[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;

	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;

	struct ioring_mmap mmap[3];
};

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

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,
	},
	{
		.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	= "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	= 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];

	/* 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) {
			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;
		}
		if (!td->o.odirect && o->uncached)
			sqe->rw_flags = RWF_UNCACHED;
		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 {
		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;
			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 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 || 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_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_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_sys_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 unsigned roundup_pow2(unsigned depth)
{
	return 1UL << __fls(depth - 1);
}

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

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