[fio.git] / engines / libaio.c

/*
 * libaio engine
 *
 * IO engine using the Linux native aio interface.
 *
 */
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <errno.h>
#include <assert.h>

#include "../fio.h"

#ifdef FIO_HAVE_LIBAIO

#define ev_to_iou(ev)	(struct io_u *) ((unsigned long) (ev)->obj)

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

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

static struct fio_option options[] = {
	{
		.name	= "userspace_reap",
		.type	= FIO_OPT_STR_SET,
		.off1	= offsetof(struct libaio_options, userspace_reap),
		.help	= "Use alternative user-space reap implementation",
	},
	{
		.name	= NULL,
	},
};

static int fio_libaio_prep(struct thread_data fio_unused *td, struct io_u *io_u)
{
	struct fio_file *f = io_u->file;

	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);
	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);
	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 = ev_to_iou(ev);

	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, 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 == 0 ? 0 : min;
	int r, events = 0;

	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, t);
		}
		if (r >= 0)
			events += r;
		else if (r == -EAGAIN)
			usleep(100);
	} while (events < min);

	return r < 0 ? r : events;
}

static int 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->iocbs_nr == (int) 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->iocbs_nr)
			return FIO_Q_BUSY;

		do_io_u_sync(td, io_u);
		return FIO_Q_COMPLETED;
	}

	if (io_u->ddir == DDIR_TRIM) {
		if (ld->iocbs_nr)
			return FIO_Q_BUSY;

		do_io_u_trim(td, io_u);
		return FIO_Q_COMPLETED;
	}

	ld->iocbs[ld->iocbs_nr] = &io_u->iocb;
	ld->io_us[ld->iocbs_nr] = io_u;
	ld->iocbs_nr++;
	return FIO_Q_QUEUED;
}

static void fio_libaio_queued(struct thread_data *td, struct io_u **io_us,
			      unsigned int nr)
{
	struct timeval 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;
	int ret;

	if (!ld->iocbs_nr)
		return 0;

	io_us = ld->io_us;
	iocbs = ld->iocbs;
	do {
		ret = io_submit(ld->aio_ctx, ld->iocbs_nr, iocbs);
		if (ret > 0) {
			fio_libaio_queued(td, io_us, ret);
			io_u_mark_submit(td, ret);
			ld->iocbs_nr -= ret;
			io_us += ret;
			iocbs += ret;
			ret = 0;
		} else if (!ret || ret == -EAGAIN || ret == -EINTR) {
			if (!ret)
				io_u_mark_submit(td, ret);
			continue;
		} else
			break;
	} while (ld->iocbs_nr);

	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) {
		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_data *ld = malloc(sizeof(*ld));
	struct libaio_options *o = td->eo;
	int err = 0;

	memset(ld, 0, 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->aio_events = malloc(td->o.iodepth * sizeof(struct io_event));
	memset(ld->aio_events, 0, td->o.iodepth * sizeof(struct io_event));
	ld->iocbs = malloc(td->o.iodepth * sizeof(struct iocb *));
	memset(ld->iocbs, 0, sizeof(struct iocb *));
	ld->io_us = malloc(td->o.iodepth * sizeof(struct io_u *));
	memset(ld->io_us, 0, td->o.iodepth * sizeof(struct io_u *));
	ld->iocbs_nr = 0;

	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),
};

#else /* FIO_HAVE_LIBAIO */

/*
 * When we have a proper configure system in place, we simply wont build
 * and install this io engine. For now install a crippled version that
 * just complains and fails to load.
 */
static int fio_libaio_init(struct thread_data fio_unused *td)
{
	log_err("fio: libaio not available\n");
	return 1;
}

static struct ioengine_ops ioengine = {
	.name		= "libaio",
	.version	= FIO_IOOPS_VERSION,
	.init		= fio_libaio_init,
};

#endif

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 <stdio.h>
	8	#include <stdlib.h>
	9	#include <unistd.h>
	10	#include <errno.h>
	11	#include <assert.h>
	12
	13	#include "../fio.h"
	14
	15	#ifdef FIO_HAVE_LIBAIO
	16
	17	#define ev_to_iou(ev) (struct io_u *) ((unsigned long) (ev)->obj)
	18
	19	struct libaio_data {
	20	io_context_t aio_ctx;
	21	struct io_event *aio_events;
	22	struct iocb **iocbs;
	23	struct io_u **io_us;
	24	int iocbs_nr;
	25	};
	26
	27	struct libaio_options {
	28	struct thread_data *td;
	29	unsigned int userspace_reap;
	30	};
	31
	32	static struct fio_option options[] = {
	33	{
	34	.name = "userspace_reap",
	35	.type = FIO_OPT_STR_SET,
	36	.off1 = offsetof(struct libaio_options, userspace_reap),
	37	.help = "Use alternative user-space reap implementation",
	38	},
	39	{
	40	.name = NULL,
	41	},
	42	};
	43
	44	static int fio_libaio_prep(struct thread_data fio_unused td, struct io_u io_u)
	45	{
	46	struct fio_file *f = io_u->file;
	47
	48	if (io_u->ddir == DDIR_READ)
	49	io_prep_pread(&io_u->iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
	50	else if (io_u->ddir == DDIR_WRITE)
	51	io_prep_pwrite(&io_u->iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
	52	else if (ddir_sync(io_u->ddir))
	53	io_prep_fsync(&io_u->iocb, f->fd);
	54
	55	return 0;
	56	}
	57
	58	static struct io_u fio_libaio_event(struct thread_data td, int event)
	59	{
	60	struct libaio_data *ld = td->io_ops->data;
	61	struct io_event *ev;
	62	struct io_u *io_u;
	63
	64	ev = ld->aio_events + event;
	65	io_u = ev_to_iou(ev);
	66
	67	if (ev->res != io_u->xfer_buflen) {
	68	if (ev->res > io_u->xfer_buflen)
	69	io_u->error = -ev->res;
	70	else
	71	io_u->resid = io_u->xfer_buflen - ev->res;
	72	} else
	73	io_u->error = 0;
	74
	75	return io_u;
	76	}
	77
	78	struct aio_ring {
	79	unsigned id; /** kernel internal index number */
	80	unsigned nr; /** number of io_events */
	81	unsigned head;
	82	unsigned tail;
	83
	84	unsigned magic;
	85	unsigned compat_features;
	86	unsigned incompat_features;
	87	unsigned header_length; /** size of aio_ring */
	88
	89	struct io_event events[0];
	90	};
	91
	92	#define AIO_RING_MAGIC 0xa10a10a1
	93
	94	static int user_io_getevents(io_context_t aio_ctx, unsigned int max,
	95	struct io_event *events)
	96	{
	97	long i = 0;
	98	unsigned head;
	99	struct aio_ring ring = (struct aio_ring) aio_ctx;
	100
	101	while (i < max) {
	102	head = ring->head;
	103
	104	if (head == ring->tail) {
	105	/* There are no more completions */
	106	break;
	107	} else {
	108	/* There is another completion to reap */
	109	events[i] = ring->events[head];
	110	read_barrier();
	111	ring->head = (head + 1) % ring->nr;
	112	i++;
	113	}
	114	}
	115
	116	return i;
	117	}
	118
	119	static int fio_libaio_getevents(struct thread_data *td, unsigned int min,
	120	unsigned int max, struct timespec *t)
	121	{
	122	struct libaio_data *ld = td->io_ops->data;
	123	struct libaio_options *o = td->eo;
	124	unsigned actual_min = td->o.iodepth_batch_complete == 0 ? 0 : min;
	125	int r, events = 0;
	126
	127	do {
	128	if (o->userspace_reap == 1
	129	&& actual_min == 0
	130	&& ((struct aio_ring *)(ld->aio_ctx))->magic
	131	== AIO_RING_MAGIC) {
	132	r = user_io_getevents(ld->aio_ctx, max,
	133	ld->aio_events + events);
	134	} else {
	135	r = io_getevents(ld->aio_ctx, actual_min,
	136	max, ld->aio_events + events, t);
	137	}
	138	if (r >= 0)
	139	events += r;
	140	else if (r == -EAGAIN)
	141	usleep(100);
	142	} while (events < min);
	143
	144	return r < 0 ? r : events;
	145	}
	146
	147	static int fio_libaio_queue(struct thread_data td, struct io_u io_u)
	148	{
	149	struct libaio_data *ld = td->io_ops->data;
	150
	151	fio_ro_check(td, io_u);
	152
	153	if (ld->iocbs_nr == (int) td->o.iodepth)
	154	return FIO_Q_BUSY;
	155
	156	/*
	157	* fsync is tricky, since it can fail and we need to do it
	158	* serialized with other io. the reason is that linux doesn't
	159	* support aio fsync yet. So return busy for the case where we
	160	* have pending io, to let fio complete those first.
	161	*/
	162	if (ddir_sync(io_u->ddir)) {
	163	if (ld->iocbs_nr)
	164	return FIO_Q_BUSY;
	165
	166	do_io_u_sync(td, io_u);
	167	return FIO_Q_COMPLETED;
	168	}
	169
	170	if (io_u->ddir == DDIR_TRIM) {
	171	if (ld->iocbs_nr)
	172	return FIO_Q_BUSY;
	173
	174	do_io_u_trim(td, io_u);
	175	return FIO_Q_COMPLETED;
	176	}
	177
	178	ld->iocbs[ld->iocbs_nr] = &io_u->iocb;
	179	ld->io_us[ld->iocbs_nr] = io_u;
	180	ld->iocbs_nr++;
	181	return FIO_Q_QUEUED;
	182	}
	183
	184	static void fio_libaio_queued(struct thread_data td, struct io_u *io_us,
	185	unsigned int nr)
	186	{
	187	struct timeval now;
	188	unsigned int i;
	189
	190	if (!fio_fill_issue_time(td))
	191	return;
	192
	193	fio_gettime(&now, NULL);
	194
	195	for (i = 0; i < nr; i++) {
	196	struct io_u *io_u = io_us[i];
	197
	198	memcpy(&io_u->issue_time, &now, sizeof(now));
	199	io_u_queued(td, io_u);
	200	}
	201	}
	202
	203	static int fio_libaio_commit(struct thread_data *td)
	204	{
	205	struct libaio_data *ld = td->io_ops->data;
	206	struct iocb **iocbs;
	207	struct io_u **io_us;
	208	int ret;
	209
	210	if (!ld->iocbs_nr)
	211	return 0;
	212
	213	io_us = ld->io_us;
	214	iocbs = ld->iocbs;
	215	do {
	216	ret = io_submit(ld->aio_ctx, ld->iocbs_nr, iocbs);
	217	if (ret > 0) {
	218	fio_libaio_queued(td, io_us, ret);
	219	io_u_mark_submit(td, ret);
	220	ld->iocbs_nr -= ret;
	221	io_us += ret;
	222	iocbs += ret;
	223	ret = 0;
	224	} else if (!ret \|\| ret == -EAGAIN \|\| ret == -EINTR) {
	225	if (!ret)
	226	io_u_mark_submit(td, ret);
	227	continue;
	228	} else
	229	break;
	230	} while (ld->iocbs_nr);
	231
	232	return ret;
	233	}
	234
	235	static int fio_libaio_cancel(struct thread_data td, struct io_u io_u)
	236	{
	237	struct libaio_data *ld = td->io_ops->data;
	238
	239	return io_cancel(ld->aio_ctx, &io_u->iocb, ld->aio_events);
	240	}
	241
	242	static void fio_libaio_cleanup(struct thread_data *td)
	243	{
	244	struct libaio_data *ld = td->io_ops->data;
	245
	246	if (ld) {
	247	io_destroy(ld->aio_ctx);
	248	free(ld->aio_events);
	249	free(ld->iocbs);
	250	free(ld->io_us);
	251	free(ld);
	252	}
	253	}
	254
	255	static int fio_libaio_init(struct thread_data *td)
	256	{
	257	struct libaio_data ld = malloc(sizeof(ld));
	258	struct libaio_options *o = td->eo;
	259	int err = 0;
	260
	261	memset(ld, 0, sizeof(*ld));
	262
	263	/*
	264	* First try passing in 0 for queue depth, since we don't
	265	* care about the user ring. If that fails, the kernel is too old
	266	* and we need the right depth.
	267	*/
	268	if (!o->userspace_reap)
	269	err = io_queue_init(INT_MAX, &ld->aio_ctx);
	270	if (o->userspace_reap \|\| err == -EINVAL)
	271	err = io_queue_init(td->o.iodepth, &ld->aio_ctx);
	272	if (err) {
	273	td_verror(td, -err, "io_queue_init");
	274	log_err("fio: check /proc/sys/fs/aio-max-nr\n");
	275	free(ld);
	276	return 1;
	277	}
	278
	279	ld->aio_events = malloc(td->o.iodepth * sizeof(struct io_event));
	280	memset(ld->aio_events, 0, td->o.iodepth * sizeof(struct io_event));
	281	ld->iocbs = malloc(td->o.iodepth * sizeof(struct iocb *));
	282	memset(ld->iocbs, 0, sizeof(struct iocb *));
	283	ld->io_us = malloc(td->o.iodepth * sizeof(struct io_u *));
	284	memset(ld->io_us, 0, td->o.iodepth * sizeof(struct io_u *));
	285	ld->iocbs_nr = 0;
	286
	287	td->io_ops->data = ld;
	288	return 0;
	289	}
	290
	291	static struct ioengine_ops ioengine = {
	292	.name = "libaio",
	293	.version = FIO_IOOPS_VERSION,
	294	.init = fio_libaio_init,
	295	.prep = fio_libaio_prep,
	296	.queue = fio_libaio_queue,
	297	.commit = fio_libaio_commit,
	298	.cancel = fio_libaio_cancel,
	299	.getevents = fio_libaio_getevents,
	300	.event = fio_libaio_event,
	301	.cleanup = fio_libaio_cleanup,
	302	.open_file = generic_open_file,
	303	.close_file = generic_close_file,
	304	.get_file_size = generic_get_file_size,
	305	.options = options,
	306	.option_struct_size = sizeof(struct libaio_options),
	307	};
	308
	309	#else /* FIO_HAVE_LIBAIO */
	310
	311	/*
	312	* When we have a proper configure system in place, we simply wont build
	313	* and install this io engine. For now install a crippled version that
	314	* just complains and fails to load.
	315	*/
	316	static int fio_libaio_init(struct thread_data fio_unused *td)
	317	{
	318	log_err("fio: libaio not available\n");
	319	return 1;
	320	}
	321
	322	static struct ioengine_ops ioengine = {
	323	.name = "libaio",
	324	.version = FIO_IOOPS_VERSION,
	325	.init = fio_libaio_init,
	326	};
	327
	328	#endif
	329
	330	static void fio_init fio_libaio_register(void)
	331	{
	332	register_ioengine(&ioengine);
	333	}
	334
	335	static void fio_exit fio_libaio_unregister(void)
	336	{
	337	unregister_ioengine(&ioengine);
	338	}