[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;
	unsigned int hipri;
};

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	unsigned int hipri;
	48	};
	49
	50	static struct fio_option options[] = {
	51	{
	52	.name = "userspace_reap",
	53	.lname = "Libaio userspace reaping",
	54	.type = FIO_OPT_STR_SET,
	55	.off1 = offsetof(struct libaio_options, userspace_reap),
	56	.help = "Use alternative user-space reap implementation",
	57	.category = FIO_OPT_C_ENGINE,
	58	.group = FIO_OPT_G_LIBAIO,
	59	},
	60	{
	61	.name = NULL,
	62	},
	63	};
	64
	65	static inline void ring_inc(struct libaio_data ld, unsigned int val,
	66	unsigned int add)
	67	{
	68	if (ld->is_pow2)
	69	val = (val + add) & (ld->entries - 1);
	70	else
	71	val = (val + add) % ld->entries;
	72	}
	73
	74	static int fio_libaio_prep(struct thread_data fio_unused td, struct io_u io_u)
	75	{
	76	struct fio_file *f = io_u->file;
	77	struct iocb *iocb = &io_u->iocb;
	78
	79	if (io_u->ddir == DDIR_READ) {
	80	io_prep_pread(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
	81	} else if (io_u->ddir == DDIR_WRITE) {
	82	io_prep_pwrite(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
	83	} else if (ddir_sync(io_u->ddir))
	84	io_prep_fsync(iocb, f->fd);
	85
	86	return 0;
	87	}
	88
	89	static struct io_u fio_libaio_event(struct thread_data td, int event)
	90	{
	91	struct libaio_data *ld = td->io_ops_data;
	92	struct io_event *ev;
	93	struct io_u *io_u;
	94
	95	ev = ld->aio_events + event;
	96	io_u = container_of(ev->obj, struct io_u, iocb);
	97
	98	if (ev->res != io_u->xfer_buflen) {
	99	if (ev->res > io_u->xfer_buflen)
	100	io_u->error = -ev->res;
	101	else
	102	io_u->resid = io_u->xfer_buflen - ev->res;
	103	} else
	104	io_u->error = 0;
	105
	106	return io_u;
	107	}
	108
	109	struct aio_ring {
	110	unsigned id; /** kernel internal index number */
	111	unsigned nr; /** number of io_events */
	112	unsigned head;
	113	unsigned tail;
	114
	115	unsigned magic;
	116	unsigned compat_features;
	117	unsigned incompat_features;
	118	unsigned header_length; /** size of aio_ring */
	119
	120	struct io_event events[0];
	121	};
	122
	123	#define AIO_RING_MAGIC 0xa10a10a1
	124
	125	static int user_io_getevents(io_context_t aio_ctx, unsigned int max,
	126	struct io_event *events)
	127	{
	128	long i = 0;
	129	unsigned head;
	130	struct aio_ring ring = (struct aio_ring) aio_ctx;
	131
	132	while (i < max) {
	133	head = ring->head;
	134
	135	if (head == ring->tail) {
	136	/* There are no more completions */
	137	break;
	138	} else {
	139	/* There is another completion to reap */
	140	events[i] = ring->events[head];
	141	read_barrier();
	142	ring->head = (head + 1) % ring->nr;
	143	i++;
	144	}
	145	}
	146
	147	return i;
	148	}
	149
	150	static int fio_libaio_getevents(struct thread_data *td, unsigned int min,
	151	unsigned int max, const struct timespec *t)
	152	{
	153	struct libaio_data *ld = td->io_ops_data;
	154	struct libaio_options *o = td->eo;
	155	unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
	156	struct timespec __lt, *lt = NULL;
	157	int r, events = 0;
	158
	159	if (t) {
	160	__lt = *t;
	161	lt = &__lt;
	162	}
	163
	164	do {
	165	if (o->userspace_reap == 1
	166	&& actual_min == 0
	167	&& ((struct aio_ring *)(ld->aio_ctx))->magic
	168	== AIO_RING_MAGIC) {
	169	r = user_io_getevents(ld->aio_ctx, max,
	170	ld->aio_events + events);
	171	} else {
	172	r = io_getevents(ld->aio_ctx, actual_min,
	173	max, ld->aio_events + events, lt);
	174	}
	175	if (r > 0)
	176	events += r;
	177	else if ((min && r == 0) \|\| r == -EAGAIN) {
	178	fio_libaio_commit(td);
	179	if (actual_min)
	180	usleep(10);
	181	} else if (r != -EINTR)
	182	break;
	183	} while (events < min);
	184
	185	return r < 0 ? r : events;
	186	}
	187
	188	static enum fio_q_status fio_libaio_queue(struct thread_data *td,
	189	struct io_u *io_u)
	190	{
	191	struct libaio_data *ld = td->io_ops_data;
	192
	193	fio_ro_check(td, io_u);
	194
	195	if (ld->queued == td->o.iodepth)
	196	return FIO_Q_BUSY;
	197
	198	/*
	199	* fsync is tricky, since it can fail and we need to do it
	200	* serialized with other io. the reason is that linux doesn't
	201	* support aio fsync yet. So return busy for the case where we
	202	* have pending io, to let fio complete those first.
	203	*/
	204	if (ddir_sync(io_u->ddir)) {
	205	if (ld->queued)
	206	return FIO_Q_BUSY;
	207
	208	do_io_u_sync(td, io_u);
	209	return FIO_Q_COMPLETED;
	210	}
	211
	212	if (io_u->ddir == DDIR_TRIM) {
	213	if (ld->queued)
	214	return FIO_Q_BUSY;
	215
	216	do_io_u_trim(td, io_u);
	217	io_u_mark_submit(td, 1);
	218	io_u_mark_complete(td, 1);
	219	return FIO_Q_COMPLETED;
	220	}
	221
	222	ld->iocbs[ld->head] = &io_u->iocb;
	223	ld->io_us[ld->head] = io_u;
	224	ring_inc(ld, &ld->head, 1);
	225	ld->queued++;
	226	return FIO_Q_QUEUED;
	227	}
	228
	229	static void fio_libaio_queued(struct thread_data td, struct io_u *io_us,
	230	unsigned int nr)
	231	{
	232	struct timespec now;
	233	unsigned int i;
	234
	235	if (!fio_fill_issue_time(td))
	236	return;
	237
	238	fio_gettime(&now, NULL);
	239
	240	for (i = 0; i < nr; i++) {
	241	struct io_u *io_u = io_us[i];
	242
	243	memcpy(&io_u->issue_time, &now, sizeof(now));
	244	io_u_queued(td, io_u);
	245	}
	246	}
	247
	248	static int fio_libaio_commit(struct thread_data *td)
	249	{
	250	struct libaio_data *ld = td->io_ops_data;
	251	struct iocb **iocbs;
	252	struct io_u **io_us;
	253	struct timespec ts;
	254	int ret, wait_start = 0;
	255
	256	if (!ld->queued)
	257	return 0;
	258
	259	do {
	260	long nr = ld->queued;
	261
	262	nr = min((unsigned int) nr, ld->entries - ld->tail);
	263	io_us = ld->io_us + ld->tail;
	264	iocbs = ld->iocbs + ld->tail;
	265
	266	ret = io_submit(ld->aio_ctx, nr, iocbs);
	267	if (ret > 0) {
	268	fio_libaio_queued(td, io_us, ret);
	269	io_u_mark_submit(td, ret);
	270
	271	ld->queued -= ret;
	272	ring_inc(ld, &ld->tail, ret);
	273	ret = 0;
	274	wait_start = 0;
	275	} else if (ret == -EINTR \|\| !ret) {
	276	if (!ret)
	277	io_u_mark_submit(td, ret);
	278	wait_start = 0;
	279	continue;
	280	} else if (ret == -EAGAIN) {
	281	/*
	282	* If we get EAGAIN, we should break out without
	283	* error and let the upper layer reap some
	284	* events for us. If we have no queued IO, we
	285	* must loop here. If we loop for more than 30s,
	286	* just error out, something must be buggy in the
	287	* IO path.
	288	*/
	289	if (ld->queued) {
	290	ret = 0;
	291	break;
	292	}
	293	if (!wait_start) {
	294	fio_gettime(&ts, NULL);
	295	wait_start = 1;
	296	} else if (mtime_since_now(&ts) > 30000) {
	297	log_err("fio: aio appears to be stalled, giving up\n");
	298	break;
	299	}
	300	usleep(1);
	301	continue;
	302	} else if (ret == -ENOMEM) {
	303	/*
	304	* If we get -ENOMEM, reap events if we can. If
	305	* we cannot, treat it as a fatal event since there's
	306	* nothing we can do about it.
	307	*/
	308	if (ld->queued)
	309	ret = 0;
	310	break;
	311	} else
	312	break;
	313	} while (ld->queued);
	314
	315	return ret;
	316	}
	317
	318	static int fio_libaio_cancel(struct thread_data td, struct io_u io_u)
	319	{
	320	struct libaio_data *ld = td->io_ops_data;
	321
	322	return io_cancel(ld->aio_ctx, &io_u->iocb, ld->aio_events);
	323	}
	324
	325	static void fio_libaio_cleanup(struct thread_data *td)
	326	{
	327	struct libaio_data *ld = td->io_ops_data;
	328
	329	if (ld) {
	330	/*
	331	* Work-around to avoid huge RCU stalls at exit time. If we
	332	* don't do this here, then it'll be torn down by exit_aio().
	333	* But for that case we can parallellize the freeing, thus
	334	* speeding it up a lot.
	335	*/
	336	if (!(td->flags & TD_F_CHILD))
	337	io_destroy(ld->aio_ctx);
	338	free(ld->aio_events);
	339	free(ld->iocbs);
	340	free(ld->io_us);
	341	free(ld);
	342	}
	343	}
	344
	345	static int fio_libaio_post_init(struct thread_data *td)
	346	{
	347	struct libaio_data *ld = td->io_ops_data;
	348	int err;
	349
	350	err = io_queue_init(td->o.iodepth, &ld->aio_ctx);
	351	if (err) {
	352	td_verror(td, -err, "io_queue_init");
	353	return 1;
	354	}
	355
	356	return 0;
	357	}
	358
	359	static int fio_libaio_init(struct thread_data *td)
	360	{
	361	struct libaio_data *ld;
	362
	363	ld = calloc(1, sizeof(*ld));
	364
	365	ld->entries = td->o.iodepth;
	366	ld->is_pow2 = is_power_of_2(ld->entries);
	367	ld->aio_events = calloc(ld->entries, sizeof(struct io_event));
	368	ld->iocbs = calloc(ld->entries, sizeof(struct iocb *));
	369	ld->io_us = calloc(ld->entries, sizeof(struct io_u *));
	370
	371	td->io_ops_data = ld;
	372	return 0;
	373	}
	374
	375	static struct ioengine_ops ioengine = {
	376	.name = "libaio",
	377	.version = FIO_IOOPS_VERSION,
	378	.flags = FIO_ASYNCIO_SYNC_TRIM,
	379	.init = fio_libaio_init,
	380	.post_init = fio_libaio_post_init,
	381	.prep = fio_libaio_prep,
	382	.queue = fio_libaio_queue,
	383	.commit = fio_libaio_commit,
	384	.cancel = fio_libaio_cancel,
	385	.getevents = fio_libaio_getevents,
	386	.event = fio_libaio_event,
	387	.cleanup = fio_libaio_cleanup,
	388	.open_file = generic_open_file,
	389	.close_file = generic_close_file,
	390	.get_file_size = generic_get_file_size,
	391	.options = options,
	392	.option_struct_size = sizeof(struct libaio_options),
	393	};
	394
	395	static void fio_init fio_libaio_register(void)
	396	{
	397	register_ioengine(&ioengine);
	398	}
	399
	400	static void fio_exit fio_libaio_unregister(void)
	401	{
	402	unregister_ioengine(&ioengine);
	403	}