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