[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 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 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;

	if (io_u->ddir == DDIR_READ || io_u->ddir == DDIR_WRITE) {
		if (o->fixedbufs) {
			if (io_u->ddir == DDIR_READ)
				sqe->opcode = IORING_OP_READ_FIXED;
			else
				sqe->opcode = IORING_OP_WRITE_FIXED;
			sqe->addr = io_u->xfer_buf;
			sqe->len = io_u->xfer_buflen;
			sqe->index = io_u->index;
		} else {
			if (io_u->ddir == DDIR_READ)
				sqe->opcode = IORING_OP_READV;
			else
				sqe->opcode = IORING_OP_WRITEV;
			sqe->addr = &ld->iovecs[io_u->index];
			sqe->len = 1;
		}
		sqe->off = io_u->offset;
	} else if (ddir_sync(io_u->ddir))
		sqe->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->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_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 iovec *vecs = NULL;
	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);
		vecs = ld->iovecs;
	}

	ret = syscall(__NR_sys_io_uring_setup, depth, vecs, depth, &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 io_u *io_u;
	int err, i;

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

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

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

	return 0;
}

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

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

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

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

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

	td->io_ops_data = ld;
	return 0;
}

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

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

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

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

static void fio_exit fio_ioring_unregister(void)
{
	unregister_ioengine(&ioengine);
}
#endif
Commit	Line	Data
52885fa2	1	/*
bffad86f	2	* io_uring engine
52885fa2	3	*
bffad86f	4	* IO engine using the new native Linux aio io_uring interface. See:
a90cd050	5	*
bffad86f	6	* http://git.kernel.dk/cgit/linux-block/log/?h=io_uring
52885fa2 JA	7	*
	8	*/
	9	#include <stdlib.h>
	10	#include <unistd.h>
	11	#include <errno.h>
52885fa2 JA	12	#include <sys/time.h>
	13	#include <sys/resource.h>
	14
	15	#include "../fio.h"
	16	#include "../lib/pow2.h"
	17	#include "../optgroup.h"
	18	#include "../lib/memalign.h"
b87aa01a	19	#include "../lib/fls.h"
52885fa2	20
bffad86f	21	#ifdef ARCH_HAVE_IOURING
52885fa2	22
57fa61f0	23	#include "../lib/types.h"
bffad86f	24	#include "../os/io_uring.h"
9a2d78b3	25
bffad86f	26	struct io_sq_ring {
e2239016 JA	27	unsigned *head;
	28	unsigned *tail;
	29	unsigned *ring_mask;
	30	unsigned *ring_entries;
	31	unsigned *flags;
	32	unsigned *array;
52885fa2 JA	33	};
52885fa2 JA	34
bffad86f	35	struct io_cq_ring {
e2239016 JA	36	unsigned *head;
	37	unsigned *tail;
	38	unsigned *ring_mask;
	39	unsigned *ring_entries;
f0403f94	40	struct io_uring_cqe *cqes;
9a2d78b3 JA	41	};
9a2d78b3 JA	42
bffad86f	43	struct ioring_mmap {
9a2d78b3 JA	44	void *ptr;
9a2d78b3 JA	45	size_t len;
52885fa2 JA	46	};
52885fa2 JA	47
bffad86f	48	struct ioring_data {
9a2d78b3 JA	49	int ring_fd;
9a2d78b3 JA	50
52885fa2 JA	51	struct io_u **io_us;
	52	struct io_u **io_u_index;
	53
bffad86f	54	struct io_sq_ring sq_ring;
f0403f94	55	struct io_uring_sqe *sqes;
9a2d78b3	56	struct iovec *iovecs;
b87aa01a	57	unsigned sq_ring_mask;
52885fa2	58
bffad86f	59	struct io_cq_ring cq_ring;
b87aa01a	60	unsigned cq_ring_mask;
52885fa2 JA	61
	62	int queued;
	63	int cq_ring_off;
b87aa01a	64	unsigned iodepth;
96563db9 JA	65
	66	uint64_t cachehit;
	67	uint64_t cachemiss;
9a2d78b3	68
bffad86f	69	struct ioring_mmap mmap[3];
52885fa2 JA	70	};
52885fa2 JA	71
bffad86f	72	struct ioring_options {
52885fa2 JA	73	void *pad;
	74	unsigned int hipri;
	75	unsigned int fixedbufs;
a90cd050 JA	76	unsigned int sqthread;
a90cd050 JA	77	unsigned int sqthread_set;
771c9901	78	unsigned int sqthread_poll;
a90cd050	79	unsigned int sqwq;
52885fa2 JA	80	};
52885fa2 JA	81
bffad86f	82	static int fio_ioring_sqthread_cb(void data, unsigned long long val)
a90cd050	83	{
bffad86f	84	struct ioring_options *o = data;
a90cd050 JA	85
	86	o->sqthread = *val;
	87	o->sqthread_set = 1;
	88	return 0;
	89	}
	90
52885fa2 JA	91	static struct fio_option options[] = {
	92	{
	93	.name = "hipri",
	94	.lname = "High Priority",
	95	.type = FIO_OPT_STR_SET,
bffad86f	96	.off1 = offsetof(struct ioring_options, hipri),
52885fa2 JA	97	.help = "Use polled IO completions",
	98	.category = FIO_OPT_C_ENGINE,
	99	.group = FIO_OPT_G_LIBAIO,
	100	},
	101	{
	102	.name = "fixedbufs",
	103	.lname = "Fixed (pre-mapped) IO buffers",
	104	.type = FIO_OPT_STR_SET,
bffad86f	105	.off1 = offsetof(struct ioring_options, fixedbufs),
52885fa2 JA	106	.help = "Pre map IO buffers",
	107	.category = FIO_OPT_C_ENGINE,
	108	.group = FIO_OPT_G_LIBAIO,
	109	},
a90cd050 JA	110	{
	111	.name = "sqthread",
	112	.lname = "Use kernel SQ thread on this CPU",
	113	.type = FIO_OPT_INT,
bffad86f	114	.cb = fio_ioring_sqthread_cb,
a90cd050 JA	115	.help = "Offload submission to kernel thread",
	116	.category = FIO_OPT_C_ENGINE,
	117	.group = FIO_OPT_G_LIBAIO,
	118	},
771c9901 JA	119	{
	120	.name = "sqthread_poll",
	121	.lname = "Kernel SQ thread should poll",
	122	.type = FIO_OPT_STR_SET,
bffad86f	123	.off1 = offsetof(struct ioring_options, sqthread_poll),
771c9901 JA	124	.help = "Used with sqthread, enables kernel side polling",
	125	.category = FIO_OPT_C_ENGINE,
	126	.group = FIO_OPT_G_LIBAIO,
	127	},
a90cd050 JA	128	{
	129	.name = "sqwq",
	130	.lname = "Offload submission to kernel workqueue",
	131	.type = FIO_OPT_STR_SET,
bffad86f	132	.off1 = offsetof(struct ioring_options, sqwq),
a90cd050 JA	133	.help = "Offload submission to kernel workqueue",
	134	.category = FIO_OPT_C_ENGINE,
	135	.group = FIO_OPT_G_LIBAIO,
	136	},
52885fa2 JA	137	{
	138	.name = NULL,
	139	},
	140	};
	141
bffad86f	142	static int io_uring_enter(struct ioring_data *ld, unsigned int to_submit,
52885fa2 JA	143	unsigned int min_complete, unsigned int flags)
52885fa2 JA	144	{
9a2d78b3 JA	145	return syscall(__NR_sys_io_uring_enter, ld->ring_fd, to_submit,
9a2d78b3 JA	146	min_complete, flags);
52885fa2 JA	147	}
52885fa2 JA	148
bffad86f	149	static int fio_ioring_prep(struct thread_data td, struct io_u io_u)
52885fa2	150	{
bffad86f	151	struct ioring_data *ld = td->io_ops_data;
cfcc8564	152	struct ioring_options *o = td->eo;
52885fa2	153	struct fio_file *f = io_u->file;
f0403f94	154	struct io_uring_sqe *sqe;
52885fa2	155
f0403f94 JA	156	sqe = &ld->sqes[io_u->index];
	157	sqe->fd = f->fd;
	158	sqe->flags = 0;
	159	sqe->ioprio = 0;
52885fa2	160
e3970057	161	if (io_u->ddir == DDIR_READ \|\| io_u->ddir == DDIR_WRITE) {
f0403f94 JA	162	if (o->fixedbufs) {
	163	if (io_u->ddir == DDIR_READ)
	164	sqe->opcode = IORING_OP_READ_FIXED;
cfcc8564	165	else
f0403f94 JA	166	sqe->opcode = IORING_OP_WRITE_FIXED;
	167	sqe->addr = io_u->xfer_buf;
	168	sqe->len = io_u->xfer_buflen;
a7086591	169	sqe->index = io_u->index;
cfcc8564	170	} else {
f0403f94 JA	171	if (io_u->ddir == DDIR_READ)
f0403f94 JA	172	sqe->opcode = IORING_OP_READV;
cfcc8564	173	else
f0403f94 JA	174	sqe->opcode = IORING_OP_WRITEV;
	175	sqe->addr = &ld->iovecs[io_u->index];
	176	sqe->len = 1;
cfcc8564	177	}
f0403f94	178	sqe->off = io_u->offset;
52885fa2	179	} else if (ddir_sync(io_u->ddir))
f0403f94	180	sqe->opcode = IORING_OP_FSYNC;
52885fa2	181
a7086591	182	sqe->data = (unsigned long) io_u;
52885fa2 JA	183	return 0;
	184	}
	185
bffad86f	186	static struct io_u fio_ioring_event(struct thread_data td, int event)
52885fa2	187	{
bffad86f	188	struct ioring_data *ld = td->io_ops_data;
f0403f94	189	struct io_uring_cqe *cqe;
52885fa2	190	struct io_u *io_u;
b87aa01a	191	unsigned index;
52885fa2	192
b87aa01a	193	index = (event + ld->cq_ring_off) & ld->cq_ring_mask;
52885fa2	194
f0403f94	195	cqe = &ld->cq_ring.cqes[index];
a7086591	196	io_u = (struct io_u *) cqe->data;
52885fa2	197
f0403f94 JA	198	if (cqe->res != io_u->xfer_buflen) {
	199	if (cqe->res > io_u->xfer_buflen)
	200	io_u->error = -cqe->res;
52885fa2	201	else
f0403f94	202	io_u->resid = io_u->xfer_buflen - cqe->res;
52885fa2 JA	203	} else
	204	io_u->error = 0;
	205
96563db9	206	if (io_u->ddir == DDIR_READ) {
f0403f94	207	if (cqe->flags & IOCQE_FLAG_CACHEHIT)
96563db9 JA	208	ld->cachehit++;
	209	else
	210	ld->cachemiss++;
	211	}
	212
52885fa2 JA	213	return io_u;
	214	}
	215
bffad86f	216	static int fio_ioring_cqring_reap(struct thread_data *td, unsigned int events,
52885fa2 JA	217	unsigned int max)
52885fa2 JA	218	{
bffad86f JA	219	struct ioring_data *ld = td->io_ops_data;
bffad86f JA	220	struct io_cq_ring *ring = &ld->cq_ring;
e2239016	221	unsigned head, reaped = 0;
52885fa2	222
9a2d78b3	223	head = *ring->head;
52885fa2 JA	224	do {
52885fa2 JA	225	read_barrier();
9a2d78b3	226	if (head == *ring->tail)
52885fa2 JA	227	break;
	228	reaped++;
	229	head++;
52885fa2 JA	230	} while (reaped + events < max);
52885fa2 JA	231
9a2d78b3	232	*ring->head = head;
52885fa2 JA	233	write_barrier();
	234	return reaped;
	235	}
	236
bffad86f JA	237	static int fio_ioring_getevents(struct thread_data *td, unsigned int min,
bffad86f JA	238	unsigned int max, const struct timespec *t)
52885fa2	239	{
bffad86f	240	struct ioring_data *ld = td->io_ops_data;
52885fa2	241	unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
bffad86f JA	242	struct ioring_options *o = td->eo;
bffad86f JA	243	struct io_cq_ring *ring = &ld->cq_ring;
b87aa01a JA	244	unsigned events = 0;
b87aa01a JA	245	int r;
52885fa2	246
9a2d78b3	247	ld->cq_ring_off = *ring->head;
52885fa2	248	do {
bffad86f	249	r = fio_ioring_cqring_reap(td, events, max);
52885fa2 JA	250	if (r) {
	251	events += r;
	252	continue;
	253	}
	254
771c9901	255	if (!o->sqthread_poll) {
9a2d78b3 JA	256	r = io_uring_enter(ld, 0, actual_min,
9a2d78b3 JA	257	IORING_ENTER_GETEVENTS);
771c9901 JA	258	if (r < 0) {
	259	if (errno == EAGAIN)
	260	continue;
9a2d78b3	261	td_verror(td, errno, "io_uring_enter");
771c9901 JA	262	break;
771c9901 JA	263	}
52885fa2 JA	264	}
	265	} while (events < min);
	266
	267	return r < 0 ? r : events;
	268	}
	269
bffad86f JA	270	static enum fio_q_status fio_ioring_queue(struct thread_data *td,
bffad86f JA	271	struct io_u *io_u)
52885fa2	272	{
bffad86f JA	273	struct ioring_data *ld = td->io_ops_data;
bffad86f JA	274	struct io_sq_ring *ring = &ld->sq_ring;
52885fa2 JA	275	unsigned tail, next_tail;
	276
	277	fio_ro_check(td, io_u);
	278
b87aa01a	279	if (ld->queued == ld->iodepth)
52885fa2 JA	280	return FIO_Q_BUSY;
52885fa2 JA	281
52885fa2 JA	282	if (io_u->ddir == DDIR_TRIM) {
	283	if (ld->queued)
	284	return FIO_Q_BUSY;
	285
	286	do_io_u_trim(td, io_u);
	287	io_u_mark_submit(td, 1);
	288	io_u_mark_complete(td, 1);
	289	return FIO_Q_COMPLETED;
	290	}
	291
9a2d78b3	292	tail = *ring->tail;
52885fa2	293	next_tail = tail + 1;
52885fa2	294	read_barrier();
9a2d78b3	295	if (next_tail == *ring->head)
52885fa2 JA	296	return FIO_Q_BUSY;
52885fa2 JA	297
b87aa01a	298	ring->array[tail & ld->sq_ring_mask] = io_u->index;
9a2d78b3	299	*ring->tail = next_tail;
52885fa2 JA	300	write_barrier();
	301
	302	ld->queued++;
	303	return FIO_Q_QUEUED;
	304	}
	305
bffad86f	306	static void fio_ioring_queued(struct thread_data *td, int start, int nr)
52885fa2	307	{
bffad86f	308	struct ioring_data *ld = td->io_ops_data;
52885fa2 JA	309	struct timespec now;
	310
	311	if (!fio_fill_issue_time(td))
	312	return;
	313
	314	fio_gettime(&now, NULL);
	315
	316	while (nr--) {
bffad86f	317	struct io_sq_ring *ring = &ld->sq_ring;
9a2d78b3	318	int index = ring->array[start & ld->sq_ring_mask];
f8289afc	319	struct io_u *io_u = ld->io_u_index[index];
52885fa2 JA	320
	321	memcpy(&io_u->issue_time, &now, sizeof(now));
	322	io_u_queued(td, io_u);
	323
	324	start++;
52885fa2 JA	325	}
	326	}
	327
bffad86f	328	static int fio_ioring_commit(struct thread_data *td)
52885fa2	329	{
bffad86f JA	330	struct ioring_data *ld = td->io_ops_data;
bffad86f JA	331	struct ioring_options *o = td->eo;
52885fa2 JA	332	int ret;
	333
	334	if (!ld->queued)
	335	return 0;
	336
771c9901 JA	337	/* Nothing to do */
771c9901 JA	338	if (o->sqthread_poll) {
bffad86f	339	struct io_sq_ring *ring = &ld->sq_ring;
4cdbc048	340
9a2d78b3 JA	341	if (*ring->flags & IORING_SQ_NEED_WAKEUP)
9a2d78b3 JA	342	io_uring_enter(ld, ld->queued, 0, 0);
771c9901 JA	343	ld->queued = 0;
	344	return 0;
	345	}
	346
52885fa2	347	do {
9a2d78b3	348	unsigned start = *ld->sq_ring.head;
52885fa2 JA	349	long nr = ld->queued;
52885fa2 JA	350
9a2d78b3	351	ret = io_uring_enter(ld, nr, 0, IORING_ENTER_GETEVENTS);
52885fa2	352	if (ret > 0) {
bffad86f	353	fio_ioring_queued(td, start, ret);
52885fa2 JA	354	io_u_mark_submit(td, ret);
	355
	356	ld->queued -= ret;
	357	ret = 0;
a90cd050 JA	358	} else if (!ret) {
a90cd050 JA	359	io_u_mark_submit(td, ret);
52885fa2	360	continue;
a90cd050 JA	361	} else {
a90cd050 JA	362	if (errno == EAGAIN) {
bffad86f	363	ret = fio_ioring_cqring_reap(td, 0, ld->queued);
a90cd050 JA	364	if (ret)
	365	continue;
	366	/* Shouldn't happen */
	367	usleep(1);
	368	continue;
52885fa2	369	}
9a2d78b3	370	td_verror(td, errno, "io_uring_enter submit");
52885fa2	371	break;
a90cd050	372	}
52885fa2 JA	373	} while (ld->queued);
	374
	375	return ret;
	376	}
	377
bffad86f	378	static void fio_ioring_unmap(struct ioring_data *ld)
52885fa2	379	{
9a2d78b3	380	int i;
52885fa2	381
9a2d78b3 JA	382	for (i = 0; i < ARRAY_SIZE(ld->mmap); i++)
	383	munmap(ld->mmap[i].ptr, ld->mmap[i].len);
	384	close(ld->ring_fd);
b87aa01a JA	385	}
b87aa01a JA	386
bffad86f	387	static void fio_ioring_cleanup(struct thread_data *td)
52885fa2	388	{
bffad86f	389	struct ioring_data *ld = td->io_ops_data;
52885fa2 JA	390
52885fa2 JA	391	if (ld) {
96563db9 JA	392	td->ts.cachehit += ld->cachehit;
	393	td->ts.cachemiss += ld->cachemiss;
	394
52885fa2	395	if (!(td->flags & TD_F_CHILD))
bffad86f	396	fio_ioring_unmap(ld);
9a2d78b3	397
52885fa2 JA	398	free(ld->io_u_index);
52885fa2 JA	399	free(ld->io_us);
9a2d78b3	400	free(ld->iovecs);
52885fa2 JA	401	free(ld);
	402	}
	403	}
	404
bffad86f	405	static int fio_ioring_mmap(struct ioring_data ld, struct io_uring_params p)
9a2d78b3	406	{
bffad86f JA	407	struct io_sq_ring *sring = &ld->sq_ring;
bffad86f JA	408	struct io_cq_ring *cring = &ld->cq_ring;
9a2d78b3 JA	409	void *ptr;
9a2d78b3 JA	410
e2239016	411	ld->mmap[0].len = p->sq_off.array + p->sq_entries * sizeof(__u32);
9a2d78b3 JA	412	ptr = mmap(0, ld->mmap[0].len, PROT_READ \| PROT_WRITE,
	413	MAP_SHARED \| MAP_POPULATE, ld->ring_fd,
	414	IORING_OFF_SQ_RING);
	415	ld->mmap[0].ptr = ptr;
	416	sring->head = ptr + p->sq_off.head;
	417	sring->tail = ptr + p->sq_off.tail;
	418	sring->ring_mask = ptr + p->sq_off.ring_mask;
	419	sring->ring_entries = ptr + p->sq_off.ring_entries;
	420	sring->flags = ptr + p->sq_off.flags;
ac122fea	421	sring->array = ptr + p->sq_off.array;
9a2d78b3 JA	422	ld->sq_ring_mask = *sring->ring_mask;
9a2d78b3 JA	423
f0403f94 JA	424	ld->mmap[1].len = p->sq_entries * sizeof(struct io_uring_sqe);
f0403f94 JA	425	ld->sqes = mmap(0, ld->mmap[1].len, PROT_READ \| PROT_WRITE,
9a2d78b3	426	MAP_SHARED \| MAP_POPULATE, ld->ring_fd,
f0403f94 JA	427	IORING_OFF_SQES);
f0403f94 JA	428	ld->mmap[1].ptr = ld->sqes;
9a2d78b3	429
f0403f94 JA	430	ld->mmap[2].len = p->cq_off.cqes +
f0403f94 JA	431	p->cq_entries * sizeof(struct io_uring_cqe);
9a2d78b3 JA	432	ptr = mmap(0, ld->mmap[2].len, PROT_READ \| PROT_WRITE,
	433	MAP_SHARED \| MAP_POPULATE, ld->ring_fd,
	434	IORING_OFF_CQ_RING);
	435	ld->mmap[2].ptr = ptr;
	436	cring->head = ptr + p->cq_off.head;
	437	cring->tail = ptr + p->cq_off.tail;
	438	cring->ring_mask = ptr + p->cq_off.ring_mask;
	439	cring->ring_entries = ptr + p->cq_off.ring_entries;
f0403f94	440	cring->cqes = ptr + p->cq_off.cqes;
9a2d78b3 JA	441	ld->cq_ring_mask = *cring->ring_mask;
	442	return 0;
	443	}
	444
bffad86f	445	static int fio_ioring_queue_init(struct thread_data *td)
52885fa2	446	{
bffad86f JA	447	struct ioring_data *ld = td->io_ops_data;
bffad86f JA	448	struct ioring_options *o = td->eo;
52885fa2	449	int depth = td->o.iodepth;
650346e1	450	struct iovec *vecs = NULL;
bffad86f	451	struct io_uring_params p;
9a2d78b3 JA	452	int ret;
	453
	454	memset(&p, 0, sizeof(p));
52885fa2 JA	455
52885fa2 JA	456	if (o->hipri)
bffad86f	457	p.flags \|= IORING_SETUP_IOPOLL;
a90cd050	458	if (o->sqthread_set) {
9a2d78b3	459	p.sq_thread_cpu = o->sqthread;
bffad86f	460	p.flags \|= IORING_SETUP_SQTHREAD;
771c9901	461	if (o->sqthread_poll)
bffad86f	462	p.flags \|= IORING_SETUP_SQPOLL;
f635f1fb JA	463	}
f635f1fb JA	464	if (o->sqwq)
bffad86f	465	p.flags \|= IORING_SETUP_SQWQ;
a90cd050	466
52885fa2 JA	467	if (o->fixedbufs) {
	468	struct rlimit rlim = {
	469	.rlim_cur = RLIM_INFINITY,
	470	.rlim_max = RLIM_INFINITY,
	471	};
	472
	473	setrlimit(RLIMIT_MEMLOCK, &rlim);
650346e1	474	vecs = ld->iovecs;
52885fa2 JA	475	}
52885fa2 JA	476
411a0821	477	ret = syscall(__NR_sys_io_uring_setup, depth, vecs, depth, &p);
9a2d78b3 JA	478	if (ret < 0)
	479	return ret;
	480
	481	ld->ring_fd = ret;
bffad86f	482	return fio_ioring_mmap(ld, &p);
52885fa2 JA	483	}
52885fa2 JA	484
bffad86f	485	static int fio_ioring_post_init(struct thread_data *td)
52885fa2	486	{
bffad86f	487	struct ioring_data *ld = td->io_ops_data;
52885fa2	488	struct io_u *io_u;
650346e1	489	int err, i;
52885fa2	490
650346e1 JA	491	for (i = 0; i < td->o.iodepth; i++) {
650346e1 JA	492	struct iovec *iov = &ld->iovecs[i];
9a2d78b3	493
650346e1 JA	494	io_u = ld->io_u_index[i];
	495	iov->iov_base = io_u->buf;
	496	iov->iov_len = td_max_bs(td);
52885fa2 JA	497	}
52885fa2 JA	498
bffad86f	499	err = fio_ioring_queue_init(td);
52885fa2	500	if (err) {
d63a472d	501	td_verror(td, errno, "io_queue_init");
52885fa2 JA	502	return 1;
	503	}
	504
	505	return 0;
	506	}
	507
9a2d78b3 JA	508	static unsigned roundup_pow2(unsigned depth)
	509	{
	510	return 1UL << __fls(depth - 1);
	511	}
	512
bffad86f	513	static int fio_ioring_init(struct thread_data *td)
52885fa2	514	{
bffad86f	515	struct ioring_data *ld;
52885fa2	516
52885fa2 JA	517	ld = calloc(1, sizeof(*ld));
52885fa2 JA	518
b87aa01a JA	519	/* ring depth must be a power-of-2 */
	520	ld->iodepth = td->o.iodepth;
	521	td->o.iodepth = roundup_pow2(td->o.iodepth);
	522
52885fa2 JA	523	/* io_u index */
	524	ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *));
	525	ld->io_us = calloc(td->o.iodepth, sizeof(struct io_u *));
650346e1	526	ld->iovecs = calloc(td->o.iodepth, sizeof(struct iovec));
52885fa2 JA	527
	528	td->io_ops_data = ld;
	529	return 0;
	530	}
	531
bffad86f	532	static int fio_ioring_io_u_init(struct thread_data td, struct io_u io_u)
52885fa2	533	{
bffad86f	534	struct ioring_data *ld = td->io_ops_data;
52885fa2 JA	535
	536	ld->io_u_index[io_u->index] = io_u;
	537	return 0;
	538	}
	539
	540	static struct ioengine_ops ioengine = {
bffad86f	541	.name = "io_uring",
52885fa2	542	.version = FIO_IOOPS_VERSION,
bffad86f JA	543	.init = fio_ioring_init,
	544	.post_init = fio_ioring_post_init,
	545	.io_u_init = fio_ioring_io_u_init,
	546	.prep = fio_ioring_prep,
	547	.queue = fio_ioring_queue,
	548	.commit = fio_ioring_commit,
	549	.getevents = fio_ioring_getevents,
	550	.event = fio_ioring_event,
	551	.cleanup = fio_ioring_cleanup,
52885fa2 JA	552	.open_file = generic_open_file,
	553	.close_file = generic_close_file,
	554	.get_file_size = generic_get_file_size,
	555	.options = options,
bffad86f	556	.option_struct_size = sizeof(struct ioring_options),
52885fa2 JA	557	};
52885fa2 JA	558
bffad86f	559	static void fio_init fio_ioring_register(void)
52885fa2	560	{
52885fa2	561	register_ioengine(&ioengine);
52885fa2 JA	562	}
52885fa2 JA	563
bffad86f	564	static void fio_exit fio_ioring_unregister(void)
52885fa2	565	{
52885fa2	566	unregister_ioengine(&ioengine);
52885fa2	567	}
1f90e9bb	568	#endif