[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;

	/*
	 * Basic ring buffer. 'head' is incremented in _queue(), and
	 * 'tail' is incremented in _commit(). We keep 'queued' so
	 * that we know if the ring is full or empty, when
	 * 'head' == 'tail'. 'entries' is the ring size, and
	 * 'is_pow2' is just an optimization to use AND instead of
	 * modulus to get the remainder on ring increment.
	 */
	int is_pow2;
	unsigned int entries;
	unsigned int queued;
	unsigned int head;
	unsigned int tail;
};

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_ENGINE,
		.group	= FIO_OPT_G_LIBAIO,
	},
	{
		.name	= NULL,
	},
};

static inline void ring_inc(struct libaio_data *ld, unsigned int *val,
			    unsigned int add)
{
	if (ld->is_pow2)
		*val = (*val + add) & (ld->entries - 1);
	else
		*val = (*val + add) % ld->entries;
}

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, const 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;
	struct timespec __lt, *lt = NULL;
	int r, events = 0;

	if (t) {
		__lt = *t;
		lt = &__lt;
	}

	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, lt);
		}
		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->queued == 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->queued)
			return FIO_Q_BUSY;

		do_io_u_sync(td, io_u);
		return FIO_Q_COMPLETED;
	}

	if (io_u->ddir == DDIR_TRIM) {
		if (ld->queued)
			return FIO_Q_BUSY;

		do_io_u_trim(td, io_u);
		return FIO_Q_COMPLETED;
	}

	ld->iocbs[ld->head] = &io_u->iocb;
	ld->io_us[ld->head] = io_u;
	ring_inc(ld, &ld->head, 1);
	ld->queued++;
	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;
	struct timeval tv;
	int ret, wait_start = 0;

	if (!ld->queued)
		return 0;

	do {
		long nr = ld->queued;

		nr = min((unsigned int) nr, ld->entries - ld->tail);
		io_us = ld->io_us + ld->tail;
		iocbs = ld->iocbs + ld->tail;

		ret = io_submit(ld->aio_ctx, nr, iocbs);
		if (ret > 0) {
			fio_libaio_queued(td, io_us, ret);
			io_u_mark_submit(td, ret);

			ld->queued -= ret;
			ring_inc(ld, &ld->tail, ret);
			ret = 0;
		} else if (ret == -EINTR || !ret) {
			if (!ret)
				io_u_mark_submit(td, ret);
			continue;
		} else if (ret == -EAGAIN) {
			/*
			 * If we get EAGAIN, we should break out without
			 * error and let the upper layer reap some
			 * events for us. If we have no queued IO, we
			 * must loop here. If we loop for more than 30s,
			 * just error out, something must be buggy in the
			 * IO path.
			 */
			if (ld->queued) {
				ret = 0;
				break;
			}
			if (!wait_start) {
				fio_gettime(&tv, NULL);
				wait_start = 0;
			} else if (mtime_since_now(&tv) > 30000) {
				log_err("fio: aio appears to be stalled, giving up\n");
				break;
			}
			usleep(1);
			continue;
		} else
			break;
	} while (ld->queued);

	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_options *o = td->eo;
	struct libaio_data *ld;
	int err = 0;

	ld = calloc(1, 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->entries = td->o.iodepth;
	ld->is_pow2 = is_power_of_2(ld->entries);
	ld->aio_events = calloc(ld->entries, sizeof(struct io_event));
	ld->iocbs = calloc(ld->entries, sizeof(struct iocb *));
	ld->io_us = calloc(ld->entries, sizeof(struct io_u *));

	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
	22	/*
	23	* Basic ring buffer. 'head' is incremented in _queue(), and
	24	* 'tail' is incremented in _commit(). We keep 'queued' so
	25	* that we know if the ring is full or empty, when
	26	* 'head' == 'tail'. 'entries' is the ring size, and
	27	* 'is_pow2' is just an optimization to use AND instead of
	28	* modulus to get the remainder on ring increment.
	29	*/
	30	int is_pow2;
	31	unsigned int entries;
	32	unsigned int queued;
	33	unsigned int head;
	34	unsigned int tail;
	35	};
	36
	37	struct libaio_options {
	38	struct thread_data *td;
	39	unsigned int userspace_reap;
	40	};
	41
	42	static struct fio_option options[] = {
	43	{
	44	.name = "userspace_reap",
	45	.lname = "Libaio userspace reaping",
	46	.type = FIO_OPT_STR_SET,
	47	.off1 = offsetof(struct libaio_options, userspace_reap),
	48	.help = "Use alternative user-space reap implementation",
	49	.category = FIO_OPT_C_ENGINE,
	50	.group = FIO_OPT_G_LIBAIO,
	51	},
	52	{
	53	.name = NULL,
	54	},
	55	};
	56
	57	static inline void ring_inc(struct libaio_data ld, unsigned int val,
	58	unsigned int add)
	59	{
	60	if (ld->is_pow2)
	61	val = (val + add) & (ld->entries - 1);
	62	else
	63	val = (val + add) % ld->entries;
	64	}
	65
	66	static int fio_libaio_prep(struct thread_data fio_unused td, struct io_u io_u)
	67	{
	68	struct fio_file *f = io_u->file;
	69
	70	if (io_u->ddir == DDIR_READ)
	71	io_prep_pread(&io_u->iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
	72	else if (io_u->ddir == DDIR_WRITE)
	73	io_prep_pwrite(&io_u->iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
	74	else if (ddir_sync(io_u->ddir))
	75	io_prep_fsync(&io_u->iocb, f->fd);
	76
	77	return 0;
	78	}
	79
	80	static struct io_u fio_libaio_event(struct thread_data td, int event)
	81	{
	82	struct libaio_data *ld = td->io_ops->data;
	83	struct io_event *ev;
	84	struct io_u *io_u;
	85
	86	ev = ld->aio_events + event;
	87	io_u = container_of(ev->obj, struct io_u, iocb);
	88
	89	if (ev->res != io_u->xfer_buflen) {
	90	if (ev->res > io_u->xfer_buflen)
	91	io_u->error = -ev->res;
	92	else
	93	io_u->resid = io_u->xfer_buflen - ev->res;
	94	} else
	95	io_u->error = 0;
	96
	97	return io_u;
	98	}
	99
	100	struct aio_ring {
	101	unsigned id; /** kernel internal index number */
	102	unsigned nr; /** number of io_events */
	103	unsigned head;
	104	unsigned tail;
	105
	106	unsigned magic;
	107	unsigned compat_features;
	108	unsigned incompat_features;
	109	unsigned header_length; /** size of aio_ring */
	110
	111	struct io_event events[0];
	112	};
	113
	114	#define AIO_RING_MAGIC 0xa10a10a1
	115
	116	static int user_io_getevents(io_context_t aio_ctx, unsigned int max,
	117	struct io_event *events)
	118	{
	119	long i = 0;
	120	unsigned head;
	121	struct aio_ring ring = (struct aio_ring) aio_ctx;
	122
	123	while (i < max) {
	124	head = ring->head;
	125
	126	if (head == ring->tail) {
	127	/* There are no more completions */
	128	break;
	129	} else {
	130	/* There is another completion to reap */
	131	events[i] = ring->events[head];
	132	read_barrier();
	133	ring->head = (head + 1) % ring->nr;
	134	i++;
	135	}
	136	}
	137
	138	return i;
	139	}
	140
	141	static int fio_libaio_getevents(struct thread_data *td, unsigned int min,
	142	unsigned int max, const struct timespec *t)
	143	{
	144	struct libaio_data *ld = td->io_ops->data;
	145	struct libaio_options *o = td->eo;
	146	unsigned actual_min = td->o.iodepth_batch_complete == 0 ? 0 : min;
	147	struct timespec __lt, *lt = NULL;
	148	int r, events = 0;
	149
	150	if (t) {
	151	__lt = *t;
	152	lt = &__lt;
	153	}
	154
	155	do {
	156	if (o->userspace_reap == 1
	157	&& actual_min == 0
	158	&& ((struct aio_ring *)(ld->aio_ctx))->magic
	159	== AIO_RING_MAGIC) {
	160	r = user_io_getevents(ld->aio_ctx, max,
	161	ld->aio_events + events);
	162	} else {
	163	r = io_getevents(ld->aio_ctx, actual_min,
	164	max, ld->aio_events + events, lt);
	165	}
	166	if (r >= 0)
	167	events += r;
	168	else if (r == -EAGAIN)
	169	usleep(100);
	170	} while (events < min);
	171
	172	return r < 0 ? r : events;
	173	}
	174
	175	static int fio_libaio_queue(struct thread_data td, struct io_u io_u)
	176	{
	177	struct libaio_data *ld = td->io_ops->data;
	178
	179	fio_ro_check(td, io_u);
	180
	181	if (ld->queued == td->o.iodepth)
	182	return FIO_Q_BUSY;
	183
	184	/*
	185	* fsync is tricky, since it can fail and we need to do it
	186	* serialized with other io. the reason is that linux doesn't
	187	* support aio fsync yet. So return busy for the case where we
	188	* have pending io, to let fio complete those first.
	189	*/
	190	if (ddir_sync(io_u->ddir)) {
	191	if (ld->queued)
	192	return FIO_Q_BUSY;
	193
	194	do_io_u_sync(td, io_u);
	195	return FIO_Q_COMPLETED;
	196	}
	197
	198	if (io_u->ddir == DDIR_TRIM) {
	199	if (ld->queued)
	200	return FIO_Q_BUSY;
	201
	202	do_io_u_trim(td, io_u);
	203	return FIO_Q_COMPLETED;
	204	}
	205
	206	ld->iocbs[ld->head] = &io_u->iocb;
	207	ld->io_us[ld->head] = io_u;
	208	ring_inc(ld, &ld->head, 1);
	209	ld->queued++;
	210	return FIO_Q_QUEUED;
	211	}
	212
	213	static void fio_libaio_queued(struct thread_data td, struct io_u *io_us,
	214	unsigned int nr)
	215	{
	216	struct timeval now;
	217	unsigned int i;
	218
	219	if (!fio_fill_issue_time(td))
	220	return;
	221
	222	fio_gettime(&now, NULL);
	223
	224	for (i = 0; i < nr; i++) {
	225	struct io_u *io_u = io_us[i];
	226
	227	memcpy(&io_u->issue_time, &now, sizeof(now));
	228	io_u_queued(td, io_u);
	229	}
	230	}
	231
	232	static int fio_libaio_commit(struct thread_data *td)
	233	{
	234	struct libaio_data *ld = td->io_ops->data;
	235	struct iocb **iocbs;
	236	struct io_u **io_us;
	237	struct timeval tv;
	238	int ret, wait_start = 0;
	239
	240	if (!ld->queued)
	241	return 0;
	242
	243	do {
	244	long nr = ld->queued;
	245
	246	nr = min((unsigned int) nr, ld->entries - ld->tail);
	247	io_us = ld->io_us + ld->tail;
	248	iocbs = ld->iocbs + ld->tail;
	249
	250	ret = io_submit(ld->aio_ctx, nr, iocbs);
	251	if (ret > 0) {
	252	fio_libaio_queued(td, io_us, ret);
	253	io_u_mark_submit(td, ret);
	254
	255	ld->queued -= ret;
	256	ring_inc(ld, &ld->tail, ret);
	257	ret = 0;
	258	} else if (ret == -EINTR \|\| !ret) {
	259	if (!ret)
	260	io_u_mark_submit(td, ret);
	261	continue;
	262	} else if (ret == -EAGAIN) {
	263	/*
	264	* If we get EAGAIN, we should break out without
	265	* error and let the upper layer reap some
	266	* events for us. If we have no queued IO, we
	267	* must loop here. If we loop for more than 30s,
	268	* just error out, something must be buggy in the
	269	* IO path.
	270	*/
	271	if (ld->queued) {
	272	ret = 0;
	273	break;
	274	}
	275	if (!wait_start) {
	276	fio_gettime(&tv, NULL);
	277	wait_start = 0;
	278	} else if (mtime_since_now(&tv) > 30000) {
	279	log_err("fio: aio appears to be stalled, giving up\n");
	280	break;
	281	}
	282	usleep(1);
	283	continue;
	284	} else
	285	break;
	286	} while (ld->queued);
	287
	288	return ret;
	289	}
	290
	291	static int fio_libaio_cancel(struct thread_data td, struct io_u io_u)
	292	{
	293	struct libaio_data *ld = td->io_ops->data;
	294
	295	return io_cancel(ld->aio_ctx, &io_u->iocb, ld->aio_events);
	296	}
	297
	298	static void fio_libaio_cleanup(struct thread_data *td)
	299	{
	300	struct libaio_data *ld = td->io_ops->data;
	301
	302	if (ld) {
	303	io_destroy(ld->aio_ctx);
	304	free(ld->aio_events);
	305	free(ld->iocbs);
	306	free(ld->io_us);
	307	free(ld);
	308	}
	309	}
	310
	311	static int fio_libaio_init(struct thread_data *td)
	312	{
	313	struct libaio_options *o = td->eo;
	314	struct libaio_data *ld;
	315	int err = 0;
	316
	317	ld = calloc(1, sizeof(*ld));
	318
	319	/*
	320	* First try passing in 0 for queue depth, since we don't
	321	* care about the user ring. If that fails, the kernel is too old
	322	* and we need the right depth.
	323	*/
	324	if (!o->userspace_reap)
	325	err = io_queue_init(INT_MAX, &ld->aio_ctx);
	326	if (o->userspace_reap \|\| err == -EINVAL)
	327	err = io_queue_init(td->o.iodepth, &ld->aio_ctx);
	328	if (err) {
	329	td_verror(td, -err, "io_queue_init");
	330	log_err("fio: check /proc/sys/fs/aio-max-nr\n");
	331	free(ld);
	332	return 1;
	333	}
	334
	335	ld->entries = td->o.iodepth;
	336	ld->is_pow2 = is_power_of_2(ld->entries);
	337	ld->aio_events = calloc(ld->entries, sizeof(struct io_event));
	338	ld->iocbs = calloc(ld->entries, sizeof(struct iocb *));
	339	ld->io_us = calloc(ld->entries, sizeof(struct io_u *));
	340
	341	td->io_ops->data = ld;
	342	return 0;
	343	}
	344
	345	static struct ioengine_ops ioengine = {
	346	.name = "libaio",
	347	.version = FIO_IOOPS_VERSION,
	348	.init = fio_libaio_init,
	349	.prep = fio_libaio_prep,
	350	.queue = fio_libaio_queue,
	351	.commit = fio_libaio_commit,
	352	.cancel = fio_libaio_cancel,
	353	.getevents = fio_libaio_getevents,
	354	.event = fio_libaio_event,
	355	.cleanup = fio_libaio_cleanup,
	356	.open_file = generic_open_file,
	357	.close_file = generic_close_file,
	358	.get_file_size = generic_get_file_size,
	359	.options = options,
	360	.option_struct_size = sizeof(struct libaio_options),
	361	};
	362
	363	static void fio_init fio_libaio_register(void)
	364	{
	365	register_ioengine(&ioengine);
	366	}
	367
	368	static void fio_exit fio_libaio_unregister(void)
	369	{
	370	unregister_ioengine(&ioengine);
	371	}