[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"
#include "../lib/roundup.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;

	int *fds;

	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;
	bool ioprio_class_set;
	bool ioprio_set;

	struct ioring_mmap mmap[3];
};

struct ioring_options {
	void *pad;
	unsigned int hipri;
	unsigned int cmdprio_percentage;
	unsigned int fixedbufs;
	unsigned int registerfiles;
	unsigned int sqpoll_thread;
	unsigned int sqpoll_set;
	unsigned int sqpoll_cpu;
	unsigned int nonvectored;
	unsigned int uncached;
	unsigned int nowait;
};

static const int ddir_to_op[2][2] = {
	{ IORING_OP_READV, IORING_OP_READ },
	{ IORING_OP_WRITEV, IORING_OP_WRITE }
};

static const int fixed_ddir_to_op[2] = {
	IORING_OP_READ_FIXED,
	IORING_OP_WRITE_FIXED
};

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_IOURING,
	},
#ifdef FIO_HAVE_IOPRIO_CLASS
	{
		.name	= "cmdprio_percentage",
		.lname	= "high priority percentage",
		.type	= FIO_OPT_INT,
		.off1	= offsetof(struct ioring_options, cmdprio_percentage),
		.minval	= 1,
		.maxval	= 100,
		.help	= "Send high priority I/O this percentage of the time",
		.category = FIO_OPT_C_ENGINE,
		.group	= FIO_OPT_G_IOURING,
	},
#else
	{
		.name	= "cmdprio_percentage",
		.lname	= "high priority percentage",
		.type	= FIO_OPT_UNSUPPORTED,
		.help	= "Your platform does not support I/O priority classes",
	},
#endif
	{
		.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_IOURING,
	},
	{
		.name	= "registerfiles",
		.lname	= "Register file set",
		.type	= FIO_OPT_STR_SET,
		.off1	= offsetof(struct ioring_options, registerfiles),
		.help	= "Pre-open/register files",
		.category = FIO_OPT_C_ENGINE,
		.group	= FIO_OPT_G_IOURING,
	},
	{
		.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_IOURING,
	},
	{
		.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_IOURING,
	},
	{
		.name	= "nonvectored",
		.lname	= "Non-vectored",
		.type	= FIO_OPT_INT,
		.off1	= offsetof(struct ioring_options, nonvectored),
		.help	= "Use non-vectored read/write commands",
		.category = FIO_OPT_C_ENGINE,
		.group	= FIO_OPT_G_IOURING,
	},
	{
		.name	= "uncached",
		.lname	= "Uncached",
		.type	= FIO_OPT_INT,
		.off1	= offsetof(struct ioring_options, uncached),
		.help	= "Use RWF_UNCACHED for buffered read/writes",
		.category = FIO_OPT_C_ENGINE,
		.group	= FIO_OPT_G_IOURING,
	},
	{
		.name	= "nowait",
		.lname	= "RWF_NOWAIT",
		.type	= FIO_OPT_BOOL,
		.off1	= offsetof(struct ioring_options, nowait),
		.help	= "Use RWF_NOWAIT for reads/writes",
		.category = FIO_OPT_C_ENGINE,
		.group	= FIO_OPT_G_IOURING,
	},
	{
		.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_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];

	/* zero out fields not used in this submission */
	memset(sqe, 0, sizeof(*sqe));

	if (o->registerfiles) {
		sqe->fd = f->engine_pos;
		sqe->flags = IOSQE_FIXED_FILE;
	} else {
		sqe->fd = f->fd;
	}

	if (io_u->ddir == DDIR_READ || io_u->ddir == DDIR_WRITE) {
		if (o->fixedbufs) {
			sqe->opcode = fixed_ddir_to_op[io_u->ddir];
			sqe->addr = (unsigned long) io_u->xfer_buf;
			sqe->len = io_u->xfer_buflen;
			sqe->buf_index = io_u->index;
		} else {
			sqe->opcode = ddir_to_op[io_u->ddir][!!o->nonvectored];
			if (o->nonvectored) {
				sqe->addr = (unsigned long)
						ld->iovecs[io_u->index].iov_base;
				sqe->len = ld->iovecs[io_u->index].iov_len;
			} else {
				sqe->addr = (unsigned long) &ld->iovecs[io_u->index];
				sqe->len = 1;
			}
		}
		if (!td->o.odirect && o->uncached)
			sqe->rw_flags = RWF_UNCACHED;
		if (o->nowait)
			sqe->rw_flags |= RWF_NOWAIT;
		if (ld->ioprio_class_set)
			sqe->ioprio = td->o.ioprio_class << 13;
		if (ld->ioprio_set)
			sqe->ioprio |= td->o.ioprio;
		sqe->off = io_u->offset;
	} else if (ddir_sync(io_u->ddir)) {
		if (io_u->ddir == DDIR_SYNC_FILE_RANGE) {
			sqe->off = f->first_write;
			sqe->len = f->last_write - f->first_write;
			sqe->sync_range_flags = td->o.sync_file_range;
			sqe->opcode = IORING_OP_SYNC_FILE_RANGE;
		} else {
			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 {
		if (head == atomic_load_acquire(ring->tail))
			break;
		reaped++;
		head++;
	} while (reaped + events < max);

	if (reaped)
		atomic_store_release(ring->head, head);

	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;
			if (actual_min != 0)
				actual_min -= r;
			continue;
		}

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

	return r < 0 ? r : events;
}

static void fio_ioring_prio_prep(struct thread_data *td, struct io_u *io_u)
{
	struct ioring_options *o = td->eo;
	struct ioring_data *ld = td->io_ops_data;
	if (rand_between(&td->prio_state, 0, 99) < o->cmdprio_percentage) {
		ld->sqes[io_u->index].ioprio = IOPRIO_CLASS_RT << IOPRIO_CLASS_SHIFT;
		io_u->flags |= IO_U_F_PRIORITY;
	}
	return;
}

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;
	struct ioring_options *o = td->eo;
	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;
	if (next_tail == atomic_load_acquire(ring->head))
		return FIO_Q_BUSY;

	if (o->cmdprio_percentage)
		fio_ioring_prio_prep(td, io_u);
	ring->array[tail & ld->sq_ring_mask] = io_u->index;
	atomic_store_release(ring->tail, next_tail);

	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 || errno == EINTR) {
				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->fds);
		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_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_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_register_files(struct thread_data *td)
{
	struct ioring_data *ld = td->io_ops_data;
	struct fio_file *f;
	unsigned int i;
	int ret;

	ld->fds = calloc(td->o.nr_files, sizeof(int));

	for_each_file(td, f, i) {
		ret = generic_open_file(td, f);
		if (ret)
			goto err;
		ld->fds[i] = f->fd;
		f->engine_pos = i;
	}

	ret = syscall(__NR_io_uring_register, ld->ring_fd,
			IORING_REGISTER_FILES, ld->fds, td->o.nr_files);
	if (ret) {
err:
		free(ld->fds);
		ld->fds = NULL;
	}

	/*
	 * Pretend the file is closed again, and really close it if we hit
	 * an error.
	 */
	for_each_file(td, f, i) {
		if (ret) {
			int fio_unused ret2;
			ret2 = generic_close_file(td, f);
		} else
			f->fd = -1;
	}

	return ret;
}

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

	if (o->registerfiles) {
		err = fio_ioring_register_files(td);
		if (err) {
			td_verror(td, errno, "ioring_register_files");
			return 1;
		}
	}

	return 0;
}

static int fio_ioring_init(struct thread_data *td)
{
	struct ioring_options *o = td->eo;
	struct ioring_data *ld;
	struct thread_options *to = &td->o;

	/* sqthread submission requires registered files */
	if (o->sqpoll_thread)
		o->registerfiles = 1;

	if (o->registerfiles && td->o.nr_files != td->o.open_files) {
		log_err("fio: io_uring registered files require nr_files to "
			"be identical to open_files\n");
		return 1;
	}

	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;

	/*
	 * Check for option conflicts
	 */
	if ((fio_option_is_set(to, ioprio) || fio_option_is_set(to, ioprio_class)) &&
			o->cmdprio_percentage != 0) {
		log_err("%s: cmdprio_percentage option and mutually exclusive "
				"prio or prioclass option is set, exiting\n", to->name);
		td_verror(td, EINVAL, "fio_io_uring_init");
		return 1;
	}

	if (fio_option_is_set(&td->o, ioprio_class))
		ld->ioprio_class_set = true;
	if (fio_option_is_set(&td->o, ioprio))
		ld->ioprio_set = true;

	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 int fio_ioring_open_file(struct thread_data *td, struct fio_file *f)
{
	struct ioring_data *ld = td->io_ops_data;
	struct ioring_options *o = td->eo;

	if (!ld || !o->registerfiles)
		return generic_open_file(td, f);

	f->fd = ld->fds[f->engine_pos];
	return 0;
}

static int fio_ioring_close_file(struct thread_data *td, struct fio_file *f)
{
	struct ioring_data *ld = td->io_ops_data;
	struct ioring_options *o = td->eo;

	if (!ld || !o->registerfiles)
		return generic_close_file(td, f);

	f->fd = -1;
	return 0;
}

static struct ioengine_ops ioengine = {
	.name			= "io_uring",
	.version		= FIO_IOOPS_VERSION,
	.flags			= FIO_ASYNCIO_SYNC_TRIM,
	.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		= fio_ioring_open_file,
	.close_file		= fio_ioring_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	#include "../lib/roundup.h"
	21
	22	#ifdef ARCH_HAVE_IOURING
	23
	24	#include "../lib/types.h"
	25	#include "../os/linux/io_uring.h"
	26
	27	struct io_sq_ring {
	28	unsigned *head;
	29	unsigned *tail;
	30	unsigned *ring_mask;
	31	unsigned *ring_entries;
	32	unsigned *flags;
	33	unsigned *array;
	34	};
	35
	36	struct io_cq_ring {
	37	unsigned *head;
	38	unsigned *tail;
	39	unsigned *ring_mask;
	40	unsigned *ring_entries;
	41	struct io_uring_cqe *cqes;
	42	};
	43
	44	struct ioring_mmap {
	45	void *ptr;
	46	size_t len;
	47	};
	48
	49	struct ioring_data {
	50	int ring_fd;
	51
	52	struct io_u **io_u_index;
	53
	54	int *fds;
	55
	56	struct io_sq_ring sq_ring;
	57	struct io_uring_sqe *sqes;
	58	struct iovec *iovecs;
	59	unsigned sq_ring_mask;
	60
	61	struct io_cq_ring cq_ring;
	62	unsigned cq_ring_mask;
	63
	64	int queued;
	65	int cq_ring_off;
	66	unsigned iodepth;
	67	bool ioprio_class_set;
	68	bool ioprio_set;
	69
	70	struct ioring_mmap mmap[3];
	71	};
	72
	73	struct ioring_options {
	74	void *pad;
	75	unsigned int hipri;
	76	unsigned int cmdprio_percentage;
	77	unsigned int fixedbufs;
	78	unsigned int registerfiles;
	79	unsigned int sqpoll_thread;
	80	unsigned int sqpoll_set;
	81	unsigned int sqpoll_cpu;
	82	unsigned int nonvectored;
	83	unsigned int uncached;
	84	unsigned int nowait;
	85	};
	86
	87	static const int ddir_to_op[2][2] = {
	88	{ IORING_OP_READV, IORING_OP_READ },
	89	{ IORING_OP_WRITEV, IORING_OP_WRITE }
	90	};
	91
	92	static const int fixed_ddir_to_op[2] = {
	93	IORING_OP_READ_FIXED,
	94	IORING_OP_WRITE_FIXED
	95	};
	96
	97	static int fio_ioring_sqpoll_cb(void data, unsigned long long val)
	98	{
	99	struct ioring_options *o = data;
	100
	101	o->sqpoll_cpu = *val;
	102	o->sqpoll_set = 1;
	103	return 0;
	104	}
	105
	106	static struct fio_option options[] = {
	107	{
	108	.name = "hipri",
	109	.lname = "High Priority",
	110	.type = FIO_OPT_STR_SET,
	111	.off1 = offsetof(struct ioring_options, hipri),
	112	.help = "Use polled IO completions",
	113	.category = FIO_OPT_C_ENGINE,
	114	.group = FIO_OPT_G_IOURING,
	115	},
	116	#ifdef FIO_HAVE_IOPRIO_CLASS
	117	{
	118	.name = "cmdprio_percentage",
	119	.lname = "high priority percentage",
	120	.type = FIO_OPT_INT,
	121	.off1 = offsetof(struct ioring_options, cmdprio_percentage),
	122	.minval = 1,
	123	.maxval = 100,
	124	.help = "Send high priority I/O this percentage of the time",
	125	.category = FIO_OPT_C_ENGINE,
	126	.group = FIO_OPT_G_IOURING,
	127	},
	128	#else
	129	{
	130	.name = "cmdprio_percentage",
	131	.lname = "high priority percentage",
	132	.type = FIO_OPT_UNSUPPORTED,
	133	.help = "Your platform does not support I/O priority classes",
	134	},
	135	#endif
	136	{
	137	.name = "fixedbufs",
	138	.lname = "Fixed (pre-mapped) IO buffers",
	139	.type = FIO_OPT_STR_SET,
	140	.off1 = offsetof(struct ioring_options, fixedbufs),
	141	.help = "Pre map IO buffers",
	142	.category = FIO_OPT_C_ENGINE,
	143	.group = FIO_OPT_G_IOURING,
	144	},
	145	{
	146	.name = "registerfiles",
	147	.lname = "Register file set",
	148	.type = FIO_OPT_STR_SET,
	149	.off1 = offsetof(struct ioring_options, registerfiles),
	150	.help = "Pre-open/register files",
	151	.category = FIO_OPT_C_ENGINE,
	152	.group = FIO_OPT_G_IOURING,
	153	},
	154	{
	155	.name = "sqthread_poll",
	156	.lname = "Kernel SQ thread polling",
	157	.type = FIO_OPT_INT,
	158	.off1 = offsetof(struct ioring_options, sqpoll_thread),
	159	.help = "Offload submission/completion to kernel thread",
	160	.category = FIO_OPT_C_ENGINE,
	161	.group = FIO_OPT_G_IOURING,
	162	},
	163	{
	164	.name = "sqthread_poll_cpu",
	165	.lname = "SQ Thread Poll CPU",
	166	.type = FIO_OPT_INT,
	167	.cb = fio_ioring_sqpoll_cb,
	168	.help = "What CPU to run SQ thread polling on",
	169	.category = FIO_OPT_C_ENGINE,
	170	.group = FIO_OPT_G_IOURING,
	171	},
	172	{
	173	.name = "nonvectored",
	174	.lname = "Non-vectored",
	175	.type = FIO_OPT_INT,
	176	.off1 = offsetof(struct ioring_options, nonvectored),
	177	.help = "Use non-vectored read/write commands",
	178	.category = FIO_OPT_C_ENGINE,
	179	.group = FIO_OPT_G_IOURING,
	180	},
	181	{
	182	.name = "uncached",
	183	.lname = "Uncached",
	184	.type = FIO_OPT_INT,
	185	.off1 = offsetof(struct ioring_options, uncached),
	186	.help = "Use RWF_UNCACHED for buffered read/writes",
	187	.category = FIO_OPT_C_ENGINE,
	188	.group = FIO_OPT_G_IOURING,
	189	},
	190	{
	191	.name = "nowait",
	192	.lname = "RWF_NOWAIT",
	193	.type = FIO_OPT_BOOL,
	194	.off1 = offsetof(struct ioring_options, nowait),
	195	.help = "Use RWF_NOWAIT for reads/writes",
	196	.category = FIO_OPT_C_ENGINE,
	197	.group = FIO_OPT_G_IOURING,
	198	},
	199	{
	200	.name = NULL,
	201	},
	202	};
	203
	204	static int io_uring_enter(struct ioring_data *ld, unsigned int to_submit,
	205	unsigned int min_complete, unsigned int flags)
	206	{
	207	return syscall(__NR_io_uring_enter, ld->ring_fd, to_submit,
	208	min_complete, flags, NULL, 0);
	209	}
	210
	211	static int fio_ioring_prep(struct thread_data td, struct io_u io_u)
	212	{
	213	struct ioring_data *ld = td->io_ops_data;
	214	struct ioring_options *o = td->eo;
	215	struct fio_file *f = io_u->file;
	216	struct io_uring_sqe *sqe;
	217
	218	sqe = &ld->sqes[io_u->index];
	219
	220	/* zero out fields not used in this submission */
	221	memset(sqe, 0, sizeof(*sqe));
	222
	223	if (o->registerfiles) {
	224	sqe->fd = f->engine_pos;
	225	sqe->flags = IOSQE_FIXED_FILE;
	226	} else {
	227	sqe->fd = f->fd;
	228	}
	229
	230	if (io_u->ddir == DDIR_READ \|\| io_u->ddir == DDIR_WRITE) {
	231	if (o->fixedbufs) {
	232	sqe->opcode = fixed_ddir_to_op[io_u->ddir];
	233	sqe->addr = (unsigned long) io_u->xfer_buf;
	234	sqe->len = io_u->xfer_buflen;
	235	sqe->buf_index = io_u->index;
	236	} else {
	237	sqe->opcode = ddir_to_op[io_u->ddir][!!o->nonvectored];
	238	if (o->nonvectored) {
	239	sqe->addr = (unsigned long)
	240	ld->iovecs[io_u->index].iov_base;
	241	sqe->len = ld->iovecs[io_u->index].iov_len;
	242	} else {
	243	sqe->addr = (unsigned long) &ld->iovecs[io_u->index];
	244	sqe->len = 1;
	245	}
	246	}
	247	if (!td->o.odirect && o->uncached)
	248	sqe->rw_flags = RWF_UNCACHED;
	249	if (o->nowait)
	250	sqe->rw_flags \|= RWF_NOWAIT;
	251	if (ld->ioprio_class_set)
	252	sqe->ioprio = td->o.ioprio_class << 13;
	253	if (ld->ioprio_set)
	254	sqe->ioprio \|= td->o.ioprio;
	255	sqe->off = io_u->offset;
	256	} else if (ddir_sync(io_u->ddir)) {
	257	if (io_u->ddir == DDIR_SYNC_FILE_RANGE) {
	258	sqe->off = f->first_write;
	259	sqe->len = f->last_write - f->first_write;
	260	sqe->sync_range_flags = td->o.sync_file_range;
	261	sqe->opcode = IORING_OP_SYNC_FILE_RANGE;
	262	} else {
	263	if (io_u->ddir == DDIR_DATASYNC)
	264	sqe->fsync_flags \|= IORING_FSYNC_DATASYNC;
	265	sqe->opcode = IORING_OP_FSYNC;
	266	}
	267	}
	268
	269	sqe->user_data = (unsigned long) io_u;
	270	return 0;
	271	}
	272
	273	static struct io_u fio_ioring_event(struct thread_data td, int event)
	274	{
	275	struct ioring_data *ld = td->io_ops_data;
	276	struct io_uring_cqe *cqe;
	277	struct io_u *io_u;
	278	unsigned index;
	279
	280	index = (event + ld->cq_ring_off) & ld->cq_ring_mask;
	281
	282	cqe = &ld->cq_ring.cqes[index];
	283	io_u = (struct io_u *) (uintptr_t) cqe->user_data;
	284
	285	if (cqe->res != io_u->xfer_buflen) {
	286	if (cqe->res > io_u->xfer_buflen)
	287	io_u->error = -cqe->res;
	288	else
	289	io_u->resid = io_u->xfer_buflen - cqe->res;
	290	} else
	291	io_u->error = 0;
	292
	293	return io_u;
	294	}
	295
	296	static int fio_ioring_cqring_reap(struct thread_data *td, unsigned int events,
	297	unsigned int max)
	298	{
	299	struct ioring_data *ld = td->io_ops_data;
	300	struct io_cq_ring *ring = &ld->cq_ring;
	301	unsigned head, reaped = 0;
	302
	303	head = *ring->head;
	304	do {
	305	if (head == atomic_load_acquire(ring->tail))
	306	break;
	307	reaped++;
	308	head++;
	309	} while (reaped + events < max);
	310
	311	if (reaped)
	312	atomic_store_release(ring->head, head);
	313
	314	return reaped;
	315	}
	316
	317	static int fio_ioring_getevents(struct thread_data *td, unsigned int min,
	318	unsigned int max, const struct timespec *t)
	319	{
	320	struct ioring_data *ld = td->io_ops_data;
	321	unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
	322	struct ioring_options *o = td->eo;
	323	struct io_cq_ring *ring = &ld->cq_ring;
	324	unsigned events = 0;
	325	int r;
	326
	327	ld->cq_ring_off = *ring->head;
	328	do {
	329	r = fio_ioring_cqring_reap(td, events, max);
	330	if (r) {
	331	events += r;
	332	if (actual_min != 0)
	333	actual_min -= r;
	334	continue;
	335	}
	336
	337	if (!o->sqpoll_thread) {
	338	r = io_uring_enter(ld, 0, actual_min,
	339	IORING_ENTER_GETEVENTS);
	340	if (r < 0) {
	341	if (errno == EAGAIN \|\| errno == EINTR)
	342	continue;
	343	td_verror(td, errno, "io_uring_enter");
	344	break;
	345	}
	346	}
	347	} while (events < min);
	348
	349	return r < 0 ? r : events;
	350	}
	351
	352	static void fio_ioring_prio_prep(struct thread_data td, struct io_u io_u)
	353	{
	354	struct ioring_options *o = td->eo;
	355	struct ioring_data *ld = td->io_ops_data;
	356	if (rand_between(&td->prio_state, 0, 99) < o->cmdprio_percentage) {
	357	ld->sqes[io_u->index].ioprio = IOPRIO_CLASS_RT << IOPRIO_CLASS_SHIFT;
	358	io_u->flags \|= IO_U_F_PRIORITY;
	359	}
	360	return;
	361	}
	362
	363	static enum fio_q_status fio_ioring_queue(struct thread_data *td,
	364	struct io_u *io_u)
	365	{
	366	struct ioring_data *ld = td->io_ops_data;
	367	struct io_sq_ring *ring = &ld->sq_ring;
	368	struct ioring_options *o = td->eo;
	369	unsigned tail, next_tail;
	370
	371	fio_ro_check(td, io_u);
	372
	373	if (ld->queued == ld->iodepth)
	374	return FIO_Q_BUSY;
	375
	376	if (io_u->ddir == DDIR_TRIM) {
	377	if (ld->queued)
	378	return FIO_Q_BUSY;
	379
	380	do_io_u_trim(td, io_u);
	381	io_u_mark_submit(td, 1);
	382	io_u_mark_complete(td, 1);
	383	return FIO_Q_COMPLETED;
	384	}
	385
	386	tail = *ring->tail;
	387	next_tail = tail + 1;
	388	if (next_tail == atomic_load_acquire(ring->head))
	389	return FIO_Q_BUSY;
	390
	391	if (o->cmdprio_percentage)
	392	fio_ioring_prio_prep(td, io_u);
	393	ring->array[tail & ld->sq_ring_mask] = io_u->index;
	394	atomic_store_release(ring->tail, next_tail);
	395
	396	ld->queued++;
	397	return FIO_Q_QUEUED;
	398	}
	399
	400	static void fio_ioring_queued(struct thread_data *td, int start, int nr)
	401	{
	402	struct ioring_data *ld = td->io_ops_data;
	403	struct timespec now;
	404
	405	if (!fio_fill_issue_time(td))
	406	return;
	407
	408	fio_gettime(&now, NULL);
	409
	410	while (nr--) {
	411	struct io_sq_ring *ring = &ld->sq_ring;
	412	int index = ring->array[start & ld->sq_ring_mask];
	413	struct io_u *io_u = ld->io_u_index[index];
	414
	415	memcpy(&io_u->issue_time, &now, sizeof(now));
	416	io_u_queued(td, io_u);
	417
	418	start++;
	419	}
	420	}
	421
	422	static int fio_ioring_commit(struct thread_data *td)
	423	{
	424	struct ioring_data *ld = td->io_ops_data;
	425	struct ioring_options *o = td->eo;
	426	int ret;
	427
	428	if (!ld->queued)
	429	return 0;
	430
	431	/*
	432	* Kernel side does submission. just need to check if the ring is
	433	* flagged as needing a kick, if so, call io_uring_enter(). This
	434	* only happens if we've been idle too long.
	435	*/
	436	if (o->sqpoll_thread) {
	437	struct io_sq_ring *ring = &ld->sq_ring;
	438
	439	read_barrier();
	440	if (*ring->flags & IORING_SQ_NEED_WAKEUP)
	441	io_uring_enter(ld, ld->queued, 0,
	442	IORING_ENTER_SQ_WAKEUP);
	443	ld->queued = 0;
	444	return 0;
	445	}
	446
	447	do {
	448	unsigned start = *ld->sq_ring.head;
	449	long nr = ld->queued;
	450
	451	ret = io_uring_enter(ld, nr, 0, IORING_ENTER_GETEVENTS);
	452	if (ret > 0) {
	453	fio_ioring_queued(td, start, ret);
	454	io_u_mark_submit(td, ret);
	455
	456	ld->queued -= ret;
	457	ret = 0;
	458	} else if (!ret) {
	459	io_u_mark_submit(td, ret);
	460	continue;
	461	} else {
	462	if (errno == EAGAIN \|\| errno == EINTR) {
	463	ret = fio_ioring_cqring_reap(td, 0, ld->queued);
	464	if (ret)
	465	continue;
	466	/* Shouldn't happen */
	467	usleep(1);
	468	continue;
	469	}
	470	td_verror(td, errno, "io_uring_enter submit");
	471	break;
	472	}
	473	} while (ld->queued);
	474
	475	return ret;
	476	}
	477
	478	static void fio_ioring_unmap(struct ioring_data *ld)
	479	{
	480	int i;
	481
	482	for (i = 0; i < ARRAY_SIZE(ld->mmap); i++)
	483	munmap(ld->mmap[i].ptr, ld->mmap[i].len);
	484	close(ld->ring_fd);
	485	}
	486
	487	static void fio_ioring_cleanup(struct thread_data *td)
	488	{
	489	struct ioring_data *ld = td->io_ops_data;
	490
	491	if (ld) {
	492	if (!(td->flags & TD_F_CHILD))
	493	fio_ioring_unmap(ld);
	494
	495	free(ld->io_u_index);
	496	free(ld->iovecs);
	497	free(ld->fds);
	498	free(ld);
	499	}
	500	}
	501
	502	static int fio_ioring_mmap(struct ioring_data ld, struct io_uring_params p)
	503	{
	504	struct io_sq_ring *sring = &ld->sq_ring;
	505	struct io_cq_ring *cring = &ld->cq_ring;
	506	void *ptr;
	507
	508	ld->mmap[0].len = p->sq_off.array + p->sq_entries * sizeof(__u32);
	509	ptr = mmap(0, ld->mmap[0].len, PROT_READ \| PROT_WRITE,
	510	MAP_SHARED \| MAP_POPULATE, ld->ring_fd,
	511	IORING_OFF_SQ_RING);
	512	ld->mmap[0].ptr = ptr;
	513	sring->head = ptr + p->sq_off.head;
	514	sring->tail = ptr + p->sq_off.tail;
	515	sring->ring_mask = ptr + p->sq_off.ring_mask;
	516	sring->ring_entries = ptr + p->sq_off.ring_entries;
	517	sring->flags = ptr + p->sq_off.flags;
	518	sring->array = ptr + p->sq_off.array;
	519	ld->sq_ring_mask = *sring->ring_mask;
	520
	521	ld->mmap[1].len = p->sq_entries * sizeof(struct io_uring_sqe);
	522	ld->sqes = mmap(0, ld->mmap[1].len, PROT_READ \| PROT_WRITE,
	523	MAP_SHARED \| MAP_POPULATE, ld->ring_fd,
	524	IORING_OFF_SQES);
	525	ld->mmap[1].ptr = ld->sqes;
	526
	527	ld->mmap[2].len = p->cq_off.cqes +
	528	p->cq_entries * sizeof(struct io_uring_cqe);
	529	ptr = mmap(0, ld->mmap[2].len, PROT_READ \| PROT_WRITE,
	530	MAP_SHARED \| MAP_POPULATE, ld->ring_fd,
	531	IORING_OFF_CQ_RING);
	532	ld->mmap[2].ptr = ptr;
	533	cring->head = ptr + p->cq_off.head;
	534	cring->tail = ptr + p->cq_off.tail;
	535	cring->ring_mask = ptr + p->cq_off.ring_mask;
	536	cring->ring_entries = ptr + p->cq_off.ring_entries;
	537	cring->cqes = ptr + p->cq_off.cqes;
	538	ld->cq_ring_mask = *cring->ring_mask;
	539	return 0;
	540	}
	541
	542	static int fio_ioring_queue_init(struct thread_data *td)
	543	{
	544	struct ioring_data *ld = td->io_ops_data;
	545	struct ioring_options *o = td->eo;
	546	int depth = td->o.iodepth;
	547	struct io_uring_params p;
	548	int ret;
	549
	550	memset(&p, 0, sizeof(p));
	551
	552	if (o->hipri)
	553	p.flags \|= IORING_SETUP_IOPOLL;
	554	if (o->sqpoll_thread) {
	555	p.flags \|= IORING_SETUP_SQPOLL;
	556	if (o->sqpoll_set) {
	557	p.flags \|= IORING_SETUP_SQ_AFF;
	558	p.sq_thread_cpu = o->sqpoll_cpu;
	559	}
	560	}
	561
	562	ret = syscall(__NR_io_uring_setup, depth, &p);
	563	if (ret < 0)
	564	return ret;
	565
	566	ld->ring_fd = ret;
	567
	568	if (o->fixedbufs) {
	569	struct rlimit rlim = {
	570	.rlim_cur = RLIM_INFINITY,
	571	.rlim_max = RLIM_INFINITY,
	572	};
	573
	574	if (setrlimit(RLIMIT_MEMLOCK, &rlim) < 0)
	575	return -1;
	576
	577	ret = syscall(__NR_io_uring_register, ld->ring_fd,
	578	IORING_REGISTER_BUFFERS, ld->iovecs, depth);
	579	if (ret < 0)
	580	return ret;
	581	}
	582
	583	return fio_ioring_mmap(ld, &p);
	584	}
	585
	586	static int fio_ioring_register_files(struct thread_data *td)
	587	{
	588	struct ioring_data *ld = td->io_ops_data;
	589	struct fio_file *f;
	590	unsigned int i;
	591	int ret;
	592
	593	ld->fds = calloc(td->o.nr_files, sizeof(int));
	594
	595	for_each_file(td, f, i) {
	596	ret = generic_open_file(td, f);
	597	if (ret)
	598	goto err;
	599	ld->fds[i] = f->fd;
	600	f->engine_pos = i;
	601	}
	602
	603	ret = syscall(__NR_io_uring_register, ld->ring_fd,
	604	IORING_REGISTER_FILES, ld->fds, td->o.nr_files);
	605	if (ret) {
	606	err:
	607	free(ld->fds);
	608	ld->fds = NULL;
	609	}
	610
	611	/*
	612	* Pretend the file is closed again, and really close it if we hit
	613	* an error.
	614	*/
	615	for_each_file(td, f, i) {
	616	if (ret) {
	617	int fio_unused ret2;
	618	ret2 = generic_close_file(td, f);
	619	} else
	620	f->fd = -1;
	621	}
	622
	623	return ret;
	624	}
	625
	626	static int fio_ioring_post_init(struct thread_data *td)
	627	{
	628	struct ioring_data *ld = td->io_ops_data;
	629	struct ioring_options *o = td->eo;
	630	struct io_u *io_u;
	631	int err, i;
	632
	633	for (i = 0; i < td->o.iodepth; i++) {
	634	struct iovec *iov = &ld->iovecs[i];
	635
	636	io_u = ld->io_u_index[i];
	637	iov->iov_base = io_u->buf;
	638	iov->iov_len = td_max_bs(td);
	639	}
	640
	641	err = fio_ioring_queue_init(td);
	642	if (err) {
	643	td_verror(td, errno, "io_queue_init");
	644	return 1;
	645	}
	646
	647	if (o->registerfiles) {
	648	err = fio_ioring_register_files(td);
	649	if (err) {
	650	td_verror(td, errno, "ioring_register_files");
	651	return 1;
	652	}
	653	}
	654
	655	return 0;
	656	}
	657
	658	static int fio_ioring_init(struct thread_data *td)
	659	{
	660	struct ioring_options *o = td->eo;
	661	struct ioring_data *ld;
	662	struct thread_options *to = &td->o;
	663
	664	/* sqthread submission requires registered files */
	665	if (o->sqpoll_thread)
	666	o->registerfiles = 1;
	667
	668	if (o->registerfiles && td->o.nr_files != td->o.open_files) {
	669	log_err("fio: io_uring registered files require nr_files to "
	670	"be identical to open_files\n");
	671	return 1;
	672	}
	673
	674	ld = calloc(1, sizeof(*ld));
	675
	676	/* ring depth must be a power-of-2 */
	677	ld->iodepth = td->o.iodepth;
	678	td->o.iodepth = roundup_pow2(td->o.iodepth);
	679
	680	/* io_u index */
	681	ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *));
	682	ld->iovecs = calloc(td->o.iodepth, sizeof(struct iovec));
	683
	684	td->io_ops_data = ld;
	685
	686	/*
	687	* Check for option conflicts
	688	*/
	689	if ((fio_option_is_set(to, ioprio) \|\| fio_option_is_set(to, ioprio_class)) &&
	690	o->cmdprio_percentage != 0) {
	691	log_err("%s: cmdprio_percentage option and mutually exclusive "
	692	"prio or prioclass option is set, exiting\n", to->name);
	693	td_verror(td, EINVAL, "fio_io_uring_init");
	694	return 1;
	695	}
	696
	697	if (fio_option_is_set(&td->o, ioprio_class))
	698	ld->ioprio_class_set = true;
	699	if (fio_option_is_set(&td->o, ioprio))
	700	ld->ioprio_set = true;
	701
	702	return 0;
	703	}
	704
	705	static int fio_ioring_io_u_init(struct thread_data td, struct io_u io_u)
	706	{
	707	struct ioring_data *ld = td->io_ops_data;
	708
	709	ld->io_u_index[io_u->index] = io_u;
	710	return 0;
	711	}
	712
	713	static int fio_ioring_open_file(struct thread_data td, struct fio_file f)
	714	{
	715	struct ioring_data *ld = td->io_ops_data;
	716	struct ioring_options *o = td->eo;
	717
	718	if (!ld \|\| !o->registerfiles)
	719	return generic_open_file(td, f);
	720
	721	f->fd = ld->fds[f->engine_pos];
	722	return 0;
	723	}
	724
	725	static int fio_ioring_close_file(struct thread_data td, struct fio_file f)
	726	{
	727	struct ioring_data *ld = td->io_ops_data;
	728	struct ioring_options *o = td->eo;
	729
	730	if (!ld \|\| !o->registerfiles)
	731	return generic_close_file(td, f);
	732
	733	f->fd = -1;
	734	return 0;
	735	}
	736
	737	static struct ioengine_ops ioengine = {
	738	.name = "io_uring",
	739	.version = FIO_IOOPS_VERSION,
	740	.flags = FIO_ASYNCIO_SYNC_TRIM,
	741	.init = fio_ioring_init,
	742	.post_init = fio_ioring_post_init,
	743	.io_u_init = fio_ioring_io_u_init,
	744	.prep = fio_ioring_prep,
	745	.queue = fio_ioring_queue,
	746	.commit = fio_ioring_commit,
	747	.getevents = fio_ioring_getevents,
	748	.event = fio_ioring_event,
	749	.cleanup = fio_ioring_cleanup,
	750	.open_file = fio_ioring_open_file,
	751	.close_file = fio_ioring_close_file,
	752	.get_file_size = generic_get_file_size,
	753	.options = options,
	754	.option_struct_size = sizeof(struct ioring_options),
	755	};
	756
	757	static void fio_init fio_ioring_register(void)
	758	{
	759	register_ioengine(&ioengine);
	760	}
	761
	762	static void fio_exit fio_ioring_unregister(void)
	763	{
	764	unregister_ioengine(&ioengine);
	765	}
	766	#endif