[fio.git] / engines / libaio.c

/*
 * libaio engine
 *
 * IO engine using the Linux native aio interface.
 *
 */
#include <stdlib.h>
#include <unistd.h>
#include <errno.h>
#include <libaio.h>
#include <sys/time.h>
#include <sys/resource.h>

#include "../fio.h"
#include "../lib/pow2.h"
#include "../optgroup.h"
#include "../lib/memalign.h"

static int fio_libaio_commit(struct thread_data *td);

struct libaio_data {
	io_context_t aio_ctx;
	struct io_event *aio_events;
	struct iocb **iocbs;
	struct io_u **io_us;

	struct io_u **io_u_index;

	/*
	 * 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 {
	void *pad;
	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;
	struct iocb *iocb = &io_u->iocb;

	if (io_u->ddir == DDIR_READ) {
		io_prep_pread(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
	} else if (io_u->ddir == DDIR_WRITE) {
		io_prep_pwrite(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
	} else if (ddir_sync(io_u->ddir))
		io_prep_fsync(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_min == 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 ((min && r == 0) || r == -EAGAIN) {
			fio_libaio_commit(td);
			if (actual_min)
				usleep(10);
		} else if (r != -EINTR)
			break;
	} while (events < min);

	return r < 0 ? r : events;
}

static enum fio_q_status 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);
		io_u_mark_submit(td, 1);
		io_u_mark_complete(td, 1);
		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 timespec 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 timespec ts;
	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;
			wait_start = 0;
		} else if (ret == -EINTR || !ret) {
			if (!ret)
				io_u_mark_submit(td, ret);
			wait_start = 0;
			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(&ts, NULL);
				wait_start = 1;
			} else if (mtime_since_now(&ts) > 30000) {
				log_err("fio: aio appears to be stalled, giving up\n");
				break;
			}
			usleep(1);
			continue;
		} else if (ret == -ENOMEM) {
			/*
			 * If we get -ENOMEM, reap events if we can. If
			 * we cannot, treat it as a fatal event since there's
			 * nothing we can do about it.
			 */
			if (ld->queued)
				ret = 0;
			break;
		} 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) {
		/*
		 * Work-around to avoid huge RCU stalls at exit time. If we
		 * don't do this here, then it'll be torn down by exit_aio().
		 * But for that case we can parallellize the freeing, thus
		 * speeding it up a lot.
		 */
		if (!(td->flags & TD_F_CHILD))
			io_destroy(ld->aio_ctx);
		free(ld->aio_events);
		free(ld->iocbs);
		free(ld->io_us);
		free(ld);
	}
}

static int fio_libaio_post_init(struct thread_data *td)
{
	struct libaio_data *ld = td->io_ops_data;
	int err;

	err = io_queue_init(td->o.iodepth, &ld->aio_ctx);
	if (err) {
		td_verror(td, -err, "io_queue_init");
		return 1;
	}

	return 0;
}

static int fio_libaio_init(struct thread_data *td)
{
	struct libaio_data *ld;

	ld = calloc(1, sizeof(*ld));

	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,
	.flags			= FIO_ASYNCIO_SYNC_TRIM,
	.init			= fio_libaio_init,
	.post_init		= fio_libaio_post_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 <stdlib.h>
	8	#include <unistd.h>
	9	#include <errno.h>
	10	#include <libaio.h>
	11	#include <sys/time.h>
	12	#include <sys/resource.h>
	13
	14	#include "../fio.h"
	15	#include "../lib/pow2.h"
	16	#include "../optgroup.h"
	17	#include "../lib/memalign.h"
	18
	19	static int fio_libaio_commit(struct thread_data *td);
	20
	21	struct libaio_data {
	22	io_context_t aio_ctx;
	23	struct io_event *aio_events;
	24	struct iocb **iocbs;
	25	struct io_u **io_us;
	26
	27	struct io_u **io_u_index;
	28
	29	/*
	30	* Basic ring buffer. 'head' is incremented in _queue(), and
	31	* 'tail' is incremented in _commit(). We keep 'queued' so
	32	* that we know if the ring is full or empty, when
	33	* 'head' == 'tail'. 'entries' is the ring size, and
	34	* 'is_pow2' is just an optimization to use AND instead of
	35	* modulus to get the remainder on ring increment.
	36	*/
	37	int is_pow2;
	38	unsigned int entries;
	39	unsigned int queued;
	40	unsigned int head;
	41	unsigned int tail;
	42	};
	43
	44	struct libaio_options {
	45	void *pad;
	46	unsigned int userspace_reap;
	47	};
	48
	49	static struct fio_option options[] = {
	50	{
	51	.name = "userspace_reap",
	52	.lname = "Libaio userspace reaping",
	53	.type = FIO_OPT_STR_SET,
	54	.off1 = offsetof(struct libaio_options, userspace_reap),
	55	.help = "Use alternative user-space reap implementation",
	56	.category = FIO_OPT_C_ENGINE,
	57	.group = FIO_OPT_G_LIBAIO,
	58	},
	59	{
	60	.name = NULL,
	61	},
	62	};
	63
	64	static inline void ring_inc(struct libaio_data ld, unsigned int val,
	65	unsigned int add)
	66	{
	67	if (ld->is_pow2)
	68	val = (val + add) & (ld->entries - 1);
	69	else
	70	val = (val + add) % ld->entries;
	71	}
	72
	73	static int fio_libaio_prep(struct thread_data fio_unused td, struct io_u io_u)
	74	{
	75	struct fio_file *f = io_u->file;
	76	struct iocb *iocb = &io_u->iocb;
	77
	78	if (io_u->ddir == DDIR_READ) {
	79	io_prep_pread(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
	80	} else if (io_u->ddir == DDIR_WRITE) {
	81	io_prep_pwrite(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
	82	} else if (ddir_sync(io_u->ddir))
	83	io_prep_fsync(iocb, f->fd);
	84
	85	return 0;
	86	}
	87
	88	static struct io_u fio_libaio_event(struct thread_data td, int event)
	89	{
	90	struct libaio_data *ld = td->io_ops_data;
	91	struct io_event *ev;
	92	struct io_u *io_u;
	93
	94	ev = ld->aio_events + event;
	95	io_u = container_of(ev->obj, struct io_u, iocb);
	96
	97	if (ev->res != io_u->xfer_buflen) {
	98	if (ev->res > io_u->xfer_buflen)
	99	io_u->error = -ev->res;
	100	else
	101	io_u->resid = io_u->xfer_buflen - ev->res;
	102	} else
	103	io_u->error = 0;
	104
	105	return io_u;
	106	}
	107
	108	struct aio_ring {
	109	unsigned id; /** kernel internal index number */
	110	unsigned nr; /** number of io_events */
	111	unsigned head;
	112	unsigned tail;
	113
	114	unsigned magic;
	115	unsigned compat_features;
	116	unsigned incompat_features;
	117	unsigned header_length; /** size of aio_ring */
	118
	119	struct io_event events[0];
	120	};
	121
	122	#define AIO_RING_MAGIC 0xa10a10a1
	123
	124	static int user_io_getevents(io_context_t aio_ctx, unsigned int max,
	125	struct io_event *events)
	126	{
	127	long i = 0;
	128	unsigned head;
	129	struct aio_ring ring = (struct aio_ring) aio_ctx;
	130
	131	while (i < max) {
	132	head = ring->head;
	133
	134	if (head == ring->tail) {
	135	/* There are no more completions */
	136	break;
	137	} else {
	138	/* There is another completion to reap */
	139	events[i] = ring->events[head];
	140	read_barrier();
	141	ring->head = (head + 1) % ring->nr;
	142	i++;
	143	}
	144	}
	145
	146	return i;
	147	}
	148
	149	static int fio_libaio_getevents(struct thread_data *td, unsigned int min,
	150	unsigned int max, const struct timespec *t)
	151	{
	152	struct libaio_data *ld = td->io_ops_data;
	153	struct libaio_options *o = td->eo;
	154	unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
	155	struct timespec __lt, *lt = NULL;
	156	int r, events = 0;
	157
	158	if (t) {
	159	__lt = *t;
	160	lt = &__lt;
	161	}
	162
	163	do {
	164	if (o->userspace_reap == 1
	165	&& actual_min == 0
	166	&& ((struct aio_ring *)(ld->aio_ctx))->magic
	167	== AIO_RING_MAGIC) {
	168	r = user_io_getevents(ld->aio_ctx, max,
	169	ld->aio_events + events);
	170	} else {
	171	r = io_getevents(ld->aio_ctx, actual_min,
	172	max, ld->aio_events + events, lt);
	173	}
	174	if (r > 0)
	175	events += r;
	176	else if ((min && r == 0) \|\| r == -EAGAIN) {
	177	fio_libaio_commit(td);
	178	if (actual_min)
	179	usleep(10);
	180	} else if (r != -EINTR)
	181	break;
	182	} while (events < min);
	183
	184	return r < 0 ? r : events;
	185	}
	186
	187	static enum fio_q_status fio_libaio_queue(struct thread_data *td,
	188	struct io_u *io_u)
	189	{
	190	struct libaio_data *ld = td->io_ops_data;
	191
	192	fio_ro_check(td, io_u);
	193
	194	if (ld->queued == td->o.iodepth)
	195	return FIO_Q_BUSY;
	196
	197	/*
	198	* fsync is tricky, since it can fail and we need to do it
	199	* serialized with other io. the reason is that linux doesn't
	200	* support aio fsync yet. So return busy for the case where we
	201	* have pending io, to let fio complete those first.
	202	*/
	203	if (ddir_sync(io_u->ddir)) {
	204	if (ld->queued)
	205	return FIO_Q_BUSY;
	206
	207	do_io_u_sync(td, io_u);
	208	return FIO_Q_COMPLETED;
	209	}
	210
	211	if (io_u->ddir == DDIR_TRIM) {
	212	if (ld->queued)
	213	return FIO_Q_BUSY;
	214
	215	do_io_u_trim(td, io_u);
	216	io_u_mark_submit(td, 1);
	217	io_u_mark_complete(td, 1);
	218	return FIO_Q_COMPLETED;
	219	}
	220
	221	ld->iocbs[ld->head] = &io_u->iocb;
	222	ld->io_us[ld->head] = io_u;
	223	ring_inc(ld, &ld->head, 1);
	224	ld->queued++;
	225	return FIO_Q_QUEUED;
	226	}
	227
	228	static void fio_libaio_queued(struct thread_data td, struct io_u *io_us,
	229	unsigned int nr)
	230	{
	231	struct timespec now;
	232	unsigned int i;
	233
	234	if (!fio_fill_issue_time(td))
	235	return;
	236
	237	fio_gettime(&now, NULL);
	238
	239	for (i = 0; i < nr; i++) {
	240	struct io_u *io_u = io_us[i];
	241
	242	memcpy(&io_u->issue_time, &now, sizeof(now));
	243	io_u_queued(td, io_u);
	244	}
	245	}
	246
	247	static int fio_libaio_commit(struct thread_data *td)
	248	{
	249	struct libaio_data *ld = td->io_ops_data;
	250	struct iocb **iocbs;
	251	struct io_u **io_us;
	252	struct timespec ts;
	253	int ret, wait_start = 0;
	254
	255	if (!ld->queued)
	256	return 0;
	257
	258	do {
	259	long nr = ld->queued;
	260
	261	nr = min((unsigned int) nr, ld->entries - ld->tail);
	262	io_us = ld->io_us + ld->tail;
	263	iocbs = ld->iocbs + ld->tail;
	264
	265	ret = io_submit(ld->aio_ctx, nr, iocbs);
	266	if (ret > 0) {
	267	fio_libaio_queued(td, io_us, ret);
	268	io_u_mark_submit(td, ret);
	269
	270	ld->queued -= ret;
	271	ring_inc(ld, &ld->tail, ret);
	272	ret = 0;
	273	wait_start = 0;
	274	} else if (ret == -EINTR \|\| !ret) {
	275	if (!ret)
	276	io_u_mark_submit(td, ret);
	277	wait_start = 0;
	278	continue;
	279	} else if (ret == -EAGAIN) {
	280	/*
	281	* If we get EAGAIN, we should break out without
	282	* error and let the upper layer reap some
	283	* events for us. If we have no queued IO, we
	284	* must loop here. If we loop for more than 30s,
	285	* just error out, something must be buggy in the
	286	* IO path.
	287	*/
	288	if (ld->queued) {
	289	ret = 0;
	290	break;
	291	}
	292	if (!wait_start) {
	293	fio_gettime(&ts, NULL);
	294	wait_start = 1;
	295	} else if (mtime_since_now(&ts) > 30000) {
	296	log_err("fio: aio appears to be stalled, giving up\n");
	297	break;
	298	}
	299	usleep(1);
	300	continue;
	301	} else if (ret == -ENOMEM) {
	302	/*
	303	* If we get -ENOMEM, reap events if we can. If
	304	* we cannot, treat it as a fatal event since there's
	305	* nothing we can do about it.
	306	*/
	307	if (ld->queued)
	308	ret = 0;
	309	break;
	310	} else
	311	break;
	312	} while (ld->queued);
	313
	314	return ret;
	315	}
	316
	317	static int fio_libaio_cancel(struct thread_data td, struct io_u io_u)
	318	{
	319	struct libaio_data *ld = td->io_ops_data;
	320
	321	return io_cancel(ld->aio_ctx, &io_u->iocb, ld->aio_events);
	322	}
	323
	324	static void fio_libaio_cleanup(struct thread_data *td)
	325	{
	326	struct libaio_data *ld = td->io_ops_data;
	327
	328	if (ld) {
	329	/*
	330	* Work-around to avoid huge RCU stalls at exit time. If we
	331	* don't do this here, then it'll be torn down by exit_aio().
	332	* But for that case we can parallellize the freeing, thus
	333	* speeding it up a lot.
	334	*/
	335	if (!(td->flags & TD_F_CHILD))
	336	io_destroy(ld->aio_ctx);
	337	free(ld->aio_events);
	338	free(ld->iocbs);
	339	free(ld->io_us);
	340	free(ld);
	341	}
	342	}
	343
	344	static int fio_libaio_post_init(struct thread_data *td)
	345	{
	346	struct libaio_data *ld = td->io_ops_data;
	347	int err;
	348
	349	err = io_queue_init(td->o.iodepth, &ld->aio_ctx);
	350	if (err) {
	351	td_verror(td, -err, "io_queue_init");
	352	return 1;
	353	}
	354
	355	return 0;
	356	}
	357
	358	static int fio_libaio_init(struct thread_data *td)
	359	{
	360	struct libaio_data *ld;
	361
	362	ld = calloc(1, sizeof(*ld));
	363
	364	ld->entries = td->o.iodepth;
	365	ld->is_pow2 = is_power_of_2(ld->entries);
	366	ld->aio_events = calloc(ld->entries, sizeof(struct io_event));
	367	ld->iocbs = calloc(ld->entries, sizeof(struct iocb *));
	368	ld->io_us = calloc(ld->entries, sizeof(struct io_u *));
	369
	370	td->io_ops_data = ld;
	371	return 0;
	372	}
	373
	374	static struct ioengine_ops ioengine = {
	375	.name = "libaio",
	376	.version = FIO_IOOPS_VERSION,
	377	.flags = FIO_ASYNCIO_SYNC_TRIM,
	378	.init = fio_libaio_init,
	379	.post_init = fio_libaio_post_init,
	380	.prep = fio_libaio_prep,
	381	.queue = fio_libaio_queue,
	382	.commit = fio_libaio_commit,
	383	.cancel = fio_libaio_cancel,
	384	.getevents = fio_libaio_getevents,
	385	.event = fio_libaio_event,
	386	.cleanup = fio_libaio_cleanup,
	387	.open_file = generic_open_file,
	388	.close_file = generic_close_file,
	389	.get_file_size = generic_get_file_size,
	390	.options = options,
	391	.option_struct_size = sizeof(struct libaio_options),
	392	};
	393
	394	static void fio_init fio_libaio_register(void)
	395	{
	396	register_ioengine(&ioengine);
	397	}
	398
	399	static void fio_exit fio_libaio_unregister(void)
	400	{
	401	unregister_ioengine(&ioengine);
	402	}