[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"

#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;

	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)
			free(ld->user_iocbs);
		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;

	return fio_libaio_old_queue_init(ld, depth, hipri, useriocb);
#else
	return fio_libaio_old_queue_init(ld, depth, hipri, useriocb);
#endif
}

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;
		void *p;

		ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *));
		size = td->o.iodepth * sizeof(struct iocb);
		if (posix_memalign(&p, page_size, size)) {
			log_err("fio: libaio iocb allocation failure\n");
			free(ld);
			return 1;
		}
		ld->user_iocbs = p;
	}

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