[fio.git] / engines / aioring.c

/*
 * aioring engine
 *
 * IO engine using the new native Linux libaio ring interface. See:
 *
 * http://git.kernel.dk/cgit/linux-block/log/?h=aio-poll
 *
 */
#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"

#ifdef ARCH_HAVE_AIORING

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

/*
 * io_setup2(2) flags
 */
#ifndef IOCTX_FLAG_IOPOLL
#define IOCTX_FLAG_IOPOLL	(1 << 0)
#endif
#ifndef IOCTX_FLAG_SCQRING
#define IOCTX_FLAG_SCQRING	(1 << 1)
#endif
#ifndef IOCTX_FLAG_FIXEDBUFS
#define IOCTX_FLAG_FIXEDBUFS	(1 << 2)
#endif
#ifndef IOCTX_FLAG_SQTHREAD
#define IOCTX_FLAG_SQTHREAD	(1 << 3)
#endif
#ifndef IOCTX_FLAG_SQWQ
#define IOCTX_FLAG_SQWQ		(1 << 4)
#endif
#ifndef IOCTX_FLAG_SQPOLL
#define IOCTX_FLAG_SQPOLL	(1 << 5)
#endif


/*
 * io_ring_enter(2) flags
 */
#ifndef IORING_FLAG_SUBMIT
#define IORING_FLAG_SUBMIT	(1 << 0)
#endif
#ifndef IORING_FLAG_GETEVENTS
#define IORING_FLAG_GETEVENTS	(1 << 1)
#endif

typedef uint64_t u64;
typedef uint32_t u32;
typedef uint16_t u16;

#define IORING_SQ_NEED_WAKEUP	(1 << 0)

struct aio_sq_ring {
	union {
		struct {
			u32 head;
			u32 tail;
			u32 nr_events;
			u16 sq_thread_cpu;
			u16 kflags;
			u64 iocbs;
		};
		u32 pad[16];
	};
	u32 array[0];
};

struct aio_cq_ring {
	union {
		struct {
			u32 head;
			u32 tail;
			u32 nr_events;
		};
		struct io_event pad;
	};
	struct io_event events[0];
};

struct aioring_data {
	io_context_t aio_ctx;
	struct io_u **io_us;
	struct io_u **io_u_index;

	struct aio_sq_ring *sq_ring;
	struct iocb *iocbs;

	struct aio_cq_ring *cq_ring;
	struct io_event *events;

	int queued;
	int cq_ring_off;
};

struct aioring_options {
	void *pad;
	unsigned int hipri;
	unsigned int fixedbufs;
	unsigned int sqthread;
	unsigned int sqthread_set;
	unsigned int sqthread_poll;
	unsigned int sqwq;
};

static int fio_aioring_sqthread_cb(void *data,
				   unsigned long long *val)
{
	struct aioring_options *o = data;

	o->sqthread = *val;
	o->sqthread_set = 1;
	return 0;
}

static struct fio_option options[] = {
	{
		.name	= "hipri",
		.lname	= "High Priority",
		.type	= FIO_OPT_STR_SET,
		.off1	= offsetof(struct aioring_options, hipri),
		.help	= "Use polled IO completions",
		.category = FIO_OPT_C_ENGINE,
		.group	= FIO_OPT_G_LIBAIO,
	},
	{
		.name	= "fixedbufs",
		.lname	= "Fixed (pre-mapped) IO buffers",
		.type	= FIO_OPT_STR_SET,
		.off1	= offsetof(struct aioring_options, fixedbufs),
		.help	= "Pre map IO buffers",
		.category = FIO_OPT_C_ENGINE,
		.group	= FIO_OPT_G_LIBAIO,
	},
	{
		.name	= "sqthread",
		.lname	= "Use kernel SQ thread on this CPU",
		.type	= FIO_OPT_INT,
		.cb	= fio_aioring_sqthread_cb,
		.help	= "Offload submission to kernel thread",
		.category = FIO_OPT_C_ENGINE,
		.group	= FIO_OPT_G_LIBAIO,
	},
	{
		.name	= "sqthread_poll",
		.lname	= "Kernel SQ thread should poll",
		.type	= FIO_OPT_STR_SET,
		.off1	= offsetof(struct aioring_options, sqthread_poll),
		.help	= "Used with sqthread, enables kernel side polling",
		.category = FIO_OPT_C_ENGINE,
		.group	= FIO_OPT_G_LIBAIO,
	},
	{
		.name	= "sqwq",
		.lname	= "Offload submission to kernel workqueue",
		.type	= FIO_OPT_STR_SET,
		.off1	= offsetof(struct aioring_options, sqwq),
		.help	= "Offload submission to kernel workqueue",
		.category = FIO_OPT_C_ENGINE,
		.group	= FIO_OPT_G_LIBAIO,
	},
	{
		.name	= NULL,
	},
};

static int io_ring_enter(io_context_t ctx, unsigned int to_submit,
			 unsigned int min_complete, unsigned int flags)
{
	return syscall(__NR_sys_io_ring_enter, ctx, to_submit, min_complete,
			flags);
}

static int fio_aioring_prep(struct thread_data *td, struct io_u *io_u)
{
	struct aioring_data *ld = td->io_ops_data;
	struct fio_file *f = io_u->file;
	struct aioring_options *o = td->eo;
	struct iocb *iocb;

	iocb = &ld->iocbs[io_u->index];

	if (io_u->ddir == DDIR_READ) {
		if (o->fixedbufs) {
			iocb->aio_fildes = f->fd;
			iocb->aio_lio_opcode = IO_CMD_PREAD;
			iocb->u.c.offset = io_u->offset;
		} else {
			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) {
		if (o->fixedbufs) {
			iocb->aio_fildes = f->fd;
			iocb->aio_lio_opcode = IO_CMD_PWRITE;
			iocb->u.c.offset = io_u->offset;
		} else {
			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);

	iocb->data = io_u;
	return 0;
}

static struct io_u *fio_aioring_event(struct thread_data *td, int event)
{
	struct aioring_data *ld = td->io_ops_data;
	struct io_event *ev;
	struct io_u *io_u;
	int index;

	index = event + ld->cq_ring_off;
	if (index >= ld->cq_ring->nr_events)
		index -= ld->cq_ring->nr_events;

	ev = &ld->cq_ring->events[index];
	io_u = ev->data;

	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;
}

static int fio_aioring_cqring_reap(struct thread_data *td, unsigned int events,
				   unsigned int max)
{
	struct aioring_data *ld = td->io_ops_data;
	struct aio_cq_ring *ring = ld->cq_ring;
	u32 head, reaped = 0;

	head = ring->head;
	do {
		read_barrier();
		if (head == ring->tail)
			break;
		reaped++;
		head++;
		if (head == ring->nr_events)
			head = 0;
	} while (reaped + events < max);

	ring->head = head;
	write_barrier();
	return reaped;
}

static int fio_aioring_getevents(struct thread_data *td, unsigned int min,
				 unsigned int max, const struct timespec *t)
{
	struct aioring_data *ld = td->io_ops_data;
	unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
	struct aioring_options *o = td->eo;
	struct aio_cq_ring *ring = ld->cq_ring;
	int r, events = 0;

	ld->cq_ring_off = ring->head;
	do {
		r = fio_aioring_cqring_reap(td, events, max);
		if (r) {
			events += r;
			continue;
		}

		if (!o->sqthread_poll) {
			r = io_ring_enter(ld->aio_ctx, 0, actual_min,
						IORING_FLAG_GETEVENTS);
			if (r < 0) {
				if (errno == EAGAIN)
					continue;
				td_verror(td, errno, "io_ring_enter get");
				break;
			}
		}
	} while (events < min);

	return r < 0 ? r : events;
}

static enum fio_q_status fio_aioring_queue(struct thread_data *td,
					   struct io_u *io_u)
{
	struct aioring_data *ld = td->io_ops_data;
	struct aio_sq_ring *ring = ld->sq_ring;
	unsigned tail, next_tail;

	fio_ro_check(td, io_u);

	if (ld->queued == td->o.iodepth)
		return FIO_Q_BUSY;

	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;
	}

	tail = ring->tail;
	next_tail = tail + 1;
	if (next_tail == ring->nr_events)
		next_tail = 0;
	read_barrier();
	if (next_tail == ring->head)
		return FIO_Q_BUSY;

	ring->array[tail] = io_u->index;
	ring->tail = next_tail;
	write_barrier();

	ld->queued++;
	return FIO_Q_QUEUED;
}

static void fio_aioring_queued(struct thread_data *td, int start, int nr)
{
	struct aioring_data *ld = td->io_ops_data;
	struct timespec now;

	if (!fio_fill_issue_time(td))
		return;

	fio_gettime(&now, NULL);

	while (nr--) {
		int index = ld->sq_ring->array[start];
		struct io_u *io_u = io_u = ld->io_u_index[index];

		memcpy(&io_u->issue_time, &now, sizeof(now));
		io_u_queued(td, io_u);

		start++;
		if (start == ld->sq_ring->nr_events)
			start = 0;
	}
}

static int fio_aioring_commit(struct thread_data *td)
{
	struct aioring_data *ld = td->io_ops_data;
	struct aioring_options *o = td->eo;
	int ret;

	if (!ld->queued)
		return 0;

	/* Nothing to do */
	if (o->sqthread_poll) {
		struct aio_sq_ring *ring = ld->sq_ring;

		if (ring->kflags & IORING_SQ_NEED_WAKEUP)
			io_ring_enter(ld->aio_ctx, ld->queued, 0, IORING_FLAG_SUBMIT);
		ld->queued = 0;
		return 0;
	}

	do {
		int start = ld->sq_ring->head;
		long nr = ld->queued;

		ret = io_ring_enter(ld->aio_ctx, nr, 0, IORING_FLAG_SUBMIT |
						IORING_FLAG_GETEVENTS);
		if (ret > 0) {
			fio_aioring_queued(td, start, ret);
			io_u_mark_submit(td, ret);

			ld->queued -= ret;
			ret = 0;
		} else if (!ret) {
			io_u_mark_submit(td, ret);
			continue;
		} else {
			if (errno == EAGAIN) {
				ret = fio_aioring_cqring_reap(td, 0, ld->queued);
				if (ret)
					continue;
				/* Shouldn't happen */
				usleep(1);
				continue;
			}
			td_verror(td, errno, "io_ring_enter sumit");
			break;
		}
	} while (ld->queued);

	return ret;
}

static size_t aioring_cq_size(struct thread_data *td)
{
	return sizeof(struct aio_cq_ring) + 2 * td->o.iodepth * sizeof(struct io_event);
}

static size_t aioring_sq_iocb(struct thread_data *td)
{
	return sizeof(struct iocb) * td->o.iodepth;
}

static size_t aioring_sq_size(struct thread_data *td)
{
	return sizeof(struct aio_sq_ring) + td->o.iodepth * sizeof(u32);
}

static void fio_aioring_cleanup(struct thread_data *td)
{
	struct aioring_data *ld = td->io_ops_data;

	if (ld) {
		/* Bump depth to match init depth */
		td->o.iodepth++;

		/*
		 * 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->io_u_index);
		free(ld->io_us);
		fio_memfree(ld->sq_ring, aioring_sq_size(td), false);
		fio_memfree(ld->iocbs, aioring_sq_iocb(td), false);
		fio_memfree(ld->cq_ring, aioring_cq_size(td), false);
		free(ld);
	}
}

static int fio_aioring_queue_init(struct thread_data *td)
{
	struct aioring_data *ld = td->io_ops_data;
	struct aioring_options *o = td->eo;
	int flags = IOCTX_FLAG_SCQRING;
	int depth = td->o.iodepth;

	if (o->hipri)
		flags |= IOCTX_FLAG_IOPOLL;
	if (o->sqthread_set) {
		ld->sq_ring->sq_thread_cpu = o->sqthread;
		flags |= IOCTX_FLAG_SQTHREAD;
		if (o->sqthread_poll)
			flags |= IOCTX_FLAG_SQPOLL;
	} else if (o->sqwq)
		flags |= IOCTX_FLAG_SQWQ;

	if (o->fixedbufs) {
		struct rlimit rlim = {
			.rlim_cur = RLIM_INFINITY,
			.rlim_max = RLIM_INFINITY,
		};

		setrlimit(RLIMIT_MEMLOCK, &rlim);
		flags |= IOCTX_FLAG_FIXEDBUFS;
	}

	return syscall(__NR_sys_io_setup2, depth, flags,
			ld->sq_ring, ld->cq_ring, &ld->aio_ctx);
}

static int fio_aioring_post_init(struct thread_data *td)
{
	struct aioring_data *ld = td->io_ops_data;
	struct aioring_options *o = td->eo;
	struct io_u *io_u;
	struct iocb *iocb;
	int err = 0;

	if (o->fixedbufs) {
		int i;

		for (i = 0; i < td->o.iodepth; i++) {
			io_u = ld->io_u_index[i];
			iocb = &ld->iocbs[i];
			iocb->u.c.buf = io_u->buf;
			iocb->u.c.nbytes = td_max_bs(td);

			if (o->hipri)
				iocb->u.c.flags |= IOCB_FLAG_HIPRI;
		}
	}

	err = fio_aioring_queue_init(td);
	if (err) {
		td_verror(td, -err, "io_queue_init");
		return 1;
	}

	/* Adjust depth back again */
	td->o.iodepth--;
	return 0;
}

static int fio_aioring_init(struct thread_data *td)
{
	struct aioring_options *o = td->eo;
	struct aioring_data *ld;

	if (o->sqthread_set && o->sqwq) {
		log_err("fio: aioring sqthread and sqwq are mutually exclusive\n");
		return 1;
	}

	/* ring needs an extra entry, add one to achieve QD set */
	td->o.iodepth++;

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

	/* io_u index */
	ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *));
	ld->io_us = calloc(td->o.iodepth, sizeof(struct io_u *));

	ld->iocbs = fio_memalign(page_size, aioring_sq_iocb(td), false);
	memset(ld->iocbs, 0, aioring_sq_iocb(td));

	ld->sq_ring = fio_memalign(page_size, aioring_sq_size(td), false);
	memset(ld->sq_ring, 0, aioring_sq_size(td));
	ld->sq_ring->nr_events = td->o.iodepth;
	ld->sq_ring->iocbs = (u64) (uintptr_t) ld->iocbs;

	ld->cq_ring = fio_memalign(page_size, aioring_cq_size(td), false);
	memset(ld->cq_ring, 0, aioring_cq_size(td));
	ld->cq_ring->nr_events = td->o.iodepth * 2;

	td->io_ops_data = ld;
	return 0;
}

static int fio_aioring_io_u_init(struct thread_data *td, struct io_u *io_u)
{
	struct aioring_data *ld = td->io_ops_data;

	ld->io_u_index[io_u->index] = io_u;
	return 0;
}

static struct ioengine_ops ioengine = {
	.name			= "aio-ring",
	.version		= FIO_IOOPS_VERSION,
	.init			= fio_aioring_init,
	.post_init		= fio_aioring_post_init,
	.io_u_init		= fio_aioring_io_u_init,
	.prep			= fio_aioring_prep,
	.queue			= fio_aioring_queue,
	.commit			= fio_aioring_commit,
	.getevents		= fio_aioring_getevents,
	.event			= fio_aioring_event,
	.cleanup		= fio_aioring_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 aioring_options),
};

static void fio_init fio_aioring_register(void)
{
	register_ioengine(&ioengine);
}

static void fio_exit fio_aioring_unregister(void)
{
	unregister_ioengine(&ioengine);
}
#endif
Commit	Line	Data
52885fa2 JA	1	/*
	2	* aioring engine
	3	*
a90cd050 JA	4	* IO engine using the new native Linux libaio ring interface. See:
	5	*
	6	* http://git.kernel.dk/cgit/linux-block/log/?h=aio-poll
52885fa2 JA	7	*
	8	*/
	9	#include <stdlib.h>
	10	#include <unistd.h>
	11	#include <errno.h>
	12	#include <libaio.h>
	13	#include <sys/time.h>
	14	#include <sys/resource.h>
	15
	16	#include "../fio.h"
	17	#include "../lib/pow2.h"
	18	#include "../optgroup.h"
	19	#include "../lib/memalign.h"
	20
a30e63cf JA	21	#ifdef ARCH_HAVE_AIORING
a30e63cf JA	22
52885fa2 JA	23	#ifndef IOCB_FLAG_HIPRI
	24	#define IOCB_FLAG_HIPRI (1 << 2)
	25	#endif
	26
	27	/*
	28	* io_setup2(2) flags
	29	*/
	30	#ifndef IOCTX_FLAG_IOPOLL
	31	#define IOCTX_FLAG_IOPOLL (1 << 0)
	32	#endif
	33	#ifndef IOCTX_FLAG_SCQRING
	34	#define IOCTX_FLAG_SCQRING (1 << 1)
	35	#endif
	36	#ifndef IOCTX_FLAG_FIXEDBUFS
	37	#define IOCTX_FLAG_FIXEDBUFS (1 << 2)
	38	#endif
	39	#ifndef IOCTX_FLAG_SQTHREAD
	40	#define IOCTX_FLAG_SQTHREAD (1 << 3)
	41	#endif
	42	#ifndef IOCTX_FLAG_SQWQ
	43	#define IOCTX_FLAG_SQWQ (1 << 4)
	44	#endif
771c9901 JA	45	#ifndef IOCTX_FLAG_SQPOLL
	46	#define IOCTX_FLAG_SQPOLL (1 << 5)
	47	#endif
	48
52885fa2 JA	49
	50	/*
	51	* io_ring_enter(2) flags
	52	*/
	53	#ifndef IORING_FLAG_SUBMIT
	54	#define IORING_FLAG_SUBMIT (1 << 0)
	55	#endif
	56	#ifndef IORING_FLAG_GETEVENTS
	57	#define IORING_FLAG_GETEVENTS (1 << 1)
	58	#endif
	59
	60	typedef uint64_t u64;
	61	typedef uint32_t u32;
	62	typedef uint16_t u16;
	63
4cdbc048 JA	64	#define IORING_SQ_NEED_WAKEUP (1 << 0)
4cdbc048 JA	65
52885fa2 JA	66	struct aio_sq_ring {
	67	union {
	68	struct {
	69	u32 head;
	70	u32 tail;
	71	u32 nr_events;
	72	u16 sq_thread_cpu;
4cdbc048	73	u16 kflags;
52885fa2 JA	74	u64 iocbs;
	75	};
	76	u32 pad[16];
	77	};
	78	u32 array[0];
	79	};
	80
	81	struct aio_cq_ring {
	82	union {
	83	struct {
	84	u32 head;
	85	u32 tail;
	86	u32 nr_events;
	87	};
	88	struct io_event pad;
	89	};
	90	struct io_event events[0];
	91	};
	92
	93	struct aioring_data {
	94	io_context_t aio_ctx;
	95	struct io_u **io_us;
	96	struct io_u **io_u_index;
	97
	98	struct aio_sq_ring *sq_ring;
	99	struct iocb *iocbs;
	100
	101	struct aio_cq_ring *cq_ring;
	102	struct io_event *events;
	103
	104	int queued;
	105	int cq_ring_off;
	106	};
	107
	108	struct aioring_options {
	109	void *pad;
	110	unsigned int hipri;
	111	unsigned int fixedbufs;
a90cd050 JA	112	unsigned int sqthread;
a90cd050 JA	113	unsigned int sqthread_set;
771c9901	114	unsigned int sqthread_poll;
a90cd050	115	unsigned int sqwq;
52885fa2 JA	116	};
52885fa2 JA	117
a90cd050 JA	118	static int fio_aioring_sqthread_cb(void *data,
	119	unsigned long long *val)
	120	{
	121	struct aioring_options *o = data;
	122
	123	o->sqthread = *val;
	124	o->sqthread_set = 1;
	125	return 0;
	126	}
	127
52885fa2 JA	128	static struct fio_option options[] = {
	129	{
	130	.name = "hipri",
	131	.lname = "High Priority",
	132	.type = FIO_OPT_STR_SET,
	133	.off1 = offsetof(struct aioring_options, hipri),
	134	.help = "Use polled IO completions",
	135	.category = FIO_OPT_C_ENGINE,
	136	.group = FIO_OPT_G_LIBAIO,
	137	},
	138	{
	139	.name = "fixedbufs",
	140	.lname = "Fixed (pre-mapped) IO buffers",
	141	.type = FIO_OPT_STR_SET,
	142	.off1 = offsetof(struct aioring_options, fixedbufs),
	143	.help = "Pre map IO buffers",
	144	.category = FIO_OPT_C_ENGINE,
	145	.group = FIO_OPT_G_LIBAIO,
	146	},
a90cd050 JA	147	{
	148	.name = "sqthread",
	149	.lname = "Use kernel SQ thread on this CPU",
	150	.type = FIO_OPT_INT,
	151	.cb = fio_aioring_sqthread_cb,
	152	.help = "Offload submission to kernel thread",
	153	.category = FIO_OPT_C_ENGINE,
	154	.group = FIO_OPT_G_LIBAIO,
	155	},
771c9901 JA	156	{
	157	.name = "sqthread_poll",
	158	.lname = "Kernel SQ thread should poll",
	159	.type = FIO_OPT_STR_SET,
	160	.off1 = offsetof(struct aioring_options, sqthread_poll),
	161	.help = "Used with sqthread, enables kernel side polling",
	162	.category = FIO_OPT_C_ENGINE,
	163	.group = FIO_OPT_G_LIBAIO,
	164	},
a90cd050 JA	165	{
	166	.name = "sqwq",
	167	.lname = "Offload submission to kernel workqueue",
	168	.type = FIO_OPT_STR_SET,
	169	.off1 = offsetof(struct aioring_options, sqwq),
	170	.help = "Offload submission to kernel workqueue",
	171	.category = FIO_OPT_C_ENGINE,
	172	.group = FIO_OPT_G_LIBAIO,
	173	},
52885fa2 JA	174	{
	175	.name = NULL,
	176	},
	177	};
	178
52885fa2 JA	179	static int io_ring_enter(io_context_t ctx, unsigned int to_submit,
	180	unsigned int min_complete, unsigned int flags)
	181	{
52885fa2 JA	182	return syscall(__NR_sys_io_ring_enter, ctx, to_submit, min_complete,
52885fa2 JA	183	flags);
52885fa2 JA	184	}
	185
	186	static int fio_aioring_prep(struct thread_data td, struct io_u io_u)
	187	{
	188	struct aioring_data *ld = td->io_ops_data;
	189	struct fio_file *f = io_u->file;
	190	struct aioring_options *o = td->eo;
	191	struct iocb *iocb;
	192
	193	iocb = &ld->iocbs[io_u->index];
	194
	195	if (io_u->ddir == DDIR_READ) {
	196	if (o->fixedbufs) {
	197	iocb->aio_fildes = f->fd;
	198	iocb->aio_lio_opcode = IO_CMD_PREAD;
	199	iocb->u.c.offset = io_u->offset;
	200	} else {
	201	io_prep_pread(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
	202	if (o->hipri)
	203	iocb->u.c.flags \|= IOCB_FLAG_HIPRI;
	204	}
	205	} else if (io_u->ddir == DDIR_WRITE) {
	206	if (o->fixedbufs) {
	207	iocb->aio_fildes = f->fd;
	208	iocb->aio_lio_opcode = IO_CMD_PWRITE;
	209	iocb->u.c.offset = io_u->offset;
	210	} else {
	211	io_prep_pwrite(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
	212	if (o->hipri)
	213	iocb->u.c.flags \|= IOCB_FLAG_HIPRI;
	214	}
	215	} else if (ddir_sync(io_u->ddir))
	216	io_prep_fsync(iocb, f->fd);
	217
	218	iocb->data = io_u;
	219	return 0;
	220	}
	221
	222	static struct io_u fio_aioring_event(struct thread_data td, int event)
	223	{
	224	struct aioring_data *ld = td->io_ops_data;
	225	struct io_event *ev;
	226	struct io_u *io_u;
	227	int index;
	228
	229	index = event + ld->cq_ring_off;
	230	if (index >= ld->cq_ring->nr_events)
	231	index -= ld->cq_ring->nr_events;
	232
	233	ev = &ld->cq_ring->events[index];
	234	io_u = ev->data;
	235
	236	if (ev->res != io_u->xfer_buflen) {
	237	if (ev->res > io_u->xfer_buflen)
	238	io_u->error = -ev->res;
	239	else
	240	io_u->resid = io_u->xfer_buflen - ev->res;
	241	} else
	242	io_u->error = 0;
	243
	244	return io_u;
	245	}
	246
	247	static int fio_aioring_cqring_reap(struct thread_data *td, unsigned int events,
248	unsigned int max)
249	{
250	struct aioring_data *ld = td->io_ops_data;
251	struct aio_cq_ring *ring = ld->cq_ring;
252	u32 head, reaped = 0;
253
254	head = ring->head;
255	do {
256	read_barrier();
257	if (head == ring->tail)
258	break;
259	reaped++;
260	head++;
261	if (head == ring->nr_events)
262	head = 0;
263	} while (reaped + events < max);
264
265	ring->head = head;
266	write_barrier();
267	return reaped;
268	}
269
270	static int fio_aioring_getevents(struct thread_data *td, unsigned int min,
271	unsigned int max, const struct timespec *t)
272	{
273	struct aioring_data *ld = td->io_ops_data;
274	unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
771c9901	275	struct aioring_options *o = td->eo;
52885fa2 JA	276	struct aio_cq_ring *ring = ld->cq_ring;
	277	int r, events = 0;
	278
	279	ld->cq_ring_off = ring->head;
	280	do {
	281	r = fio_aioring_cqring_reap(td, events, max);
	282	if (r) {
	283	events += r;
	284	continue;
	285	}
	286
771c9901 JA	287	if (!o->sqthread_poll) {
	288	r = io_ring_enter(ld->aio_ctx, 0, actual_min,
	289	IORING_FLAG_GETEVENTS);
	290	if (r < 0) {
	291	if (errno == EAGAIN)
	292	continue;
	293	td_verror(td, errno, "io_ring_enter get");
	294	break;
	295	}
52885fa2 JA	296	}
	297	} while (events < min);
	298
	299	return r < 0 ? r : events;
	300	}
	301
	302	static enum fio_q_status fio_aioring_queue(struct thread_data *td,
	303	struct io_u *io_u)
	304	{
	305	struct aioring_data *ld = td->io_ops_data;
	306	struct aio_sq_ring *ring = ld->sq_ring;
	307	unsigned tail, next_tail;
	308
	309	fio_ro_check(td, io_u);
	310
	311	if (ld->queued == td->o.iodepth)
	312	return FIO_Q_BUSY;
	313
52885fa2 JA	314	if (io_u->ddir == DDIR_TRIM) {
	315	if (ld->queued)
	316	return FIO_Q_BUSY;
	317
	318	do_io_u_trim(td, io_u);
	319	io_u_mark_submit(td, 1);
	320	io_u_mark_complete(td, 1);
	321	return FIO_Q_COMPLETED;
	322	}
	323
	324	tail = ring->tail;
	325	next_tail = tail + 1;
	326	if (next_tail == ring->nr_events)
	327	next_tail = 0;
	328	read_barrier();
	329	if (next_tail == ring->head)
	330	return FIO_Q_BUSY;
	331
	332	ring->array[tail] = io_u->index;
	333	ring->tail = next_tail;
	334	write_barrier();
	335
	336	ld->queued++;
	337	return FIO_Q_QUEUED;
	338	}
	339
	340	static void fio_aioring_queued(struct thread_data *td, int start, int nr)
	341	{
	342	struct aioring_data *ld = td->io_ops_data;
	343	struct timespec now;
	344
	345	if (!fio_fill_issue_time(td))
	346	return;
	347
	348	fio_gettime(&now, NULL);
	349
	350	while (nr--) {
	351	int index = ld->sq_ring->array[start];
	352	struct io_u *io_u = io_u = ld->io_u_index[index];
	353
	354	memcpy(&io_u->issue_time, &now, sizeof(now));
	355	io_u_queued(td, io_u);
	356
	357	start++;
	358	if (start == ld->sq_ring->nr_events)
	359	start = 0;
	360	}
	361	}
	362
	363	static int fio_aioring_commit(struct thread_data *td)
	364	{
	365	struct aioring_data *ld = td->io_ops_data;
771c9901	366	struct aioring_options *o = td->eo;
52885fa2 JA	367	int ret;
	368
	369	if (!ld->queued)
	370	return 0;
	371
771c9901 JA	372	/* Nothing to do */
771c9901 JA	373	if (o->sqthread_poll) {
4cdbc048 JA	374	struct aio_sq_ring *ring = ld->sq_ring;
	375
	376	if (ring->kflags & IORING_SQ_NEED_WAKEUP)
	377	io_ring_enter(ld->aio_ctx, ld->queued, 0, IORING_FLAG_SUBMIT);
771c9901 JA	378	ld->queued = 0;
	379	return 0;
	380	}
	381
52885fa2 JA	382	do {
	383	int start = ld->sq_ring->head;
	384	long nr = ld->queued;
	385
	386	ret = io_ring_enter(ld->aio_ctx, nr, 0, IORING_FLAG_SUBMIT \|
	387	IORING_FLAG_GETEVENTS);
52885fa2 JA	388	if (ret > 0) {
	389	fio_aioring_queued(td, start, ret);
	390	io_u_mark_submit(td, ret);
	391
	392	ld->queued -= ret;
	393	ret = 0;
a90cd050 JA	394	} else if (!ret) {
a90cd050 JA	395	io_u_mark_submit(td, ret);
52885fa2	396	continue;
a90cd050 JA	397	} else {
	398	if (errno == EAGAIN) {
	399	ret = fio_aioring_cqring_reap(td, 0, ld->queued);
	400	if (ret)
	401	continue;
	402	/* Shouldn't happen */
	403	usleep(1);
	404	continue;
52885fa2	405	}
a90cd050	406	td_verror(td, errno, "io_ring_enter sumit");
52885fa2	407	break;
a90cd050	408	}
52885fa2 JA	409	} while (ld->queued);
	410
	411	return ret;
	412	}
	413
	414	static size_t aioring_cq_size(struct thread_data *td)
	415	{
	416	return sizeof(struct aio_cq_ring) + 2 * td->o.iodepth * sizeof(struct io_event);
	417	}
	418
	419	static size_t aioring_sq_iocb(struct thread_data *td)
	420	{
	421	return sizeof(struct iocb) * td->o.iodepth;
	422	}
	423
	424	static size_t aioring_sq_size(struct thread_data *td)
	425	{
	426	return sizeof(struct aio_sq_ring) + td->o.iodepth * sizeof(u32);
	427	}
	428
	429	static void fio_aioring_cleanup(struct thread_data *td)
	430	{
	431	struct aioring_data *ld = td->io_ops_data;
	432
	433	if (ld) {
a90cd050 JA	434	/* Bump depth to match init depth */
	435	td->o.iodepth++;
	436
52885fa2 JA	437	/*
	438	* Work-around to avoid huge RCU stalls at exit time. If we
	439	* don't do this here, then it'll be torn down by exit_aio().
	440	* But for that case we can parallellize the freeing, thus
	441	* speeding it up a lot.
	442	*/
	443	if (!(td->flags & TD_F_CHILD))
	444	io_destroy(ld->aio_ctx);
	445	free(ld->io_u_index);
	446	free(ld->io_us);
	447	fio_memfree(ld->sq_ring, aioring_sq_size(td), false);
	448	fio_memfree(ld->iocbs, aioring_sq_iocb(td), false);
	449	fio_memfree(ld->cq_ring, aioring_cq_size(td), false);
	450	free(ld);
	451	}
	452	}
	453
	454	static int fio_aioring_queue_init(struct thread_data *td)
	455	{
52885fa2 JA	456	struct aioring_data *ld = td->io_ops_data;
	457	struct aioring_options *o = td->eo;
	458	int flags = IOCTX_FLAG_SCQRING;
	459	int depth = td->o.iodepth;
	460
	461	if (o->hipri)
	462	flags \|= IOCTX_FLAG_IOPOLL;
a90cd050 JA	463	if (o->sqthread_set) {
	464	ld->sq_ring->sq_thread_cpu = o->sqthread;
	465	flags \|= IOCTX_FLAG_SQTHREAD;
771c9901 JA	466	if (o->sqthread_poll)
771c9901 JA	467	flags \|= IOCTX_FLAG_SQPOLL;
a90cd050 JA	468	} else if (o->sqwq)
	469	flags \|= IOCTX_FLAG_SQWQ;
	470
52885fa2 JA	471	if (o->fixedbufs) {
	472	struct rlimit rlim = {
	473	.rlim_cur = RLIM_INFINITY,
	474	.rlim_max = RLIM_INFINITY,
	475	};
	476
	477	setrlimit(RLIMIT_MEMLOCK, &rlim);
	478	flags \|= IOCTX_FLAG_FIXEDBUFS;
	479	}
	480
	481	return syscall(__NR_sys_io_setup2, depth, flags,
	482	ld->sq_ring, ld->cq_ring, &ld->aio_ctx);
52885fa2 JA	483	}
	484
	485	static int fio_aioring_post_init(struct thread_data *td)
	486	{
	487	struct aioring_data *ld = td->io_ops_data;
	488	struct aioring_options *o = td->eo;
	489	struct io_u *io_u;
	490	struct iocb *iocb;
	491	int err = 0;
	492
	493	if (o->fixedbufs) {
	494	int i;
	495
	496	for (i = 0; i < td->o.iodepth; i++) {
	497	io_u = ld->io_u_index[i];
	498	iocb = &ld->iocbs[i];
	499	iocb->u.c.buf = io_u->buf;
	500	iocb->u.c.nbytes = td_max_bs(td);
	501
	502	if (o->hipri)
	503	iocb->u.c.flags \|= IOCB_FLAG_HIPRI;
	504	}
	505	}
	506
	507	err = fio_aioring_queue_init(td);
	508	if (err) {
	509	td_verror(td, -err, "io_queue_init");
	510	return 1;
	511	}
	512
a90cd050 JA	513	/* Adjust depth back again */
a90cd050 JA	514	td->o.iodepth--;
52885fa2 JA	515	return 0;
	516	}
	517
	518	static int fio_aioring_init(struct thread_data *td)
	519	{
a90cd050	520	struct aioring_options *o = td->eo;
52885fa2 JA	521	struct aioring_data *ld;
52885fa2 JA	522
a90cd050 JA	523	if (o->sqthread_set && o->sqwq) {
	524	log_err("fio: aioring sqthread and sqwq are mutually exclusive\n");
	525	return 1;
	526	}
	527
3b617146 JA	528	/* ring needs an extra entry, add one to achieve QD set */
3b617146 JA	529	td->o.iodepth++;
52885fa2 JA	530
	531	ld = calloc(1, sizeof(*ld));
	532
	533	/* io_u index */
	534	ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *));
	535	ld->io_us = calloc(td->o.iodepth, sizeof(struct io_u *));
	536
	537	ld->iocbs = fio_memalign(page_size, aioring_sq_iocb(td), false);
	538	memset(ld->iocbs, 0, aioring_sq_iocb(td));
	539
	540	ld->sq_ring = fio_memalign(page_size, aioring_sq_size(td), false);
	541	memset(ld->sq_ring, 0, aioring_sq_size(td));
	542	ld->sq_ring->nr_events = td->o.iodepth;
1f90e9bb	543	ld->sq_ring->iocbs = (u64) (uintptr_t) ld->iocbs;
52885fa2 JA	544
	545	ld->cq_ring = fio_memalign(page_size, aioring_cq_size(td), false);
	546	memset(ld->cq_ring, 0, aioring_cq_size(td));
	547	ld->cq_ring->nr_events = td->o.iodepth * 2;
	548
	549	td->io_ops_data = ld;
	550	return 0;
	551	}
	552
	553	static int fio_aioring_io_u_init(struct thread_data td, struct io_u io_u)
	554	{
	555	struct aioring_data *ld = td->io_ops_data;
	556
	557	ld->io_u_index[io_u->index] = io_u;
	558	return 0;
	559	}
	560
	561	static struct ioengine_ops ioengine = {
	562	.name = "aio-ring",
	563	.version = FIO_IOOPS_VERSION,
	564	.init = fio_aioring_init,
	565	.post_init = fio_aioring_post_init,
	566	.io_u_init = fio_aioring_io_u_init,
	567	.prep = fio_aioring_prep,
	568	.queue = fio_aioring_queue,
	569	.commit = fio_aioring_commit,
	570	.getevents = fio_aioring_getevents,
	571	.event = fio_aioring_event,
	572	.cleanup = fio_aioring_cleanup,
	573	.open_file = generic_open_file,
	574	.close_file = generic_close_file,
	575	.get_file_size = generic_get_file_size,
	576	.options = options,
	577	.option_struct_size = sizeof(struct aioring_options),
	578	};
	579
	580	static void fio_init fio_aioring_register(void)
	581	{
52885fa2	582	register_ioengine(&ioengine);
52885fa2 JA	583	}
	584
	585	static void fio_exit fio_aioring_unregister(void)
	586	{
52885fa2	587	unregister_ioengine(&ioengine);
52885fa2	588	}
1f90e9bb	589	#endif