[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;
	} 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,
	.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
	1	/*
	2	* libaio engine
	3	*
	4	* IO engine using the Linux native aio interface.
	5	*
	6	*/
	7	#include <stdlib.h>
	8	#include <unistd.h>
	9	#include <errno.h>
	10	#include <libaio.h>
	11	#include <sys/time.h>
	12	#include <sys/resource.h>
	13
	14	#include "../fio.h"
	15	#include "../lib/pow2.h"
	16	#include "../optgroup.h"
	17	#include "../lib/memalign.h"
	18	#include "cmdprio.h"
	19
	20	/* Should be defined in newest aio_abi.h */
	21	#ifndef IOCB_FLAG_IOPRIO
	22	#define IOCB_FLAG_IOPRIO (1 << 1)
	23	#endif
	24
	25	/* Hack for libaio < 0.3.111 */
	26	#ifndef CONFIG_LIBAIO_RW_FLAGS
	27	#define aio_rw_flags __pad2
	28	#endif
	29
	30	static int fio_libaio_commit(struct thread_data *td);
	31	static int fio_libaio_init(struct thread_data *td);
	32
	33	struct libaio_data {
	34	io_context_t aio_ctx;
	35	struct io_event *aio_events;
	36	struct iocb **iocbs;
	37	struct io_u **io_us;
	38
	39	struct io_u **io_u_index;
	40
	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;
	54
	55	struct cmdprio cmdprio;
	56	};
	57
	58	struct libaio_options {
	59	struct thread_data *td;
	60	unsigned int userspace_reap;
	61	struct cmdprio_options cmdprio_options;
	62	unsigned int nowait;
	63	};
	64
	65	static struct fio_option options[] = {
	66	{
	67	.name = "userspace_reap",
	68	.lname = "Libaio userspace reaping",
	69	.type = FIO_OPT_STR_SET,
	70	.off1 = offsetof(struct libaio_options, userspace_reap),
	71	.help = "Use alternative user-space reap implementation",
	72	.category = FIO_OPT_C_ENGINE,
	73	.group = FIO_OPT_G_LIBAIO,
	74	},
	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	},
	84	CMDPRIO_OPTIONS(struct libaio_options, FIO_OPT_G_LIBAIO),
	85	{
	86	.name = NULL,
	87	},
	88	};
	89
	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
	99	static int fio_libaio_prep(struct thread_data td, struct io_u io_u)
	100	{
	101	struct libaio_options *o = td->eo;
	102	struct fio_file *f = io_u->file;
	103	struct iocb *iocb = &io_u->iocb;
	104
	105	if (io_u->ddir == DDIR_READ) {
	106	io_prep_pread(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
	107	if (o->nowait)
	108	iocb->aio_rw_flags \|= RWF_NOWAIT;
	109	} else if (io_u->ddir == DDIR_WRITE) {
	110	io_prep_pwrite(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
	111	if (o->nowait)
	112	iocb->aio_rw_flags \|= RWF_NOWAIT;
	113	} else if (ddir_sync(io_u->ddir))
	114	io_prep_fsync(iocb, f->fd);
	115
	116	return 0;
	117	}
	118
	119	static inline void fio_libaio_cmdprio_prep(struct thread_data *td,
	120	struct io_u *io_u)
	121	{
	122	struct libaio_data *ld = td->io_ops_data;
	123	struct cmdprio *cmdprio = &ld->cmdprio;
	124
	125	if (fio_cmdprio_set_ioprio(td, cmdprio, io_u)) {
	126	io_u->iocb.aio_reqprio = io_u->ioprio;
	127	io_u->iocb.u.c.flags \|= IOCB_FLAG_IOPRIO;
	128	}
	129	}
	130
	131	static struct io_u fio_libaio_event(struct thread_data td, int event)
	132	{
	133	struct libaio_data *ld = td->io_ops_data;
	134	struct io_event *ev;
	135	struct io_u *io_u;
	136
	137	ev = ld->aio_events + event;
	138	io_u = container_of(ev->obj, struct io_u, iocb);
	139
	140	if (ev->res != io_u->xfer_buflen) {
	141	if (ev->res > io_u->xfer_buflen)
	142	io_u->error = -ev->res;
	143	else
	144	io_u->resid = io_u->xfer_buflen - ev->res;
	145	} else
	146	io_u->error = 0;
	147
	148	return io_u;
	149	}
	150
	151	struct aio_ring {
	152	unsigned id; /** kernel internal index number */
	153	unsigned nr; /** number of io_events */
	154	unsigned head;
	155	unsigned tail;
	156
	157	unsigned magic;
	158	unsigned compat_features;
	159	unsigned incompat_features;
	160	unsigned header_length; /** size of aio_ring */
	161
	162	struct io_event events[0];
	163	};
	164
	165	#define AIO_RING_MAGIC 0xa10a10a1
	166
	167	static int user_io_getevents(io_context_t aio_ctx, unsigned int max,
	168	struct io_event *events)
	169	{
	170	long i = 0;
	171	unsigned head;
	172	struct aio_ring ring = (struct aio_ring) aio_ctx;
	173
	174	while (i < max) {
	175	head = ring->head;
	176
	177	if (head == ring->tail) {
	178	/* There are no more completions */
	179	break;
	180	} else {
	181	/* There is another completion to reap */
	182	events[i] = ring->events[head];
	183	atomic_store_release(&ring->head,
	184	(head + 1) % ring->nr);
	185	i++;
	186	}
	187	}
	188
	189	return i;
	190	}
	191
	192	static int fio_libaio_getevents(struct thread_data *td, unsigned int min,
	193	unsigned int max, const struct timespec *t)
	194	{
	195	struct libaio_data *ld = td->io_ops_data;
	196	struct libaio_options *o = td->eo;
	197	unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
	198	struct timespec __lt, *lt = NULL;
	199	int r, events = 0;
	200
	201	if (t) {
	202	__lt = *t;
	203	lt = &__lt;
	204	}
	205
	206	do {
	207	if (o->userspace_reap == 1
	208	&& actual_min == 0
	209	&& ((struct aio_ring *)(ld->aio_ctx))->magic
	210	== AIO_RING_MAGIC) {
	211	r = user_io_getevents(ld->aio_ctx, max - events,
	212	ld->aio_events + events);
	213	} else {
	214	r = io_getevents(ld->aio_ctx, actual_min,
	215	max - events, ld->aio_events + events, lt);
	216	}
	217	if (r > 0) {
	218	events += r;
	219	actual_min -= min((unsigned int)events, actual_min);
	220	}
	221	else if ((min && r == 0) \|\| r == -EAGAIN) {
	222	fio_libaio_commit(td);
	223	if (actual_min)
	224	usleep(10);
	225	} else if (r != -EINTR)
	226	break;
	227	} while (events < min);
	228
	229	return r < 0 ? r : events;
	230	}
	231
	232	static enum fio_q_status fio_libaio_queue(struct thread_data *td,
	233	struct io_u *io_u)
	234	{
	235	struct libaio_data *ld = td->io_ops_data;
	236
	237	fio_ro_check(td, io_u);
	238
	239	if (ld->queued == td->o.iodepth)
	240	return FIO_Q_BUSY;
	241
	242	/*
	243	* fsync is tricky, since it can fail and we need to do it
	244	* serialized with other io. the reason is that linux doesn't
	245	* support aio fsync yet. So return busy for the case where we
	246	* have pending io, to let fio complete those first.
	247	*/
	248	if (ddir_sync(io_u->ddir)) {
	249	if (ld->queued)
	250	return FIO_Q_BUSY;
	251
	252	do_io_u_sync(td, io_u);
	253	return FIO_Q_COMPLETED;
	254	}
	255
	256	if (io_u->ddir == DDIR_TRIM) {
	257	if (ld->queued)
	258	return FIO_Q_BUSY;
	259
	260	do_io_u_trim(td, io_u);
	261	io_u_mark_submit(td, 1);
	262	io_u_mark_complete(td, 1);
	263	return FIO_Q_COMPLETED;
	264	}
	265
	266	if (ld->cmdprio.mode != CMDPRIO_MODE_NONE)
	267	fio_libaio_cmdprio_prep(td, io_u);
	268
	269	ld->iocbs[ld->head] = &io_u->iocb;
	270	ld->io_us[ld->head] = io_u;
	271	ring_inc(ld, &ld->head, 1);
	272	ld->queued++;
	273	return FIO_Q_QUEUED;
	274	}
	275
	276	static void fio_libaio_queued(struct thread_data td, struct io_u *io_us,
	277	unsigned int nr)
	278	{
	279	struct timespec now;
	280	unsigned int i;
	281
	282	if (!fio_fill_issue_time(td))
	283	return;
	284
	285	fio_gettime(&now, NULL);
	286
	287	for (i = 0; i < nr; i++) {
	288	struct io_u *io_u = io_us[i];
	289
	290	memcpy(&io_u->issue_time, &now, sizeof(now));
	291	io_u_queued(td, io_u);
	292	}
	293
	294	/*
	295	* only used for iolog
	296	*/
	297	if (td->o.read_iolog_file)
	298	memcpy(&td->last_issue, &now, sizeof(now));
	299	}
	300
	301	static int fio_libaio_commit(struct thread_data *td)
	302	{
	303	struct libaio_data *ld = td->io_ops_data;
	304	struct iocb **iocbs;
	305	struct io_u **io_us;
	306	struct timespec ts;
	307	int ret, wait_start = 0;
	308
	309	if (!ld->queued)
	310	return 0;
	311
	312	do {
	313	long nr = ld->queued;
	314
	315	nr = min((unsigned int) nr, ld->entries - ld->tail);
	316	io_us = ld->io_us + ld->tail;
	317	iocbs = ld->iocbs + ld->tail;
	318
	319	ret = io_submit(ld->aio_ctx, nr, iocbs);
	320	if (ret > 0) {
	321	fio_libaio_queued(td, io_us, ret);
	322	io_u_mark_submit(td, ret);
	323
	324	ld->queued -= ret;
	325	ring_inc(ld, &ld->tail, ret);
	326	ret = 0;
	327	wait_start = 0;
	328	} else if (ret == -EINTR \|\| !ret) {
	329	if (!ret)
	330	io_u_mark_submit(td, ret);
	331	wait_start = 0;
	332	continue;
	333	} else if (ret == -EAGAIN) {
	334	/*
	335	* If we get EAGAIN, we should break out without
	336	* error and let the upper layer reap some
	337	* events for us. If we have no queued IO, we
	338	* must loop here. If we loop for more than 30s,
	339	* just error out, something must be buggy in the
	340	* IO path.
	341	*/
	342	if (ld->queued) {
	343	ret = 0;
	344	break;
	345	}
	346	if (!wait_start) {
	347	fio_gettime(&ts, NULL);
	348	wait_start = 1;
	349	} else if (mtime_since_now(&ts) > 30000) {
	350	log_err("fio: aio appears to be stalled, giving up\n");
	351	break;
	352	}
	353	usleep(1);
	354	continue;
	355	} else if (ret == -ENOMEM) {
	356	/*
	357	* If we get -ENOMEM, reap events if we can. If
	358	* we cannot, treat it as a fatal event since there's
	359	* nothing we can do about it.
	360	*/
	361	if (ld->queued)
	362	ret = 0;
	363	break;
	364	} else
	365	break;
	366	} while (ld->queued);
	367
	368	return ret;
	369	}
	370
	371	static int fio_libaio_cancel(struct thread_data td, struct io_u io_u)
	372	{
	373	struct libaio_data *ld = td->io_ops_data;
	374
	375	return io_cancel(ld->aio_ctx, &io_u->iocb, ld->aio_events);
	376	}
	377
	378	static void fio_libaio_cleanup(struct thread_data *td)
	379	{
	380	struct libaio_data *ld = td->io_ops_data;
	381
	382	if (ld) {
	383	/*
	384	* Work-around to avoid huge RCU stalls at exit time. If we
	385	* don't do this here, then it'll be torn down by exit_aio().
	386	* But for that case we can parallellize the freeing, thus
	387	* speeding it up a lot.
	388	*/
	389	if (!(td->flags & TD_F_CHILD))
	390	io_destroy(ld->aio_ctx);
	391
	392	fio_cmdprio_cleanup(&ld->cmdprio);
	393	free(ld->aio_events);
	394	free(ld->iocbs);
	395	free(ld->io_us);
	396	free(ld);
	397	}
	398	}
	399
	400	static int fio_libaio_post_init(struct thread_data *td)
	401	{
	402	struct libaio_data *ld = td->io_ops_data;
	403	int err;
	404
	405	err = io_queue_init(td->o.iodepth, &ld->aio_ctx);
	406	if (err) {
	407	td_verror(td, -err, "io_queue_init");
	408	return 1;
	409	}
	410
	411	return 0;
	412	}
	413
	414	static int fio_libaio_init(struct thread_data *td)
	415	{
	416	struct libaio_data *ld;
	417	struct libaio_options *o = td->eo;
	418	int ret;
	419
	420	ld = calloc(1, sizeof(*ld));
	421
	422	ld->entries = td->o.iodepth;
	423	ld->is_pow2 = is_power_of_2(ld->entries);
	424	ld->aio_events = calloc(ld->entries, sizeof(struct io_event));
	425	ld->iocbs = calloc(ld->entries, sizeof(struct iocb *));
	426	ld->io_us = calloc(ld->entries, sizeof(struct io_u *));
	427
	428	td->io_ops_data = ld;
	429
	430	ret = fio_cmdprio_init(td, &ld->cmdprio, &o->cmdprio_options);
	431	if (ret) {
	432	td_verror(td, EINVAL, "fio_libaio_init");
	433	return 1;
	434	}
	435
	436	return 0;
	437	}
	438
	439	FIO_STATIC struct ioengine_ops ioengine = {
	440	.name = "libaio",
	441	.version = FIO_IOOPS_VERSION,
	442	.flags = FIO_ASYNCIO_SYNC_TRIM \|
	443	FIO_ASYNCIO_SETS_ISSUE_TIME,
	444	.init = fio_libaio_init,
	445	.post_init = fio_libaio_post_init,
	446	.prep = fio_libaio_prep,
	447	.queue = fio_libaio_queue,
	448	.commit = fio_libaio_commit,
	449	.cancel = fio_libaio_cancel,
	450	.getevents = fio_libaio_getevents,
	451	.event = fio_libaio_event,
	452	.cleanup = fio_libaio_cleanup,
	453	.open_file = generic_open_file,
	454	.close_file = generic_close_file,
	455	.get_file_size = generic_get_file_size,
	456	.options = options,
	457	.option_struct_size = sizeof(struct libaio_options),
	458	};
	459
	460	static void fio_init fio_libaio_register(void)
	461	{
	462	register_ioengine(&ioengine);
	463	}
	464
	465	static void fio_exit fio_libaio_unregister(void)
	466	{
	467	unregister_ioengine(&ioengine);
	468	}