[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_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 should poll",
		.type	= FIO_OPT_INT,
		.cb	= fio_ioring_sqpoll_cb,
		.help	= "Offload submission to kernel thread",
		.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 (io_u->ddir == DDIR_READ)
			sqe->opcode = IORING_OP_READV;
		else
			sqe->opcode = IORING_OP_WRITEV;

		if (o->fixedbufs) {
			sqe->flags |= IOSQE_FIXED_BUFFER;
			sqe->addr = io_u->xfer_buf;
			sqe->len = io_u->xfer_buflen;
			sqe->buf_index = io_u->index;
		} else {
			sqe->addr = &ld->iovecs[io_u->index];
			sqe->len = 1;
		}
		sqe->off = io_u->offset;
	} else if (ddir_sync(io_u->ddir))
		sqe->opcode = IORING_OP_FSYNC;

	sqe->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->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_set) {
			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;

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