[fio.git] / engines / io_uring.c

/*
 * io_uring engine
 *
 * IO engine using the new native Linux aio io_uring interface. See:
 *
 * http://git.kernel.dk/cgit/linux-block/log/?h=io_uring
 *
 */
#include <stdlib.h>
#include <unistd.h>
#include <errno.h>
#include <sys/time.h>
#include <sys/resource.h>

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

#ifdef ARCH_HAVE_IOURING

#include "../lib/types.h"
#include "../os/linux/io_uring.h"

struct io_sq_ring {
	unsigned *head;
	unsigned *tail;
	unsigned *ring_mask;
	unsigned *ring_entries;
	unsigned *flags;
	unsigned *array;
};

struct io_cq_ring {
	unsigned *head;
	unsigned *tail;
	unsigned *ring_mask;
	unsigned *ring_entries;
	struct io_uring_cqe *cqes;
};

struct ioring_mmap {
	void *ptr;
	size_t len;
};

struct ioring_data {
	int ring_fd;

	struct io_u **io_u_index;

	struct io_sq_ring sq_ring;
	struct io_uring_sqe *sqes;
	struct iovec *iovecs;
	unsigned sq_ring_mask;

	struct io_cq_ring cq_ring;
	unsigned cq_ring_mask;

	int queued;
	int cq_ring_off;
	unsigned iodepth;

	struct ioring_mmap mmap[3];
};

struct ioring_options {
	void *pad;
	unsigned int hipri;
	unsigned int fixedbufs;
	unsigned int sqpoll_thread;
	unsigned int sqpoll_set;
	unsigned int sqpoll_cpu;
};

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

	o->sqpoll_cpu = *val;
	o->sqpoll_set = 1;
	return 0;
}

static struct fio_option options[] = {
	{
		.name	= "hipri",
		.lname	= "High Priority",
		.type	= FIO_OPT_STR_SET,
		.off1	= offsetof(struct ioring_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 ioring_options, fixedbufs),
		.help	= "Pre map IO buffers",
		.category = FIO_OPT_C_ENGINE,
		.group	= FIO_OPT_G_LIBAIO,
	},
	{
		.name	= "sqthread_poll",
		.lname	= "Kernel SQ thread polling",
		.type	= FIO_OPT_INT,
		.off1	= offsetof(struct ioring_options, sqpoll_thread),
		.help	= "Offload submission/completion to kernel thread",
		.category = FIO_OPT_C_ENGINE,
		.group	= FIO_OPT_G_LIBAIO,
	},
	{
		.name	= "sqthread_poll_cpu",
		.lname	= "SQ Thread Poll CPU",
		.type	= FIO_OPT_INT,
		.cb	= fio_ioring_sqpoll_cb,
		.help	= "What CPU to run SQ thread polling on",
		.category = FIO_OPT_C_ENGINE,
		.group	= FIO_OPT_G_LIBAIO,
	},
	{
		.name	= NULL,
	},
};

static int io_uring_enter(struct ioring_data *ld, unsigned int to_submit,
			 unsigned int min_complete, unsigned int flags)
{
	return syscall(__NR_sys_io_uring_enter, ld->ring_fd, to_submit,
			min_complete, flags, NULL, 0);
}

static int fio_ioring_prep(struct thread_data *td, struct io_u *io_u)
{
	struct ioring_data *ld = td->io_ops_data;
	struct ioring_options *o = td->eo;
	struct fio_file *f = io_u->file;
	struct io_uring_sqe *sqe;

	sqe = &ld->sqes[io_u->index];
	sqe->fd = f->fd;
	sqe->flags = 0;
	sqe->ioprio = 0;
	sqe->buf_index = 0;

	if (io_u->ddir == DDIR_READ || io_u->ddir == DDIR_WRITE) {
		if (o->fixedbufs) {
			if (io_u->ddir == DDIR_READ)
				sqe->opcode = IORING_OP_READ_FIXED;
			else
				sqe->opcode = IORING_OP_WRITE_FIXED;
			sqe->addr = (unsigned long) io_u->xfer_buf;
			sqe->len = io_u->xfer_buflen;
			sqe->buf_index = io_u->index;
		} else {
			if (io_u->ddir == DDIR_READ)
				sqe->opcode = IORING_OP_READV;
			else
				sqe->opcode = IORING_OP_WRITEV;
			sqe->addr = (unsigned long) &ld->iovecs[io_u->index];
			sqe->len = 1;
		}
		sqe->off = io_u->offset;
	} else if (ddir_sync(io_u->ddir)) {
		sqe->fsync_flags = 0;
		if (io_u->ddir == DDIR_DATASYNC)
			sqe->fsync_flags |= IORING_FSYNC_DATASYNC;
		sqe->opcode = IORING_OP_FSYNC;
	}

	sqe->user_data = (unsigned long) io_u;
	return 0;
}

static struct io_u *fio_ioring_event(struct thread_data *td, int event)
{
	struct ioring_data *ld = td->io_ops_data;
	struct io_uring_cqe *cqe;
	struct io_u *io_u;
	unsigned index;

	index = (event + ld->cq_ring_off) & ld->cq_ring_mask;

	cqe = &ld->cq_ring.cqes[index];
	io_u = (struct io_u *) (uintptr_t) cqe->user_data;

	if (cqe->res != io_u->xfer_buflen) {
		if (cqe->res > io_u->xfer_buflen)
			io_u->error = -cqe->res;
		else
			io_u->resid = io_u->xfer_buflen - cqe->res;
	} else
		io_u->error = 0;

	return io_u;
}

static int fio_ioring_cqring_reap(struct thread_data *td, unsigned int events,
				   unsigned int max)
{
	struct ioring_data *ld = td->io_ops_data;
	struct io_cq_ring *ring = &ld->cq_ring;
	unsigned head, reaped = 0;

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

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

static int fio_ioring_getevents(struct thread_data *td, unsigned int min,
				unsigned int max, const struct timespec *t)
{
	struct ioring_data *ld = td->io_ops_data;
	unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
	struct ioring_options *o = td->eo;
	struct io_cq_ring *ring = &ld->cq_ring;
	unsigned events = 0;
	int r;

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

		if (!o->sqpoll_thread) {
			r = io_uring_enter(ld, 0, actual_min,
						IORING_ENTER_GETEVENTS);
			if (r < 0) {
				if (errno == EAGAIN)
					continue;
				td_verror(td, errno, "io_uring_enter");
				break;
			}
		}
	} while (events < min);

	return r < 0 ? r : events;
}

static enum fio_q_status fio_ioring_queue(struct thread_data *td,
					  struct io_u *io_u)
{
	struct ioring_data *ld = td->io_ops_data;
	struct io_sq_ring *ring = &ld->sq_ring;
	unsigned tail, next_tail;

	fio_ro_check(td, io_u);

	if (ld->queued == ld->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;
	read_barrier();
	if (next_tail == *ring->head)
		return FIO_Q_BUSY;

	/* ensure sqe stores are ordered with tail update */
	write_barrier();
	ring->array[tail & ld->sq_ring_mask] = io_u->index;
	*ring->tail = next_tail;
	write_barrier();

	ld->queued++;
	return FIO_Q_QUEUED;
}

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

	if (!fio_fill_issue_time(td))
		return;

	fio_gettime(&now, NULL);

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

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

		start++;
	}
}

static int fio_ioring_commit(struct thread_data *td)
{
	struct ioring_data *ld = td->io_ops_data;
	struct ioring_options *o = td->eo;
	int ret;

	if (!ld->queued)
		return 0;

	/*
	 * Kernel side does submission. just need to check if the ring is
	 * flagged as needing a kick, if so, call io_uring_enter(). This
	 * only happens if we've been idle too long.
	 */
	if (o->sqpoll_thread) {
		struct io_sq_ring *ring = &ld->sq_ring;

		read_barrier();
		if (*ring->flags & IORING_SQ_NEED_WAKEUP)
			io_uring_enter(ld, ld->queued, 0,
					IORING_ENTER_SQ_WAKEUP);
		ld->queued = 0;
		return 0;
	}

	do {
		unsigned start = *ld->sq_ring.head;
		long nr = ld->queued;

		ret = io_uring_enter(ld, nr, 0, IORING_ENTER_GETEVENTS);
		if (ret > 0) {
			fio_ioring_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_ioring_cqring_reap(td, 0, ld->queued);
				if (ret)
					continue;
				/* Shouldn't happen */
				usleep(1);
				continue;
			}
			td_verror(td, errno, "io_uring_enter submit");
			break;
		}
	} while (ld->queued);

	return ret;
}

static void fio_ioring_unmap(struct ioring_data *ld)
{
	int i;

	for (i = 0; i < ARRAY_SIZE(ld->mmap); i++)
		munmap(ld->mmap[i].ptr, ld->mmap[i].len);
	close(ld->ring_fd);
}

static void fio_ioring_cleanup(struct thread_data *td)
{
	struct ioring_data *ld = td->io_ops_data;

	if (ld) {
		if (!(td->flags & TD_F_CHILD))
			fio_ioring_unmap(ld);

		free(ld->io_u_index);
		free(ld->iovecs);
		free(ld);
	}
}

static int fio_ioring_mmap(struct ioring_data *ld, struct io_uring_params *p)
{
	struct io_sq_ring *sring = &ld->sq_ring;
	struct io_cq_ring *cring = &ld->cq_ring;
	void *ptr;

	ld->mmap[0].len = p->sq_off.array + p->sq_entries * sizeof(__u32);
	ptr = mmap(0, ld->mmap[0].len, PROT_READ | PROT_WRITE,
			MAP_SHARED | MAP_POPULATE, ld->ring_fd,
			IORING_OFF_SQ_RING);
	ld->mmap[0].ptr = ptr;
	sring->head = ptr + p->sq_off.head;
	sring->tail = ptr + p->sq_off.tail;
	sring->ring_mask = ptr + p->sq_off.ring_mask;
	sring->ring_entries = ptr + p->sq_off.ring_entries;
	sring->flags = ptr + p->sq_off.flags;
	sring->array = ptr + p->sq_off.array;
	ld->sq_ring_mask = *sring->ring_mask;

	ld->mmap[1].len = p->sq_entries * sizeof(struct io_uring_sqe);
	ld->sqes = mmap(0, ld->mmap[1].len, PROT_READ | PROT_WRITE,
				MAP_SHARED | MAP_POPULATE, ld->ring_fd,
				IORING_OFF_SQES);
	ld->mmap[1].ptr = ld->sqes;

	ld->mmap[2].len = p->cq_off.cqes +
				p->cq_entries * sizeof(struct io_uring_cqe);
	ptr = mmap(0, ld->mmap[2].len, PROT_READ | PROT_WRITE,
			MAP_SHARED | MAP_POPULATE, ld->ring_fd,
			IORING_OFF_CQ_RING);
	ld->mmap[2].ptr = ptr;
	cring->head = ptr + p->cq_off.head;
	cring->tail = ptr + p->cq_off.tail;
	cring->ring_mask = ptr + p->cq_off.ring_mask;
	cring->ring_entries = ptr + p->cq_off.ring_entries;
	cring->cqes = ptr + p->cq_off.cqes;
	ld->cq_ring_mask = *cring->ring_mask;
	return 0;
}

static int fio_ioring_queue_init(struct thread_data *td)
{
	struct ioring_data *ld = td->io_ops_data;
	struct ioring_options *o = td->eo;
	int depth = td->o.iodepth;
	struct io_uring_params p;
	int ret;

	memset(&p, 0, sizeof(p));

	if (o->hipri)
		p.flags |= IORING_SETUP_IOPOLL;
	if (o->sqpoll_thread) {
		p.flags |= IORING_SETUP_SQPOLL;
		if (o->sqpoll_set) {
			p.flags |= IORING_SETUP_SQ_AFF;
			p.sq_thread_cpu = o->sqpoll_cpu;
		}
	}

	ret = syscall(__NR_sys_io_uring_setup, depth, &p);
	if (ret < 0)
		return ret;

	ld->ring_fd = ret;

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

		if (setrlimit(RLIMIT_MEMLOCK, &rlim) < 0)
			return -1;

		ret = syscall(__NR_sys_io_uring_register, ld->ring_fd,
				IORING_REGISTER_BUFFERS, ld->iovecs, depth);
		if (ret < 0)
			return ret;
	}

	return fio_ioring_mmap(ld, &p);
}

static int fio_ioring_post_init(struct thread_data *td)
{
	struct ioring_data *ld = td->io_ops_data;
	struct io_u *io_u;
	int err, i;

	for (i = 0; i < td->o.iodepth; i++) {
		struct iovec *iov = &ld->iovecs[i];

		io_u = ld->io_u_index[i];
		iov->iov_base = io_u->buf;
		iov->iov_len = td_max_bs(td);
	}

	err = fio_ioring_queue_init(td);
	if (err) {
		td_verror(td, errno, "io_queue_init");
		return 1;
	}

	return 0;
}

static unsigned roundup_pow2(unsigned depth)
{
	return 1UL << __fls(depth - 1);
}

static int fio_ioring_init(struct thread_data *td)
{
	struct ioring_data *ld;

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

	/* ring depth must be a power-of-2 */
	ld->iodepth = td->o.iodepth;
	td->o.iodepth = roundup_pow2(td->o.iodepth);

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

	td->io_ops_data = ld;
	return 0;
}

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

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

static struct ioengine_ops ioengine = {
	.name			= "io_uring",
	.version		= FIO_IOOPS_VERSION,
	.init			= fio_ioring_init,
	.post_init		= fio_ioring_post_init,
	.io_u_init		= fio_ioring_io_u_init,
	.prep			= fio_ioring_prep,
	.queue			= fio_ioring_queue,
	.commit			= fio_ioring_commit,
	.getevents		= fio_ioring_getevents,
	.event			= fio_ioring_event,
	.cleanup		= fio_ioring_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 ioring_options),
};

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

static void fio_exit fio_ioring_unregister(void)
{
	unregister_ioengine(&ioengine);
}
#endif
Commit	Line	Data
	1	/*
	2	* io_uring engine
	3	*
	4	* IO engine using the new native Linux aio io_uring interface. See:
	5	*
	6	* http://git.kernel.dk/cgit/linux-block/log/?h=io_uring
	7	*
	8	*/
	9	#include <stdlib.h>
	10	#include <unistd.h>
	11	#include <errno.h>
	12	#include <sys/time.h>
	13	#include <sys/resource.h>
	14
	15	#include "../fio.h"
	16	#include "../lib/pow2.h"
	17	#include "../optgroup.h"
	18	#include "../lib/memalign.h"
	19	#include "../lib/fls.h"
	20
	21	#ifdef ARCH_HAVE_IOURING
	22
	23	#include "../lib/types.h"
	24	#include "../os/linux/io_uring.h"
	25
	26	struct io_sq_ring {
	27	unsigned *head;
	28	unsigned *tail;
	29	unsigned *ring_mask;
	30	unsigned *ring_entries;
	31	unsigned *flags;
	32	unsigned *array;
	33	};
	34
	35	struct io_cq_ring {
	36	unsigned *head;
	37	unsigned *tail;
	38	unsigned *ring_mask;
	39	unsigned *ring_entries;
	40	struct io_uring_cqe *cqes;
	41	};
	42
	43	struct ioring_mmap {
	44	void *ptr;
	45	size_t len;
	46	};
	47
	48	struct ioring_data {
	49	int ring_fd;
	50
	51	struct io_u **io_u_index;
	52
	53	struct io_sq_ring sq_ring;
	54	struct io_uring_sqe *sqes;
	55	struct iovec *iovecs;
	56	unsigned sq_ring_mask;
	57
	58	struct io_cq_ring cq_ring;
	59	unsigned cq_ring_mask;
	60
	61	int queued;
	62	int cq_ring_off;
	63	unsigned iodepth;
	64
	65	struct ioring_mmap mmap[3];
	66	};
	67
	68	struct ioring_options {
	69	void *pad;
	70	unsigned int hipri;
	71	unsigned int fixedbufs;
	72	unsigned int sqpoll_thread;
	73	unsigned int sqpoll_set;
	74	unsigned int sqpoll_cpu;
	75	};
	76
	77	static int fio_ioring_sqpoll_cb(void data, unsigned long long val)
	78	{
	79	struct ioring_options *o = data;
	80
	81	o->sqpoll_cpu = *val;
	82	o->sqpoll_set = 1;
	83	return 0;
	84	}
	85
	86	static struct fio_option options[] = {
	87	{
	88	.name = "hipri",
	89	.lname = "High Priority",
	90	.type = FIO_OPT_STR_SET,
	91	.off1 = offsetof(struct ioring_options, hipri),
	92	.help = "Use polled IO completions",
	93	.category = FIO_OPT_C_ENGINE,
	94	.group = FIO_OPT_G_LIBAIO,
	95	},
	96	{
	97	.name = "fixedbufs",
	98	.lname = "Fixed (pre-mapped) IO buffers",
	99	.type = FIO_OPT_STR_SET,
	100	.off1 = offsetof(struct ioring_options, fixedbufs),
	101	.help = "Pre map IO buffers",
	102	.category = FIO_OPT_C_ENGINE,
	103	.group = FIO_OPT_G_LIBAIO,
	104	},
	105	{
	106	.name = "sqthread_poll",
	107	.lname = "Kernel SQ thread polling",
	108	.type = FIO_OPT_INT,
	109	.off1 = offsetof(struct ioring_options, sqpoll_thread),
	110	.help = "Offload submission/completion to kernel thread",
	111	.category = FIO_OPT_C_ENGINE,
	112	.group = FIO_OPT_G_LIBAIO,
	113	},
	114	{
	115	.name = "sqthread_poll_cpu",
	116	.lname = "SQ Thread Poll CPU",
	117	.type = FIO_OPT_INT,
	118	.cb = fio_ioring_sqpoll_cb,
	119	.help = "What CPU to run SQ thread polling on",
	120	.category = FIO_OPT_C_ENGINE,
	121	.group = FIO_OPT_G_LIBAIO,
	122	},
	123	{
	124	.name = NULL,
	125	},
	126	};
	127
	128	static int io_uring_enter(struct ioring_data *ld, unsigned int to_submit,
	129	unsigned int min_complete, unsigned int flags)
	130	{
	131	return syscall(__NR_sys_io_uring_enter, ld->ring_fd, to_submit,
	132	min_complete, flags, NULL, 0);
	133	}
	134
	135	static int fio_ioring_prep(struct thread_data td, struct io_u io_u)
	136	{
	137	struct ioring_data *ld = td->io_ops_data;
	138	struct ioring_options *o = td->eo;
	139	struct fio_file *f = io_u->file;
	140	struct io_uring_sqe *sqe;
	141
	142	sqe = &ld->sqes[io_u->index];
	143	sqe->fd = f->fd;
	144	sqe->flags = 0;
	145	sqe->ioprio = 0;
	146	sqe->buf_index = 0;
	147
	148	if (io_u->ddir == DDIR_READ \|\| io_u->ddir == DDIR_WRITE) {
	149	if (o->fixedbufs) {
	150	if (io_u->ddir == DDIR_READ)
	151	sqe->opcode = IORING_OP_READ_FIXED;
	152	else
	153	sqe->opcode = IORING_OP_WRITE_FIXED;
	154	sqe->addr = (unsigned long) io_u->xfer_buf;
	155	sqe->len = io_u->xfer_buflen;
	156	sqe->buf_index = io_u->index;
	157	} else {
	158	if (io_u->ddir == DDIR_READ)
	159	sqe->opcode = IORING_OP_READV;
	160	else
	161	sqe->opcode = IORING_OP_WRITEV;
	162	sqe->addr = (unsigned long) &ld->iovecs[io_u->index];
	163	sqe->len = 1;
	164	}
	165	sqe->off = io_u->offset;
	166	} else if (ddir_sync(io_u->ddir)) {
	167	sqe->fsync_flags = 0;
	168	if (io_u->ddir == DDIR_DATASYNC)
	169	sqe->fsync_flags \|= IORING_FSYNC_DATASYNC;
	170	sqe->opcode = IORING_OP_FSYNC;
	171	}
	172
	173	sqe->user_data = (unsigned long) io_u;
	174	return 0;
	175	}
	176
	177	static struct io_u fio_ioring_event(struct thread_data td, int event)
	178	{
	179	struct ioring_data *ld = td->io_ops_data;
	180	struct io_uring_cqe *cqe;
	181	struct io_u *io_u;
	182	unsigned index;
	183
	184	index = (event + ld->cq_ring_off) & ld->cq_ring_mask;
	185
	186	cqe = &ld->cq_ring.cqes[index];
	187	io_u = (struct io_u *) (uintptr_t) cqe->user_data;
	188
	189	if (cqe->res != io_u->xfer_buflen) {
	190	if (cqe->res > io_u->xfer_buflen)
	191	io_u->error = -cqe->res;
	192	else
	193	io_u->resid = io_u->xfer_buflen - cqe->res;
	194	} else
	195	io_u->error = 0;
	196
	197	return io_u;
	198	}
	199
	200	static int fio_ioring_cqring_reap(struct thread_data *td, unsigned int events,
	201	unsigned int max)
	202	{
	203	struct ioring_data *ld = td->io_ops_data;
	204	struct io_cq_ring *ring = &ld->cq_ring;
	205	unsigned head, reaped = 0;
	206
	207	head = *ring->head;
	208	do {
	209	read_barrier();
	210	if (head == *ring->tail)
	211	break;
	212	reaped++;
	213	head++;
	214	} while (reaped + events < max);
	215
	216	*ring->head = head;
	217	write_barrier();
	218	return reaped;
	219	}
	220
	221	static int fio_ioring_getevents(struct thread_data *td, unsigned int min,
	222	unsigned int max, const struct timespec *t)
	223	{
	224	struct ioring_data *ld = td->io_ops_data;
	225	unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
	226	struct ioring_options *o = td->eo;
	227	struct io_cq_ring *ring = &ld->cq_ring;
	228	unsigned events = 0;
	229	int r;
	230
	231	ld->cq_ring_off = *ring->head;
	232	do {
	233	r = fio_ioring_cqring_reap(td, events, max);
	234	if (r) {
	235	events += r;
	236	continue;
	237	}
	238
	239	if (!o->sqpoll_thread) {
	240	r = io_uring_enter(ld, 0, actual_min,
	241	IORING_ENTER_GETEVENTS);
	242	if (r < 0) {
	243	if (errno == EAGAIN)
	244	continue;
	245	td_verror(td, errno, "io_uring_enter");
	246	break;
	247	}
	248	}
	249	} while (events < min);
	250
	251	return r < 0 ? r : events;
	252	}
	253
	254	static enum fio_q_status fio_ioring_queue(struct thread_data *td,
	255	struct io_u *io_u)
	256	{
	257	struct ioring_data *ld = td->io_ops_data;
	258	struct io_sq_ring *ring = &ld->sq_ring;
	259	unsigned tail, next_tail;
	260
	261	fio_ro_check(td, io_u);
	262
	263	if (ld->queued == ld->iodepth)
	264	return FIO_Q_BUSY;
	265
	266	if (io_u->ddir == DDIR_TRIM) {
	267	if (ld->queued)
	268	return FIO_Q_BUSY;
	269
	270	do_io_u_trim(td, io_u);
	271	io_u_mark_submit(td, 1);
	272	io_u_mark_complete(td, 1);
	273	return FIO_Q_COMPLETED;
	274	}
	275
	276	tail = *ring->tail;
	277	next_tail = tail + 1;
	278	read_barrier();
	279	if (next_tail == *ring->head)
	280	return FIO_Q_BUSY;
	281
	282	/* ensure sqe stores are ordered with tail update */
	283	write_barrier();
	284	ring->array[tail & ld->sq_ring_mask] = io_u->index;
	285	*ring->tail = next_tail;
	286	write_barrier();
	287
	288	ld->queued++;
	289	return FIO_Q_QUEUED;
	290	}
	291
	292	static void fio_ioring_queued(struct thread_data *td, int start, int nr)
	293	{
	294	struct ioring_data *ld = td->io_ops_data;
	295	struct timespec now;
	296
	297	if (!fio_fill_issue_time(td))
	298	return;
	299
	300	fio_gettime(&now, NULL);
	301
	302	while (nr--) {
	303	struct io_sq_ring *ring = &ld->sq_ring;
	304	int index = ring->array[start & ld->sq_ring_mask];
	305	struct io_u *io_u = ld->io_u_index[index];
	306
	307	memcpy(&io_u->issue_time, &now, sizeof(now));
	308	io_u_queued(td, io_u);
	309
	310	start++;
	311	}
	312	}
	313
	314	static int fio_ioring_commit(struct thread_data *td)
	315	{
	316	struct ioring_data *ld = td->io_ops_data;
	317	struct ioring_options *o = td->eo;
	318	int ret;
	319
	320	if (!ld->queued)
	321	return 0;
	322
	323	/*
	324	* Kernel side does submission. just need to check if the ring is
	325	* flagged as needing a kick, if so, call io_uring_enter(). This
	326	* only happens if we've been idle too long.
	327	*/
	328	if (o->sqpoll_thread) {
	329	struct io_sq_ring *ring = &ld->sq_ring;
	330
	331	read_barrier();
	332	if (*ring->flags & IORING_SQ_NEED_WAKEUP)
	333	io_uring_enter(ld, ld->queued, 0,
	334	IORING_ENTER_SQ_WAKEUP);
	335	ld->queued = 0;
	336	return 0;
	337	}
	338
	339	do {
	340	unsigned start = *ld->sq_ring.head;
	341	long nr = ld->queued;
	342
	343	ret = io_uring_enter(ld, nr, 0, IORING_ENTER_GETEVENTS);
	344	if (ret > 0) {
	345	fio_ioring_queued(td, start, ret);
	346	io_u_mark_submit(td, ret);
	347
	348	ld->queued -= ret;
	349	ret = 0;
	350	} else if (!ret) {
	351	io_u_mark_submit(td, ret);
	352	continue;
	353	} else {
	354	if (errno == EAGAIN) {
	355	ret = fio_ioring_cqring_reap(td, 0, ld->queued);
	356	if (ret)
	357	continue;
	358	/* Shouldn't happen */
	359	usleep(1);
	360	continue;
	361	}
	362	td_verror(td, errno, "io_uring_enter submit");
	363	break;
	364	}
	365	} while (ld->queued);
	366
	367	return ret;
	368	}
	369
	370	static void fio_ioring_unmap(struct ioring_data *ld)
	371	{
	372	int i;
	373
	374	for (i = 0; i < ARRAY_SIZE(ld->mmap); i++)
	375	munmap(ld->mmap[i].ptr, ld->mmap[i].len);
	376	close(ld->ring_fd);
	377	}
	378
	379	static void fio_ioring_cleanup(struct thread_data *td)
	380	{
	381	struct ioring_data *ld = td->io_ops_data;
	382
	383	if (ld) {
	384	if (!(td->flags & TD_F_CHILD))
	385	fio_ioring_unmap(ld);
	386
	387	free(ld->io_u_index);
	388	free(ld->iovecs);
	389	free(ld);
	390	}
	391	}
	392
	393	static int fio_ioring_mmap(struct ioring_data ld, struct io_uring_params p)
	394	{
	395	struct io_sq_ring *sring = &ld->sq_ring;
	396	struct io_cq_ring *cring = &ld->cq_ring;
	397	void *ptr;
	398
	399	ld->mmap[0].len = p->sq_off.array + p->sq_entries * sizeof(__u32);
	400	ptr = mmap(0, ld->mmap[0].len, PROT_READ \| PROT_WRITE,
	401	MAP_SHARED \| MAP_POPULATE, ld->ring_fd,
	402	IORING_OFF_SQ_RING);
	403	ld->mmap[0].ptr = ptr;
	404	sring->head = ptr + p->sq_off.head;
	405	sring->tail = ptr + p->sq_off.tail;
	406	sring->ring_mask = ptr + p->sq_off.ring_mask;
	407	sring->ring_entries = ptr + p->sq_off.ring_entries;
	408	sring->flags = ptr + p->sq_off.flags;
	409	sring->array = ptr + p->sq_off.array;
	410	ld->sq_ring_mask = *sring->ring_mask;
	411
	412	ld->mmap[1].len = p->sq_entries * sizeof(struct io_uring_sqe);
	413	ld->sqes = mmap(0, ld->mmap[1].len, PROT_READ \| PROT_WRITE,
	414	MAP_SHARED \| MAP_POPULATE, ld->ring_fd,
	415	IORING_OFF_SQES);
	416	ld->mmap[1].ptr = ld->sqes;
	417
	418	ld->mmap[2].len = p->cq_off.cqes +
	419	p->cq_entries * sizeof(struct io_uring_cqe);
	420	ptr = mmap(0, ld->mmap[2].len, PROT_READ \| PROT_WRITE,
	421	MAP_SHARED \| MAP_POPULATE, ld->ring_fd,
	422	IORING_OFF_CQ_RING);
	423	ld->mmap[2].ptr = ptr;
	424	cring->head = ptr + p->cq_off.head;
	425	cring->tail = ptr + p->cq_off.tail;
	426	cring->ring_mask = ptr + p->cq_off.ring_mask;
	427	cring->ring_entries = ptr + p->cq_off.ring_entries;
	428	cring->cqes = ptr + p->cq_off.cqes;
	429	ld->cq_ring_mask = *cring->ring_mask;
	430	return 0;
	431	}
	432
	433	static int fio_ioring_queue_init(struct thread_data *td)
	434	{
	435	struct ioring_data *ld = td->io_ops_data;
	436	struct ioring_options *o = td->eo;
	437	int depth = td->o.iodepth;
	438	struct io_uring_params p;
	439	int ret;
	440
	441	memset(&p, 0, sizeof(p));
	442
	443	if (o->hipri)
	444	p.flags \|= IORING_SETUP_IOPOLL;
	445	if (o->sqpoll_thread) {
	446	p.flags \|= IORING_SETUP_SQPOLL;
	447	if (o->sqpoll_set) {
	448	p.flags \|= IORING_SETUP_SQ_AFF;
	449	p.sq_thread_cpu = o->sqpoll_cpu;
	450	}
	451	}
	452
	453	ret = syscall(__NR_sys_io_uring_setup, depth, &p);
	454	if (ret < 0)
	455	return ret;
	456
	457	ld->ring_fd = ret;
	458
	459	if (o->fixedbufs) {
	460	struct rlimit rlim = {
	461	.rlim_cur = RLIM_INFINITY,
	462	.rlim_max = RLIM_INFINITY,
	463	};
	464
	465	if (setrlimit(RLIMIT_MEMLOCK, &rlim) < 0)
	466	return -1;
	467
	468	ret = syscall(__NR_sys_io_uring_register, ld->ring_fd,
	469	IORING_REGISTER_BUFFERS, ld->iovecs, depth);
	470	if (ret < 0)
	471	return ret;
	472	}
	473
	474	return fio_ioring_mmap(ld, &p);
	475	}
	476
	477	static int fio_ioring_post_init(struct thread_data *td)
	478	{
	479	struct ioring_data *ld = td->io_ops_data;
	480	struct io_u *io_u;
	481	int err, i;
	482
	483	for (i = 0; i < td->o.iodepth; i++) {
	484	struct iovec *iov = &ld->iovecs[i];
	485
	486	io_u = ld->io_u_index[i];
	487	iov->iov_base = io_u->buf;
	488	iov->iov_len = td_max_bs(td);
	489	}
	490
	491	err = fio_ioring_queue_init(td);
	492	if (err) {
	493	td_verror(td, errno, "io_queue_init");
	494	return 1;
	495	}
	496
	497	return 0;
	498	}
	499
	500	static unsigned roundup_pow2(unsigned depth)
	501	{
	502	return 1UL << __fls(depth - 1);
	503	}
	504
	505	static int fio_ioring_init(struct thread_data *td)
	506	{
	507	struct ioring_data *ld;
	508
	509	ld = calloc(1, sizeof(*ld));
	510
	511	/* ring depth must be a power-of-2 */
	512	ld->iodepth = td->o.iodepth;
	513	td->o.iodepth = roundup_pow2(td->o.iodepth);
	514
	515	/* io_u index */
	516	ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *));
	517	ld->iovecs = calloc(td->o.iodepth, sizeof(struct iovec));
	518
	519	td->io_ops_data = ld;
	520	return 0;
	521	}
	522
	523	static int fio_ioring_io_u_init(struct thread_data td, struct io_u io_u)
	524	{
	525	struct ioring_data *ld = td->io_ops_data;
	526
	527	ld->io_u_index[io_u->index] = io_u;
	528	return 0;
	529	}
	530
	531	static struct ioengine_ops ioengine = {
	532	.name = "io_uring",
	533	.version = FIO_IOOPS_VERSION,
	534	.init = fio_ioring_init,
	535	.post_init = fio_ioring_post_init,
	536	.io_u_init = fio_ioring_io_u_init,
	537	.prep = fio_ioring_prep,
	538	.queue = fio_ioring_queue,
	539	.commit = fio_ioring_commit,
	540	.getevents = fio_ioring_getevents,
	541	.event = fio_ioring_event,
	542	.cleanup = fio_ioring_cleanup,
	543	.open_file = generic_open_file,
	544	.close_file = generic_close_file,
	545	.get_file_size = generic_get_file_size,
	546	.options = options,
	547	.option_struct_size = sizeof(struct ioring_options),
	548	};
	549
	550	static void fio_init fio_ioring_register(void)
	551	{
	552	register_ioengine(&ioengine);
	553	}
	554
	555	static void fio_exit fio_ioring_unregister(void)
	556	{
	557	unregister_ioengine(&ioengine);
	558	}
	559	#endif