[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/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_us;
	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;

	uint64_t cachehit;
	uint64_t cachemiss;

	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);
}

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 = 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 = &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 *) 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;

	if (io_u->ddir == DDIR_READ) {
		if (cqe->flags & IOCQE_FLAG_CACHEHIT)
			ld->cachehit++;
		else
			ld->cachemiss++;
	}

	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;

	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, 0);
		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) {
		td->ts.cachehit += ld->cachehit;
		td->ts.cachemiss += ld->cachemiss;

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

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

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

		setrlimit(RLIMIT_MEMLOCK, &rlim);
	}

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

	ld->ring_fd = ret;

	if (o->fixedbufs) {
		struct io_uring_register_buffers reg = {
			.iovecs = ld->iovecs,
			.nr_iovecs = depth
		};

		ret = syscall(__NR_sys_io_uring_register, ld->ring_fd,
				IORING_REGISTER_BUFFERS, &reg);
		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->io_us = 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
52885fa2	1	/*
bffad86f	2	* io_uring engine
52885fa2	3	*
bffad86f	4	* IO engine using the new native Linux aio io_uring interface. See:
a90cd050	5	*
bffad86f	6	* http://git.kernel.dk/cgit/linux-block/log/?h=io_uring
52885fa2 JA	7	*
	8	*/
	9	#include <stdlib.h>
	10	#include <unistd.h>
	11	#include <errno.h>
52885fa2 JA	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"
b87aa01a	19	#include "../lib/fls.h"
52885fa2	20
bffad86f	21	#ifdef ARCH_HAVE_IOURING
52885fa2	22
57fa61f0	23	#include "../lib/types.h"
bffad86f	24	#include "../os/io_uring.h"
9a2d78b3	25
bffad86f	26	struct io_sq_ring {
e2239016 JA	27	unsigned *head;
	28	unsigned *tail;
	29	unsigned *ring_mask;
	30	unsigned *ring_entries;
	31	unsigned *flags;
	32	unsigned *array;
52885fa2 JA	33	};
52885fa2 JA	34
bffad86f	35	struct io_cq_ring {
e2239016 JA	36	unsigned *head;
	37	unsigned *tail;
	38	unsigned *ring_mask;
	39	unsigned *ring_entries;
f0403f94	40	struct io_uring_cqe *cqes;
9a2d78b3 JA	41	};
9a2d78b3 JA	42
bffad86f	43	struct ioring_mmap {
9a2d78b3 JA	44	void *ptr;
9a2d78b3 JA	45	size_t len;
52885fa2 JA	46	};
52885fa2 JA	47
bffad86f	48	struct ioring_data {
9a2d78b3 JA	49	int ring_fd;
9a2d78b3 JA	50
52885fa2 JA	51	struct io_u **io_us;
	52	struct io_u **io_u_index;
	53
bffad86f	54	struct io_sq_ring sq_ring;
f0403f94	55	struct io_uring_sqe *sqes;
9a2d78b3	56	struct iovec *iovecs;
b87aa01a	57	unsigned sq_ring_mask;
52885fa2	58
bffad86f	59	struct io_cq_ring cq_ring;
b87aa01a	60	unsigned cq_ring_mask;
52885fa2 JA	61
	62	int queued;
	63	int cq_ring_off;
b87aa01a	64	unsigned iodepth;
96563db9 JA	65
	66	uint64_t cachehit;
	67	uint64_t cachemiss;
9a2d78b3	68
bffad86f	69	struct ioring_mmap mmap[3];
52885fa2 JA	70	};
52885fa2 JA	71
bffad86f	72	struct ioring_options {
52885fa2 JA	73	void *pad;
	74	unsigned int hipri;
	75	unsigned int fixedbufs;
3d7d00a3	76	unsigned int sqpoll_thread;
2ea53ca3 JA	77	unsigned int sqpoll_set;
2ea53ca3 JA	78	unsigned int sqpoll_cpu;
52885fa2 JA	79	};
52885fa2 JA	80
2ea53ca3	81	static int fio_ioring_sqpoll_cb(void data, unsigned long long val)
a90cd050	82	{
bffad86f	83	struct ioring_options *o = data;
a90cd050	84
2ea53ca3 JA	85	o->sqpoll_cpu = *val;
2ea53ca3 JA	86	o->sqpoll_set = 1;
a90cd050 JA	87	return 0;
	88	}
	89
52885fa2 JA	90	static struct fio_option options[] = {
	91	{
	92	.name = "hipri",
	93	.lname = "High Priority",
	94	.type = FIO_OPT_STR_SET,
bffad86f	95	.off1 = offsetof(struct ioring_options, hipri),
52885fa2 JA	96	.help = "Use polled IO completions",
	97	.category = FIO_OPT_C_ENGINE,
	98	.group = FIO_OPT_G_LIBAIO,
	99	},
	100	{
	101	.name = "fixedbufs",
	102	.lname = "Fixed (pre-mapped) IO buffers",
	103	.type = FIO_OPT_STR_SET,
bffad86f	104	.off1 = offsetof(struct ioring_options, fixedbufs),
52885fa2 JA	105	.help = "Pre map IO buffers",
	106	.category = FIO_OPT_C_ENGINE,
	107	.group = FIO_OPT_G_LIBAIO,
	108	},
771c9901 JA	109	{
771c9901 JA	110	.name = "sqthread_poll",
3d7d00a3 JA	111	.lname = "Kernel SQ thread polling",
	112	.type = FIO_OPT_INT,
	113	.off1 = offsetof(struct ioring_options, sqpoll_thread),
	114	.help = "Offload submission/completion to kernel thread",
	115	.category = FIO_OPT_C_ENGINE,
	116	.group = FIO_OPT_G_LIBAIO,
	117	},
	118	{
	119	.name = "sqthread_poll_cpu",
	120	.lname = "SQ Thread Poll CPU",
2ea53ca3 JA	121	.type = FIO_OPT_INT,
2ea53ca3 JA	122	.cb = fio_ioring_sqpoll_cb,
3d7d00a3	123	.help = "What CPU to run SQ thread polling on",
a90cd050 JA	124	.category = FIO_OPT_C_ENGINE,
	125	.group = FIO_OPT_G_LIBAIO,
	126	},
52885fa2 JA	127	{
	128	.name = NULL,
	129	},
	130	};
	131
bffad86f	132	static int io_uring_enter(struct ioring_data *ld, unsigned int to_submit,
52885fa2 JA	133	unsigned int min_complete, unsigned int flags)
52885fa2 JA	134	{
9a2d78b3 JA	135	return syscall(__NR_sys_io_uring_enter, ld->ring_fd, to_submit,
9a2d78b3 JA	136	min_complete, flags);
52885fa2 JA	137	}
52885fa2 JA	138
bffad86f	139	static int fio_ioring_prep(struct thread_data td, struct io_u io_u)
52885fa2	140	{
bffad86f	141	struct ioring_data *ld = td->io_ops_data;
cfcc8564	142	struct ioring_options *o = td->eo;
52885fa2	143	struct fio_file *f = io_u->file;
f0403f94	144	struct io_uring_sqe *sqe;
52885fa2	145
f0403f94 JA	146	sqe = &ld->sqes[io_u->index];
	147	sqe->fd = f->fd;
	148	sqe->flags = 0;
	149	sqe->ioprio = 0;
2ea53ca3	150	sqe->buf_index = 0;
52885fa2	151
e3970057	152	if (io_u->ddir == DDIR_READ \|\| io_u->ddir == DDIR_WRITE) {
f0403f94	153	if (o->fixedbufs) {
48e698fa JA	154	if (io_u->ddir == DDIR_READ)
	155	sqe->opcode = IORING_OP_READ_FIXED;
	156	else
	157	sqe->opcode = IORING_OP_WRITE_FIXED;
f0403f94 JA	158	sqe->addr = io_u->xfer_buf;
f0403f94 JA	159	sqe->len = io_u->xfer_buflen;
2ea53ca3	160	sqe->buf_index = io_u->index;
cfcc8564	161	} else {
48e698fa JA	162	if (io_u->ddir == DDIR_READ)
	163	sqe->opcode = IORING_OP_READV;
	164	else
	165	sqe->opcode = IORING_OP_WRITEV;
f0403f94 JA	166	sqe->addr = &ld->iovecs[io_u->index];
f0403f94 JA	167	sqe->len = 1;
cfcc8564	168	}
f0403f94	169	sqe->off = io_u->offset;
48e698fa JA	170	} else if (ddir_sync(io_u->ddir)) {
	171	sqe->fsync_flags = 0;
	172	if (io_u->ddir == DDIR_DATASYNC)
	173	sqe->fsync_flags \|= IORING_FSYNC_DATASYNC;
f0403f94	174	sqe->opcode = IORING_OP_FSYNC;
48e698fa	175	}
52885fa2	176
48e698fa	177	sqe->user_data = (unsigned long) io_u;
52885fa2 JA	178	return 0;
	179	}
	180
bffad86f	181	static struct io_u fio_ioring_event(struct thread_data td, int event)
52885fa2	182	{
bffad86f	183	struct ioring_data *ld = td->io_ops_data;
f0403f94	184	struct io_uring_cqe *cqe;
52885fa2	185	struct io_u *io_u;
b87aa01a	186	unsigned index;
52885fa2	187
b87aa01a	188	index = (event + ld->cq_ring_off) & ld->cq_ring_mask;
52885fa2	189
f0403f94	190	cqe = &ld->cq_ring.cqes[index];
48e698fa	191	io_u = (struct io_u *) cqe->user_data;
52885fa2	192
f0403f94 JA	193	if (cqe->res != io_u->xfer_buflen) {
	194	if (cqe->res > io_u->xfer_buflen)
	195	io_u->error = -cqe->res;
52885fa2	196	else
f0403f94	197	io_u->resid = io_u->xfer_buflen - cqe->res;
52885fa2 JA	198	} else
	199	io_u->error = 0;
	200
96563db9	201	if (io_u->ddir == DDIR_READ) {
f0403f94	202	if (cqe->flags & IOCQE_FLAG_CACHEHIT)
96563db9 JA	203	ld->cachehit++;
	204	else
	205	ld->cachemiss++;
	206	}
	207
52885fa2 JA	208	return io_u;
	209	}
	210
bffad86f	211	static int fio_ioring_cqring_reap(struct thread_data *td, unsigned int events,
52885fa2 JA	212	unsigned int max)
52885fa2 JA	213	{
bffad86f JA	214	struct ioring_data *ld = td->io_ops_data;
bffad86f JA	215	struct io_cq_ring *ring = &ld->cq_ring;
e2239016	216	unsigned head, reaped = 0;
52885fa2	217
9a2d78b3	218	head = *ring->head;
52885fa2 JA	219	do {
52885fa2 JA	220	read_barrier();
9a2d78b3	221	if (head == *ring->tail)
52885fa2 JA	222	break;
	223	reaped++;
	224	head++;
52885fa2 JA	225	} while (reaped + events < max);
52885fa2 JA	226
9a2d78b3	227	*ring->head = head;
52885fa2 JA	228	write_barrier();
	229	return reaped;
	230	}
	231
bffad86f JA	232	static int fio_ioring_getevents(struct thread_data *td, unsigned int min,
bffad86f JA	233	unsigned int max, const struct timespec *t)
52885fa2	234	{
bffad86f	235	struct ioring_data *ld = td->io_ops_data;
52885fa2	236	unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
bffad86f JA	237	struct ioring_options *o = td->eo;
bffad86f JA	238	struct io_cq_ring *ring = &ld->cq_ring;
b87aa01a JA	239	unsigned events = 0;
b87aa01a JA	240	int r;
52885fa2	241
9a2d78b3	242	ld->cq_ring_off = *ring->head;
52885fa2	243	do {
bffad86f	244	r = fio_ioring_cqring_reap(td, events, max);
52885fa2 JA	245	if (r) {
	246	events += r;
	247	continue;
	248	}
	249
3d7d00a3	250	if (!o->sqpoll_thread) {
9a2d78b3 JA	251	r = io_uring_enter(ld, 0, actual_min,
9a2d78b3 JA	252	IORING_ENTER_GETEVENTS);
771c9901 JA	253	if (r < 0) {
	254	if (errno == EAGAIN)
	255	continue;
9a2d78b3	256	td_verror(td, errno, "io_uring_enter");
771c9901 JA	257	break;
771c9901 JA	258	}
52885fa2 JA	259	}
	260	} while (events < min);
	261
	262	return r < 0 ? r : events;
	263	}
	264
bffad86f JA	265	static enum fio_q_status fio_ioring_queue(struct thread_data *td,
bffad86f JA	266	struct io_u *io_u)
52885fa2	267	{
bffad86f JA	268	struct ioring_data *ld = td->io_ops_data;
bffad86f JA	269	struct io_sq_ring *ring = &ld->sq_ring;
52885fa2 JA	270	unsigned tail, next_tail;
	271
	272	fio_ro_check(td, io_u);
	273
b87aa01a	274	if (ld->queued == ld->iodepth)
52885fa2 JA	275	return FIO_Q_BUSY;
52885fa2 JA	276
52885fa2 JA	277	if (io_u->ddir == DDIR_TRIM) {
	278	if (ld->queued)
	279	return FIO_Q_BUSY;
	280
	281	do_io_u_trim(td, io_u);
	282	io_u_mark_submit(td, 1);
	283	io_u_mark_complete(td, 1);
	284	return FIO_Q_COMPLETED;
	285	}
	286
9a2d78b3	287	tail = *ring->tail;
52885fa2	288	next_tail = tail + 1;
52885fa2	289	read_barrier();
9a2d78b3	290	if (next_tail == *ring->head)
52885fa2 JA	291	return FIO_Q_BUSY;
52885fa2 JA	292
b87aa01a	293	ring->array[tail & ld->sq_ring_mask] = io_u->index;
9a2d78b3	294	*ring->tail = next_tail;
52885fa2 JA	295	write_barrier();
	296
	297	ld->queued++;
	298	return FIO_Q_QUEUED;
	299	}
	300
bffad86f	301	static void fio_ioring_queued(struct thread_data *td, int start, int nr)
52885fa2	302	{
bffad86f	303	struct ioring_data *ld = td->io_ops_data;
52885fa2 JA	304	struct timespec now;
	305
	306	if (!fio_fill_issue_time(td))
	307	return;
	308
	309	fio_gettime(&now, NULL);
	310
	311	while (nr--) {
bffad86f	312	struct io_sq_ring *ring = &ld->sq_ring;
9a2d78b3	313	int index = ring->array[start & ld->sq_ring_mask];
f8289afc	314	struct io_u *io_u = ld->io_u_index[index];
52885fa2 JA	315
	316	memcpy(&io_u->issue_time, &now, sizeof(now));
	317	io_u_queued(td, io_u);
	318
	319	start++;
52885fa2 JA	320	}
	321	}
	322
bffad86f	323	static int fio_ioring_commit(struct thread_data *td)
52885fa2	324	{
bffad86f JA	325	struct ioring_data *ld = td->io_ops_data;
bffad86f JA	326	struct ioring_options *o = td->eo;
52885fa2 JA	327	int ret;
	328
	329	if (!ld->queued)
	330	return 0;
	331
3d7d00a3 JA	332	/*
	333	* Kernel side does submission. just need to check if the ring is
	334	* flagged as needing a kick, if so, call io_uring_enter(). This
	335	* only happens if we've been idle too long.
	336	*/
	337	if (o->sqpoll_thread) {
bffad86f	338	struct io_sq_ring *ring = &ld->sq_ring;
4cdbc048	339
2ea53ca3	340	read_barrier();
9a2d78b3 JA	341	if (*ring->flags & IORING_SQ_NEED_WAKEUP)
9a2d78b3 JA	342	io_uring_enter(ld, ld->queued, 0, 0);
771c9901 JA	343	ld->queued = 0;
	344	return 0;
	345	}
	346
52885fa2	347	do {
9a2d78b3	348	unsigned start = *ld->sq_ring.head;
52885fa2 JA	349	long nr = ld->queued;
52885fa2 JA	350
9a2d78b3	351	ret = io_uring_enter(ld, nr, 0, IORING_ENTER_GETEVENTS);
52885fa2	352	if (ret > 0) {
bffad86f	353	fio_ioring_queued(td, start, ret);
52885fa2 JA	354	io_u_mark_submit(td, ret);
	355
	356	ld->queued -= ret;
	357	ret = 0;
a90cd050 JA	358	} else if (!ret) {
a90cd050 JA	359	io_u_mark_submit(td, ret);
52885fa2	360	continue;
a90cd050 JA	361	} else {
a90cd050 JA	362	if (errno == EAGAIN) {
bffad86f	363	ret = fio_ioring_cqring_reap(td, 0, ld->queued);
a90cd050 JA	364	if (ret)
	365	continue;
	366	/* Shouldn't happen */
	367	usleep(1);
	368	continue;
52885fa2	369	}
9a2d78b3	370	td_verror(td, errno, "io_uring_enter submit");
52885fa2	371	break;
a90cd050	372	}
52885fa2 JA	373	} while (ld->queued);
	374
	375	return ret;
	376	}
	377
bffad86f	378	static void fio_ioring_unmap(struct ioring_data *ld)
52885fa2	379	{
9a2d78b3	380	int i;
52885fa2	381
9a2d78b3 JA	382	for (i = 0; i < ARRAY_SIZE(ld->mmap); i++)
	383	munmap(ld->mmap[i].ptr, ld->mmap[i].len);
	384	close(ld->ring_fd);
b87aa01a JA	385	}
b87aa01a JA	386
bffad86f	387	static void fio_ioring_cleanup(struct thread_data *td)
52885fa2	388	{
bffad86f	389	struct ioring_data *ld = td->io_ops_data;
52885fa2 JA	390
52885fa2 JA	391	if (ld) {
96563db9 JA	392	td->ts.cachehit += ld->cachehit;
	393	td->ts.cachemiss += ld->cachemiss;
	394
52885fa2	395	if (!(td->flags & TD_F_CHILD))
bffad86f	396	fio_ioring_unmap(ld);
9a2d78b3	397
52885fa2 JA	398	free(ld->io_u_index);
52885fa2 JA	399	free(ld->io_us);
9a2d78b3	400	free(ld->iovecs);
52885fa2 JA	401	free(ld);
	402	}
	403	}
	404
bffad86f	405	static int fio_ioring_mmap(struct ioring_data ld, struct io_uring_params p)
9a2d78b3	406	{
bffad86f JA	407	struct io_sq_ring *sring = &ld->sq_ring;
bffad86f JA	408	struct io_cq_ring *cring = &ld->cq_ring;
9a2d78b3 JA	409	void *ptr;
9a2d78b3 JA	410
e2239016	411	ld->mmap[0].len = p->sq_off.array + p->sq_entries * sizeof(__u32);
9a2d78b3 JA	412	ptr = mmap(0, ld->mmap[0].len, PROT_READ \| PROT_WRITE,
	413	MAP_SHARED \| MAP_POPULATE, ld->ring_fd,
	414	IORING_OFF_SQ_RING);
	415	ld->mmap[0].ptr = ptr;
	416	sring->head = ptr + p->sq_off.head;
	417	sring->tail = ptr + p->sq_off.tail;
	418	sring->ring_mask = ptr + p->sq_off.ring_mask;
	419	sring->ring_entries = ptr + p->sq_off.ring_entries;
	420	sring->flags = ptr + p->sq_off.flags;
ac122fea	421	sring->array = ptr + p->sq_off.array;
9a2d78b3 JA	422	ld->sq_ring_mask = *sring->ring_mask;
9a2d78b3 JA	423
f0403f94 JA	424	ld->mmap[1].len = p->sq_entries * sizeof(struct io_uring_sqe);
f0403f94 JA	425	ld->sqes = mmap(0, ld->mmap[1].len, PROT_READ \| PROT_WRITE,
9a2d78b3	426	MAP_SHARED \| MAP_POPULATE, ld->ring_fd,
f0403f94 JA	427	IORING_OFF_SQES);
f0403f94 JA	428	ld->mmap[1].ptr = ld->sqes;
9a2d78b3	429
f0403f94 JA	430	ld->mmap[2].len = p->cq_off.cqes +
f0403f94 JA	431	p->cq_entries * sizeof(struct io_uring_cqe);
9a2d78b3 JA	432	ptr = mmap(0, ld->mmap[2].len, PROT_READ \| PROT_WRITE,
	433	MAP_SHARED \| MAP_POPULATE, ld->ring_fd,
	434	IORING_OFF_CQ_RING);
	435	ld->mmap[2].ptr = ptr;
	436	cring->head = ptr + p->cq_off.head;
	437	cring->tail = ptr + p->cq_off.tail;
	438	cring->ring_mask = ptr + p->cq_off.ring_mask;
	439	cring->ring_entries = ptr + p->cq_off.ring_entries;
f0403f94	440	cring->cqes = ptr + p->cq_off.cqes;
9a2d78b3 JA	441	ld->cq_ring_mask = *cring->ring_mask;
	442	return 0;
	443	}
	444
bffad86f	445	static int fio_ioring_queue_init(struct thread_data *td)
52885fa2	446	{
bffad86f JA	447	struct ioring_data *ld = td->io_ops_data;
bffad86f JA	448	struct ioring_options *o = td->eo;
52885fa2	449	int depth = td->o.iodepth;
bffad86f	450	struct io_uring_params p;
9a2d78b3 JA	451	int ret;
	452
	453	memset(&p, 0, sizeof(p));
52885fa2 JA	454
52885fa2 JA	455	if (o->hipri)
bffad86f	456	p.flags \|= IORING_SETUP_IOPOLL;
3d7d00a3 JA	457	if (o->sqpoll_thread) {
	458	p.flags \|= IORING_SETUP_SQPOLL;
	459	if (o->sqpoll_set) {
	460	p.flags \|= IORING_SETUP_SQ_AFF;
	461	p.sq_thread_cpu = o->sqpoll_cpu;
	462	}
f635f1fb	463	}
a90cd050	464
52885fa2 JA	465	if (o->fixedbufs) {
	466	struct rlimit rlim = {
	467	.rlim_cur = RLIM_INFINITY,
	468	.rlim_max = RLIM_INFINITY,
	469	};
	470
	471	setrlimit(RLIMIT_MEMLOCK, &rlim);
52885fa2 JA	472	}
52885fa2 JA	473
2ea53ca3	474	ret = syscall(__NR_sys_io_uring_setup, depth, &p);
9a2d78b3 JA	475	if (ret < 0)
	476	return ret;
	477
	478	ld->ring_fd = ret;
2ea53ca3 JA	479
	480	if (o->fixedbufs) {
	481	struct io_uring_register_buffers reg = {
	482	.iovecs = ld->iovecs,
	483	.nr_iovecs = depth
	484	};
	485
	486	ret = syscall(__NR_sys_io_uring_register, ld->ring_fd,
	487	IORING_REGISTER_BUFFERS, &reg);
	488	if (ret < 0)
	489	return ret;
	490	}
	491
bffad86f	492	return fio_ioring_mmap(ld, &p);
52885fa2 JA	493	}
52885fa2 JA	494
bffad86f	495	static int fio_ioring_post_init(struct thread_data *td)
52885fa2	496	{
bffad86f	497	struct ioring_data *ld = td->io_ops_data;
52885fa2	498	struct io_u *io_u;
650346e1	499	int err, i;
52885fa2	500
650346e1 JA	501	for (i = 0; i < td->o.iodepth; i++) {
650346e1 JA	502	struct iovec *iov = &ld->iovecs[i];
9a2d78b3	503
650346e1 JA	504	io_u = ld->io_u_index[i];
	505	iov->iov_base = io_u->buf;
	506	iov->iov_len = td_max_bs(td);
52885fa2 JA	507	}
52885fa2 JA	508
bffad86f	509	err = fio_ioring_queue_init(td);
52885fa2	510	if (err) {
d63a472d	511	td_verror(td, errno, "io_queue_init");
52885fa2 JA	512	return 1;
	513	}
	514
	515	return 0;
	516	}
	517
9a2d78b3 JA	518	static unsigned roundup_pow2(unsigned depth)
	519	{
	520	return 1UL << __fls(depth - 1);
	521	}
	522
bffad86f	523	static int fio_ioring_init(struct thread_data *td)
52885fa2	524	{
bffad86f	525	struct ioring_data *ld;
52885fa2	526
52885fa2 JA	527	ld = calloc(1, sizeof(*ld));
52885fa2 JA	528
b87aa01a JA	529	/* ring depth must be a power-of-2 */
	530	ld->iodepth = td->o.iodepth;
	531	td->o.iodepth = roundup_pow2(td->o.iodepth);
	532
52885fa2 JA	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 *));
650346e1	536	ld->iovecs = calloc(td->o.iodepth, sizeof(struct iovec));
52885fa2 JA	537
	538	td->io_ops_data = ld;
	539	return 0;
	540	}
	541
bffad86f	542	static int fio_ioring_io_u_init(struct thread_data td, struct io_u io_u)
52885fa2	543	{
bffad86f	544	struct ioring_data *ld = td->io_ops_data;
52885fa2 JA	545
	546	ld->io_u_index[io_u->index] = io_u;
	547	return 0;
	548	}
	549
	550	static struct ioengine_ops ioengine = {
bffad86f	551	.name = "io_uring",
52885fa2	552	.version = FIO_IOOPS_VERSION,
bffad86f JA	553	.init = fio_ioring_init,
	554	.post_init = fio_ioring_post_init,
	555	.io_u_init = fio_ioring_io_u_init,
	556	.prep = fio_ioring_prep,
	557	.queue = fio_ioring_queue,
	558	.commit = fio_ioring_commit,
	559	.getevents = fio_ioring_getevents,
	560	.event = fio_ioring_event,
	561	.cleanup = fio_ioring_cleanup,
52885fa2 JA	562	.open_file = generic_open_file,
	563	.close_file = generic_close_file,
	564	.get_file_size = generic_get_file_size,
	565	.options = options,
bffad86f	566	.option_struct_size = sizeof(struct ioring_options),
52885fa2 JA	567	};
52885fa2 JA	568
bffad86f	569	static void fio_init fio_ioring_register(void)
52885fa2	570	{
52885fa2	571	register_ioengine(&ioengine);
52885fa2 JA	572	}
52885fa2 JA	573
bffad86f	574	static void fio_exit fio_ioring_unregister(void)
52885fa2	575	{
52885fa2	576	unregister_ioengine(&ioengine);
52885fa2	577	}
1f90e9bb	578	#endif