[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 "../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_event *events;
};

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_iocb *iocbs;
	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 sqthread;
	unsigned int sqthread_set;
	unsigned int sqthread_poll;
	unsigned int sqwq;
};

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

	o->sqthread = *val;
	o->sqthread_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",
		.lname	= "Use kernel SQ thread on this CPU",
		.type	= FIO_OPT_INT,
		.cb	= fio_ioring_sqthread_cb,
		.help	= "Offload submission to kernel thread",
		.category = FIO_OPT_C_ENGINE,
		.group	= FIO_OPT_G_LIBAIO,
	},
	{
		.name	= "sqthread_poll",
		.lname	= "Kernel SQ thread should poll",
		.type	= FIO_OPT_STR_SET,
		.off1	= offsetof(struct ioring_options, sqthread_poll),
		.help	= "Used with sqthread, enables kernel side polling",
		.category = FIO_OPT_C_ENGINE,
		.group	= FIO_OPT_G_LIBAIO,
	},
	{
		.name	= "sqwq",
		.lname	= "Offload submission to kernel workqueue",
		.type	= FIO_OPT_STR_SET,
		.off1	= offsetof(struct ioring_options, sqwq),
		.help	= "Offload submission to kernel workqueue",
		.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 fio_file *f = io_u->file;
	struct io_uring_iocb *iocb;

	iocb = &ld->iocbs[io_u->index];
	iocb->fd = f->fd;
	iocb->flags = 0;
	iocb->ioprio = 0;

	if (io_u->ddir == DDIR_READ || io_u->ddir == DDIR_WRITE) {
		if (io_u->ddir == DDIR_READ)
			iocb->opcode = IORING_OP_READ;
		else
			iocb->opcode = IORING_OP_WRITE;
		iocb->off = io_u->offset;
		iocb->addr = io_u->xfer_buf;
		iocb->len = io_u->xfer_buflen;
	} else if (ddir_sync(io_u->ddir))
		iocb->opcode = IORING_OP_FSYNC;

	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_event *ev;
	struct io_u *io_u;
	unsigned index;

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

	ev = &ld->cq_ring.events[index];
	io_u = ld->io_u_index[ev->index];

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

	if (io_u->ddir == DDIR_READ) {
		if (ev->flags & IOEV_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->sqthread_poll) {
			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->sqthread_poll) {
		struct io_sq_ring *ring = &ld->sq_ring;

		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_iocb);
	ld->iocbs = mmap(0, ld->mmap[1].len, PROT_READ | PROT_WRITE,
				MAP_SHARED | MAP_POPULATE, ld->ring_fd,
				IORING_OFF_IOCB);
	ld->mmap[1].ptr = ld->iocbs;

	ld->mmap[2].len = p->cq_off.events +
				p->cq_entries * sizeof(struct io_uring_event);
	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->events = ptr + p->cq_off.events;
	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->sqthread_set) {
		p.sq_thread_cpu = o->sqthread;
		p.flags |= IORING_SETUP_SQTHREAD;
		if (o->sqthread_poll)
			p.flags |= IORING_SETUP_SQPOLL;
	}
	if (o->sqwq)
		p.flags |= IORING_SETUP_SQWQ;

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

		setrlimit(RLIMIT_MEMLOCK, &rlim);
		p.flags |= IORING_SETUP_FIXEDBUFS;
	}

	ret = syscall(__NR_sys_io_uring_setup, depth, ld->iovecs, &p);
	if (ret < 0)
		return ret;

	ld->ring_fd = 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 ioring_options *o = td->eo;
	struct io_u *io_u;
	int err;

	if (o->fixedbufs) {
		int i;

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

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

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

	return 0;
}

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

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

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

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

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

	ld->iovecs = calloc(td->o.iodepth, sizeof(struct iovec));

	td->io_ops_data = ld;
	return 0;
}

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

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

static struct ioengine_ops ioengine = {
	.name			= "io_uring",
	.version		= FIO_IOOPS_VERSION,
	.init			= fio_ioring_init,
	.post_init		= fio_ioring_post_init,
	.io_u_init		= fio_ioring_io_u_init,
	.prep			= fio_ioring_prep,
	.queue			= fio_ioring_queue,
	.commit			= fio_ioring_commit,
	.getevents		= fio_ioring_getevents,
	.event			= fio_ioring_event,
	.cleanup		= fio_ioring_cleanup,
	.open_file		= generic_open_file,
	.close_file		= generic_close_file,
	.get_file_size		= generic_get_file_size,
	.options		= options,
	.option_struct_size	= sizeof(struct ioring_options),
};

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

static void fio_exit fio_ioring_unregister(void)
{
	unregister_ioengine(&ioengine);
}
#endif
Commit	Line	Data
52885fa2	1	/*
bffad86f	2	* io_uring engine
52885fa2	3	*
bffad86f	4	* IO engine using the new native Linux aio io_uring interface. See:
a90cd050	5	*
bffad86f	6	* http://git.kernel.dk/cgit/linux-block/log/?h=io_uring
52885fa2 JA	7	*
	8	*/
	9	#include <stdlib.h>
	10	#include <unistd.h>
	11	#include <errno.h>
52885fa2 JA	12	#include <sys/time.h>
	13	#include <sys/resource.h>
	14
	15	#include "../fio.h"
	16	#include "../lib/pow2.h"
	17	#include "../optgroup.h"
	18	#include "../lib/memalign.h"
b87aa01a	19	#include "../lib/fls.h"
52885fa2	20
bffad86f	21	#ifdef ARCH_HAVE_IOURING
52885fa2	22
bffad86f	23	#include "../os/io_uring.h"
9a2d78b3	24
bffad86f	25	struct io_sq_ring {
e2239016 JA	26	unsigned *head;
	27	unsigned *tail;
	28	unsigned *ring_mask;
	29	unsigned *ring_entries;
	30	unsigned *flags;
	31	unsigned *array;
52885fa2 JA	32	};
52885fa2 JA	33
bffad86f	34	struct io_cq_ring {
e2239016 JA	35	unsigned *head;
	36	unsigned *tail;
	37	unsigned *ring_mask;
	38	unsigned *ring_entries;
bffad86f	39	struct io_uring_event *events;
9a2d78b3 JA	40	};
9a2d78b3 JA	41
bffad86f	42	struct ioring_mmap {
9a2d78b3 JA	43	void *ptr;
9a2d78b3 JA	44	size_t len;
52885fa2 JA	45	};
52885fa2 JA	46
bffad86f	47	struct ioring_data {
9a2d78b3 JA	48	int ring_fd;
9a2d78b3 JA	49
52885fa2 JA	50	struct io_u **io_us;
	51	struct io_u **io_u_index;
	52
bffad86f JA	53	struct io_sq_ring sq_ring;
bffad86f JA	54	struct io_uring_iocb *iocbs;
9a2d78b3	55	struct iovec *iovecs;
b87aa01a	56	unsigned sq_ring_mask;
52885fa2	57
bffad86f	58	struct io_cq_ring cq_ring;
b87aa01a	59	unsigned cq_ring_mask;
52885fa2 JA	60
	61	int queued;
	62	int cq_ring_off;
b87aa01a	63	unsigned iodepth;
96563db9 JA	64
	65	uint64_t cachehit;
	66	uint64_t cachemiss;
9a2d78b3	67
bffad86f	68	struct ioring_mmap mmap[3];
52885fa2 JA	69	};
52885fa2 JA	70
bffad86f	71	struct ioring_options {
52885fa2 JA	72	void *pad;
	73	unsigned int hipri;
	74	unsigned int fixedbufs;
a90cd050 JA	75	unsigned int sqthread;
a90cd050 JA	76	unsigned int sqthread_set;
771c9901	77	unsigned int sqthread_poll;
a90cd050	78	unsigned int sqwq;
52885fa2 JA	79	};
52885fa2 JA	80
bffad86f	81	static int fio_ioring_sqthread_cb(void data, unsigned long long val)
a90cd050	82	{
bffad86f	83	struct ioring_options *o = data;
a90cd050 JA	84
	85	o->sqthread = *val;
	86	o->sqthread_set = 1;
	87	return 0;
	88	}
	89
52885fa2 JA	90	static struct fio_option options[] = {
	91	{
	92	.name = "hipri",
	93	.lname = "High Priority",
	94	.type = FIO_OPT_STR_SET,
bffad86f	95	.off1 = offsetof(struct ioring_options, hipri),
52885fa2 JA	96	.help = "Use polled IO completions",
	97	.category = FIO_OPT_C_ENGINE,
	98	.group = FIO_OPT_G_LIBAIO,
	99	},
	100	{
	101	.name = "fixedbufs",
	102	.lname = "Fixed (pre-mapped) IO buffers",
	103	.type = FIO_OPT_STR_SET,
bffad86f	104	.off1 = offsetof(struct ioring_options, fixedbufs),
52885fa2 JA	105	.help = "Pre map IO buffers",
	106	.category = FIO_OPT_C_ENGINE,
	107	.group = FIO_OPT_G_LIBAIO,
	108	},
a90cd050 JA	109	{
	110	.name = "sqthread",
	111	.lname = "Use kernel SQ thread on this CPU",
	112	.type = FIO_OPT_INT,
bffad86f	113	.cb = fio_ioring_sqthread_cb,
a90cd050 JA	114	.help = "Offload submission to kernel thread",
	115	.category = FIO_OPT_C_ENGINE,
	116	.group = FIO_OPT_G_LIBAIO,
	117	},
771c9901 JA	118	{
	119	.name = "sqthread_poll",
	120	.lname = "Kernel SQ thread should poll",
	121	.type = FIO_OPT_STR_SET,
bffad86f	122	.off1 = offsetof(struct ioring_options, sqthread_poll),
771c9901 JA	123	.help = "Used with sqthread, enables kernel side polling",
	124	.category = FIO_OPT_C_ENGINE,
	125	.group = FIO_OPT_G_LIBAIO,
	126	},
a90cd050 JA	127	{
	128	.name = "sqwq",
	129	.lname = "Offload submission to kernel workqueue",
	130	.type = FIO_OPT_STR_SET,
bffad86f	131	.off1 = offsetof(struct ioring_options, sqwq),
a90cd050 JA	132	.help = "Offload submission to kernel workqueue",
	133	.category = FIO_OPT_C_ENGINE,
	134	.group = FIO_OPT_G_LIBAIO,
	135	},
52885fa2 JA	136	{
	137	.name = NULL,
	138	},
	139	};
	140
bffad86f	141	static int io_uring_enter(struct ioring_data *ld, unsigned int to_submit,
52885fa2 JA	142	unsigned int min_complete, unsigned int flags)
52885fa2 JA	143	{
9a2d78b3 JA	144	return syscall(__NR_sys_io_uring_enter, ld->ring_fd, to_submit,
9a2d78b3 JA	145	min_complete, flags);
52885fa2 JA	146	}
52885fa2 JA	147
bffad86f	148	static int fio_ioring_prep(struct thread_data td, struct io_u io_u)
52885fa2	149	{
bffad86f	150	struct ioring_data *ld = td->io_ops_data;
52885fa2	151	struct fio_file *f = io_u->file;
bffad86f	152	struct io_uring_iocb *iocb;
52885fa2 JA	153
52885fa2 JA	154	iocb = &ld->iocbs[io_u->index];
bffad86f JA	155	iocb->fd = f->fd;
	156	iocb->flags = 0;
	157	iocb->ioprio = 0;
52885fa2	158
e3970057 JA	159	if (io_u->ddir == DDIR_READ \|\| io_u->ddir == DDIR_WRITE) {
e3970057 JA	160	if (io_u->ddir == DDIR_READ)
bffad86f	161	iocb->opcode = IORING_OP_READ;
e3970057	162	else
bffad86f JA	163	iocb->opcode = IORING_OP_WRITE;
	164	iocb->off = io_u->offset;
	165	iocb->addr = io_u->xfer_buf;
	166	iocb->len = io_u->xfer_buflen;
52885fa2	167	} else if (ddir_sync(io_u->ddir))
bffad86f	168	iocb->opcode = IORING_OP_FSYNC;
52885fa2	169
52885fa2 JA	170	return 0;
	171	}
	172
bffad86f	173	static struct io_u fio_ioring_event(struct thread_data td, int event)
52885fa2	174	{
bffad86f JA	175	struct ioring_data *ld = td->io_ops_data;
bffad86f JA	176	struct io_uring_event *ev;
52885fa2	177	struct io_u *io_u;
b87aa01a	178	unsigned index;
52885fa2	179
b87aa01a	180	index = (event + ld->cq_ring_off) & ld->cq_ring_mask;
52885fa2	181
9a2d78b3	182	ev = &ld->cq_ring.events[index];
bffad86f	183	io_u = ld->io_u_index[ev->index];
52885fa2 JA	184
	185	if (ev->res != io_u->xfer_buflen) {
	186	if (ev->res > io_u->xfer_buflen)
	187	io_u->error = -ev->res;
	188	else
	189	io_u->resid = io_u->xfer_buflen - ev->res;
	190	} else
	191	io_u->error = 0;
	192
96563db9	193	if (io_u->ddir == DDIR_READ) {
bffad86f	194	if (ev->flags & IOEV_FLAG_CACHEHIT)
96563db9 JA	195	ld->cachehit++;
	196	else
	197	ld->cachemiss++;
	198	}
	199
52885fa2 JA	200	return io_u;
	201	}
	202
bffad86f	203	static int fio_ioring_cqring_reap(struct thread_data *td, unsigned int events,
52885fa2 JA	204	unsigned int max)
52885fa2 JA	205	{
bffad86f JA	206	struct ioring_data *ld = td->io_ops_data;
bffad86f JA	207	struct io_cq_ring *ring = &ld->cq_ring;
e2239016	208	unsigned head, reaped = 0;
52885fa2	209
9a2d78b3	210	head = *ring->head;
52885fa2 JA	211	do {
52885fa2 JA	212	read_barrier();
9a2d78b3	213	if (head == *ring->tail)
52885fa2 JA	214	break;
	215	reaped++;
	216	head++;
52885fa2 JA	217	} while (reaped + events < max);
52885fa2 JA	218
9a2d78b3	219	*ring->head = head;
52885fa2 JA	220	write_barrier();
	221	return reaped;
	222	}
	223
bffad86f JA	224	static int fio_ioring_getevents(struct thread_data *td, unsigned int min,
bffad86f JA	225	unsigned int max, const struct timespec *t)
52885fa2	226	{
bffad86f	227	struct ioring_data *ld = td->io_ops_data;
52885fa2	228	unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
bffad86f JA	229	struct ioring_options *o = td->eo;
bffad86f JA	230	struct io_cq_ring *ring = &ld->cq_ring;
b87aa01a JA	231	unsigned events = 0;
b87aa01a JA	232	int r;
52885fa2	233
9a2d78b3	234	ld->cq_ring_off = *ring->head;
52885fa2	235	do {
bffad86f	236	r = fio_ioring_cqring_reap(td, events, max);
52885fa2 JA	237	if (r) {
	238	events += r;
	239	continue;
	240	}
	241
771c9901	242	if (!o->sqthread_poll) {
9a2d78b3 JA	243	r = io_uring_enter(ld, 0, actual_min,
9a2d78b3 JA	244	IORING_ENTER_GETEVENTS);
771c9901 JA	245	if (r < 0) {
	246	if (errno == EAGAIN)
	247	continue;
9a2d78b3	248	td_verror(td, errno, "io_uring_enter");
771c9901 JA	249	break;
771c9901 JA	250	}
52885fa2 JA	251	}
	252	} while (events < min);
	253
	254	return r < 0 ? r : events;
	255	}
	256
bffad86f JA	257	static enum fio_q_status fio_ioring_queue(struct thread_data *td,
bffad86f JA	258	struct io_u *io_u)
52885fa2	259	{
bffad86f JA	260	struct ioring_data *ld = td->io_ops_data;
bffad86f JA	261	struct io_sq_ring *ring = &ld->sq_ring;
52885fa2 JA	262	unsigned tail, next_tail;
	263
	264	fio_ro_check(td, io_u);
	265
b87aa01a	266	if (ld->queued == ld->iodepth)
52885fa2 JA	267	return FIO_Q_BUSY;
52885fa2 JA	268
52885fa2 JA	269	if (io_u->ddir == DDIR_TRIM) {
	270	if (ld->queued)
	271	return FIO_Q_BUSY;
	272
	273	do_io_u_trim(td, io_u);
	274	io_u_mark_submit(td, 1);
	275	io_u_mark_complete(td, 1);
	276	return FIO_Q_COMPLETED;
	277	}
	278
9a2d78b3	279	tail = *ring->tail;
52885fa2	280	next_tail = tail + 1;
52885fa2	281	read_barrier();
9a2d78b3	282	if (next_tail == *ring->head)
52885fa2 JA	283	return FIO_Q_BUSY;
52885fa2 JA	284
b87aa01a	285	ring->array[tail & ld->sq_ring_mask] = io_u->index;
9a2d78b3	286	*ring->tail = next_tail;
52885fa2 JA	287	write_barrier();
	288
	289	ld->queued++;
	290	return FIO_Q_QUEUED;
	291	}
	292
bffad86f	293	static void fio_ioring_queued(struct thread_data *td, int start, int nr)
52885fa2	294	{
bffad86f	295	struct ioring_data *ld = td->io_ops_data;
52885fa2 JA	296	struct timespec now;
	297
	298	if (!fio_fill_issue_time(td))
	299	return;
	300
	301	fio_gettime(&now, NULL);
	302
	303	while (nr--) {
bffad86f	304	struct io_sq_ring *ring = &ld->sq_ring;
9a2d78b3	305	int index = ring->array[start & ld->sq_ring_mask];
f8289afc	306	struct io_u *io_u = ld->io_u_index[index];
52885fa2 JA	307
	308	memcpy(&io_u->issue_time, &now, sizeof(now));
	309	io_u_queued(td, io_u);
	310
	311	start++;
52885fa2 JA	312	}
	313	}
	314
bffad86f	315	static int fio_ioring_commit(struct thread_data *td)
52885fa2	316	{
bffad86f JA	317	struct ioring_data *ld = td->io_ops_data;
bffad86f JA	318	struct ioring_options *o = td->eo;
52885fa2 JA	319	int ret;
	320
	321	if (!ld->queued)
	322	return 0;
	323
771c9901 JA	324	/* Nothing to do */
771c9901 JA	325	if (o->sqthread_poll) {
bffad86f	326	struct io_sq_ring *ring = &ld->sq_ring;
4cdbc048	327
9a2d78b3 JA	328	if (*ring->flags & IORING_SQ_NEED_WAKEUP)
9a2d78b3 JA	329	io_uring_enter(ld, ld->queued, 0, 0);
771c9901 JA	330	ld->queued = 0;
	331	return 0;
	332	}
	333
52885fa2	334	do {
9a2d78b3	335	unsigned start = *ld->sq_ring.head;
52885fa2 JA	336	long nr = ld->queued;
52885fa2 JA	337
9a2d78b3	338	ret = io_uring_enter(ld, nr, 0, IORING_ENTER_GETEVENTS);
52885fa2	339	if (ret > 0) {
bffad86f	340	fio_ioring_queued(td, start, ret);
52885fa2 JA	341	io_u_mark_submit(td, ret);
	342
	343	ld->queued -= ret;
	344	ret = 0;
a90cd050 JA	345	} else if (!ret) {
a90cd050 JA	346	io_u_mark_submit(td, ret);
52885fa2	347	continue;
a90cd050 JA	348	} else {
a90cd050 JA	349	if (errno == EAGAIN) {
bffad86f	350	ret = fio_ioring_cqring_reap(td, 0, ld->queued);
a90cd050 JA	351	if (ret)
	352	continue;
	353	/* Shouldn't happen */
	354	usleep(1);
	355	continue;
52885fa2	356	}
9a2d78b3	357	td_verror(td, errno, "io_uring_enter submit");
52885fa2	358	break;
a90cd050	359	}
52885fa2 JA	360	} while (ld->queued);
	361
	362	return ret;
	363	}
	364
bffad86f	365	static void fio_ioring_unmap(struct ioring_data *ld)
52885fa2	366	{
9a2d78b3	367	int i;
52885fa2	368
9a2d78b3 JA	369	for (i = 0; i < ARRAY_SIZE(ld->mmap); i++)
	370	munmap(ld->mmap[i].ptr, ld->mmap[i].len);
	371	close(ld->ring_fd);
b87aa01a JA	372	}
b87aa01a JA	373
bffad86f	374	static void fio_ioring_cleanup(struct thread_data *td)
52885fa2	375	{
bffad86f	376	struct ioring_data *ld = td->io_ops_data;
52885fa2 JA	377
52885fa2 JA	378	if (ld) {
96563db9 JA	379	td->ts.cachehit += ld->cachehit;
	380	td->ts.cachemiss += ld->cachemiss;
	381
52885fa2	382	if (!(td->flags & TD_F_CHILD))
bffad86f	383	fio_ioring_unmap(ld);
9a2d78b3	384
52885fa2 JA	385	free(ld->io_u_index);
52885fa2 JA	386	free(ld->io_us);
9a2d78b3	387	free(ld->iovecs);
52885fa2 JA	388	free(ld);
	389	}
	390	}
	391
bffad86f	392	static int fio_ioring_mmap(struct ioring_data ld, struct io_uring_params p)
9a2d78b3	393	{
bffad86f JA	394	struct io_sq_ring *sring = &ld->sq_ring;
bffad86f JA	395	struct io_cq_ring *cring = &ld->cq_ring;
9a2d78b3 JA	396	void *ptr;
9a2d78b3 JA	397
e2239016	398	ld->mmap[0].len = p->sq_off.array + p->sq_entries * sizeof(__u32);
9a2d78b3 JA	399	ptr = mmap(0, ld->mmap[0].len, PROT_READ \| PROT_WRITE,
	400	MAP_SHARED \| MAP_POPULATE, ld->ring_fd,
	401	IORING_OFF_SQ_RING);
	402	ld->mmap[0].ptr = ptr;
	403	sring->head = ptr + p->sq_off.head;
	404	sring->tail = ptr + p->sq_off.tail;
	405	sring->ring_mask = ptr + p->sq_off.ring_mask;
	406	sring->ring_entries = ptr + p->sq_off.ring_entries;
	407	sring->flags = ptr + p->sq_off.flags;
ac122fea	408	sring->array = ptr + p->sq_off.array;
9a2d78b3 JA	409	ld->sq_ring_mask = *sring->ring_mask;
9a2d78b3 JA	410
bffad86f	411	ld->mmap[1].len = p->sq_entries * sizeof(struct io_uring_iocb);
9a2d78b3 JA	412	ld->iocbs = mmap(0, ld->mmap[1].len, PROT_READ \| PROT_WRITE,
	413	MAP_SHARED \| MAP_POPULATE, ld->ring_fd,
	414	IORING_OFF_IOCB);
	415	ld->mmap[1].ptr = ld->iocbs;
	416
ac122fea	417	ld->mmap[2].len = p->cq_off.events +
bffad86f	418	p->cq_entries * sizeof(struct io_uring_event);
9a2d78b3 JA	419	ptr = mmap(0, ld->mmap[2].len, PROT_READ \| PROT_WRITE,
	420	MAP_SHARED \| MAP_POPULATE, ld->ring_fd,
	421	IORING_OFF_CQ_RING);
	422	ld->mmap[2].ptr = ptr;
	423	cring->head = ptr + p->cq_off.head;
	424	cring->tail = ptr + p->cq_off.tail;
	425	cring->ring_mask = ptr + p->cq_off.ring_mask;
	426	cring->ring_entries = ptr + p->cq_off.ring_entries;
ac122fea	427	cring->events = ptr + p->cq_off.events;
9a2d78b3 JA	428	ld->cq_ring_mask = *cring->ring_mask;
	429	return 0;
	430	}
	431
bffad86f	432	static int fio_ioring_queue_init(struct thread_data *td)
52885fa2	433	{
bffad86f JA	434	struct ioring_data *ld = td->io_ops_data;
bffad86f JA	435	struct ioring_options *o = td->eo;
52885fa2	436	int depth = td->o.iodepth;
bffad86f	437	struct io_uring_params p;
9a2d78b3 JA	438	int ret;
	439
	440	memset(&p, 0, sizeof(p));
52885fa2 JA	441
52885fa2 JA	442	if (o->hipri)
bffad86f	443	p.flags \|= IORING_SETUP_IOPOLL;
a90cd050	444	if (o->sqthread_set) {
9a2d78b3	445	p.sq_thread_cpu = o->sqthread;
bffad86f	446	p.flags \|= IORING_SETUP_SQTHREAD;
771c9901	447	if (o->sqthread_poll)
bffad86f	448	p.flags \|= IORING_SETUP_SQPOLL;
f635f1fb JA	449	}
f635f1fb JA	450	if (o->sqwq)
bffad86f	451	p.flags \|= IORING_SETUP_SQWQ;
a90cd050	452
52885fa2 JA	453	if (o->fixedbufs) {
	454	struct rlimit rlim = {
	455	.rlim_cur = RLIM_INFINITY,
	456	.rlim_max = RLIM_INFINITY,
	457	};
	458
	459	setrlimit(RLIMIT_MEMLOCK, &rlim);
bffad86f	460	p.flags \|= IORING_SETUP_FIXEDBUFS;
52885fa2 JA	461	}
52885fa2 JA	462
9a2d78b3 JA	463	ret = syscall(__NR_sys_io_uring_setup, depth, ld->iovecs, &p);
	464	if (ret < 0)
	465	return ret;
	466
	467	ld->ring_fd = ret;
bffad86f	468	return fio_ioring_mmap(ld, &p);
52885fa2 JA	469	}
52885fa2 JA	470
bffad86f	471	static int fio_ioring_post_init(struct thread_data *td)
52885fa2	472	{
bffad86f JA	473	struct ioring_data *ld = td->io_ops_data;
bffad86f JA	474	struct ioring_options *o = td->eo;
52885fa2	475	struct io_u *io_u;
9a2d78b3	476	int err;
52885fa2 JA	477
	478	if (o->fixedbufs) {
	479	int i;
	480
	481	for (i = 0; i < td->o.iodepth; i++) {
9a2d78b3 JA	482	struct iovec *iov = &ld->iovecs[i];
9a2d78b3 JA	483
52885fa2	484	io_u = ld->io_u_index[i];
9a2d78b3 JA	485	iov->iov_base = io_u->buf;
9a2d78b3 JA	486	iov->iov_len = td_max_bs(td);
52885fa2 JA	487	}
	488	}
	489
bffad86f	490	err = fio_ioring_queue_init(td);
52885fa2	491	if (err) {
d63a472d	492	td_verror(td, errno, "io_queue_init");
52885fa2 JA	493	return 1;
	494	}
	495
	496	return 0;
	497	}
	498
9a2d78b3 JA	499	static unsigned roundup_pow2(unsigned depth)
	500	{
	501	return 1UL << __fls(depth - 1);
	502	}
	503
bffad86f	504	static int fio_ioring_init(struct thread_data *td)
52885fa2	505	{
bffad86f	506	struct ioring_data *ld;
52885fa2	507
52885fa2 JA	508	ld = calloc(1, sizeof(*ld));
52885fa2 JA	509
b87aa01a JA	510	/* ring depth must be a power-of-2 */
	511	ld->iodepth = td->o.iodepth;
	512	td->o.iodepth = roundup_pow2(td->o.iodepth);
	513
52885fa2 JA	514	/* io_u index */
	515	ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *));
	516	ld->io_us = calloc(td->o.iodepth, sizeof(struct io_u *));
	517
9a2d78b3	518	ld->iovecs = calloc(td->o.iodepth, sizeof(struct iovec));
52885fa2 JA	519
	520	td->io_ops_data = ld;
	521	return 0;
	522	}
	523
bffad86f	524	static int fio_ioring_io_u_init(struct thread_data td, struct io_u io_u)
52885fa2	525	{
bffad86f	526	struct ioring_data *ld = td->io_ops_data;
52885fa2 JA	527
	528	ld->io_u_index[io_u->index] = io_u;
	529	return 0;
	530	}
	531
	532	static struct ioengine_ops ioengine = {
bffad86f	533	.name = "io_uring",
52885fa2	534	.version = FIO_IOOPS_VERSION,
bffad86f JA	535	.init = fio_ioring_init,
	536	.post_init = fio_ioring_post_init,
	537	.io_u_init = fio_ioring_io_u_init,
	538	.prep = fio_ioring_prep,
	539	.queue = fio_ioring_queue,
	540	.commit = fio_ioring_commit,
	541	.getevents = fio_ioring_getevents,
	542	.event = fio_ioring_event,
	543	.cleanup = fio_ioring_cleanup,
52885fa2 JA	544	.open_file = generic_open_file,
	545	.close_file = generic_close_file,
	546	.get_file_size = generic_get_file_size,
	547	.options = options,
bffad86f	548	.option_struct_size = sizeof(struct ioring_options),
52885fa2 JA	549	};
52885fa2 JA	550
bffad86f	551	static void fio_init fio_ioring_register(void)
52885fa2	552	{
52885fa2	553	register_ioengine(&ioengine);
52885fa2 JA	554	}
52885fa2 JA	555
bffad86f	556	static void fio_exit fio_ioring_unregister(void)
52885fa2	557	{
52885fa2	558	unregister_ioengine(&ioengine);
52885fa2	559	}
1f90e9bb	560	#endif