[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;
};

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;
	unsigned actual_min = td->o.iodepth_batch_complete == 0 ? 0 : min;
	int r, events = 0;

	do {
		if (td->o.userspace_libaio_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));
	int err;

	memset(ld, 0, sizeof(*ld));

	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,
};

#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
2866c82d	1	/*
da751ca9 JA	2	* libaio engine
	3	*
	4	* IO engine using the Linux native aio interface.
2866c82d JA	5	*
	6	*/
	7	#include <stdio.h>
	8	#include <stdlib.h>
	9	#include <unistd.h>
	10	#include <errno.h>
	11	#include <assert.h>
5f350952 JA	12
5f350952 JA	13	#include "../fio.h"
2866c82d	14
34cfcdaf JA	15	#ifdef FIO_HAVE_LIBAIO
34cfcdaf JA	16
2866c82d JA	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;
755200a3	22	struct iocb **iocbs;
7e77dd02	23	struct io_u **io_us;
755200a3	24	int iocbs_nr;
2866c82d JA	25	};
2866c82d JA	26
7a16dd02	27	static int fio_libaio_prep(struct thread_data fio_unused td, struct io_u io_u)
2866c82d	28	{
53cdc686 JA	29	struct fio_file *f = io_u->file;
53cdc686 JA	30
2866c82d	31	if (io_u->ddir == DDIR_READ)
cec6b55d	32	io_prep_pread(&io_u->iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
87dc1ab1	33	else if (io_u->ddir == DDIR_WRITE)
cec6b55d	34	io_prep_pwrite(&io_u->iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
5f9099ea	35	else if (ddir_sync(io_u->ddir))
87dc1ab1	36	io_prep_fsync(&io_u->iocb, f->fd);
2866c82d JA	37
	38	return 0;
	39	}
	40
	41	static struct io_u fio_libaio_event(struct thread_data td, int event)
	42	{
	43	struct libaio_data *ld = td->io_ops->data;
f423479d JA	44	struct io_event *ev;
f423479d JA	45	struct io_u *io_u;
2866c82d	46
f423479d JA	47	ev = ld->aio_events + event;
	48	io_u = ev_to_iou(ev);
	49
	50	if (ev->res != io_u->xfer_buflen) {
	51	if (ev->res > io_u->xfer_buflen)
	52	io_u->error = -ev->res;
	53	else
	54	io_u->resid = io_u->xfer_buflen - ev->res;
	55	} else
	56	io_u->error = 0;
	57
	58	return io_u;
2866c82d JA	59	}
2866c82d JA	60
675012f0 DE	61	struct aio_ring {
	62	unsigned id; /** kernel internal index number */
	63	unsigned nr; /** number of io_events */
	64	unsigned head;
	65	unsigned tail;
	66
	67	unsigned magic;
	68	unsigned compat_features;
	69	unsigned incompat_features;
	70	unsigned header_length; /** size of aio_ring */
	71
	72	struct io_event events[0];
	73	};
	74
	75	#define AIO_RING_MAGIC 0xa10a10a1
	76
	77	static int user_io_getevents(io_context_t aio_ctx, unsigned int max,
	78	struct io_event *events)
	79	{
	80	long i = 0;
	81	unsigned head;
	82	struct aio_ring ring = (struct aio_ring)aio_ctx;
	83
	84	while (i < max) {
	85	head = ring->head;
	86
	87	if (head == ring->tail) {
	88	/* There are no more completions */
	89	break;
	90	} else {
	91	/* There is another completion to reap */
	92	events[i] = ring->events[head];
	93	read_barrier();
	94	ring->head = (head + 1) % ring->nr;
	95	i++;
	96	}
	97	}
	98
	99	return i;
	100	}
	101
e7d2e616 JA	102	static int fio_libaio_getevents(struct thread_data *td, unsigned int min,
e7d2e616 JA	103	unsigned int max, struct timespec *t)
2866c82d JA	104	{
2866c82d JA	105	struct libaio_data *ld = td->io_ops->data;
0b7fdba7 DE	106	unsigned actual_min = td->o.iodepth_batch_complete == 0 ? 0 : min;
0b7fdba7 DE	107	int r, events = 0;
2866c82d JA	108
2866c82d JA	109	do {
675012f0 DE	110	if (td->o.userspace_libaio_reap == 1
	111	&& actual_min == 0
	112	&& ((struct aio_ring *)(ld->aio_ctx))->magic
	113	== AIO_RING_MAGIC) {
	114	r = user_io_getevents(ld->aio_ctx, max,
	115	ld->aio_events + events);
	116	} else {
	117	r = io_getevents(ld->aio_ctx, actual_min,
	118	max, ld->aio_events + events, t);
	119	}
0b7fdba7 DE	120	if (r >= 0)
	121	events += r;
	122	else if (r == -EAGAIN)
2866c82d	123	usleep(100);
0b7fdba7 DE	124	} while (events < min);
	125
	126	return r < 0 ? r : events;
2866c82d JA	127	}
	128
	129	static int fio_libaio_queue(struct thread_data td, struct io_u io_u)
	130	{
	131	struct libaio_data *ld = td->io_ops->data;
2866c82d	132
7101d9c2 JA	133	fio_ro_check(td, io_u);
7101d9c2 JA	134
2dc1bbeb	135	if (ld->iocbs_nr == (int) td->o.iodepth)
755200a3 JA	136	return FIO_Q_BUSY;
	137
	138	/*
	139	* fsync is tricky, since it can fail and we need to do it
	140	* serialized with other io. the reason is that linux doesn't
	141	* support aio fsync yet. So return busy for the case where we
	142	* have pending io, to let fio complete those first.
	143	*/
f011531e	144	if (ddir_sync(io_u->ddir)) {
755200a3 JA	145	if (ld->iocbs_nr)
755200a3 JA	146	return FIO_Q_BUSY;
5f9099ea	147
f011531e	148	do_io_u_sync(td, io_u);
755200a3 JA	149	return FIO_Q_COMPLETED;
	150	}
	151
a5f3027c JA	152	if (io_u->ddir == DDIR_TRIM) {
	153	if (ld->iocbs_nr)
	154	return FIO_Q_BUSY;
	155
	156	do_io_u_trim(td, io_u);
	157	return FIO_Q_COMPLETED;
	158	}
	159
755200a3	160	ld->iocbs[ld->iocbs_nr] = &io_u->iocb;
7e77dd02	161	ld->io_us[ld->iocbs_nr] = io_u;
755200a3 JA	162	ld->iocbs_nr++;
	163	return FIO_Q_QUEUED;
	164	}
	165
7e77dd02 JA	166	static void fio_libaio_queued(struct thread_data td, struct io_u *io_us,
	167	unsigned int nr)
	168	{
	169	struct timeval now;
	170	unsigned int i;
	171
12d9d841 JA	172	if (!fio_fill_issue_time(td))
	173	return;
	174
7e77dd02 JA	175	fio_gettime(&now, NULL);
	176
	177	for (i = 0; i < nr; i++) {
	178	struct io_u *io_u = io_us[i];
	179
	180	memcpy(&io_u->issue_time, &now, sizeof(now));
	181	io_u_queued(td, io_u);
	182	}
	183	}
	184
755200a3 JA	185	static int fio_libaio_commit(struct thread_data *td)
	186	{
	187	struct libaio_data *ld = td->io_ops->data;
	188	struct iocb **iocbs;
7e77dd02	189	struct io_u **io_us;
5e00c2c4	190	int ret;
755200a3 JA	191
	192	if (!ld->iocbs_nr)
	193	return 0;
	194
7e77dd02	195	io_us = ld->io_us;
755200a3	196	iocbs = ld->iocbs;
2866c82d	197	do {
5e00c2c4 JA	198	ret = io_submit(ld->aio_ctx, ld->iocbs_nr, iocbs);
5e00c2c4 JA	199	if (ret > 0) {
7e77dd02	200	fio_libaio_queued(td, io_us, ret);
838bc709	201	io_u_mark_submit(td, ret);
5e00c2c4	202	ld->iocbs_nr -= ret;
7e77dd02	203	io_us += ret;
755200a3	204	iocbs += ret;
5e00c2c4	205	ret = 0;
838bc709 JA	206	} else if (!ret \|\| ret == -EAGAIN \|\| ret == -EINTR) {
	207	if (!ret)
	208	io_u_mark_submit(td, ret);
2866c82d	209	continue;
838bc709	210	} else
2866c82d	211	break;
5e00c2c4	212	} while (ld->iocbs_nr);
2866c82d	213
36167d82	214	return ret;
2866c82d JA	215	}
	216
	217	static int fio_libaio_cancel(struct thread_data td, struct io_u io_u)
	218	{
	219	struct libaio_data *ld = td->io_ops->data;
	220
	221	return io_cancel(ld->aio_ctx, &io_u->iocb, ld->aio_events);
	222	}
	223
	224	static void fio_libaio_cleanup(struct thread_data *td)
	225	{
	226	struct libaio_data *ld = td->io_ops->data;
	227
	228	if (ld) {
	229	io_destroy(ld->aio_ctx);
7e77dd02 JA	230	free(ld->aio_events);
	231	free(ld->iocbs);
	232	free(ld->io_us);
2866c82d	233	free(ld);
2866c82d JA	234	}
	235	}
	236
	237	static int fio_libaio_init(struct thread_data *td)
	238	{
	239	struct libaio_data ld = malloc(sizeof(ld));
c1db2dce	240	int err;
2866c82d JA	241
2866c82d JA	242	memset(ld, 0, sizeof(*ld));
c1db2dce JA	243
	244	err = io_queue_init(td->o.iodepth, &ld->aio_ctx);
	245	if (err) {
	246	td_verror(td, -err, "io_queue_init");
75de55ac	247	log_err("fio: check /proc/sys/fs/aio-max-nr\n");
cb781c75	248	free(ld);
2866c82d JA	249	return 1;
	250	}
	251
2dc1bbeb JA	252	ld->aio_events = malloc(td->o.iodepth * sizeof(struct io_event));
	253	memset(ld->aio_events, 0, td->o.iodepth * sizeof(struct io_event));
	254	ld->iocbs = malloc(td->o.iodepth * sizeof(struct iocb *));
755200a3	255	memset(ld->iocbs, 0, sizeof(struct iocb *));
2dc1bbeb JA	256	ld->io_us = malloc(td->o.iodepth * sizeof(struct io_u *));
2dc1bbeb JA	257	memset(ld->io_us, 0, td->o.iodepth * sizeof(struct io_u *));
755200a3 JA	258	ld->iocbs_nr = 0;
755200a3 JA	259
2866c82d JA	260	td->io_ops->data = ld;
	261	return 0;
	262	}
	263
5f350952	264	static struct ioengine_ops ioengine = {
2866c82d JA	265	.name = "libaio",
	266	.version = FIO_IOOPS_VERSION,
	267	.init = fio_libaio_init,
	268	.prep = fio_libaio_prep,
	269	.queue = fio_libaio_queue,
755200a3	270	.commit = fio_libaio_commit,
2866c82d JA	271	.cancel = fio_libaio_cancel,
	272	.getevents = fio_libaio_getevents,
	273	.event = fio_libaio_event,
	274	.cleanup = fio_libaio_cleanup,
b5af8293 JA	275	.open_file = generic_open_file,
b5af8293 JA	276	.close_file = generic_close_file,
df9c26b1	277	.get_file_size = generic_get_file_size,
2866c82d	278	};
34cfcdaf JA	279
	280	#else /* FIO_HAVE_LIBAIO */
	281
	282	/*
	283	* When we have a proper configure system in place, we simply wont build
	284	* and install this io engine. For now install a crippled version that
	285	* just complains and fails to load.
	286	*/
	287	static int fio_libaio_init(struct thread_data fio_unused *td)
	288	{
a3edaf76	289	log_err("fio: libaio not available\n");
34cfcdaf JA	290	return 1;
	291	}
	292
5f350952	293	static struct ioengine_ops ioengine = {
34cfcdaf JA	294	.name = "libaio",
	295	.version = FIO_IOOPS_VERSION,
	296	.init = fio_libaio_init,
	297	};
	298
	299	#endif
5f350952 JA	300
	301	static void fio_init fio_libaio_register(void)
	302	{
	303	register_ioengine(&ioengine);
	304	}
	305
	306	static void fio_exit fio_libaio_unregister(void)
	307	{
	308	unregister_ioengine(&ioengine);
	309	}