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