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