[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"

#ifndef IOCB_FLAG_HIPRI
#define IOCB_FLAG_HIPRI	(1 << 2)
#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;

	/*
	 * 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;

	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.u.c.flags |= IOCB_FLAG_HIPRI;
	} 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.u.c.flags |= IOCB_FLAG_HIPRI;
	} 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
ee636f3f JA	16	#ifndef IOCB_FLAG_HIPRI
	17	#define IOCB_FLAG_HIPRI (1 << 2)
	18	#endif
a391d73d	19
0fc2e103 JA	20	static int fio_libaio_commit(struct thread_data *td);
0fc2e103 JA	21
2866c82d JA	22	struct libaio_data {
	23	io_context_t aio_ctx;
	24	struct io_event *aio_events;
755200a3	25	struct iocb **iocbs;
7e77dd02	26	struct io_u **io_us;
acd45e02 JA	27
	28	/*
	29	* Basic ring buffer. 'head' is incremented in _queue(), and
	30	* 'tail' is incremented in _commit(). We keep 'queued' so
	31	* that we know if the ring is full or empty, when
	32	* 'head' == 'tail'. 'entries' is the ring size, and
	33	* 'is_pow2' is just an optimization to use AND instead of
	34	* modulus to get the remainder on ring increment.
	35	*/
	36	int is_pow2;
	37	unsigned int entries;
	38	unsigned int queued;
	39	unsigned int head;
	40	unsigned int tail;
2866c82d JA	41	};
2866c82d JA	42
de890a1e	43	struct libaio_options {
a1f871c7	44	void *pad;
de890a1e	45	unsigned int userspace_reap;
a391d73d	46	unsigned int hipri;
de890a1e SL	47	};
	48
	49	static struct fio_option options[] = {
	50	{
	51	.name = "userspace_reap",
e8b0e958	52	.lname = "Libaio userspace reaping",
de890a1e SL	53	.type = FIO_OPT_STR_SET,
	54	.off1 = offsetof(struct libaio_options, userspace_reap),
	55	.help = "Use alternative user-space reap implementation",
e90a0adf	56	.category = FIO_OPT_C_ENGINE,
5a3cd5f3	57	.group = FIO_OPT_G_LIBAIO,
de890a1e	58	},
a391d73d JA	59	{
a391d73d JA	60	.name = "hipri",
ee636f3f	61	.lname = "High Priority",
a391d73d JA	62	.type = FIO_OPT_STR_SET,
a391d73d JA	63	.off1 = offsetof(struct libaio_options, hipri),
6acc46d8	64	.help = "Use polled IO completions",
a391d73d JA	65	.category = FIO_OPT_C_ENGINE,
	66	.group = FIO_OPT_G_LIBAIO,
	67	},
de890a1e SL	68	{
	69	.name = NULL,
	70	},
	71	};
	72
acd45e02 JA	73	static inline void ring_inc(struct libaio_data ld, unsigned int val,
	74	unsigned int add)
	75	{
	76	if (ld->is_pow2)
	77	val = (val + add) & (ld->entries - 1);
	78	else
	79	val = (val + add) % ld->entries;
	80	}
	81
7a16dd02	82	static int fio_libaio_prep(struct thread_data fio_unused td, struct io_u io_u)
2866c82d	83	{
53cdc686	84	struct fio_file *f = io_u->file;
a391d73d	85	struct libaio_options *o = td->eo;
53cdc686	86
a391d73d	87	if (io_u->ddir == DDIR_READ) {
cec6b55d	88	io_prep_pread(&io_u->iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
a391d73d	89	if (o->hipri)
ee636f3f	90	io_u->iocb.u.c.flags \|= IOCB_FLAG_HIPRI;
a391d73d	91	} else if (io_u->ddir == DDIR_WRITE) {
cec6b55d	92	io_prep_pwrite(&io_u->iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
a391d73d	93	if (o->hipri)
ee636f3f	94	io_u->iocb.u.c.flags \|= IOCB_FLAG_HIPRI;
a391d73d	95	} else if (ddir_sync(io_u->ddir))
87dc1ab1	96	io_prep_fsync(&io_u->iocb, f->fd);
2866c82d JA	97
	98	return 0;
	99	}
	100
	101	static struct io_u fio_libaio_event(struct thread_data td, int event)
	102	{
565e784d	103	struct libaio_data *ld = td->io_ops_data;
f423479d JA	104	struct io_event *ev;
f423479d JA	105	struct io_u *io_u;
2866c82d	106
f423479d	107	ev = ld->aio_events + event;
0cc6ee59	108	io_u = container_of(ev->obj, struct io_u, iocb);
f423479d JA	109
	110	if (ev->res != io_u->xfer_buflen) {
	111	if (ev->res > io_u->xfer_buflen)
	112	io_u->error = -ev->res;
	113	else
	114	io_u->resid = io_u->xfer_buflen - ev->res;
	115	} else
	116	io_u->error = 0;
	117
	118	return io_u;
2866c82d JA	119	}
2866c82d JA	120
675012f0 DE	121	struct aio_ring {
	122	unsigned id; /** kernel internal index number */
	123	unsigned nr; /** number of io_events */
	124	unsigned head;
	125	unsigned tail;
c44b1ff5	126
675012f0 DE	127	unsigned magic;
	128	unsigned compat_features;
	129	unsigned incompat_features;
	130	unsigned header_length; /** size of aio_ring */
	131
	132	struct io_event events[0];
	133	};
	134
	135	#define AIO_RING_MAGIC 0xa10a10a1
	136
	137	static int user_io_getevents(io_context_t aio_ctx, unsigned int max,
c44b1ff5	138	struct io_event *events)
675012f0 DE	139	{
	140	long i = 0;
	141	unsigned head;
c44b1ff5	142	struct aio_ring ring = (struct aio_ring) aio_ctx;
675012f0 DE	143
	144	while (i < max) {
	145	head = ring->head;
	146
	147	if (head == ring->tail) {
	148	/* There are no more completions */
	149	break;
	150	} else {
	151	/* There is another completion to reap */
	152	events[i] = ring->events[head];
	153	read_barrier();
c44b1ff5	154	ring->head = (head + 1) % ring->nr;
675012f0 DE	155	i++;
	156	}
	157	}
	158
	159	return i;
	160	}
	161
e7d2e616	162	static int fio_libaio_getevents(struct thread_data *td, unsigned int min,
1f440ece	163	unsigned int max, const struct timespec *t)
2866c82d	164	{
565e784d	165	struct libaio_data *ld = td->io_ops_data;
de890a1e	166	struct libaio_options *o = td->eo;
82407585	167	unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
1f440ece	168	struct timespec __lt, *lt = NULL;
0b7fdba7	169	int r, events = 0;
2866c82d	170
1f440ece JA	171	if (t) {
	172	__lt = *t;
	173	lt = &__lt;
	174	}
	175
2866c82d	176	do {
de890a1e	177	if (o->userspace_reap == 1
675012f0 DE	178	&& actual_min == 0
	179	&& ((struct aio_ring *)(ld->aio_ctx))->magic
	180	== AIO_RING_MAGIC) {
	181	r = user_io_getevents(ld->aio_ctx, max,
	182	ld->aio_events + events);
	183	} else {
	184	r = io_getevents(ld->aio_ctx, actual_min,
1f440ece	185	max, ld->aio_events + events, lt);
675012f0	186	}
3441a52d	187	if (r > 0)
0b7fdba7	188	events += r;
3441a52d	189	else if ((min && r == 0) \|\| r == -EAGAIN) {
0fc2e103	190	fio_libaio_commit(td);
6347e43d JA	191	if (actual_min)
6347e43d JA	192	usleep(10);
0fc2e103	193	} else if (r != -EINTR)
a31dc2dd	194	break;
0b7fdba7 DE	195	} while (events < min);
	196
	197	return r < 0 ? r : events;
2866c82d JA	198	}
2866c82d JA	199
2e4ef4fb JA	200	static enum fio_q_status fio_libaio_queue(struct thread_data *td,
2e4ef4fb JA	201	struct io_u *io_u)
2866c82d	202	{
565e784d	203	struct libaio_data *ld = td->io_ops_data;
2866c82d	204
7101d9c2 JA	205	fio_ro_check(td, io_u);
7101d9c2 JA	206
acd45e02	207	if (ld->queued == td->o.iodepth)
755200a3 JA	208	return FIO_Q_BUSY;
	209
	210	/*
	211	* fsync is tricky, since it can fail and we need to do it
	212	* serialized with other io. the reason is that linux doesn't
	213	* support aio fsync yet. So return busy for the case where we
	214	* have pending io, to let fio complete those first.
	215	*/
f011531e	216	if (ddir_sync(io_u->ddir)) {
acd45e02	217	if (ld->queued)
755200a3	218	return FIO_Q_BUSY;
5f9099ea	219
f011531e	220	do_io_u_sync(td, io_u);
755200a3 JA	221	return FIO_Q_COMPLETED;
	222	}
	223
a5f3027c	224	if (io_u->ddir == DDIR_TRIM) {
acd45e02	225	if (ld->queued)
a5f3027c JA	226	return FIO_Q_BUSY;
	227
	228	do_io_u_trim(td, io_u);
c0681c9d VF	229	io_u_mark_submit(td, 1);
c0681c9d VF	230	io_u_mark_complete(td, 1);
a5f3027c JA	231	return FIO_Q_COMPLETED;
	232	}
	233
acd45e02 JA	234	ld->iocbs[ld->head] = &io_u->iocb;
	235	ld->io_us[ld->head] = io_u;
	236	ring_inc(ld, &ld->head, 1);
	237	ld->queued++;
755200a3 JA	238	return FIO_Q_QUEUED;
	239	}
	240
7e77dd02 JA	241	static void fio_libaio_queued(struct thread_data td, struct io_u *io_us,
	242	unsigned int nr)
	243	{
8b6a404c	244	struct timespec now;
7e77dd02 JA	245	unsigned int i;
7e77dd02 JA	246
12d9d841 JA	247	if (!fio_fill_issue_time(td))
	248	return;
	249
7e77dd02 JA	250	fio_gettime(&now, NULL);
	251
	252	for (i = 0; i < nr; i++) {
	253	struct io_u *io_u = io_us[i];
	254
	255	memcpy(&io_u->issue_time, &now, sizeof(now));
	256	io_u_queued(td, io_u);
	257	}
	258	}
	259
755200a3 JA	260	static int fio_libaio_commit(struct thread_data *td)
755200a3 JA	261	{
565e784d	262	struct libaio_data *ld = td->io_ops_data;
755200a3	263	struct iocb **iocbs;
7e77dd02	264	struct io_u **io_us;
8b6a404c	265	struct timespec ts;
a120ca7f	266	int ret, wait_start = 0;
755200a3	267
acd45e02	268	if (!ld->queued)
755200a3 JA	269	return 0;
755200a3 JA	270
2866c82d	271	do {
acd45e02 JA	272	long nr = ld->queued;
	273
	274	nr = min((unsigned int) nr, ld->entries - ld->tail);
	275	io_us = ld->io_us + ld->tail;
	276	iocbs = ld->iocbs + ld->tail;
	277
	278	ret = io_submit(ld->aio_ctx, nr, iocbs);
5e00c2c4	279	if (ret > 0) {
7e77dd02	280	fio_libaio_queued(td, io_us, ret);
838bc709	281	io_u_mark_submit(td, ret);
acd45e02 JA	282
	283	ld->queued -= ret;
	284	ring_inc(ld, &ld->tail, ret);
5e00c2c4	285	ret = 0;
e3b4e568	286	wait_start = 0;
acd45e02	287	} else if (ret == -EINTR \|\| !ret) {
838bc709 JA	288	if (!ret)
838bc709 JA	289	io_u_mark_submit(td, ret);
e3b4e568	290	wait_start = 0;
2866c82d	291	continue;
acd45e02 JA	292	} else if (ret == -EAGAIN) {
	293	/*
	294	* If we get EAGAIN, we should break out without
	295	* error and let the upper layer reap some
a120ca7f JA	296	* events for us. If we have no queued IO, we
	297	* must loop here. If we loop for more than 30s,
	298	* just error out, something must be buggy in the
	299	* IO path.
acd45e02	300	*/
a120ca7f JA	301	if (ld->queued) {
	302	ret = 0;
	303	break;
	304	}
	305	if (!wait_start) {
8b6a404c	306	fio_gettime(&ts, NULL);
d36b072d	307	wait_start = 1;
8b6a404c	308	} else if (mtime_since_now(&ts) > 30000) {
a120ca7f JA	309	log_err("fio: aio appears to be stalled, giving up\n");
	310	break;
	311	}
	312	usleep(1);
	313	continue;
a31dc2dd JA	314	} else if (ret == -ENOMEM) {
	315	/*
	316	* If we get -ENOMEM, reap events if we can. If
	317	* we cannot, treat it as a fatal event since there's
	318	* nothing we can do about it.
	319	*/
	320	if (ld->queued)
	321	ret = 0;
	322	break;
838bc709	323	} else
2866c82d	324	break;
2c3a4ae9	325	} while (ld->queued);
2866c82d	326
36167d82	327	return ret;
2866c82d JA	328	}
	329
	330	static int fio_libaio_cancel(struct thread_data td, struct io_u io_u)
	331	{
565e784d	332	struct libaio_data *ld = td->io_ops_data;
2866c82d JA	333
	334	return io_cancel(ld->aio_ctx, &io_u->iocb, ld->aio_events);
	335	}
	336
	337	static void fio_libaio_cleanup(struct thread_data *td)
	338	{
565e784d	339	struct libaio_data *ld = td->io_ops_data;
2866c82d JA	340
2866c82d JA	341	if (ld) {
f24c2649 JA	342	/*
	343	* Work-around to avoid huge RCU stalls at exit time. If we
	344	* don't do this here, then it'll be torn down by exit_aio().
	345	* But for that case we can parallellize the freeing, thus
	346	* speeding it up a lot.
	347	*/
	348	if (!(td->flags & TD_F_CHILD))
	349	io_destroy(ld->aio_ctx);
7e77dd02 JA	350	free(ld->aio_events);
	351	free(ld->iocbs);
	352	free(ld->io_us);
2866c82d	353	free(ld);
2866c82d JA	354	}
	355	}
	356
	357	static int fio_libaio_init(struct thread_data *td)
	358	{
c90e1017	359	struct libaio_options *o = td->eo;
acd45e02	360	struct libaio_data *ld;
c90e1017	361	int err = 0;
2866c82d	362
acd45e02	363	ld = calloc(1, sizeof(*ld));
c1db2dce	364
c90e1017 JA	365	/*
	366	* First try passing in 0 for queue depth, since we don't
	367	* care about the user ring. If that fails, the kernel is too old
	368	* and we need the right depth.
	369	*/
	370	if (!o->userspace_reap)
c224ec05	371	err = io_queue_init(INT_MAX, &ld->aio_ctx);
c90e1017 JA	372	if (o->userspace_reap \|\| err == -EINVAL)
c90e1017 JA	373	err = io_queue_init(td->o.iodepth, &ld->aio_ctx);
c1db2dce JA	374	if (err) {
c1db2dce JA	375	td_verror(td, -err, "io_queue_init");
75de55ac	376	log_err("fio: check /proc/sys/fs/aio-max-nr\n");
cb781c75	377	free(ld);
2866c82d JA	378	return 1;
	379	}
	380
acd45e02 JA	381	ld->entries = td->o.iodepth;
	382	ld->is_pow2 = is_power_of_2(ld->entries);
	383	ld->aio_events = calloc(ld->entries, sizeof(struct io_event));
	384	ld->iocbs = calloc(ld->entries, sizeof(struct iocb *));
	385	ld->io_us = calloc(ld->entries, sizeof(struct io_u *));
755200a3	386
565e784d	387	td->io_ops_data = ld;
2866c82d JA	388	return 0;
	389	}
	390
5f350952	391	static struct ioengine_ops ioengine = {
de890a1e SL	392	.name = "libaio",
	393	.version = FIO_IOOPS_VERSION,
	394	.init = fio_libaio_init,
	395	.prep = fio_libaio_prep,
	396	.queue = fio_libaio_queue,
	397	.commit = fio_libaio_commit,
	398	.cancel = fio_libaio_cancel,
	399	.getevents = fio_libaio_getevents,
	400	.event = fio_libaio_event,
	401	.cleanup = fio_libaio_cleanup,
	402	.open_file = generic_open_file,
	403	.close_file = generic_close_file,
	404	.get_file_size = generic_get_file_size,
	405	.options = options,
	406	.option_struct_size = sizeof(struct libaio_options),
2866c82d	407	};
34cfcdaf	408
5f350952 JA	409	static void fio_init fio_libaio_register(void)
	410	{
	411	register_ioengine(&ioengine);
	412	}
	413
	414	static void fio_exit fio_libaio_unregister(void)
	415	{
	416	unregister_ioengine(&ioengine);
	417	}