[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_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_thread;
	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 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_LIBAIO,
	},
	{
		.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_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 (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->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 = &ld->iovecs[io_u->index];
			sqe->len = 1;
		}
		sqe->off = io_u->offset;
	} else if (ddir_sync(io_u->ddir)) {
		sqe->fsync_flags = 0;
		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 *) 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;

	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_thread) {
			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;

	/*
	 * 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, 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->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_thread) {
		p.flags |= IORING_SETUP_SQPOLL;
		if (o->sqpoll_set) {
			p.flags |= 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->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_u_index;
	52
	53	struct io_sq_ring sq_ring;
	54	struct io_uring_sqe *sqes;
	55	struct iovec *iovecs;
	56	unsigned sq_ring_mask;
	57
	58	struct io_cq_ring cq_ring;
	59	unsigned cq_ring_mask;
	60
	61	int queued;
	62	int cq_ring_off;
	63	unsigned iodepth;
	64
	65	uint64_t cachehit;
	66	uint64_t cachemiss;
	67
	68	struct ioring_mmap mmap[3];
	69	};
	70
	71	struct ioring_options {
	72	void *pad;
	73	unsigned int hipri;
	74	unsigned int fixedbufs;
	75	unsigned int sqpoll_thread;
	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 polling",
	111	.type = FIO_OPT_INT,
	112	.off1 = offsetof(struct ioring_options, sqpoll_thread),
	113	.help = "Offload submission/completion to kernel thread",
	114	.category = FIO_OPT_C_ENGINE,
	115	.group = FIO_OPT_G_LIBAIO,
	116	},
	117	{
	118	.name = "sqthread_poll_cpu",
	119	.lname = "SQ Thread Poll CPU",
	120	.type = FIO_OPT_INT,
	121	.cb = fio_ioring_sqpoll_cb,
	122	.help = "What CPU to run SQ thread polling on",
	123	.category = FIO_OPT_C_ENGINE,
	124	.group = FIO_OPT_G_LIBAIO,
	125	},
	126	{
	127	.name = NULL,
	128	},
	129	};
	130
	131	static int io_uring_enter(struct ioring_data *ld, unsigned int to_submit,
	132	unsigned int min_complete, unsigned int flags)
	133	{
	134	return syscall(__NR_sys_io_uring_enter, ld->ring_fd, to_submit,
	135	min_complete, flags);
	136	}
	137
	138	static int fio_ioring_prep(struct thread_data td, struct io_u io_u)
	139	{
	140	struct ioring_data *ld = td->io_ops_data;
	141	struct ioring_options *o = td->eo;
	142	struct fio_file *f = io_u->file;
	143	struct io_uring_sqe *sqe;
	144
	145	sqe = &ld->sqes[io_u->index];
	146	sqe->fd = f->fd;
	147	sqe->flags = 0;
	148	sqe->ioprio = 0;
	149	sqe->buf_index = 0;
	150
	151	if (io_u->ddir == DDIR_READ \|\| io_u->ddir == DDIR_WRITE) {
	152	if (o->fixedbufs) {
	153	if (io_u->ddir == DDIR_READ)
	154	sqe->opcode = IORING_OP_READ_FIXED;
	155	else
	156	sqe->opcode = IORING_OP_WRITE_FIXED;
	157	sqe->addr = io_u->xfer_buf;
	158	sqe->len = io_u->xfer_buflen;
	159	sqe->buf_index = io_u->index;
	160	} else {
	161	if (io_u->ddir == DDIR_READ)
	162	sqe->opcode = IORING_OP_READV;
	163	else
	164	sqe->opcode = IORING_OP_WRITEV;
	165	sqe->addr = &ld->iovecs[io_u->index];
	166	sqe->len = 1;
	167	}
	168	sqe->off = io_u->offset;
	169	} else if (ddir_sync(io_u->ddir)) {
	170	sqe->fsync_flags = 0;
	171	if (io_u->ddir == DDIR_DATASYNC)
	172	sqe->fsync_flags \|= IORING_FSYNC_DATASYNC;
	173	sqe->opcode = IORING_OP_FSYNC;
	174	}
	175
	176	sqe->user_data = (unsigned long) io_u;
	177	return 0;
	178	}
	179
	180	static struct io_u fio_ioring_event(struct thread_data td, int event)
	181	{
	182	struct ioring_data *ld = td->io_ops_data;
	183	struct io_uring_cqe *cqe;
	184	struct io_u *io_u;
	185	unsigned index;
	186
	187	index = (event + ld->cq_ring_off) & ld->cq_ring_mask;
	188
	189	cqe = &ld->cq_ring.cqes[index];
	190	io_u = (struct io_u *) cqe->user_data;
	191
	192	if (cqe->res != io_u->xfer_buflen) {
	193	if (cqe->res > io_u->xfer_buflen)
	194	io_u->error = -cqe->res;
	195	else
	196	io_u->resid = io_u->xfer_buflen - cqe->res;
	197	} else
	198	io_u->error = 0;
	199
	200	if (io_u->ddir == DDIR_READ) {
	201	if (cqe->flags & IOCQE_FLAG_CACHEHIT)
	202	ld->cachehit++;
	203	else
	204	ld->cachemiss++;
	205	}
	206
	207	return io_u;
	208	}
	209
	210	static int fio_ioring_cqring_reap(struct thread_data *td, unsigned int events,
	211	unsigned int max)
	212	{
	213	struct ioring_data *ld = td->io_ops_data;
	214	struct io_cq_ring *ring = &ld->cq_ring;
	215	unsigned head, reaped = 0;
	216
	217	head = *ring->head;
	218	do {
	219	read_barrier();
	220	if (head == *ring->tail)
	221	break;
	222	reaped++;
	223	head++;
	224	} while (reaped + events < max);
	225
	226	*ring->head = head;
	227	write_barrier();
	228	return reaped;
	229	}
	230
	231	static int fio_ioring_getevents(struct thread_data *td, unsigned int min,
	232	unsigned int max, const struct timespec *t)
	233	{
	234	struct ioring_data *ld = td->io_ops_data;
	235	unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
	236	struct ioring_options *o = td->eo;
	237	struct io_cq_ring *ring = &ld->cq_ring;
	238	unsigned events = 0;
	239	int r;
	240
	241	ld->cq_ring_off = *ring->head;
	242	do {
	243	r = fio_ioring_cqring_reap(td, events, max);
	244	if (r) {
	245	events += r;
	246	continue;
	247	}
	248
	249	if (!o->sqpoll_thread) {
	250	r = io_uring_enter(ld, 0, actual_min,
	251	IORING_ENTER_GETEVENTS);
	252	if (r < 0) {
	253	if (errno == EAGAIN)
	254	continue;
	255	td_verror(td, errno, "io_uring_enter");
	256	break;
	257	}
	258	}
	259	} while (events < min);
	260
	261	return r < 0 ? r : events;
	262	}
	263
	264	static enum fio_q_status fio_ioring_queue(struct thread_data *td,
	265	struct io_u *io_u)
	266	{
	267	struct ioring_data *ld = td->io_ops_data;
	268	struct io_sq_ring *ring = &ld->sq_ring;
	269	unsigned tail, next_tail;
	270
	271	fio_ro_check(td, io_u);
	272
	273	if (ld->queued == ld->iodepth)
	274	return FIO_Q_BUSY;
	275
	276	if (io_u->ddir == DDIR_TRIM) {
	277	if (ld->queued)
	278	return FIO_Q_BUSY;
	279
	280	do_io_u_trim(td, io_u);
	281	io_u_mark_submit(td, 1);
	282	io_u_mark_complete(td, 1);
	283	return FIO_Q_COMPLETED;
	284	}
	285
	286	tail = *ring->tail;
	287	next_tail = tail + 1;
	288	read_barrier();
	289	if (next_tail == *ring->head)
	290	return FIO_Q_BUSY;
	291
	292	ring->array[tail & ld->sq_ring_mask] = io_u->index;
	293	*ring->tail = next_tail;
	294	write_barrier();
	295
	296	ld->queued++;
	297	return FIO_Q_QUEUED;
	298	}
	299
	300	static void fio_ioring_queued(struct thread_data *td, int start, int nr)
	301	{
	302	struct ioring_data *ld = td->io_ops_data;
	303	struct timespec now;
	304
	305	if (!fio_fill_issue_time(td))
	306	return;
	307
	308	fio_gettime(&now, NULL);
	309
	310	while (nr--) {
	311	struct io_sq_ring *ring = &ld->sq_ring;
	312	int index = ring->array[start & ld->sq_ring_mask];
	313	struct io_u *io_u = ld->io_u_index[index];
	314
	315	memcpy(&io_u->issue_time, &now, sizeof(now));
	316	io_u_queued(td, io_u);
	317
	318	start++;
	319	}
	320	}
	321
	322	static int fio_ioring_commit(struct thread_data *td)
	323	{
	324	struct ioring_data *ld = td->io_ops_data;
	325	struct ioring_options *o = td->eo;
	326	int ret;
	327
	328	if (!ld->queued)
	329	return 0;
	330
	331	/*
	332	* Kernel side does submission. just need to check if the ring is
	333	* flagged as needing a kick, if so, call io_uring_enter(). This
	334	* only happens if we've been idle too long.
	335	*/
	336	if (o->sqpoll_thread) {
	337	struct io_sq_ring *ring = &ld->sq_ring;
	338
	339	read_barrier();
	340	if (*ring->flags & IORING_SQ_NEED_WAKEUP)
	341	io_uring_enter(ld, ld->queued, 0, 0);
	342	ld->queued = 0;
	343	return 0;
	344	}
	345
	346	do {
	347	unsigned start = *ld->sq_ring.head;
	348	long nr = ld->queued;
	349
	350	ret = io_uring_enter(ld, nr, 0, IORING_ENTER_GETEVENTS);
	351	if (ret > 0) {
	352	fio_ioring_queued(td, start, ret);
	353	io_u_mark_submit(td, ret);
	354
	355	ld->queued -= ret;
	356	ret = 0;
	357	} else if (!ret) {
	358	io_u_mark_submit(td, ret);
	359	continue;
	360	} else {
	361	if (errno == EAGAIN) {
	362	ret = fio_ioring_cqring_reap(td, 0, ld->queued);
	363	if (ret)
	364	continue;
	365	/* Shouldn't happen */
	366	usleep(1);
	367	continue;
	368	}
	369	td_verror(td, errno, "io_uring_enter submit");
	370	break;
	371	}
	372	} while (ld->queued);
	373
	374	return ret;
	375	}
	376
	377	static void fio_ioring_unmap(struct ioring_data *ld)
	378	{
	379	int i;
	380
	381	for (i = 0; i < ARRAY_SIZE(ld->mmap); i++)
	382	munmap(ld->mmap[i].ptr, ld->mmap[i].len);
	383	close(ld->ring_fd);
	384	}
	385
	386	static void fio_ioring_cleanup(struct thread_data *td)
	387	{
	388	struct ioring_data *ld = td->io_ops_data;
	389
	390	if (ld) {
	391	td->ts.cachehit += ld->cachehit;
	392	td->ts.cachemiss += ld->cachemiss;
	393
	394	if (!(td->flags & TD_F_CHILD))
	395	fio_ioring_unmap(ld);
	396
	397	free(ld->io_u_index);
	398	free(ld->iovecs);
	399	free(ld);
	400	}
	401	}
	402
	403	static int fio_ioring_mmap(struct ioring_data ld, struct io_uring_params p)
	404	{
	405	struct io_sq_ring *sring = &ld->sq_ring;
	406	struct io_cq_ring *cring = &ld->cq_ring;
	407	void *ptr;
	408
	409	ld->mmap[0].len = p->sq_off.array + p->sq_entries * sizeof(__u32);
	410	ptr = mmap(0, ld->mmap[0].len, PROT_READ \| PROT_WRITE,
	411	MAP_SHARED \| MAP_POPULATE, ld->ring_fd,
	412	IORING_OFF_SQ_RING);
	413	ld->mmap[0].ptr = ptr;
	414	sring->head = ptr + p->sq_off.head;
	415	sring->tail = ptr + p->sq_off.tail;
	416	sring->ring_mask = ptr + p->sq_off.ring_mask;
	417	sring->ring_entries = ptr + p->sq_off.ring_entries;
	418	sring->flags = ptr + p->sq_off.flags;
	419	sring->array = ptr + p->sq_off.array;
	420	ld->sq_ring_mask = *sring->ring_mask;
	421
	422	ld->mmap[1].len = p->sq_entries * sizeof(struct io_uring_sqe);
	423	ld->sqes = mmap(0, ld->mmap[1].len, PROT_READ \| PROT_WRITE,
	424	MAP_SHARED \| MAP_POPULATE, ld->ring_fd,
	425	IORING_OFF_SQES);
	426	ld->mmap[1].ptr = ld->sqes;
	427
	428	ld->mmap[2].len = p->cq_off.cqes +
	429	p->cq_entries * sizeof(struct io_uring_cqe);
	430	ptr = mmap(0, ld->mmap[2].len, PROT_READ \| PROT_WRITE,
	431	MAP_SHARED \| MAP_POPULATE, ld->ring_fd,
	432	IORING_OFF_CQ_RING);
	433	ld->mmap[2].ptr = ptr;
	434	cring->head = ptr + p->cq_off.head;
	435	cring->tail = ptr + p->cq_off.tail;
	436	cring->ring_mask = ptr + p->cq_off.ring_mask;
	437	cring->ring_entries = ptr + p->cq_off.ring_entries;
	438	cring->cqes = ptr + p->cq_off.cqes;
	439	ld->cq_ring_mask = *cring->ring_mask;
	440	return 0;
	441	}
	442
	443	static int fio_ioring_queue_init(struct thread_data *td)
	444	{
	445	struct ioring_data *ld = td->io_ops_data;
	446	struct ioring_options *o = td->eo;
	447	int depth = td->o.iodepth;
	448	struct io_uring_params p;
	449	int ret;
	450
	451	memset(&p, 0, sizeof(p));
	452
	453	if (o->hipri)
	454	p.flags \|= IORING_SETUP_IOPOLL;
	455	if (o->sqpoll_thread) {
	456	p.flags \|= IORING_SETUP_SQPOLL;
	457	if (o->sqpoll_set) {
	458	p.flags \|= IORING_SETUP_SQ_AFF;
	459	p.sq_thread_cpu = o->sqpoll_cpu;
	460	}
	461	}
	462
	463	if (o->fixedbufs) {
	464	struct rlimit rlim = {
	465	.rlim_cur = RLIM_INFINITY,
	466	.rlim_max = RLIM_INFINITY,
	467	};
	468
	469	setrlimit(RLIMIT_MEMLOCK, &rlim);
	470	}
	471
	472	ret = syscall(__NR_sys_io_uring_setup, depth, &p);
	473	if (ret < 0)
	474	return ret;
	475
	476	ld->ring_fd = ret;
	477
	478	if (o->fixedbufs) {
	479	struct io_uring_register_buffers reg = {
	480	.iovecs = ld->iovecs,
	481	.nr_iovecs = depth
	482	};
	483
	484	ret = syscall(__NR_sys_io_uring_register, ld->ring_fd,
	485	IORING_REGISTER_BUFFERS, &reg);
	486	if (ret < 0)
	487	return ret;
	488	}
	489
	490	return fio_ioring_mmap(ld, &p);
	491	}
	492
	493	static int fio_ioring_post_init(struct thread_data *td)
	494	{
	495	struct ioring_data *ld = td->io_ops_data;
	496	struct io_u *io_u;
	497	int err, i;
	498
	499	for (i = 0; i < td->o.iodepth; i++) {
	500	struct iovec *iov = &ld->iovecs[i];
	501
	502	io_u = ld->io_u_index[i];
	503	iov->iov_base = io_u->buf;
	504	iov->iov_len = td_max_bs(td);
	505	}
	506
	507	err = fio_ioring_queue_init(td);
	508	if (err) {
	509	td_verror(td, errno, "io_queue_init");
	510	return 1;
	511	}
	512
	513	return 0;
	514	}
	515
	516	static unsigned roundup_pow2(unsigned depth)
	517	{
	518	return 1UL << __fls(depth - 1);
	519	}
	520
	521	static int fio_ioring_init(struct thread_data *td)
	522	{
	523	struct ioring_data *ld;
	524
	525	ld = calloc(1, sizeof(*ld));
	526
	527	/* ring depth must be a power-of-2 */
	528	ld->iodepth = td->o.iodepth;
	529	td->o.iodepth = roundup_pow2(td->o.iodepth);
	530
	531	/* io_u index */
	532	ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *));
	533	ld->iovecs = calloc(td->o.iodepth, sizeof(struct iovec));
	534
	535	td->io_ops_data = ld;
	536	return 0;
	537	}
	538
	539	static int fio_ioring_io_u_init(struct thread_data td, struct io_u io_u)
	540	{
	541	struct ioring_data *ld = td->io_ops_data;
	542
	543	ld->io_u_index[io_u->index] = io_u;
	544	return 0;
	545	}
	546
	547	static struct ioengine_ops ioengine = {
	548	.name = "io_uring",
	549	.version = FIO_IOOPS_VERSION,
	550	.init = fio_ioring_init,
	551	.post_init = fio_ioring_post_init,
	552	.io_u_init = fio_ioring_io_u_init,
	553	.prep = fio_ioring_prep,
	554	.queue = fio_ioring_queue,
	555	.commit = fio_ioring_commit,
	556	.getevents = fio_ioring_getevents,
	557	.event = fio_ioring_event,
	558	.cleanup = fio_ioring_cleanup,
	559	.open_file = generic_open_file,
	560	.close_file = generic_close_file,
	561	.get_file_size = generic_get_file_size,
	562	.options = options,
	563	.option_struct_size = sizeof(struct ioring_options),
	564	};
	565
	566	static void fio_init fio_ioring_register(void)
	567	{
	568	register_ioengine(&ioengine);
	569	}
	570
	571	static void fio_exit fio_ioring_unregister(void)
	572	{
	573	unregister_ioengine(&ioengine);
	574	}
	575	#endif