld) {
f24c2649 JA	341	/*
	342	* Work-around to avoid huge RCU stalls at exit time. If we
	343	* don't do this here, then it'll be torn down by exit_aio().
	344	* But for that case we can parallellize the freeing, thus
	345	* speeding it up a lot.
	346	*/
	347	if (!(td->flags & TD_F_CHILD))
	348	io_destroy(ld->aio_ctx);
7e77dd02 JA	349	free(ld->aio_events);
	350	free(ld->iocbs);
	351	free(ld->io_us);
2866c82d	352	free(ld);
2866c82d JA	353	}
	354	}
	355
	356	static int fio_libaio_init(struct thread_data *td)
	357	{
c90e1017	358	struct libaio_options *o = td->eo;
acd45e02	359	struct libaio_data *ld;
c90e1017	360	int err = 0;
2866c82d	361
acd45e02	362	ld = calloc(1, sizeof(*ld));
c1db2dce	363
c90e1017 JA	364	/*
	365	* First try passing in 0 for queue depth, since we don't
	366	* care about the user ring. If that fails, the kernel is too old
	367	* and we need the right depth.
	368	*/
	369	if (!o->userspace_reap)
c224ec05	370	err = io_queue_init(INT_MAX, &ld->aio_ctx);
c90e1017 JA	371	if (o->userspace_reap \|\| err == -EINVAL)
c90e1017 JA	372	err = io_queue_init(td->o.iodepth, &ld->aio_ctx);
c1db2dce JA	373	if (err) {
c1db2dce JA	374	td_verror(td, -err, "io_queue_init");
75de55ac	375	log_err("fio: check /proc/sys/fs/aio-max-nr\n");
cb781c75	376	free(ld);
2866c82d JA	377	return 1;
	378	}
	379
acd45e02 JA	380	ld->entries = td->o.iodepth;
	381	ld->is_pow2 = is_power_of_2(ld->entries);
	382	ld->aio_events = calloc(ld->entries, sizeof(struct io_event));
	383	ld->iocbs = calloc(ld->entries, sizeof(struct iocb *));
	384	ld->io_us = calloc(ld->entries, sizeof(struct io_u *));
755200a3	385
565e784d	386	td->io_ops_data = ld;
2866c82d JA	387	return 0;
	388	}
	389
5f350952	390	static struct ioengine_ops ioengine = {
de890a1e SL	391	.name = "libaio",
	392	.version = FIO_IOOPS_VERSION,
	393	.init = fio_libaio_init,
	394	.prep = fio_libaio_prep,
	395	.queue = fio_libaio_queue,
	396	.commit = fio_libaio_commit,
	397	.cancel = fio_libaio_cancel,
	398	.getevents = fio_libaio_getevents,
	399	.event = fio_libaio_event,
	400	.cleanup = fio_libaio_cleanup,
	401	.open_file = generic_open_file,
	402	.close_file = generic_close_file,
	403	.get_file_size = generic_get_file_size,
	404	.options = options,
	405	.option_struct_size = sizeof(struct libaio_options),
2866c82d	406	};
34cfcdaf	407
5f350952 JA	408	static void fio_init fio_libaio_register(void)
	409	{
	410	register_ioengine(&ioengine);
	411	}
	412
	413	static void fio_exit fio_libaio_unregister(void)
	414	{
	415	unregister_ioengine(&ioengine);
	416	}