[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"
#include "cmdprio.h"

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

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

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

struct libaio_data {
	io_context_t aio_ctx;
	struct io_event *aio_events;
	struct iocb **iocbs;
	struct io_u **io_us;

	struct io_u **io_u_index;

	/*
	 * Basic ring buffer. 'head' is incremented in _queue(), and
	 * 'tail' is incremented in _commit(). We keep 'queued' so
	 * that we know if the ring is full or empty, when
	 * 'head' == 'tail'. 'entries' is the ring size, and
	 * 'is_pow2' is just an optimization to use AND instead of
	 * modulus to get the remainder on ring increment.
	 */
	int is_pow2;
	unsigned int entries;
	unsigned int queued;
	unsigned int head;
	unsigned int tail;

	struct cmdprio cmdprio;
};

struct libaio_options {
	struct thread_data *td;
	unsigned int userspace_reap;
	struct cmdprio_options cmdprio_options;
	unsigned int nowait;
};

static struct fio_option options[] = {
	{
		.name	= "userspace_reap",
		.lname	= "Libaio userspace reaping",
		.type	= FIO_OPT_STR_SET,
		.off1	= offsetof(struct libaio_options, userspace_reap),
		.help	= "Use alternative user-space reap implementation",
		.category = FIO_OPT_C_ENGINE,
		.group	= FIO_OPT_G_LIBAIO,
	},
	{
		.name	= "nowait",
		.lname	= "RWF_NOWAIT",
		.type	= FIO_OPT_BOOL,
		.off1	= offsetof(struct libaio_options, nowait),
		.help	= "Set RWF_NOWAIT for reads/writes",
		.category = FIO_OPT_C_ENGINE,
		.group	= FIO_OPT_G_LIBAIO,
	},
	CMDPRIO_OPTIONS(struct libaio_options, FIO_OPT_G_LIBAIO),
	{
		.name	= NULL,
	},
};

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

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

	if (io_u->ddir == DDIR_READ) {
		io_prep_pread(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
		if (o->nowait)
			iocb->aio_rw_flags |= RWF_NOWAIT;
	} else if (io_u->ddir == DDIR_WRITE) {
		io_prep_pwrite(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
		if (o->nowait)
			iocb->aio_rw_flags |= RWF_NOWAIT;
#ifdef FIO_HAVE_RWF_ATOMIC
		if (td->o.oatomic)
			iocb->aio_rw_flags |= RWF_ATOMIC;
#endif
	} else if (ddir_sync(io_u->ddir))
		io_prep_fsync(iocb, f->fd);

	return 0;
}

static inline void fio_libaio_cmdprio_prep(struct thread_data *td,
					   struct io_u *io_u)
{
	struct libaio_data *ld = td->io_ops_data;
	struct cmdprio *cmdprio = &ld->cmdprio;

	if (fio_cmdprio_set_ioprio(td, cmdprio, io_u)) {
		io_u->iocb.aio_reqprio = io_u->ioprio;
		io_u->iocb.u.c.flags |= IOCB_FLAG_IOPRIO;
	}
}

static struct io_u *fio_libaio_event(struct thread_data *td, int event)
{
	struct libaio_data *ld = td->io_ops_data;
	struct io_event *ev;
	struct io_u *io_u;

	ev = ld->aio_events + event;
	io_u = container_of(ev->obj, struct io_u, iocb);

	if (ev->res != io_u->xfer_buflen) {
		if (ev->res > io_u->xfer_buflen)
			io_u->error = -ev->res;
		else
			io_u->resid = io_u->xfer_buflen - ev->res;
	} else
		io_u->error = 0;

	return io_u;
}

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

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

	struct io_event events[0];
};

#define AIO_RING_MAGIC	0xa10a10a1

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

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

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

	return i;
}

static int fio_libaio_getevents(struct thread_data *td, unsigned int min,
				unsigned int max, const struct timespec *t)
{
	struct libaio_data *ld = td->io_ops_data;
	struct libaio_options *o = td->eo;
	unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
	struct timespec __lt, *lt = NULL;
	int r, events = 0;

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

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

	if (ld->cmdprio.mode != CMDPRIO_MODE_NONE)
		fio_libaio_cmdprio_prep(td, io_u);

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

static void fio_libaio_queued(struct thread_data *td, struct io_u **io_us,
			      unsigned int nr)
{
	struct timespec now;
	unsigned int i;

	if (!fio_fill_issue_time(td))
		return;

	fio_gettime(&now, NULL);

	for (i = 0; i < nr; i++) {
		struct io_u *io_u = io_us[i];

		memcpy(&io_u->issue_time, &now, sizeof(now));
		io_u_queued(td, io_u);
	}

	/*
	 * only used for iolog
	 */
	if (td->o.read_iolog_file)
		memcpy(&td->last_issue, &now, sizeof(now));
}

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);

		fio_cmdprio_cleanup(&ld->cmdprio);
		free(ld->aio_events);
		free(ld->iocbs);
		free(ld->io_us);
		free(ld);
	}
}

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

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

	return 0;
}

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

	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;

	ret = fio_cmdprio_init(td, &ld->cmdprio, &o->cmdprio_options);
	if (ret) {
		td_verror(td, EINVAL, "fio_libaio_init");
		return 1;
	}

	return 0;
}

FIO_STATIC struct ioengine_ops ioengine = {
	.name			= "libaio",
	.version		= FIO_IOOPS_VERSION,
	.flags			= FIO_ASYNCIO_SYNC_TRIM |
					FIO_ASYNCIO_SETS_ISSUE_TIME |
					FIO_ATOMICWRITES,
	.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"
e9f6567a	18	#include "cmdprio.h"
2866c82d	19
b2a432bf PC	20	/* Should be defined in newest aio_abi.h */
	21	#ifndef IOCB_FLAG_IOPRIO
	22	#define IOCB_FLAG_IOPRIO (1 << 1)
	23	#endif
	24
7d42e66e KK	25	/* Hack for libaio < 0.3.111 */
	26	#ifndef CONFIG_LIBAIO_RW_FLAGS
	27	#define aio_rw_flags __pad2
	28	#endif
	29
0fc2e103	30	static int fio_libaio_commit(struct thread_data *td);
b2a432bf	31	static int fio_libaio_init(struct thread_data *td);
0fc2e103	32
2866c82d JA	33	struct libaio_data {
	34	io_context_t aio_ctx;
	35	struct io_event *aio_events;
755200a3	36	struct iocb **iocbs;
7e77dd02	37	struct io_u **io_us;
acd45e02	38
3c3168e9 JA	39	struct io_u **io_u_index;
3c3168e9 JA	40
acd45e02 JA	41	/*
	42	* Basic ring buffer. 'head' is incremented in _queue(), and
	43	* 'tail' is incremented in _commit(). We keep 'queued' so
	44	* that we know if the ring is full or empty, when
	45	* 'head' == 'tail'. 'entries' is the ring size, and
	46	* 'is_pow2' is just an optimization to use AND instead of
	47	* modulus to get the remainder on ring increment.
	48	*/
	49	int is_pow2;
	50	unsigned int entries;
	51	unsigned int queued;
	52	unsigned int head;
	53	unsigned int tail;
d6cbeab4 NC	54
d6cbeab4 NC	55	struct cmdprio cmdprio;
2866c82d JA	56	};
2866c82d JA	57
de890a1e	58	struct libaio_options {
a48f0cc7	59	struct thread_data *td;
de890a1e	60	unsigned int userspace_reap;
d6cbeab4	61	struct cmdprio_options cmdprio_options;
7d42e66e	62	unsigned int nowait;
de890a1e SL	63	};
	64
	65	static struct fio_option options[] = {
	66	{
	67	.name = "userspace_reap",
e8b0e958	68	.lname = "Libaio userspace reaping",
de890a1e SL	69	.type = FIO_OPT_STR_SET,
	70	.off1 = offsetof(struct libaio_options, userspace_reap),
	71	.help = "Use alternative user-space reap implementation",
e90a0adf	72	.category = FIO_OPT_C_ENGINE,
5a3cd5f3	73	.group = FIO_OPT_G_LIBAIO,
de890a1e	74	},
7d42e66e KK	75	{
	76	.name = "nowait",
	77	.lname = "RWF_NOWAIT",
	78	.type = FIO_OPT_BOOL,
	79	.off1 = offsetof(struct libaio_options, nowait),
	80	.help = "Set RWF_NOWAIT for reads/writes",
	81	.category = FIO_OPT_C_ENGINE,
	82	.group = FIO_OPT_G_LIBAIO,
	83	},
2838f77a	84	CMDPRIO_OPTIONS(struct libaio_options, FIO_OPT_G_LIBAIO),
de890a1e SL	85	{
	86	.name = NULL,
	87	},
	88	};
	89
acd45e02 JA	90	static inline void ring_inc(struct libaio_data ld, unsigned int val,
	91	unsigned int add)
	92	{
	93	if (ld->is_pow2)
	94	val = (val + add) & (ld->entries - 1);
	95	else
	96	val = (val + add) % ld->entries;
	97	}
	98
7d42e66e	99	static int fio_libaio_prep(struct thread_data td, struct io_u io_u)
2866c82d	100	{
7d42e66e	101	struct libaio_options *o = td->eo;
53cdc686	102	struct fio_file *f = io_u->file;
ad4e9298	103	struct iocb *iocb = &io_u->iocb;
53cdc686	104
a391d73d	105	if (io_u->ddir == DDIR_READ) {
702906e9	106	io_prep_pread(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
7d42e66e KK	107	if (o->nowait)
7d42e66e KK	108	iocb->aio_rw_flags \|= RWF_NOWAIT;
a391d73d	109	} else if (io_u->ddir == DDIR_WRITE) {
702906e9	110	io_prep_pwrite(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
7d42e66e KK	111	if (o->nowait)
7d42e66e KK	112	iocb->aio_rw_flags \|= RWF_NOWAIT;
a79319fe JG	113	#ifdef FIO_HAVE_RWF_ATOMIC
	114	if (td->o.oatomic)
	115	iocb->aio_rw_flags \|= RWF_ATOMIC;
	116	#endif
a391d73d	117	} else if (ddir_sync(io_u->ddir))
3c3168e9	118	io_prep_fsync(iocb, f->fd);
2866c82d JA	119
	120	return 0;
	121	}
	122
127715b6 NC	123	static inline void fio_libaio_cmdprio_prep(struct thread_data *td,
127715b6 NC	124	struct io_u *io_u)
b2a432bf	125	{
d6cbeab4 NC	126	struct libaio_data *ld = td->io_ops_data;
d6cbeab4 NC	127	struct cmdprio *cmdprio = &ld->cmdprio;
127715b6 NC	128
	129	if (fio_cmdprio_set_ioprio(td, cmdprio, io_u)) {
	130	io_u->iocb.aio_reqprio = io_u->ioprio;
b2a432bf	131	io_u->iocb.u.c.flags \|= IOCB_FLAG_IOPRIO;
b2a432bf	132	}
b2a432bf PC	133	}
b2a432bf PC	134
2866c82d JA	135	static struct io_u fio_libaio_event(struct thread_data td, int event)
2866c82d JA	136	{
565e784d	137	struct libaio_data *ld = td->io_ops_data;
f423479d JA	138	struct io_event *ev;
f423479d JA	139	struct io_u *io_u;
2866c82d	140
f423479d	141	ev = ld->aio_events + event;
702906e9	142	io_u = container_of(ev->obj, struct io_u, iocb);
f423479d JA	143
	144	if (ev->res != io_u->xfer_buflen) {
	145	if (ev->res > io_u->xfer_buflen)
	146	io_u->error = -ev->res;
	147	else
	148	io_u->resid = io_u->xfer_buflen - ev->res;
	149	} else
	150	io_u->error = 0;
	151
	152	return io_u;
2866c82d JA	153	}
2866c82d JA	154
675012f0 DE	155	struct aio_ring {
	156	unsigned id; /** kernel internal index number */
	157	unsigned nr; /** number of io_events */
	158	unsigned head;
	159	unsigned tail;
c44b1ff5	160
675012f0 DE	161	unsigned magic;
	162	unsigned compat_features;
	163	unsigned incompat_features;
	164	unsigned header_length; /** size of aio_ring */
	165
	166	struct io_event events[0];
	167	};
	168
	169	#define AIO_RING_MAGIC 0xa10a10a1
	170
	171	static int user_io_getevents(io_context_t aio_ctx, unsigned int max,
c44b1ff5	172	struct io_event *events)
675012f0 DE	173	{
	174	long i = 0;
	175	unsigned head;
c44b1ff5	176	struct aio_ring ring = (struct aio_ring) aio_ctx;
675012f0 DE	177
	178	while (i < max) {
	179	head = ring->head;
	180
	181	if (head == ring->tail) {
	182	/* There are no more completions */
	183	break;
	184	} else {
	185	/* There is another completion to reap */
	186	events[i] = ring->events[head];
d473a06d BVA	187	atomic_store_release(&ring->head,
d473a06d BVA	188	(head + 1) % ring->nr);
675012f0 DE	189	i++;
	190	}
	191	}
	192
	193	return i;
	194	}
	195
e7d2e616	196	static int fio_libaio_getevents(struct thread_data *td, unsigned int min,
1f440ece	197	unsigned int max, const struct timespec *t)
2866c82d	198	{
565e784d	199	struct libaio_data *ld = td->io_ops_data;
de890a1e	200	struct libaio_options *o = td->eo;
82407585	201	unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
1f440ece	202	struct timespec __lt, *lt = NULL;
0b7fdba7	203	int r, events = 0;
2866c82d	204
1f440ece JA	205	if (t) {
	206	__lt = *t;
	207	lt = &__lt;
	208	}
	209
2866c82d	210	do {
de890a1e	211	if (o->userspace_reap == 1
675012f0 DE	212	&& actual_min == 0
	213	&& ((struct aio_ring *)(ld->aio_ctx))->magic
	214	== AIO_RING_MAGIC) {
a8f6714e	215	r = user_io_getevents(ld->aio_ctx, max - events,
675012f0 DE	216	ld->aio_events + events);
	217	} else {
	218	r = io_getevents(ld->aio_ctx, actual_min,
a8f6714e	219	max - events, ld->aio_events + events, lt);
675012f0	220	}
a8f6714e	221	if (r > 0) {
0b7fdba7	222	events += r;
a51fe82b	223	actual_min -= min((unsigned int)events, actual_min);
a8f6714e	224	}
3441a52d	225	else if ((min && r == 0) \|\| r == -EAGAIN) {
0fc2e103	226	fio_libaio_commit(td);
6347e43d JA	227	if (actual_min)
6347e43d JA	228	usleep(10);
0fc2e103	229	} else if (r != -EINTR)
a31dc2dd	230	break;
0b7fdba7 DE	231	} while (events < min);
	232
	233	return r < 0 ? r : events;
2866c82d JA	234	}
2866c82d JA	235
2e4ef4fb JA	236	static enum fio_q_status fio_libaio_queue(struct thread_data *td,
2e4ef4fb JA	237	struct io_u *io_u)
2866c82d	238	{
565e784d	239	struct libaio_data *ld = td->io_ops_data;
2866c82d	240
7101d9c2 JA	241	fio_ro_check(td, io_u);
7101d9c2 JA	242
acd45e02	243	if (ld->queued == td->o.iodepth)
755200a3 JA	244	return FIO_Q_BUSY;
	245
	246	/*
	247	* fsync is tricky, since it can fail and we need to do it
	248	* serialized with other io. the reason is that linux doesn't
	249	* support aio fsync yet. So return busy for the case where we
	250	* have pending io, to let fio complete those first.
	251	*/
f011531e	252	if (ddir_sync(io_u->ddir)) {
acd45e02	253	if (ld->queued)
755200a3	254	return FIO_Q_BUSY;
5f9099ea	255
f011531e	256	do_io_u_sync(td, io_u);
755200a3 JA	257	return FIO_Q_COMPLETED;
	258	}
	259
a5f3027c	260	if (io_u->ddir == DDIR_TRIM) {
acd45e02	261	if (ld->queued)
a5f3027c JA	262	return FIO_Q_BUSY;
	263
	264	do_io_u_trim(td, io_u);
c0681c9d VF	265	io_u_mark_submit(td, 1);
c0681c9d VF	266	io_u_mark_complete(td, 1);
a5f3027c JA	267	return FIO_Q_COMPLETED;
	268	}
	269
d6cbeab4	270	if (ld->cmdprio.mode != CMDPRIO_MODE_NONE)
ff00f247	271	fio_libaio_cmdprio_prep(td, io_u);
b2a432bf	272
702906e9	273	ld->iocbs[ld->head] = &io_u->iocb;
acd45e02 JA	274	ld->io_us[ld->head] = io_u;
	275	ring_inc(ld, &ld->head, 1);
	276	ld->queued++;
755200a3 JA	277	return FIO_Q_QUEUED;
	278	}
	279
7e77dd02 JA	280	static void fio_libaio_queued(struct thread_data td, struct io_u *io_us,
	281	unsigned int nr)
	282	{
8b6a404c	283	struct timespec now;
7e77dd02 JA	284	unsigned int i;
7e77dd02 JA	285
12d9d841 JA	286	if (!fio_fill_issue_time(td))
	287	return;
	288
7e77dd02 JA	289	fio_gettime(&now, NULL);
	290
	291	for (i = 0; i < nr; i++) {
	292	struct io_u *io_u = io_us[i];
	293
	294	memcpy(&io_u->issue_time, &now, sizeof(now));
	295	io_u_queued(td, io_u);
	296	}
39f56400 VF	297
	298	/*
	299	* only used for iolog
	300	*/
	301	if (td->o.read_iolog_file)
	302	memcpy(&td->last_issue, &now, sizeof(now));
7e77dd02 JA	303	}
7e77dd02 JA	304
755200a3 JA	305	static int fio_libaio_commit(struct thread_data *td)
755200a3 JA	306	{
565e784d	307	struct libaio_data *ld = td->io_ops_data;
755200a3	308	struct iocb **iocbs;
7e77dd02	309	struct io_u **io_us;
8b6a404c	310	struct timespec ts;
a120ca7f	311	int ret, wait_start = 0;
755200a3	312
acd45e02	313	if (!ld->queued)
755200a3 JA	314	return 0;
755200a3 JA	315
2866c82d	316	do {
acd45e02 JA	317	long nr = ld->queued;
	318
	319	nr = min((unsigned int) nr, ld->entries - ld->tail);
	320	io_us = ld->io_us + ld->tail;
	321	iocbs = ld->iocbs + ld->tail;
	322
	323	ret = io_submit(ld->aio_ctx, nr, iocbs);
5e00c2c4	324	if (ret > 0) {
7e77dd02	325	fio_libaio_queued(td, io_us, ret);
838bc709	326	io_u_mark_submit(td, ret);
acd45e02 JA	327
	328	ld->queued -= ret;
	329	ring_inc(ld, &ld->tail, ret);
5e00c2c4	330	ret = 0;
e3b4e568	331	wait_start = 0;
acd45e02	332	} else if (ret == -EINTR \|\| !ret) {
838bc709 JA	333	if (!ret)
838bc709 JA	334	io_u_mark_submit(td, ret);
e3b4e568	335	wait_start = 0;
2866c82d	336	continue;
acd45e02 JA	337	} else if (ret == -EAGAIN) {
	338	/*
	339	* If we get EAGAIN, we should break out without
	340	* error and let the upper layer reap some
a120ca7f JA	341	* events for us. If we have no queued IO, we
	342	* must loop here. If we loop for more than 30s,
	343	* just error out, something must be buggy in the
	344	* IO path.
acd45e02	345	*/
a120ca7f JA	346	if (ld->queued) {
	347	ret = 0;
	348	break;
	349	}
	350	if (!wait_start) {
8b6a404c	351	fio_gettime(&ts, NULL);
d36b072d	352	wait_start = 1;
8b6a404c	353	} else if (mtime_since_now(&ts) > 30000) {
a120ca7f JA	354	log_err("fio: aio appears to be stalled, giving up\n");
	355	break;
	356	}
	357	usleep(1);
	358	continue;
a31dc2dd JA	359	} else if (ret == -ENOMEM) {
	360	/*
	361	* If we get -ENOMEM, reap events if we can. If
	362	* we cannot, treat it as a fatal event since there's
	363	* nothing we can do about it.
	364	*/
	365	if (ld->queued)
	366	ret = 0;
	367	break;
838bc709	368	} else
2866c82d	369	break;
2c3a4ae9	370	} while (ld->queued);
2866c82d	371
36167d82	372	return ret;
2866c82d JA	373	}
	374
	375	static int fio_libaio_cancel(struct thread_data td, struct io_u io_u)
	376	{
565e784d	377	struct libaio_data *ld = td->io_ops_data;
2866c82d JA	378
	379	return io_cancel(ld->aio_ctx, &io_u->iocb, ld->aio_events);
	380	}
	381
	382	static void fio_libaio_cleanup(struct thread_data *td)
	383	{
565e784d	384	struct libaio_data *ld = td->io_ops_data;
2866c82d JA	385
2866c82d JA	386	if (ld) {
f24c2649 JA	387	/*
	388	* Work-around to avoid huge RCU stalls at exit time. If we
	389	* don't do this here, then it'll be torn down by exit_aio().
	390	* But for that case we can parallellize the freeing, thus
	391	* speeding it up a lot.
	392	*/
	393	if (!(td->flags & TD_F_CHILD))
	394	io_destroy(ld->aio_ctx);
d6cbeab4 NC	395
d6cbeab4 NC	396	fio_cmdprio_cleanup(&ld->cmdprio);
7e77dd02 JA	397	free(ld->aio_events);
	398	free(ld->iocbs);
	399	free(ld->io_us);
2866c82d	400	free(ld);
2866c82d JA	401	}
	402	}
	403
2041bd34 JA	404	static int fio_libaio_post_init(struct thread_data *td)
	405	{
	406	struct libaio_data *ld = td->io_ops_data;
ad4e9298	407	int err;
2041bd34	408
ad4e9298	409	err = io_queue_init(td->o.iodepth, &ld->aio_ctx);
2041bd34 JA	410	if (err) {
	411	td_verror(td, -err, "io_queue_init");
	412	return 1;
	413	}
	414
	415	return 0;
ebec344d JA	416	}
ebec344d JA	417
2866c82d JA	418	static int fio_libaio_init(struct thread_data *td)
2866c82d JA	419	{
acd45e02	420	struct libaio_data *ld;
b2a432bf	421	struct libaio_options *o = td->eo;
e9f6567a	422	int ret;
2866c82d	423
acd45e02	424	ld = calloc(1, sizeof(*ld));
c1db2dce	425
acd45e02 JA	426	ld->entries = td->o.iodepth;
	427	ld->is_pow2 = is_power_of_2(ld->entries);
	428	ld->aio_events = calloc(ld->entries, sizeof(struct io_event));
	429	ld->iocbs = calloc(ld->entries, sizeof(struct iocb *));
	430	ld->io_us = calloc(ld->entries, sizeof(struct io_u *));
755200a3	431
565e784d	432	td->io_ops_data = ld;
e9f6567a	433
d6cbeab4	434	ret = fio_cmdprio_init(td, &ld->cmdprio, &o->cmdprio_options);
e9f6567a	435	if (ret) {
b2a432bf PC	436	td_verror(td, EINVAL, "fio_libaio_init");
	437	return 1;
	438	}
e9f6567a	439
2866c82d JA	440	return 0;
	441	}
	442
5a8a6a03	443	FIO_STATIC struct ioengine_ops ioengine = {
de890a1e SL	444	.name = "libaio",
de890a1e SL	445	.version = FIO_IOOPS_VERSION,
4e7e7898	446	.flags = FIO_ASYNCIO_SYNC_TRIM \|
a79319fe JG	447	FIO_ASYNCIO_SETS_ISSUE_TIME \|
a79319fe JG	448	FIO_ATOMICWRITES,
de890a1e	449	.init = fio_libaio_init,
2041bd34	450	.post_init = fio_libaio_post_init,
de890a1e SL	451	.prep = fio_libaio_prep,
	452	.queue = fio_libaio_queue,
	453	.commit = fio_libaio_commit,
	454	.cancel = fio_libaio_cancel,
	455	.getevents = fio_libaio_getevents,
	456	.event = fio_libaio_event,
	457	.cleanup = fio_libaio_cleanup,
	458	.open_file = generic_open_file,
	459	.close_file = generic_close_file,
	460	.get_file_size = generic_get_file_size,
	461	.options = options,
	462	.option_struct_size = sizeof(struct libaio_options),
2866c82d	463	};
34cfcdaf	464
5f350952 JA	465	static void fio_init fio_libaio_register(void)
	466	{
	467	register_ioengine(&ioengine);
	468	}
	469
	470	static void fio_exit fio_libaio_unregister(void)
	471	{
	472	unregister_ioengine(&ioengine);
	473	}