[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
#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;

	/*
	 * 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_queue_init(struct libaio_data *ld, unsigned int depth,
				 bool hipri)
{
#ifdef __NR_sys_io_setup2
	int flags = 0;

	if (hipri)
		flags = IOCTX_FLAG_IOPOLL;

	return syscall(__NR_sys_io_setup2, depth, flags, &ld->aio_ctx);
#else
	if (hipri) {
		log_err("fio: polled aio not available on your platform\n");
		return 1;
	}
	return io_queue_init(depth, &ld->aio_ctx);
#endif
}

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.
	 */
	err = fio_libaio_queue_init(ld, td->o.iodepth, o->hipri);
	if (err) {
		td_verror(td, -err, "io_queue_init");
		log_err("fio: check /proc/sys/fs/aio-max-nr\n");
		free(ld);
		return 1;
	}

	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
ebec344d JA	19	#ifndef IOCTX_FLAG_IOPOLL
	20	#define IOCTX_FLAG_IOPOLL (1 << 0)
	21	#endif
a391d73d	22
0fc2e103 JA	23	static int fio_libaio_commit(struct thread_data *td);
0fc2e103 JA	24
2866c82d JA	25	struct libaio_data {
	26	io_context_t aio_ctx;
	27	struct io_event *aio_events;
755200a3	28	struct iocb **iocbs;
7e77dd02	29	struct io_u **io_us;
acd45e02 JA	30
	31	/*
	32	* Basic ring buffer. 'head' is incremented in _queue(), and
	33	* 'tail' is incremented in _commit(). We keep 'queued' so
	34	* that we know if the ring is full or empty, when
	35	* 'head' == 'tail'. 'entries' is the ring size, and
	36	* 'is_pow2' is just an optimization to use AND instead of
	37	* modulus to get the remainder on ring increment.
	38	*/
	39	int is_pow2;
	40	unsigned int entries;
	41	unsigned int queued;
	42	unsigned int head;
	43	unsigned int tail;
2866c82d JA	44	};
2866c82d JA	45
de890a1e	46	struct libaio_options {
a1f871c7	47	void *pad;
de890a1e	48	unsigned int userspace_reap;
a391d73d	49	unsigned int hipri;
de890a1e SL	50	};
	51
	52	static struct fio_option options[] = {
	53	{
	54	.name = "userspace_reap",
e8b0e958	55	.lname = "Libaio userspace reaping",
de890a1e SL	56	.type = FIO_OPT_STR_SET,
	57	.off1 = offsetof(struct libaio_options, userspace_reap),
	58	.help = "Use alternative user-space reap implementation",
e90a0adf	59	.category = FIO_OPT_C_ENGINE,
5a3cd5f3	60	.group = FIO_OPT_G_LIBAIO,
de890a1e	61	},
a391d73d JA	62	{
a391d73d JA	63	.name = "hipri",
ee636f3f	64	.lname = "High Priority",
a391d73d JA	65	.type = FIO_OPT_STR_SET,
a391d73d JA	66	.off1 = offsetof(struct libaio_options, hipri),
6acc46d8	67	.help = "Use polled IO completions",
a391d73d JA	68	.category = FIO_OPT_C_ENGINE,
	69	.group = FIO_OPT_G_LIBAIO,
	70	},
de890a1e SL	71	{
	72	.name = NULL,
	73	},
	74	};
	75
acd45e02 JA	76	static inline void ring_inc(struct libaio_data ld, unsigned int val,
	77	unsigned int add)
	78	{
	79	if (ld->is_pow2)
	80	val = (val + add) & (ld->entries - 1);
	81	else
	82	val = (val + add) % ld->entries;
	83	}
	84
7a16dd02	85	static int fio_libaio_prep(struct thread_data fio_unused td, struct io_u io_u)
2866c82d	86	{
53cdc686	87	struct fio_file *f = io_u->file;
a391d73d	88	struct libaio_options *o = td->eo;
53cdc686	89
a391d73d	90	if (io_u->ddir == DDIR_READ) {
cec6b55d	91	io_prep_pread(&io_u->iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
a391d73d	92	if (o->hipri)
ee636f3f	93	io_u->iocb.u.c.flags \|= IOCB_FLAG_HIPRI;
a391d73d	94	} else if (io_u->ddir == DDIR_WRITE) {
cec6b55d	95	io_prep_pwrite(&io_u->iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
a391d73d	96	if (o->hipri)
ee636f3f	97	io_u->iocb.u.c.flags \|= IOCB_FLAG_HIPRI;
a391d73d	98	} else if (ddir_sync(io_u->ddir))
87dc1ab1	99	io_prep_fsync(&io_u->iocb, f->fd);
2866c82d JA	100
	101	return 0;
	102	}
	103
	104	static struct io_u fio_libaio_event(struct thread_data td, int event)
	105	{
565e784d	106	struct libaio_data *ld = td->io_ops_data;
f423479d JA	107	struct io_event *ev;
f423479d JA	108	struct io_u *io_u;
2866c82d	109
f423479d	110	ev = ld->aio_events + event;
0cc6ee59	111	io_u = container_of(ev->obj, struct io_u, iocb);
f423479d JA	112
	113	if (ev->res != io_u->xfer_buflen) {
	114	if (ev->res > io_u->xfer_buflen)
	115	io_u->error = -ev->res;
	116	else
	117	io_u->resid = io_u->xfer_buflen - ev->res;
	118	} else
	119	io_u->error = 0;
	120
	121	return io_u;
2866c82d JA	122	}
2866c82d JA	123
675012f0 DE	124	struct aio_ring {
	125	unsigned id; /** kernel internal index number */
	126	unsigned nr; /** number of io_events */
	127	unsigned head;
	128	unsigned tail;
c44b1ff5	129
675012f0 DE	130	unsigned magic;
	131	unsigned compat_features;
	132	unsigned incompat_features;
	133	unsigned header_length; /** size of aio_ring */
	134
	135	struct io_event events[0];
	136	};
	137
	138	#define AIO_RING_MAGIC 0xa10a10a1
	139
	140	static int user_io_getevents(io_context_t aio_ctx, unsigned int max,
c44b1ff5	141	struct io_event *events)
675012f0 DE	142	{
	143	long i = 0;
	144	unsigned head;
c44b1ff5	145	struct aio_ring ring = (struct aio_ring) aio_ctx;
675012f0 DE	146
	147	while (i < max) {
	148	head = ring->head;
	149
	150	if (head == ring->tail) {
	151	/* There are no more completions */
	152	break;
	153	} else {
	154	/* There is another completion to reap */
	155	events[i] = ring->events[head];
	156	read_barrier();
c44b1ff5	157	ring->head = (head + 1) % ring->nr;
675012f0 DE	158	i++;
	159	}
	160	}
	161
	162	return i;
	163	}
	164
e7d2e616	165	static int fio_libaio_getevents(struct thread_data *td, unsigned int min,
1f440ece	166	unsigned int max, const struct timespec *t)
2866c82d	167	{
565e784d	168	struct libaio_data *ld = td->io_ops_data;
de890a1e	169	struct libaio_options *o = td->eo;
82407585	170	unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
1f440ece	171	struct timespec __lt, *lt = NULL;
0b7fdba7	172	int r, events = 0;
2866c82d	173
1f440ece JA	174	if (t) {
	175	__lt = *t;
	176	lt = &__lt;
	177	}
	178
2866c82d	179	do {
de890a1e	180	if (o->userspace_reap == 1
675012f0 DE	181	&& actual_min == 0
	182	&& ((struct aio_ring *)(ld->aio_ctx))->magic
	183	== AIO_RING_MAGIC) {
	184	r = user_io_getevents(ld->aio_ctx, max,
	185	ld->aio_events + events);
	186	} else {
	187	r = io_getevents(ld->aio_ctx, actual_min,
1f440ece	188	max, ld->aio_events + events, lt);
675012f0	189	}
3441a52d	190	if (r > 0)
0b7fdba7	191	events += r;
3441a52d	192	else if ((min && r == 0) \|\| r == -EAGAIN) {
0fc2e103	193	fio_libaio_commit(td);
6347e43d JA	194	if (actual_min)
6347e43d JA	195	usleep(10);
0fc2e103	196	} else if (r != -EINTR)
a31dc2dd	197	break;
0b7fdba7 DE	198	} while (events < min);
	199
	200	return r < 0 ? r : events;
2866c82d JA	201	}
2866c82d JA	202
2e4ef4fb JA	203	static enum fio_q_status fio_libaio_queue(struct thread_data *td,
2e4ef4fb JA	204	struct io_u *io_u)
2866c82d	205	{
565e784d	206	struct libaio_data *ld = td->io_ops_data;
2866c82d	207
7101d9c2 JA	208	fio_ro_check(td, io_u);
7101d9c2 JA	209
acd45e02	210	if (ld->queued == td->o.iodepth)
755200a3 JA	211	return FIO_Q_BUSY;
	212
	213	/*
	214	* fsync is tricky, since it can fail and we need to do it
	215	* serialized with other io. the reason is that linux doesn't
	216	* support aio fsync yet. So return busy for the case where we
	217	* have pending io, to let fio complete those first.
	218	*/
f011531e	219	if (ddir_sync(io_u->ddir)) {
acd45e02	220	if (ld->queued)
755200a3	221	return FIO_Q_BUSY;
5f9099ea	222
f011531e	223	do_io_u_sync(td, io_u);
755200a3 JA	224	return FIO_Q_COMPLETED;
	225	}
	226
a5f3027c	227	if (io_u->ddir == DDIR_TRIM) {
acd45e02	228	if (ld->queued)
a5f3027c JA	229	return FIO_Q_BUSY;
	230
	231	do_io_u_trim(td, io_u);
c0681c9d VF	232	io_u_mark_submit(td, 1);
c0681c9d VF	233	io_u_mark_complete(td, 1);
a5f3027c JA	234	return FIO_Q_COMPLETED;
	235	}
	236
acd45e02 JA	237	ld->iocbs[ld->head] = &io_u->iocb;
	238	ld->io_us[ld->head] = io_u;
	239	ring_inc(ld, &ld->head, 1);
	240	ld->queued++;
755200a3 JA	241	return FIO_Q_QUEUED;
	242	}
	243
7e77dd02 JA	244	static void fio_libaio_queued(struct thread_data td, struct io_u *io_us,
	245	unsigned int nr)
	246	{
8b6a404c	247	struct timespec now;
7e77dd02 JA	248	unsigned int i;
7e77dd02 JA	249
12d9d841 JA	250	if (!fio_fill_issue_time(td))
	251	return;
	252
7e77dd02 JA	253	fio_gettime(&now, NULL);
	254
	255	for (i = 0; i < nr; i++) {
	256	struct io_u *io_u = io_us[i];
	257
	258	memcpy(&io_u->issue_time, &now, sizeof(now));
	259	io_u_queued(td, io_u);
	260	}
	261	}
	262
755200a3 JA	263	static int fio_libaio_commit(struct thread_data *td)
755200a3 JA	264	{
565e784d	265	struct libaio_data *ld = td->io_ops_data;
755200a3	266	struct iocb **iocbs;
7e77dd02	267	struct io_u **io_us;
8b6a404c	268	struct timespec ts;
a120ca7f	269	int ret, wait_start = 0;
755200a3	270
acd45e02	271	if (!ld->queued)
755200a3 JA	272	return 0;
755200a3 JA	273
2866c82d	274	do {
acd45e02 JA	275	long nr = ld->queued;
	276
	277	nr = min((unsigned int) nr, ld->entries - ld->tail);
	278	io_us = ld->io_us + ld->tail;
	279	iocbs = ld->iocbs + ld->tail;
	280
	281	ret = io_submit(ld->aio_ctx, nr, iocbs);
5e00c2c4	282	if (ret > 0) {
7e77dd02	283	fio_libaio_queued(td, io_us, ret);
838bc709	284	io_u_mark_submit(td, ret);
acd45e02 JA	285
	286	ld->queued -= ret;
	287	ring_inc(ld, &ld->tail, ret);
5e00c2c4	288	ret = 0;
e3b4e568	289	wait_start = 0;
acd45e02	290	} else if (ret == -EINTR \|\| !ret) {
838bc709 JA	291	if (!ret)
838bc709 JA	292	io_u_mark_submit(td, ret);
e3b4e568	293	wait_start = 0;
2866c82d	294	continue;
acd45e02 JA	295	} else if (ret == -EAGAIN) {
	296	/*
	297	* If we get EAGAIN, we should break out without
	298	* error and let the upper layer reap some
a120ca7f JA	299	* events for us. If we have no queued IO, we
	300	* must loop here. If we loop for more than 30s,
	301	* just error out, something must be buggy in the
	302	* IO path.
acd45e02	303	*/
a120ca7f JA	304	if (ld->queued) {
	305	ret = 0;
	306	break;
	307	}
	308	if (!wait_start) {
8b6a404c	309	fio_gettime(&ts, NULL);
d36b072d	310	wait_start = 1;
8b6a404c	311	} else if (mtime_since_now(&ts) > 30000) {
a120ca7f JA	312	log_err("fio: aio appears to be stalled, giving up\n");
	313	break;
	314	}
	315	usleep(1);
	316	continue;
a31dc2dd JA	317	} else if (ret == -ENOMEM) {
	318	/*
	319	* If we get -ENOMEM, reap events if we can. If
	320	* we cannot, treat it as a fatal event since there's
	321	* nothing we can do about it.
	322	*/
	323	if (ld->queued)
	324	ret = 0;
	325	break;
838bc709	326	} else
2866c82d	327	break;
2c3a4ae9	328	} while (ld->queued);
2866c82d	329
36167d82	330	return ret;
2866c82d JA	331	}
	332
	333	static int fio_libaio_cancel(struct thread_data td, struct io_u io_u)
	334	{
565e784d	335	struct libaio_data *ld = td->io_ops_data;
2866c82d JA	336
	337	return io_cancel(ld->aio_ctx, &io_u->iocb, ld->aio_events);
	338	}
	339
	340	static void fio_libaio_cleanup(struct thread_data *td)
	341	{
565e784d	342	struct libaio_data *ld = td->io_ops_data;
2866c82d JA	343
2866c82d JA	344	if (ld) {
f24c2649 JA	345	/*
	346	* Work-around to avoid huge RCU stalls at exit time. If we
	347	* don't do this here, then it'll be torn down by exit_aio().
	348	* But for that case we can parallellize the freeing, thus
	349	* speeding it up a lot.
	350	*/
	351	if (!(td->flags & TD_F_CHILD))
	352	io_destroy(ld->aio_ctx);
7e77dd02 JA	353	free(ld->aio_events);
	354	free(ld->iocbs);
	355	free(ld->io_us);
2866c82d	356	free(ld);
2866c82d JA	357	}
	358	}
	359
ebec344d JA	360	static int fio_libaio_queue_init(struct libaio_data *ld, unsigned int depth,
	361	bool hipri)
	362	{
	363	#ifdef __NR_sys_io_setup2
a1b006fe	364	int flags = 0;
ebec344d JA	365
	366	if (hipri)
	367	flags = IOCTX_FLAG_IOPOLL;
	368
	369	return syscall(__NR_sys_io_setup2, depth, flags, &ld->aio_ctx);
	370	#else
	371	if (hipri) {
	372	log_err("fio: polled aio not available on your platform\n");
	373	return 1;
	374	}
	375	return io_queue_init(depth, &ld->aio_ctx);
	376	#endif
	377	}
	378
2866c82d JA	379	static int fio_libaio_init(struct thread_data *td)
2866c82d JA	380	{
c90e1017	381	struct libaio_options *o = td->eo;
acd45e02	382	struct libaio_data *ld;
c90e1017	383	int err = 0;
2866c82d	384
acd45e02	385	ld = calloc(1, sizeof(*ld));
c1db2dce	386
c90e1017 JA	387	/*
	388	* First try passing in 0 for queue depth, since we don't
	389	* care about the user ring. If that fails, the kernel is too old
	390	* and we need the right depth.
	391	*/
a1b006fe	392	err = fio_libaio_queue_init(ld, td->o.iodepth, o->hipri);
c1db2dce JA	393	if (err) {
c1db2dce JA	394	td_verror(td, -err, "io_queue_init");
75de55ac	395	log_err("fio: check /proc/sys/fs/aio-max-nr\n");
cb781c75	396	free(ld);
2866c82d JA	397	return 1;
	398	}
	399
acd45e02 JA	400	ld->entries = td->o.iodepth;
	401	ld->is_pow2 = is_power_of_2(ld->entries);
	402	ld->aio_events = calloc(ld->entries, sizeof(struct io_event));
	403	ld->iocbs = calloc(ld->entries, sizeof(struct iocb *));
	404	ld->io_us = calloc(ld->entries, sizeof(struct io_u *));
755200a3	405
565e784d	406	td->io_ops_data = ld;
2866c82d JA	407	return 0;
	408	}
	409
5f350952	410	static struct ioengine_ops ioengine = {
de890a1e SL	411	.name = "libaio",
	412	.version = FIO_IOOPS_VERSION,
	413	.init = fio_libaio_init,
	414	.prep = fio_libaio_prep,
	415	.queue = fio_libaio_queue,
	416	.commit = fio_libaio_commit,
	417	.cancel = fio_libaio_cancel,
	418	.getevents = fio_libaio_getevents,
	419	.event = fio_libaio_event,
	420	.cleanup = fio_libaio_cleanup,
	421	.open_file = generic_open_file,
	422	.close_file = generic_close_file,
	423	.get_file_size = generic_get_file_size,
	424	.options = options,
	425	.option_struct_size = sizeof(struct libaio_options),
2866c82d	426	};
34cfcdaf	427
5f350952 JA	428	static void fio_init fio_libaio_register(void)
	429	{
	430	register_ioengine(&ioengine);
	431	}
	432
	433	static void fio_exit fio_libaio_unregister(void)
	434	{
	435	unregister_ioengine(&ioengine);
	436	}