[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 <libaio.h>

#include "../fio.h"

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 = 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, 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),
};

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>
67bf9823	12	#include <libaio.h>
5f350952 JA	13
5f350952 JA	14	#include "../fio.h"
2866c82d	15
2866c82d JA	16	struct libaio_data {
	17	io_context_t aio_ctx;
	18	struct io_event *aio_events;
755200a3	19	struct iocb **iocbs;
7e77dd02	20	struct io_u **io_us;
755200a3	21	int iocbs_nr;
2866c82d JA	22	};
2866c82d JA	23
de890a1e SL	24	struct libaio_options {
	25	struct thread_data *td;
	26	unsigned int userspace_reap;
	27	};
	28
	29	static struct fio_option options[] = {
	30	{
	31	.name = "userspace_reap",
	32	.type = FIO_OPT_STR_SET,
	33	.off1 = offsetof(struct libaio_options, userspace_reap),
	34	.help = "Use alternative user-space reap implementation",
	35	},
	36	{
	37	.name = NULL,
	38	},
	39	};
	40
7a16dd02	41	static int fio_libaio_prep(struct thread_data fio_unused td, struct io_u io_u)
2866c82d	42	{
53cdc686 JA	43	struct fio_file *f = io_u->file;
53cdc686 JA	44
2866c82d	45	if (io_u->ddir == DDIR_READ)
cec6b55d	46	io_prep_pread(&io_u->iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
87dc1ab1	47	else if (io_u->ddir == DDIR_WRITE)
cec6b55d	48	io_prep_pwrite(&io_u->iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
5f9099ea	49	else if (ddir_sync(io_u->ddir))
87dc1ab1	50	io_prep_fsync(&io_u->iocb, f->fd);
2866c82d JA	51
	52	return 0;
	53	}
	54
	55	static struct io_u fio_libaio_event(struct thread_data td, int event)
	56	{
	57	struct libaio_data *ld = td->io_ops->data;
f423479d JA	58	struct io_event *ev;
f423479d JA	59	struct io_u *io_u;
2866c82d	60
f423479d	61	ev = ld->aio_events + event;
0cc6ee59	62	io_u = container_of(ev->obj, struct io_u, iocb);
f423479d JA	63
	64	if (ev->res != io_u->xfer_buflen) {
	65	if (ev->res > io_u->xfer_buflen)
	66	io_u->error = -ev->res;
	67	else
	68	io_u->resid = io_u->xfer_buflen - ev->res;
	69	} else
	70	io_u->error = 0;
	71
	72	return io_u;
2866c82d JA	73	}
2866c82d JA	74
675012f0 DE	75	struct aio_ring {
	76	unsigned id; /** kernel internal index number */
	77	unsigned nr; /** number of io_events */
	78	unsigned head;
	79	unsigned tail;
c44b1ff5	80
675012f0 DE	81	unsigned magic;
	82	unsigned compat_features;
	83	unsigned incompat_features;
	84	unsigned header_length; /** size of aio_ring */
	85
	86	struct io_event events[0];
	87	};
	88
	89	#define AIO_RING_MAGIC 0xa10a10a1
	90
	91	static int user_io_getevents(io_context_t aio_ctx, unsigned int max,
c44b1ff5	92	struct io_event *events)
675012f0 DE	93	{
	94	long i = 0;
	95	unsigned head;
c44b1ff5	96	struct aio_ring ring = (struct aio_ring) aio_ctx;
675012f0 DE	97
	98	while (i < max) {
	99	head = ring->head;
	100
	101	if (head == ring->tail) {
	102	/* There are no more completions */
	103	break;
	104	} else {
	105	/* There is another completion to reap */
	106	events[i] = ring->events[head];
	107	read_barrier();
c44b1ff5	108	ring->head = (head + 1) % ring->nr;
675012f0 DE	109	i++;
	110	}
	111	}
	112
	113	return i;
	114	}
	115
e7d2e616 JA	116	static int fio_libaio_getevents(struct thread_data *td, unsigned int min,
e7d2e616 JA	117	unsigned int max, struct timespec *t)
2866c82d JA	118	{
2866c82d JA	119	struct libaio_data *ld = td->io_ops->data;
de890a1e	120	struct libaio_options *o = td->eo;
0b7fdba7 DE	121	unsigned actual_min = td->o.iodepth_batch_complete == 0 ? 0 : min;
0b7fdba7 DE	122	int r, events = 0;
2866c82d JA	123
2866c82d JA	124	do {
de890a1e	125	if (o->userspace_reap == 1
675012f0 DE	126	&& actual_min == 0
	127	&& ((struct aio_ring *)(ld->aio_ctx))->magic
	128	== AIO_RING_MAGIC) {
	129	r = user_io_getevents(ld->aio_ctx, max,
	130	ld->aio_events + events);
	131	} else {
	132	r = io_getevents(ld->aio_ctx, actual_min,
	133	max, ld->aio_events + events, t);
	134	}
0b7fdba7 DE	135	if (r >= 0)
	136	events += r;
	137	else if (r == -EAGAIN)
2866c82d	138	usleep(100);
0b7fdba7 DE	139	} while (events < min);
	140
	141	return r < 0 ? r : events;
2866c82d JA	142	}
	143
	144	static int fio_libaio_queue(struct thread_data td, struct io_u io_u)
	145	{
	146	struct libaio_data *ld = td->io_ops->data;
2866c82d	147
7101d9c2 JA	148	fio_ro_check(td, io_u);
7101d9c2 JA	149
2dc1bbeb	150	if (ld->iocbs_nr == (int) td->o.iodepth)
755200a3 JA	151	return FIO_Q_BUSY;
	152
	153	/*
	154	* fsync is tricky, since it can fail and we need to do it
	155	* serialized with other io. the reason is that linux doesn't
	156	* support aio fsync yet. So return busy for the case where we
	157	* have pending io, to let fio complete those first.
	158	*/
f011531e	159	if (ddir_sync(io_u->ddir)) {
755200a3 JA	160	if (ld->iocbs_nr)
755200a3 JA	161	return FIO_Q_BUSY;
5f9099ea	162
f011531e	163	do_io_u_sync(td, io_u);
755200a3 JA	164	return FIO_Q_COMPLETED;
	165	}
	166
a5f3027c JA	167	if (io_u->ddir == DDIR_TRIM) {
	168	if (ld->iocbs_nr)
	169	return FIO_Q_BUSY;
	170
	171	do_io_u_trim(td, io_u);
	172	return FIO_Q_COMPLETED;
	173	}
	174
755200a3	175	ld->iocbs[ld->iocbs_nr] = &io_u->iocb;
7e77dd02	176	ld->io_us[ld->iocbs_nr] = io_u;
755200a3 JA	177	ld->iocbs_nr++;
	178	return FIO_Q_QUEUED;
	179	}
	180
7e77dd02 JA	181	static void fio_libaio_queued(struct thread_data td, struct io_u *io_us,
	182	unsigned int nr)
	183	{
	184	struct timeval now;
	185	unsigned int i;
	186
12d9d841 JA	187	if (!fio_fill_issue_time(td))
	188	return;
	189
7e77dd02 JA	190	fio_gettime(&now, NULL);
	191
	192	for (i = 0; i < nr; i++) {
	193	struct io_u *io_u = io_us[i];
	194
	195	memcpy(&io_u->issue_time, &now, sizeof(now));
	196	io_u_queued(td, io_u);
	197	}
	198	}
	199
755200a3 JA	200	static int fio_libaio_commit(struct thread_data *td)
	201	{
	202	struct libaio_data *ld = td->io_ops->data;
	203	struct iocb **iocbs;
7e77dd02	204	struct io_u **io_us;
5e00c2c4	205	int ret;
755200a3 JA	206
	207	if (!ld->iocbs_nr)
	208	return 0;
	209
7e77dd02	210	io_us = ld->io_us;
755200a3	211	iocbs = ld->iocbs;
2866c82d	212	do {
5e00c2c4 JA	213	ret = io_submit(ld->aio_ctx, ld->iocbs_nr, iocbs);
5e00c2c4 JA	214	if (ret > 0) {
7e77dd02	215	fio_libaio_queued(td, io_us, ret);
838bc709	216	io_u_mark_submit(td, ret);
5e00c2c4	217	ld->iocbs_nr -= ret;
7e77dd02	218	io_us += ret;
755200a3	219	iocbs += ret;
5e00c2c4	220	ret = 0;
838bc709 JA	221	} else if (!ret \|\| ret == -EAGAIN \|\| ret == -EINTR) {
	222	if (!ret)
	223	io_u_mark_submit(td, ret);
2866c82d	224	continue;
838bc709	225	} else
2866c82d	226	break;
5e00c2c4	227	} while (ld->iocbs_nr);
2866c82d	228
36167d82	229	return ret;
2866c82d JA	230	}
	231
	232	static int fio_libaio_cancel(struct thread_data td, struct io_u io_u)
	233	{
	234	struct libaio_data *ld = td->io_ops->data;
	235
	236	return io_cancel(ld->aio_ctx, &io_u->iocb, ld->aio_events);
	237	}
	238
	239	static void fio_libaio_cleanup(struct thread_data *td)
	240	{
	241	struct libaio_data *ld = td->io_ops->data;
	242
	243	if (ld) {
	244	io_destroy(ld->aio_ctx);
7e77dd02 JA	245	free(ld->aio_events);
	246	free(ld->iocbs);
	247	free(ld->io_us);
2866c82d	248	free(ld);
2866c82d JA	249	}
	250	}
	251
	252	static int fio_libaio_init(struct thread_data *td)
	253	{
	254	struct libaio_data ld = malloc(sizeof(ld));
c90e1017 JA	255	struct libaio_options *o = td->eo;
c90e1017 JA	256	int err = 0;
2866c82d JA	257
2866c82d JA	258	memset(ld, 0, sizeof(*ld));
c1db2dce	259
c90e1017 JA	260	/*
	261	* First try passing in 0 for queue depth, since we don't
	262	* care about the user ring. If that fails, the kernel is too old
	263	* and we need the right depth.
	264	*/
	265	if (!o->userspace_reap)
c224ec05	266	err = io_queue_init(INT_MAX, &ld->aio_ctx);
c90e1017 JA	267	if (o->userspace_reap \|\| err == -EINVAL)
c90e1017 JA	268	err = io_queue_init(td->o.iodepth, &ld->aio_ctx);
c1db2dce JA	269	if (err) {
c1db2dce JA	270	td_verror(td, -err, "io_queue_init");
75de55ac	271	log_err("fio: check /proc/sys/fs/aio-max-nr\n");
cb781c75	272	free(ld);
2866c82d JA	273	return 1;
	274	}
	275
2dc1bbeb JA	276	ld->aio_events = malloc(td->o.iodepth * sizeof(struct io_event));
	277	memset(ld->aio_events, 0, td->o.iodepth * sizeof(struct io_event));
	278	ld->iocbs = malloc(td->o.iodepth * sizeof(struct iocb *));
755200a3	279	memset(ld->iocbs, 0, sizeof(struct iocb *));
2dc1bbeb JA	280	ld->io_us = malloc(td->o.iodepth * sizeof(struct io_u *));
2dc1bbeb JA	281	memset(ld->io_us, 0, td->o.iodepth * sizeof(struct io_u *));
755200a3 JA	282	ld->iocbs_nr = 0;
755200a3 JA	283
2866c82d JA	284	td->io_ops->data = ld;
	285	return 0;
	286	}
	287
5f350952	288	static struct ioengine_ops ioengine = {
de890a1e SL	289	.name = "libaio",
	290	.version = FIO_IOOPS_VERSION,
	291	.init = fio_libaio_init,
	292	.prep = fio_libaio_prep,
	293	.queue = fio_libaio_queue,
	294	.commit = fio_libaio_commit,
	295	.cancel = fio_libaio_cancel,
	296	.getevents = fio_libaio_getevents,
	297	.event = fio_libaio_event,
	298	.cleanup = fio_libaio_cleanup,
	299	.open_file = generic_open_file,
	300	.close_file = generic_close_file,
	301	.get_file_size = generic_get_file_size,
	302	.options = options,
	303	.option_struct_size = sizeof(struct libaio_options),
2866c82d	304	};
34cfcdaf	305
5f350952 JA	306	static void fio_init fio_libaio_register(void)
	307	{
	308	register_ioengine(&ioengine);
	309	}
	310
	311	static void fio_exit fio_libaio_unregister(void)
	312	{
	313	unregister_ioengine(&ioengine);
	314	}