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

#ifndef IOCB_FLAG_HIPRI
#define IOCB_FLAG_HIPRI	(1 << 2)
#endif

#ifndef IOCTX_FLAG_USERIOCB
#define IOCTX_FLAG_USERIOCB	(1 << 0)
#endif
#ifndef IOCTX_FLAG_IOPOLL
#define IOCTX_FLAG_IOPOLL	(1 << 1)
#endif

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 iocb *user_iocbs;
	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;
	unsigned int useriocb;
};

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	= "hipri",
		.lname	= "High Priority",
		.type	= FIO_OPT_STR_SET,
		.off1	= offsetof(struct libaio_options, hipri),
		.help	= "Use polled IO completions",
		.category = FIO_OPT_C_ENGINE,
		.group	= FIO_OPT_G_LIBAIO,
	},
	{
		.name	= "useriocb",
		.lname	= "User IOCBs",
		.type	= FIO_OPT_STR_SET,
		.off1	= offsetof(struct libaio_options, useriocb),
		.help	= "Use user mapped IOCBs",
		.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 libaio_data *ld = td->io_ops_data;
	struct fio_file *f = io_u->file;
	struct libaio_options *o = td->eo;
	struct iocb *iocb;

	if (o->useriocb)
		iocb = &ld->user_iocbs[io_u->index];
	else
		iocb = &io_u->iocb;

	iocb->u.c.flags = 0;

	if (io_u->ddir == DDIR_READ) {
		io_prep_pread(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
		if (o->hipri)
			iocb->u.c.flags |= IOCB_FLAG_HIPRI;
	} else if (io_u->ddir == DDIR_WRITE) {
		io_prep_pwrite(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
		if (o->hipri)
			iocb->u.c.flags |= IOCB_FLAG_HIPRI;
	} 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 libaio_options *o = td->eo;
	struct io_event *ev;
	struct io_u *io_u;

	ev = ld->aio_events + event;
	if (o->useriocb) {
		int index = (int) (uintptr_t) ev->obj;
		io_u = ld->io_u_index[index];
	} else
		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;
	struct libaio_options *o = td->eo;

	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 (o->useriocb)
		ld->iocbs[ld->head] = (struct iocb *) (uintptr_t) io_u->index;
	else
		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);
		if (ld->user_iocbs) {
			size_t size = td->o.iodepth * sizeof(struct iocb);
			fio_memfree(ld->user_iocbs, size, false);
		}
		free(ld);
	}
}

static int fio_libaio_old_queue_init(struct libaio_data *ld, unsigned int depth,
				     bool hipri, bool useriocb)
{
	if (hipri) {
		log_err("fio: polled aio not available on your platform\n");
		return 1;
	}
	if (useriocb) {
		log_err("fio: user mapped iocbs not available on your platform\n");
		return 1;
	}

	return io_queue_init(depth, &ld->aio_ctx);
}

static int fio_libaio_queue_init(struct libaio_data *ld, unsigned int depth,
				 bool hipri, bool useriocb)
{
#ifdef __NR_sys_io_setup2
	int ret, flags = 0;

	if (hipri)
		flags |= IOCTX_FLAG_IOPOLL;
	if (useriocb)
		flags |= IOCTX_FLAG_USERIOCB;

	ret = syscall(__NR_sys_io_setup2, depth, flags, ld->user_iocbs,
			&ld->aio_ctx);
	if (!ret)
		return 0;
	/* fall through to old syscall */
#endif
	return fio_libaio_old_queue_init(ld, depth, hipri, useriocb);
}

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));

	if (o->useriocb) {
		size_t size;

		ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *));
		size = td->o.iodepth * sizeof(struct iocb);
		ld->user_iocbs = fio_memalign(page_size, size, false);
		memset(ld->user_iocbs, 0, size);
	}

	/*
	 * 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.
	 */
	err = fio_libaio_queue_init(ld, td->o.iodepth, o->hipri, o->useriocb);
	if (err) {
		td_verror(td, -err, "io_queue_init");
		log_err("fio: check /proc/sys/fs/aio-max-nr\n");
		if (ld->user_iocbs) {
			size_t size = td->o.iodepth * sizeof(struct iocb);
			fio_memfree(ld->user_iocbs, size, false);
		}
		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 int fio_libaio_io_u_init(struct thread_data *td, struct io_u *io_u)
{
	struct libaio_options *o = td->eo;

	if (o->useriocb) {
		struct libaio_data *ld = td->io_ops_data;

		ld->io_u_index[io_u->index] = io_u;
	}

	return 0;
}

static struct ioengine_ops ioengine = {
	.name			= "libaio",
	.version		= FIO_IOOPS_VERSION,
	.init			= fio_libaio_init,
	.io_u_init		= fio_libaio_io_u_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>
5f350952 JA	11
5f350952 JA	12	#include "../fio.h"
0f38bbef	13	#include "../lib/pow2.h"
d220c761	14	#include "../optgroup.h"
90a8c9b2	15	#include "../lib/memalign.h"
2866c82d	16
ee636f3f JA	17	#ifndef IOCB_FLAG_HIPRI
	18	#define IOCB_FLAG_HIPRI (1 << 2)
	19	#endif
2da0f5d5	20
3c3168e9	21	#ifndef IOCTX_FLAG_USERIOCB
2da0f5d5 JA	22	#define IOCTX_FLAG_USERIOCB (1 << 0)
	23	#endif
	24	#ifndef IOCTX_FLAG_IOPOLL
	25	#define IOCTX_FLAG_IOPOLL (1 << 1)
3c3168e9	26	#endif
a391d73d	27
0fc2e103 JA	28	static int fio_libaio_commit(struct thread_data *td);
0fc2e103 JA	29
2866c82d JA	30	struct libaio_data {
	31	io_context_t aio_ctx;
	32	struct io_event *aio_events;
755200a3	33	struct iocb **iocbs;
7e77dd02	34	struct io_u **io_us;
acd45e02	35
3c3168e9 JA	36	struct iocb *user_iocbs;
	37	struct io_u **io_u_index;
	38
acd45e02 JA	39	/*
	40	* Basic ring buffer. 'head' is incremented in _queue(), and
	41	* 'tail' is incremented in _commit(). We keep 'queued' so
	42	* that we know if the ring is full or empty, when
	43	* 'head' == 'tail'. 'entries' is the ring size, and
	44	* 'is_pow2' is just an optimization to use AND instead of
	45	* modulus to get the remainder on ring increment.
	46	*/
	47	int is_pow2;
	48	unsigned int entries;
	49	unsigned int queued;
	50	unsigned int head;
	51	unsigned int tail;
2866c82d JA	52	};
2866c82d JA	53
de890a1e	54	struct libaio_options {
a1f871c7	55	void *pad;
de890a1e	56	unsigned int userspace_reap;
a391d73d	57	unsigned int hipri;
3c3168e9	58	unsigned int useriocb;
de890a1e SL	59	};
	60
	61	static struct fio_option options[] = {
	62	{
	63	.name = "userspace_reap",
e8b0e958	64	.lname = "Libaio userspace reaping",
de890a1e SL	65	.type = FIO_OPT_STR_SET,
	66	.off1 = offsetof(struct libaio_options, userspace_reap),
	67	.help = "Use alternative user-space reap implementation",
e90a0adf	68	.category = FIO_OPT_C_ENGINE,
5a3cd5f3	69	.group = FIO_OPT_G_LIBAIO,
de890a1e	70	},
a391d73d JA	71	{
a391d73d JA	72	.name = "hipri",
ee636f3f	73	.lname = "High Priority",
a391d73d JA	74	.type = FIO_OPT_STR_SET,
a391d73d JA	75	.off1 = offsetof(struct libaio_options, hipri),
6acc46d8	76	.help = "Use polled IO completions",
a391d73d JA	77	.category = FIO_OPT_C_ENGINE,
	78	.group = FIO_OPT_G_LIBAIO,
	79	},
3c3168e9 JA	80	{
	81	.name = "useriocb",
	82	.lname = "User IOCBs",
	83	.type = FIO_OPT_STR_SET,
	84	.off1 = offsetof(struct libaio_options, useriocb),
	85	.help = "Use user mapped IOCBs",
	86	.category = FIO_OPT_C_ENGINE,
	87	.group = FIO_OPT_G_LIBAIO,
	88	},
de890a1e SL	89	{
	90	.name = NULL,
	91	},
	92	};
	93
acd45e02 JA	94	static inline void ring_inc(struct libaio_data ld, unsigned int val,
	95	unsigned int add)
	96	{
	97	if (ld->is_pow2)
	98	val = (val + add) & (ld->entries - 1);
	99	else
	100	val = (val + add) % ld->entries;
	101	}
	102
7a16dd02	103	static int fio_libaio_prep(struct thread_data fio_unused td, struct io_u io_u)
2866c82d	104	{
3c3168e9	105	struct libaio_data *ld = td->io_ops_data;
53cdc686	106	struct fio_file *f = io_u->file;
a391d73d	107	struct libaio_options *o = td->eo;
3c3168e9 JA	108	struct iocb *iocb;
	109
	110	if (o->useriocb)
	111	iocb = &ld->user_iocbs[io_u->index];
	112	else
	113	iocb = &io_u->iocb;
53cdc686	114
90a8c9b2 JA	115	iocb->u.c.flags = 0;
90a8c9b2 JA	116
a391d73d	117	if (io_u->ddir == DDIR_READ) {
3c3168e9	118	io_prep_pread(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
a391d73d	119	if (o->hipri)
3c3168e9	120	iocb->u.c.flags \|= IOCB_FLAG_HIPRI;
a391d73d	121	} else if (io_u->ddir == DDIR_WRITE) {
3c3168e9	122	io_prep_pwrite(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
a391d73d	123	if (o->hipri)
3c3168e9	124	iocb->u.c.flags \|= IOCB_FLAG_HIPRI;
a391d73d	125	} else if (ddir_sync(io_u->ddir))
3c3168e9	126	io_prep_fsync(iocb, f->fd);
2866c82d JA	127
	128	return 0;
	129	}
	130
	131	static struct io_u fio_libaio_event(struct thread_data td, int event)
	132	{
565e784d	133	struct libaio_data *ld = td->io_ops_data;
3c3168e9	134	struct libaio_options *o = td->eo;
f423479d JA	135	struct io_event *ev;
f423479d JA	136	struct io_u *io_u;
2866c82d	137
f423479d	138	ev = ld->aio_events + event;
3c3168e9 JA	139	if (o->useriocb) {
	140	int index = (int) (uintptr_t) ev->obj;
	141	io_u = ld->io_u_index[index];
	142	} else
	143	io_u = container_of(ev->obj, struct io_u, iocb);
f423479d JA	144
	145	if (ev->res != io_u->xfer_buflen) {
	146	if (ev->res > io_u->xfer_buflen)
	147	io_u->error = -ev->res;
	148	else
	149	io_u->resid = io_u->xfer_buflen - ev->res;
	150	} else
	151	io_u->error = 0;
	152
	153	return io_u;
2866c82d JA	154	}
2866c82d JA	155
675012f0 DE	156	struct aio_ring {
	157	unsigned id; /** kernel internal index number */
	158	unsigned nr; /** number of io_events */
	159	unsigned head;
	160	unsigned tail;
c44b1ff5	161
675012f0 DE	162	unsigned magic;
	163	unsigned compat_features;
	164	unsigned incompat_features;
	165	unsigned header_length; /** size of aio_ring */
	166
	167	struct io_event events[0];
	168	};
	169
	170	#define AIO_RING_MAGIC 0xa10a10a1
	171
	172	static int user_io_getevents(io_context_t aio_ctx, unsigned int max,
c44b1ff5	173	struct io_event *events)
675012f0 DE	174	{
	175	long i = 0;
	176	unsigned head;
c44b1ff5	177	struct aio_ring ring = (struct aio_ring) aio_ctx;
675012f0 DE	178
	179	while (i < max) {
	180	head = ring->head;
	181
	182	if (head == ring->tail) {
	183	/* There are no more completions */
	184	break;
	185	} else {
	186	/* There is another completion to reap */
	187	events[i] = ring->events[head];
	188	read_barrier();
c44b1ff5	189	ring->head = (head + 1) % ring->nr;
675012f0 DE	190	i++;
	191	}
	192	}
	193
	194	return i;
	195	}
	196
e7d2e616	197	static int fio_libaio_getevents(struct thread_data *td, unsigned int min,
1f440ece	198	unsigned int max, const struct timespec *t)
2866c82d	199	{
565e784d	200	struct libaio_data *ld = td->io_ops_data;
de890a1e	201	struct libaio_options *o = td->eo;
82407585	202	unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
1f440ece	203	struct timespec __lt, *lt = NULL;
0b7fdba7	204	int r, events = 0;
2866c82d	205
1f440ece JA	206	if (t) {
	207	__lt = *t;
	208	lt = &__lt;
	209	}
	210
2866c82d	211	do {
de890a1e	212	if (o->userspace_reap == 1
675012f0 DE	213	&& actual_min == 0
	214	&& ((struct aio_ring *)(ld->aio_ctx))->magic
	215	== AIO_RING_MAGIC) {
	216	r = user_io_getevents(ld->aio_ctx, max,
	217	ld->aio_events + events);
	218	} else {
	219	r = io_getevents(ld->aio_ctx, actual_min,
1f440ece	220	max, ld->aio_events + events, lt);
675012f0	221	}
3441a52d	222	if (r > 0)
0b7fdba7	223	events += r;
3441a52d	224	else if ((min && r == 0) \|\| r == -EAGAIN) {
0fc2e103	225	fio_libaio_commit(td);
6347e43d JA	226	if (actual_min)
6347e43d JA	227	usleep(10);
0fc2e103	228	} else if (r != -EINTR)
a31dc2dd	229	break;
0b7fdba7 DE	230	} while (events < min);
	231
	232	return r < 0 ? r : events;
2866c82d JA	233	}
2866c82d JA	234
2e4ef4fb JA	235	static enum fio_q_status fio_libaio_queue(struct thread_data *td,
2e4ef4fb JA	236	struct io_u *io_u)
2866c82d	237	{
565e784d	238	struct libaio_data *ld = td->io_ops_data;
3c3168e9	239	struct libaio_options *o = td->eo;
2866c82d	240
7101d9c2 JA	241	fio_ro_check(td, io_u);
7101d9c2 JA	242
acd45e02	243	if (ld->queued == td->o.iodepth)
755200a3 JA	244	return FIO_Q_BUSY;
	245
	246	/*
	247	* fsync is tricky, since it can fail and we need to do it
	248	* serialized with other io. the reason is that linux doesn't
	249	* support aio fsync yet. So return busy for the case where we
	250	* have pending io, to let fio complete those first.
	251	*/
f011531e	252	if (ddir_sync(io_u->ddir)) {
acd45e02	253	if (ld->queued)
755200a3	254	return FIO_Q_BUSY;
5f9099ea	255
f011531e	256	do_io_u_sync(td, io_u);
755200a3 JA	257	return FIO_Q_COMPLETED;
	258	}
	259
a5f3027c	260	if (io_u->ddir == DDIR_TRIM) {
acd45e02	261	if (ld->queued)
a5f3027c JA	262	return FIO_Q_BUSY;
	263
	264	do_io_u_trim(td, io_u);
c0681c9d VF	265	io_u_mark_submit(td, 1);
c0681c9d VF	266	io_u_mark_complete(td, 1);
a5f3027c JA	267	return FIO_Q_COMPLETED;
	268	}
	269
3c3168e9 JA	270	if (o->useriocb)
	271	ld->iocbs[ld->head] = (struct iocb *) (uintptr_t) io_u->index;
	272	else
	273	ld->iocbs[ld->head] = &io_u->iocb;
	274
acd45e02 JA	275	ld->io_us[ld->head] = io_u;
	276	ring_inc(ld, &ld->head, 1);
	277	ld->queued++;
755200a3 JA	278	return FIO_Q_QUEUED;
	279	}
	280
7e77dd02 JA	281	static void fio_libaio_queued(struct thread_data td, struct io_u *io_us,
	282	unsigned int nr)
	283	{
8b6a404c	284	struct timespec now;
7e77dd02 JA	285	unsigned int i;
7e77dd02 JA	286
12d9d841 JA	287	if (!fio_fill_issue_time(td))
	288	return;
	289
7e77dd02 JA	290	fio_gettime(&now, NULL);
	291
	292	for (i = 0; i < nr; i++) {
	293	struct io_u *io_u = io_us[i];
	294
	295	memcpy(&io_u->issue_time, &now, sizeof(now));
	296	io_u_queued(td, io_u);
	297	}
	298	}
	299
755200a3 JA	300	static int fio_libaio_commit(struct thread_data *td)
755200a3 JA	301	{
565e784d	302	struct libaio_data *ld = td->io_ops_data;
755200a3	303	struct iocb **iocbs;
7e77dd02	304	struct io_u **io_us;
8b6a404c	305	struct timespec ts;
a120ca7f	306	int ret, wait_start = 0;
755200a3	307
acd45e02	308	if (!ld->queued)
755200a3 JA	309	return 0;
755200a3 JA	310
2866c82d	311	do {
acd45e02 JA	312	long nr = ld->queued;
	313
	314	nr = min((unsigned int) nr, ld->entries - ld->tail);
	315	io_us = ld->io_us + ld->tail;
	316	iocbs = ld->iocbs + ld->tail;
	317
	318	ret = io_submit(ld->aio_ctx, nr, iocbs);
5e00c2c4	319	if (ret > 0) {
7e77dd02	320	fio_libaio_queued(td, io_us, ret);
838bc709	321	io_u_mark_submit(td, ret);
acd45e02 JA	322
	323	ld->queued -= ret;
	324	ring_inc(ld, &ld->tail, ret);
5e00c2c4	325	ret = 0;
e3b4e568	326	wait_start = 0;
acd45e02	327	} else if (ret == -EINTR \|\| !ret) {
838bc709 JA	328	if (!ret)
838bc709 JA	329	io_u_mark_submit(td, ret);
e3b4e568	330	wait_start = 0;
2866c82d	331	continue;
acd45e02 JA	332	} else if (ret == -EAGAIN) {
	333	/*
	334	* If we get EAGAIN, we should break out without
	335	* error and let the upper layer reap some
a120ca7f JA	336	* events for us. If we have no queued IO, we
	337	* must loop here. If we loop for more than 30s,
	338	* just error out, something must be buggy in the
	339	* IO path.
acd45e02	340	*/
a120ca7f JA	341	if (ld->queued) {
	342	ret = 0;
	343	break;
	344	}
	345	if (!wait_start) {
8b6a404c	346	fio_gettime(&ts, NULL);
d36b072d	347	wait_start = 1;
8b6a404c	348	} else if (mtime_since_now(&ts) > 30000) {
a120ca7f JA	349	log_err("fio: aio appears to be stalled, giving up\n");
	350	break;
	351	}
	352	usleep(1);
	353	continue;
a31dc2dd JA	354	} else if (ret == -ENOMEM) {
	355	/*
	356	* If we get -ENOMEM, reap events if we can. If
	357	* we cannot, treat it as a fatal event since there's
	358	* nothing we can do about it.
	359	*/
	360	if (ld->queued)
	361	ret = 0;
	362	break;
838bc709	363	} else
2866c82d	364	break;
2c3a4ae9	365	} while (ld->queued);
2866c82d	366
36167d82	367	return ret;
2866c82d JA	368	}
	369
	370	static int fio_libaio_cancel(struct thread_data td, struct io_u io_u)
	371	{
565e784d	372	struct libaio_data *ld = td->io_ops_data;
2866c82d JA	373
	374	return io_cancel(ld->aio_ctx, &io_u->iocb, ld->aio_events);
	375	}
	376
	377	static void fio_libaio_cleanup(struct thread_data *td)
	378	{
565e784d	379	struct libaio_data *ld = td->io_ops_data;
2866c82d JA	380
2866c82d JA	381	if (ld) {
f24c2649 JA	382	/*
	383	* Work-around to avoid huge RCU stalls at exit time. If we
	384	* don't do this here, then it'll be torn down by exit_aio().
	385	* But for that case we can parallellize the freeing, thus
	386	* speeding it up a lot.
	387	*/
	388	if (!(td->flags & TD_F_CHILD))
	389	io_destroy(ld->aio_ctx);
7e77dd02 JA	390	free(ld->aio_events);
	391	free(ld->iocbs);
	392	free(ld->io_us);
90a8c9b2 JA	393	if (ld->user_iocbs) {
	394	size_t size = td->o.iodepth * sizeof(struct iocb);
	395	fio_memfree(ld->user_iocbs, size, false);
	396	}
2866c82d	397	free(ld);
2866c82d JA	398	}
	399	}
	400
e65ef959 JA	401	static int fio_libaio_old_queue_init(struct libaio_data *ld, unsigned int depth,
e65ef959 JA	402	bool hipri, bool useriocb)
ebec344d	403	{
ebec344d JA	404	if (hipri) {
	405	log_err("fio: polled aio not available on your platform\n");
	406	return 1;
	407	}
3c3168e9 JA	408	if (useriocb) {
	409	log_err("fio: user mapped iocbs not available on your platform\n");
	410	return 1;
	411	}
	412
ebec344d	413	return io_queue_init(depth, &ld->aio_ctx);
e65ef959 JA	414	}
	415
	416	static int fio_libaio_queue_init(struct libaio_data *ld, unsigned int depth,
	417	bool hipri, bool useriocb)
	418	{
	419	#ifdef __NR_sys_io_setup2
	420	int ret, flags = 0;
	421
	422	if (hipri)
	423	flags \|= IOCTX_FLAG_IOPOLL;
	424	if (useriocb)
	425	flags \|= IOCTX_FLAG_USERIOCB;
	426
	427	ret = syscall(__NR_sys_io_setup2, depth, flags, ld->user_iocbs,
	428	&ld->aio_ctx);
	429	if (!ret)
	430	return 0;
0fcbc009	431	/* fall through to old syscall */
ebec344d	432	#endif
0fcbc009	433	return fio_libaio_old_queue_init(ld, depth, hipri, useriocb);
ebec344d JA	434	}
ebec344d JA	435
2866c82d JA	436	static int fio_libaio_init(struct thread_data *td)
2866c82d JA	437	{
c90e1017	438	struct libaio_options *o = td->eo;
acd45e02	439	struct libaio_data *ld;
c90e1017	440	int err = 0;
2866c82d	441
acd45e02	442	ld = calloc(1, sizeof(*ld));
c1db2dce	443
3c3168e9 JA	444	if (o->useriocb) {
3c3168e9 JA	445	size_t size;
3c3168e9 JA	446
	447	ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *));
	448	size = td->o.iodepth * sizeof(struct iocb);
90a8c9b2 JA	449	ld->user_iocbs = fio_memalign(page_size, size, false);
90a8c9b2 JA	450	memset(ld->user_iocbs, 0, size);
3c3168e9 JA	451	}
3c3168e9 JA	452
c90e1017 JA	453	/*
	454	* First try passing in 0 for queue depth, since we don't
	455	* care about the user ring. If that fails, the kernel is too old
	456	* and we need the right depth.
	457	*/
3c3168e9	458	err = fio_libaio_queue_init(ld, td->o.iodepth, o->hipri, o->useriocb);
c1db2dce JA	459	if (err) {
c1db2dce JA	460	td_verror(td, -err, "io_queue_init");
75de55ac	461	log_err("fio: check /proc/sys/fs/aio-max-nr\n");
90a8c9b2 JA	462	if (ld->user_iocbs) {
	463	size_t size = td->o.iodepth * sizeof(struct iocb);
	464	fio_memfree(ld->user_iocbs, size, false);
	465	}
cb781c75	466	free(ld);
2866c82d JA	467	return 1;
	468	}
	469
acd45e02 JA	470	ld->entries = td->o.iodepth;
	471	ld->is_pow2 = is_power_of_2(ld->entries);
	472	ld->aio_events = calloc(ld->entries, sizeof(struct io_event));
	473	ld->iocbs = calloc(ld->entries, sizeof(struct iocb *));
	474	ld->io_us = calloc(ld->entries, sizeof(struct io_u *));
755200a3	475
565e784d	476	td->io_ops_data = ld;
2866c82d JA	477	return 0;
	478	}
	479
3c3168e9 JA	480	static int fio_libaio_io_u_init(struct thread_data td, struct io_u io_u)
	481	{
	482	struct libaio_options *o = td->eo;
	483
	484	if (o->useriocb) {
	485	struct libaio_data *ld = td->io_ops_data;
	486
	487	ld->io_u_index[io_u->index] = io_u;
	488	}
	489
	490	return 0;
	491	}
	492
5f350952	493	static struct ioengine_ops ioengine = {
de890a1e SL	494	.name = "libaio",
	495	.version = FIO_IOOPS_VERSION,
	496	.init = fio_libaio_init,
3c3168e9	497	.io_u_init = fio_libaio_io_u_init,
de890a1e SL	498	.prep = fio_libaio_prep,
	499	.queue = fio_libaio_queue,
	500	.commit = fio_libaio_commit,
	501	.cancel = fio_libaio_cancel,
	502	.getevents = fio_libaio_getevents,
	503	.event = fio_libaio_event,
	504	.cleanup = fio_libaio_cleanup,
	505	.open_file = generic_open_file,
	506	.close_file = generic_close_file,
	507	.get_file_size = generic_get_file_size,
	508	.options = options,
	509	.option_struct_size = sizeof(struct libaio_options),
2866c82d	510	};
34cfcdaf	511
5f350952 JA	512	static void fio_init fio_libaio_register(void)
	513	{
	514	register_ioengine(&ioengine);
	515	}
	516
	517	static void fio_exit fio_libaio_unregister(void)
	518	{
	519	unregister_ioengine(&ioengine);
	520	}