[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"
#include "cmdprio.h"

/* Should be defined in newest aio_abi.h */
#ifndef IOCB_FLAG_IOPRIO
#define IOCB_FLAG_IOPRIO    (1 << 1)
#endif

/* Hack for libaio < 0.3.111 */
#ifndef CONFIG_LIBAIO_RW_FLAGS
#define aio_rw_flags __pad2
#endif

static int fio_libaio_commit(struct thread_data *td);
static int fio_libaio_init(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;
	struct iovec *iovecs;		/* for vectored requests */

	/*
	 * 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 cmdprio cmdprio;
};

struct libaio_options {
	struct thread_data *td;
	unsigned int userspace_reap;
	struct cmdprio_options cmdprio_options;
	unsigned int nowait;
	unsigned int vectored;
};

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	= "nowait",
		.lname	= "RWF_NOWAIT",
		.type	= FIO_OPT_BOOL,
		.off1	= offsetof(struct libaio_options, nowait),
		.help	= "Set RWF_NOWAIT for reads/writes",
		.category = FIO_OPT_C_ENGINE,
		.group	= FIO_OPT_G_LIBAIO,
	},
	{
		.name	= "libaio_vectored",
		.lname	= "Use libaio preadv,pwritev",
		.type	= FIO_OPT_BOOL,
		.off1	= offsetof(struct libaio_options, vectored),
		.help	= "Use libaio {preadv,pwritev} instead of libaio {pread,pwrite}",
		.category = FIO_OPT_C_ENGINE,
		.group	= FIO_OPT_G_LIBAIO,
	},

	CMDPRIO_OPTIONS(struct libaio_options, 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 *td, struct io_u *io_u)
{
	struct libaio_options *o = td->eo;
	struct fio_file *f = io_u->file;
	struct iocb *iocb = &io_u->iocb;
	struct libaio_data *ld = td->io_ops_data;

	if (io_u->ddir == DDIR_READ) {
		if (o->vectored) {
			struct iovec *iov = &ld->iovecs[io_u->index];

			iov->iov_base = io_u->xfer_buf;
			iov->iov_len = (size_t)io_u->xfer_buflen;
			io_prep_preadv(iocb, f->fd, iov, 1, io_u->offset);
		} else {
			io_prep_pread(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen,
						  io_u->offset);
		}
		if (o->nowait)
			iocb->aio_rw_flags |= RWF_NOWAIT;
	} else if (io_u->ddir == DDIR_WRITE) {
		if (o->vectored) {
			struct iovec *iov = &ld->iovecs[io_u->index];

			iov->iov_base = io_u->xfer_buf;
			iov->iov_len = (size_t)io_u->xfer_buflen;
			io_prep_pwritev(iocb, f->fd, iov, 1, io_u->offset);
		} else {
			io_prep_pwrite(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen,
						   io_u->offset);
		}
		if (o->nowait)
			iocb->aio_rw_flags |= RWF_NOWAIT;
#ifdef FIO_HAVE_RWF_ATOMIC
		if (td->o.oatomic)
			iocb->aio_rw_flags |= RWF_ATOMIC;
#endif
	} else if (ddir_sync(io_u->ddir))
		io_prep_fsync(iocb, f->fd);

	return 0;
}

static inline void fio_libaio_cmdprio_prep(struct thread_data *td,
					   struct io_u *io_u)
{
	struct libaio_data *ld = td->io_ops_data;
	struct cmdprio *cmdprio = &ld->cmdprio;

	if (fio_cmdprio_set_ioprio(td, cmdprio, io_u)) {
		io_u->iocb.aio_reqprio = io_u->ioprio;
		io_u->iocb.u.c.flags |= IOCB_FLAG_IOPRIO;
	}
}

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];
			atomic_store_release(&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 - events,
				ld->aio_events + events);
		} else {
			r = io_getevents(ld->aio_ctx, actual_min,
				max - events, ld->aio_events + events, lt);
		}
		if (r > 0) {
			events += r;
			actual_min -= min((unsigned int)events, actual_min);
		}
		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;
	}

	if (ld->cmdprio.mode != CMDPRIO_MODE_NONE)
		fio_libaio_cmdprio_prep(td, io_u);

	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);
	}

	/*
	 * only used for iolog
	 */
	if (td->o.read_iolog_file)
		memcpy(&td->last_issue, &now, sizeof(now));
}

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);

		fio_cmdprio_cleanup(&ld->cmdprio);
		free(ld->iovecs);
		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;
	struct libaio_options *o = td->eo;
	int ret;

	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 *));
	ld->iovecs = calloc(ld->entries, sizeof(ld->iovecs[0]));

	td->io_ops_data = ld;

	ret = fio_cmdprio_init(td, &ld->cmdprio, &o->cmdprio_options);
	if (ret) {
		td_verror(td, EINVAL, "fio_libaio_init");
		return 1;
	}

	return 0;
}

FIO_STATIC struct ioengine_ops ioengine = {
	.name			= "libaio",
	.version		= FIO_IOOPS_VERSION,
	.flags			= FIO_ASYNCIO_SYNC_TRIM |
					FIO_ASYNCIO_SETS_ISSUE_TIME |
					FIO_ATOMICWRITES,
	.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
2866c82d	1	/*
da751ca9 JA	2	* libaio engine
	3	*
	4	* IO engine using the Linux native aio interface.
2866c82d JA	5	*
2866c82d JA	6	*/
2866c82d JA	7	#include <stdlib.h>
	8	#include <unistd.h>
	9	#include <errno.h>
67bf9823	10	#include <libaio.h>
46961748 JA	11	#include <sys/time.h>
46961748 JA	12	#include <sys/resource.h>
5f350952 JA	13
5f350952 JA	14	#include "../fio.h"
0f38bbef	15	#include "../lib/pow2.h"
d220c761	16	#include "../optgroup.h"
90a8c9b2	17	#include "../lib/memalign.h"
e9f6567a	18	#include "cmdprio.h"
2866c82d	19
b2a432bf PC	20	/* Should be defined in newest aio_abi.h */
	21	#ifndef IOCB_FLAG_IOPRIO
	22	#define IOCB_FLAG_IOPRIO (1 << 1)
	23	#endif
	24
7d42e66e KK	25	/* Hack for libaio < 0.3.111 */
	26	#ifndef CONFIG_LIBAIO_RW_FLAGS
	27	#define aio_rw_flags __pad2
	28	#endif
	29
0fc2e103	30	static int fio_libaio_commit(struct thread_data *td);
b2a432bf	31	static int fio_libaio_init(struct thread_data *td);
0fc2e103	32
2866c82d JA	33	struct libaio_data {
	34	io_context_t aio_ctx;
	35	struct io_event *aio_events;
755200a3	36	struct iocb **iocbs;
7e77dd02	37	struct io_u **io_us;
acd45e02	38
3c3168e9	39	struct io_u **io_u_index;
1ceaf63b	40	struct iovec iovecs; / for vectored requests */
3c3168e9	41
acd45e02 JA	42	/*
	43	* Basic ring buffer. 'head' is incremented in _queue(), and
	44	* 'tail' is incremented in _commit(). We keep 'queued' so
	45	* that we know if the ring is full or empty, when
	46	* 'head' == 'tail'. 'entries' is the ring size, and
	47	* 'is_pow2' is just an optimization to use AND instead of
	48	* modulus to get the remainder on ring increment.
	49	*/
	50	int is_pow2;
	51	unsigned int entries;
	52	unsigned int queued;
	53	unsigned int head;
	54	unsigned int tail;
d6cbeab4 NC	55
d6cbeab4 NC	56	struct cmdprio cmdprio;
2866c82d JA	57	};
2866c82d JA	58
de890a1e	59	struct libaio_options {
a48f0cc7	60	struct thread_data *td;
de890a1e	61	unsigned int userspace_reap;
d6cbeab4	62	struct cmdprio_options cmdprio_options;
7d42e66e	63	unsigned int nowait;
1ceaf63b	64	unsigned int vectored;
de890a1e SL	65	};
	66
	67	static struct fio_option options[] = {
	68	{
	69	.name = "userspace_reap",
e8b0e958	70	.lname = "Libaio userspace reaping",
de890a1e SL	71	.type = FIO_OPT_STR_SET,
	72	.off1 = offsetof(struct libaio_options, userspace_reap),
	73	.help = "Use alternative user-space reap implementation",
e90a0adf	74	.category = FIO_OPT_C_ENGINE,
5a3cd5f3	75	.group = FIO_OPT_G_LIBAIO,
de890a1e	76	},
7d42e66e KK	77	{
	78	.name = "nowait",
	79	.lname = "RWF_NOWAIT",
	80	.type = FIO_OPT_BOOL,
	81	.off1 = offsetof(struct libaio_options, nowait),
	82	.help = "Set RWF_NOWAIT for reads/writes",
	83	.category = FIO_OPT_C_ENGINE,
	84	.group = FIO_OPT_G_LIBAIO,
	85	},
1ceaf63b RHI	86	{
	87	.name = "libaio_vectored",
	88	.lname = "Use libaio preadv,pwritev",
	89	.type = FIO_OPT_BOOL,
	90	.off1 = offsetof(struct libaio_options, vectored),
	91	.help = "Use libaio {preadv,pwritev} instead of libaio {pread,pwrite}",
	92	.category = FIO_OPT_C_ENGINE,
	93	.group = FIO_OPT_G_LIBAIO,
	94	},
	95
2838f77a	96	CMDPRIO_OPTIONS(struct libaio_options, FIO_OPT_G_LIBAIO),
de890a1e SL	97	{
	98	.name = NULL,
	99	},
	100	};
	101
acd45e02 JA	102	static inline void ring_inc(struct libaio_data ld, unsigned int val,
	103	unsigned int add)
	104	{
	105	if (ld->is_pow2)
	106	val = (val + add) & (ld->entries - 1);
	107	else
	108	val = (val + add) % ld->entries;
	109	}
	110
7d42e66e	111	static int fio_libaio_prep(struct thread_data td, struct io_u io_u)
2866c82d	112	{
7d42e66e	113	struct libaio_options *o = td->eo;
53cdc686	114	struct fio_file *f = io_u->file;
ad4e9298	115	struct iocb *iocb = &io_u->iocb;
1ceaf63b	116	struct libaio_data *ld = td->io_ops_data;
53cdc686	117
a391d73d	118	if (io_u->ddir == DDIR_READ) {
1ceaf63b RHI	119	if (o->vectored) {
	120	struct iovec *iov = &ld->iovecs[io_u->index];
	121
	122	iov->iov_base = io_u->xfer_buf;
	123	iov->iov_len = (size_t)io_u->xfer_buflen;
	124	io_prep_preadv(iocb, f->fd, iov, 1, io_u->offset);
	125	} else {
	126	io_prep_pread(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen,
	127	io_u->offset);
	128	}
7d42e66e KK	129	if (o->nowait)
7d42e66e KK	130	iocb->aio_rw_flags \|= RWF_NOWAIT;
a391d73d	131	} else if (io_u->ddir == DDIR_WRITE) {
1ceaf63b RHI	132	if (o->vectored) {
	133	struct iovec *iov = &ld->iovecs[io_u->index];
	134
	135	iov->iov_base = io_u->xfer_buf;
	136	iov->iov_len = (size_t)io_u->xfer_buflen;
	137	io_prep_pwritev(iocb, f->fd, iov, 1, io_u->offset);
	138	} else {
	139	io_prep_pwrite(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen,
	140	io_u->offset);
	141	}
7d42e66e KK	142	if (o->nowait)
7d42e66e KK	143	iocb->aio_rw_flags \|= RWF_NOWAIT;
a79319fe JG	144	#ifdef FIO_HAVE_RWF_ATOMIC
	145	if (td->o.oatomic)
	146	iocb->aio_rw_flags \|= RWF_ATOMIC;
	147	#endif
a391d73d	148	} else if (ddir_sync(io_u->ddir))
3c3168e9	149	io_prep_fsync(iocb, f->fd);
2866c82d JA	150
	151	return 0;
	152	}
	153
127715b6 NC	154	static inline void fio_libaio_cmdprio_prep(struct thread_data *td,
127715b6 NC	155	struct io_u *io_u)
b2a432bf	156	{
d6cbeab4 NC	157	struct libaio_data *ld = td->io_ops_data;
d6cbeab4 NC	158	struct cmdprio *cmdprio = &ld->cmdprio;
127715b6 NC	159
	160	if (fio_cmdprio_set_ioprio(td, cmdprio, io_u)) {
	161	io_u->iocb.aio_reqprio = io_u->ioprio;
b2a432bf	162	io_u->iocb.u.c.flags \|= IOCB_FLAG_IOPRIO;
b2a432bf	163	}
b2a432bf PC	164	}
b2a432bf PC	165
2866c82d JA	166	static struct io_u fio_libaio_event(struct thread_data td, int event)
2866c82d JA	167	{
565e784d	168	struct libaio_data *ld = td->io_ops_data;
f423479d JA	169	struct io_event *ev;
f423479d JA	170	struct io_u *io_u;
2866c82d	171
f423479d	172	ev = ld->aio_events + event;
702906e9	173	io_u = container_of(ev->obj, struct io_u, iocb);
f423479d JA	174
	175	if (ev->res != io_u->xfer_buflen) {
	176	if (ev->res > io_u->xfer_buflen)
	177	io_u->error = -ev->res;
	178	else
	179	io_u->resid = io_u->xfer_buflen - ev->res;
	180	} else
	181	io_u->error = 0;
	182
	183	return io_u;
2866c82d JA	184	}
2866c82d JA	185
675012f0 DE	186	struct aio_ring {
	187	unsigned id; /** kernel internal index number */
	188	unsigned nr; /** number of io_events */
	189	unsigned head;
	190	unsigned tail;
c44b1ff5	191
675012f0 DE	192	unsigned magic;
	193	unsigned compat_features;
	194	unsigned incompat_features;
	195	unsigned header_length; /** size of aio_ring */
	196
	197	struct io_event events[0];
	198	};
	199
	200	#define AIO_RING_MAGIC 0xa10a10a1
	201
	202	static int user_io_getevents(io_context_t aio_ctx, unsigned int max,
c44b1ff5	203	struct io_event *events)
675012f0 DE	204	{
	205	long i = 0;
	206	unsigned head;
c44b1ff5	207	struct aio_ring ring = (struct aio_ring) aio_ctx;
675012f0 DE	208
	209	while (i < max) {
	210	head = ring->head;
	211
	212	if (head == ring->tail) {
	213	/* There are no more completions */
	214	break;
	215	} else {
	216	/* There is another completion to reap */
	217	events[i] = ring->events[head];
d473a06d BVA	218	atomic_store_release(&ring->head,
d473a06d BVA	219	(head + 1) % ring->nr);
675012f0 DE	220	i++;
	221	}
	222	}
	223
	224	return i;
	225	}
	226
e7d2e616	227	static int fio_libaio_getevents(struct thread_data *td, unsigned int min,
1f440ece	228	unsigned int max, const struct timespec *t)
2866c82d	229	{
565e784d	230	struct libaio_data *ld = td->io_ops_data;
de890a1e	231	struct libaio_options *o = td->eo;
82407585	232	unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
1f440ece	233	struct timespec __lt, *lt = NULL;
0b7fdba7	234	int r, events = 0;
2866c82d	235
1f440ece JA	236	if (t) {
	237	__lt = *t;
	238	lt = &__lt;
	239	}
	240
2866c82d	241	do {
de890a1e	242	if (o->userspace_reap == 1
675012f0 DE	243	&& actual_min == 0
	244	&& ((struct aio_ring *)(ld->aio_ctx))->magic
	245	== AIO_RING_MAGIC) {
a8f6714e	246	r = user_io_getevents(ld->aio_ctx, max - events,
675012f0 DE	247	ld->aio_events + events);
	248	} else {
	249	r = io_getevents(ld->aio_ctx, actual_min,
a8f6714e	250	max - events, ld->aio_events + events, lt);
675012f0	251	}
a8f6714e	252	if (r > 0) {
0b7fdba7	253	events += r;
a51fe82b	254	actual_min -= min((unsigned int)events, actual_min);
a8f6714e	255	}
3441a52d	256	else if ((min && r == 0) \|\| r == -EAGAIN) {
0fc2e103	257	fio_libaio_commit(td);
6347e43d JA	258	if (actual_min)
6347e43d JA	259	usleep(10);
0fc2e103	260	} else if (r != -EINTR)
a31dc2dd	261	break;
0b7fdba7 DE	262	} while (events < min);
	263
	264	return r < 0 ? r : events;
2866c82d JA	265	}
2866c82d JA	266
2e4ef4fb JA	267	static enum fio_q_status fio_libaio_queue(struct thread_data *td,
2e4ef4fb JA	268	struct io_u *io_u)
2866c82d	269	{
565e784d	270	struct libaio_data *ld = td->io_ops_data;
2866c82d	271
7101d9c2 JA	272	fio_ro_check(td, io_u);
7101d9c2 JA	273
acd45e02	274	if (ld->queued == td->o.iodepth)
755200a3 JA	275	return FIO_Q_BUSY;
	276
	277	/*
	278	* fsync is tricky, since it can fail and we need to do it
	279	* serialized with other io. the reason is that linux doesn't
	280	* support aio fsync yet. So return busy for the case where we
	281	* have pending io, to let fio complete those first.
	282	*/
f011531e	283	if (ddir_sync(io_u->ddir)) {
acd45e02	284	if (ld->queued)
755200a3	285	return FIO_Q_BUSY;
5f9099ea	286
f011531e	287	do_io_u_sync(td, io_u);
755200a3 JA	288	return FIO_Q_COMPLETED;
	289	}
	290
a5f3027c	291	if (io_u->ddir == DDIR_TRIM) {
acd45e02	292	if (ld->queued)
a5f3027c JA	293	return FIO_Q_BUSY;
	294
	295	do_io_u_trim(td, io_u);
c0681c9d VF	296	io_u_mark_submit(td, 1);
c0681c9d VF	297	io_u_mark_complete(td, 1);
a5f3027c JA	298	return FIO_Q_COMPLETED;
	299	}
	300
d6cbeab4	301	if (ld->cmdprio.mode != CMDPRIO_MODE_NONE)
ff00f247	302	fio_libaio_cmdprio_prep(td, io_u);
b2a432bf	303
702906e9	304	ld->iocbs[ld->head] = &io_u->iocb;
acd45e02 JA	305	ld->io_us[ld->head] = io_u;
	306	ring_inc(ld, &ld->head, 1);
	307	ld->queued++;
755200a3 JA	308	return FIO_Q_QUEUED;
	309	}
	310
7e77dd02 JA	311	static void fio_libaio_queued(struct thread_data td, struct io_u *io_us,
	312	unsigned int nr)
	313	{
8b6a404c	314	struct timespec now;
7e77dd02 JA	315	unsigned int i;
7e77dd02 JA	316
12d9d841 JA	317	if (!fio_fill_issue_time(td))
	318	return;
	319
7e77dd02 JA	320	fio_gettime(&now, NULL);
	321
	322	for (i = 0; i < nr; i++) {
	323	struct io_u *io_u = io_us[i];
	324
	325	memcpy(&io_u->issue_time, &now, sizeof(now));
	326	io_u_queued(td, io_u);
	327	}
39f56400 VF	328
	329	/*
	330	* only used for iolog
	331	*/
	332	if (td->o.read_iolog_file)
	333	memcpy(&td->last_issue, &now, sizeof(now));
7e77dd02 JA	334	}
7e77dd02 JA	335
755200a3 JA	336	static int fio_libaio_commit(struct thread_data *td)
755200a3 JA	337	{
565e784d	338	struct libaio_data *ld = td->io_ops_data;
755200a3	339	struct iocb **iocbs;
7e77dd02	340	struct io_u **io_us;
8b6a404c	341	struct timespec ts;
a120ca7f	342	int ret, wait_start = 0;
755200a3	343
acd45e02	344	if (!ld->queued)
755200a3 JA	345	return 0;
755200a3 JA	346
2866c82d	347	do {
acd45e02 JA	348	long nr = ld->queued;
	349
	350	nr = min((unsigned int) nr, ld->entries - ld->tail);
	351	io_us = ld->io_us + ld->tail;
	352	iocbs = ld->iocbs + ld->tail;
	353
	354	ret = io_submit(ld->aio_ctx, nr, iocbs);
5e00c2c4	355	if (ret > 0) {
7e77dd02	356	fio_libaio_queued(td, io_us, ret);
838bc709	357	io_u_mark_submit(td, ret);
acd45e02 JA	358
	359	ld->queued -= ret;
	360	ring_inc(ld, &ld->tail, ret);
5e00c2c4	361	ret = 0;
e3b4e568	362	wait_start = 0;
acd45e02	363	} else if (ret == -EINTR \|\| !ret) {
838bc709 JA	364	if (!ret)
838bc709 JA	365	io_u_mark_submit(td, ret);
e3b4e568	366	wait_start = 0;
2866c82d	367	continue;
acd45e02 JA	368	} else if (ret == -EAGAIN) {
	369	/*
	370	* If we get EAGAIN, we should break out without
	371	* error and let the upper layer reap some
a120ca7f JA	372	* events for us. If we have no queued IO, we
	373	* must loop here. If we loop for more than 30s,
	374	* just error out, something must be buggy in the
	375	* IO path.
acd45e02	376	*/
a120ca7f JA	377	if (ld->queued) {
	378	ret = 0;
	379	break;
	380	}
	381	if (!wait_start) {
8b6a404c	382	fio_gettime(&ts, NULL);
d36b072d	383	wait_start = 1;
8b6a404c	384	} else if (mtime_since_now(&ts) > 30000) {
a120ca7f JA	385	log_err("fio: aio appears to be stalled, giving up\n");
	386	break;
	387	}
	388	usleep(1);
	389	continue;
a31dc2dd JA	390	} else if (ret == -ENOMEM) {
	391	/*
	392	* If we get -ENOMEM, reap events if we can. If
	393	* we cannot, treat it as a fatal event since there's
	394	* nothing we can do about it.
	395	*/
	396	if (ld->queued)
	397	ret = 0;
	398	break;
838bc709	399	} else
2866c82d	400	break;
2c3a4ae9	401	} while (ld->queued);
2866c82d	402
36167d82	403	return ret;
2866c82d JA	404	}
	405
	406	static int fio_libaio_cancel(struct thread_data td, struct io_u io_u)
	407	{
565e784d	408	struct libaio_data *ld = td->io_ops_data;
2866c82d JA	409
	410	return io_cancel(ld->aio_ctx, &io_u->iocb, ld->aio_events);
	411	}
	412
	413	static void fio_libaio_cleanup(struct thread_data *td)
	414	{
565e784d	415	struct libaio_data *ld = td->io_ops_data;
2866c82d JA	416
2866c82d JA	417	if (ld) {
f24c2649 JA	418	/*
	419	* Work-around to avoid huge RCU stalls at exit time. If we
	420	* don't do this here, then it'll be torn down by exit_aio().
	421	* But for that case we can parallellize the freeing, thus
	422	* speeding it up a lot.
	423	*/
	424	if (!(td->flags & TD_F_CHILD))
	425	io_destroy(ld->aio_ctx);
d6cbeab4 NC	426
d6cbeab4 NC	427	fio_cmdprio_cleanup(&ld->cmdprio);
1ceaf63b	428	free(ld->iovecs);
7e77dd02 JA	429	free(ld->aio_events);
	430	free(ld->iocbs);
	431	free(ld->io_us);
2866c82d	432	free(ld);
2866c82d JA	433	}
	434	}
	435
2041bd34 JA	436	static int fio_libaio_post_init(struct thread_data *td)
	437	{
	438	struct libaio_data *ld = td->io_ops_data;
ad4e9298	439	int err;
2041bd34	440
ad4e9298	441	err = io_queue_init(td->o.iodepth, &ld->aio_ctx);
2041bd34 JA	442	if (err) {
	443	td_verror(td, -err, "io_queue_init");
	444	return 1;
	445	}
	446
	447	return 0;
ebec344d JA	448	}
ebec344d JA	449
2866c82d JA	450	static int fio_libaio_init(struct thread_data *td)
2866c82d JA	451	{
acd45e02	452	struct libaio_data *ld;
b2a432bf	453	struct libaio_options *o = td->eo;
e9f6567a	454	int ret;
2866c82d	455
acd45e02	456	ld = calloc(1, sizeof(*ld));
c1db2dce	457
acd45e02 JA	458	ld->entries = td->o.iodepth;
	459	ld->is_pow2 = is_power_of_2(ld->entries);
	460	ld->aio_events = calloc(ld->entries, sizeof(struct io_event));
	461	ld->iocbs = calloc(ld->entries, sizeof(struct iocb *));
	462	ld->io_us = calloc(ld->entries, sizeof(struct io_u *));
1ceaf63b	463	ld->iovecs = calloc(ld->entries, sizeof(ld->iovecs[0]));
755200a3	464
565e784d	465	td->io_ops_data = ld;
e9f6567a	466
d6cbeab4	467	ret = fio_cmdprio_init(td, &ld->cmdprio, &o->cmdprio_options);
e9f6567a	468	if (ret) {
b2a432bf PC	469	td_verror(td, EINVAL, "fio_libaio_init");
	470	return 1;
	471	}
e9f6567a	472
2866c82d JA	473	return 0;
	474	}
	475
5a8a6a03	476	FIO_STATIC struct ioengine_ops ioengine = {
de890a1e SL	477	.name = "libaio",
de890a1e SL	478	.version = FIO_IOOPS_VERSION,
4e7e7898	479	.flags = FIO_ASYNCIO_SYNC_TRIM \|
a79319fe JG	480	FIO_ASYNCIO_SETS_ISSUE_TIME \|
a79319fe JG	481	FIO_ATOMICWRITES,
de890a1e	482	.init = fio_libaio_init,
2041bd34	483	.post_init = fio_libaio_post_init,
de890a1e SL	484	.prep = fio_libaio_prep,
	485	.queue = fio_libaio_queue,
	486	.commit = fio_libaio_commit,
	487	.cancel = fio_libaio_cancel,
	488	.getevents = fio_libaio_getevents,
	489	.event = fio_libaio_event,
	490	.cleanup = fio_libaio_cleanup,
	491	.open_file = generic_open_file,
	492	.close_file = generic_close_file,
	493	.get_file_size = generic_get_file_size,
	494	.options = options,
	495	.option_struct_size = sizeof(struct libaio_options),
2866c82d	496	};
34cfcdaf	497
5f350952 JA	498	static void fio_init fio_libaio_register(void)
	499	{
	500	register_ioengine(&ioengine);
	501	}
	502
	503	static void fio_exit fio_libaio_unregister(void)
	504	{
	505	unregister_ioengine(&ioengine);
	506	}