[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;
		} 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;

	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 = ld->io_u_index[cqe->index];

	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 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);
	}

	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_options *o = td->eo;
	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 *));

	if (o->fixedbufs)
		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;
cfcc8564	169	} else {
f0403f94 JA	170	if (io_u->ddir == DDIR_READ)
f0403f94 JA	171	sqe->opcode = IORING_OP_READV;
cfcc8564	172	else
f0403f94 JA	173	sqe->opcode = IORING_OP_WRITEV;
	174	sqe->addr = &ld->iovecs[io_u->index];
	175	sqe->len = 1;
cfcc8564	176	}
f0403f94	177	sqe->off = io_u->offset;
52885fa2	178	} else if (ddir_sync(io_u->ddir))
f0403f94	179	sqe->opcode = IORING_OP_FSYNC;
52885fa2	180
52885fa2 JA	181	return 0;
	182	}
	183
bffad86f	184	static struct io_u fio_ioring_event(struct thread_data td, int event)
52885fa2	185	{
bffad86f	186	struct ioring_data *ld = td->io_ops_data;
f0403f94	187	struct io_uring_cqe *cqe;
52885fa2	188	struct io_u *io_u;
b87aa01a	189	unsigned index;
52885fa2	190
b87aa01a	191	index = (event + ld->cq_ring_off) & ld->cq_ring_mask;
52885fa2	192
f0403f94 JA	193	cqe = &ld->cq_ring.cqes[index];
f0403f94 JA	194	io_u = ld->io_u_index[cqe->index];
52885fa2	195
f0403f94 JA	196	if (cqe->res != io_u->xfer_buflen) {
	197	if (cqe->res > io_u->xfer_buflen)
	198	io_u->error = -cqe->res;
52885fa2	199	else
f0403f94	200	io_u->resid = io_u->xfer_buflen - cqe->res;
52885fa2 JA	201	} else
	202	io_u->error = 0;
	203
96563db9	204	if (io_u->ddir == DDIR_READ) {
f0403f94	205	if (cqe->flags & IOCQE_FLAG_CACHEHIT)
96563db9 JA	206	ld->cachehit++;
	207	else
	208	ld->cachemiss++;
	209	}
	210
52885fa2 JA	211	return io_u;
	212	}
	213
bffad86f	214	static int fio_ioring_cqring_reap(struct thread_data *td, unsigned int events,
52885fa2 JA	215	unsigned int max)
52885fa2 JA	216	{
bffad86f JA	217	struct ioring_data *ld = td->io_ops_data;
bffad86f JA	218	struct io_cq_ring *ring = &ld->cq_ring;
e2239016	219	unsigned head, reaped = 0;
52885fa2	220
9a2d78b3	221	head = *ring->head;
52885fa2 JA	222	do {
52885fa2 JA	223	read_barrier();
9a2d78b3	224	if (head == *ring->tail)
52885fa2 JA	225	break;
	226	reaped++;
	227	head++;
52885fa2 JA	228	} while (reaped + events < max);
52885fa2 JA	229
9a2d78b3	230	*ring->head = head;
52885fa2 JA	231	write_barrier();
	232	return reaped;
	233	}
	234
bffad86f JA	235	static int fio_ioring_getevents(struct thread_data *td, unsigned int min,
bffad86f JA	236	unsigned int max, const struct timespec *t)
52885fa2	237	{
bffad86f	238	struct ioring_data *ld = td->io_ops_data;
52885fa2	239	unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
bffad86f JA	240	struct ioring_options *o = td->eo;
bffad86f JA	241	struct io_cq_ring *ring = &ld->cq_ring;
b87aa01a JA	242	unsigned events = 0;
b87aa01a JA	243	int r;
52885fa2	244
9a2d78b3	245	ld->cq_ring_off = *ring->head;
52885fa2	246	do {
bffad86f	247	r = fio_ioring_cqring_reap(td, events, max);
52885fa2 JA	248	if (r) {
	249	events += r;
	250	continue;
	251	}
	252
771c9901	253	if (!o->sqthread_poll) {
9a2d78b3 JA	254	r = io_uring_enter(ld, 0, actual_min,
9a2d78b3 JA	255	IORING_ENTER_GETEVENTS);
771c9901 JA	256	if (r < 0) {
	257	if (errno == EAGAIN)
	258	continue;
9a2d78b3	259	td_verror(td, errno, "io_uring_enter");
771c9901 JA	260	break;
771c9901 JA	261	}
52885fa2 JA	262	}
	263	} while (events < min);
	264
	265	return r < 0 ? r : events;
	266	}
	267
bffad86f JA	268	static enum fio_q_status fio_ioring_queue(struct thread_data *td,
bffad86f JA	269	struct io_u *io_u)
52885fa2	270	{
bffad86f JA	271	struct ioring_data *ld = td->io_ops_data;
bffad86f JA	272	struct io_sq_ring *ring = &ld->sq_ring;
52885fa2 JA	273	unsigned tail, next_tail;
	274
	275	fio_ro_check(td, io_u);
	276
b87aa01a	277	if (ld->queued == ld->iodepth)
52885fa2 JA	278	return FIO_Q_BUSY;
52885fa2 JA	279
52885fa2 JA	280	if (io_u->ddir == DDIR_TRIM) {
	281	if (ld->queued)
	282	return FIO_Q_BUSY;
	283
	284	do_io_u_trim(td, io_u);
	285	io_u_mark_submit(td, 1);
	286	io_u_mark_complete(td, 1);
	287	return FIO_Q_COMPLETED;
	288	}
	289
9a2d78b3	290	tail = *ring->tail;
52885fa2	291	next_tail = tail + 1;
52885fa2	292	read_barrier();
9a2d78b3	293	if (next_tail == *ring->head)
52885fa2 JA	294	return FIO_Q_BUSY;
52885fa2 JA	295
b87aa01a	296	ring->array[tail & ld->sq_ring_mask] = io_u->index;
9a2d78b3	297	*ring->tail = next_tail;
52885fa2 JA	298	write_barrier();
	299
	300	ld->queued++;
	301	return FIO_Q_QUEUED;
	302	}
	303
bffad86f	304	static void fio_ioring_queued(struct thread_data *td, int start, int nr)
52885fa2	305	{
bffad86f	306	struct ioring_data *ld = td->io_ops_data;
52885fa2 JA	307	struct timespec now;
	308
	309	if (!fio_fill_issue_time(td))
	310	return;
	311
	312	fio_gettime(&now, NULL);
	313
	314	while (nr--) {
bffad86f	315	struct io_sq_ring *ring = &ld->sq_ring;
9a2d78b3	316	int index = ring->array[start & ld->sq_ring_mask];
f8289afc	317	struct io_u *io_u = ld->io_u_index[index];
52885fa2 JA	318
	319	memcpy(&io_u->issue_time, &now, sizeof(now));
	320	io_u_queued(td, io_u);
	321
	322	start++;
52885fa2 JA	323	}
	324	}
	325
bffad86f	326	static int fio_ioring_commit(struct thread_data *td)
52885fa2	327	{
bffad86f JA	328	struct ioring_data *ld = td->io_ops_data;
bffad86f JA	329	struct ioring_options *o = td->eo;
52885fa2 JA	330	int ret;
	331
	332	if (!ld->queued)
	333	return 0;
	334
771c9901 JA	335	/* Nothing to do */
771c9901 JA	336	if (o->sqthread_poll) {
bffad86f	337	struct io_sq_ring *ring = &ld->sq_ring;
4cdbc048	338
9a2d78b3 JA	339	if (*ring->flags & IORING_SQ_NEED_WAKEUP)
9a2d78b3 JA	340	io_uring_enter(ld, ld->queued, 0, 0);
771c9901 JA	341	ld->queued = 0;
	342	return 0;
	343	}
	344
52885fa2	345	do {
9a2d78b3	346	unsigned start = *ld->sq_ring.head;
52885fa2 JA	347	long nr = ld->queued;
52885fa2 JA	348
9a2d78b3	349	ret = io_uring_enter(ld, nr, 0, IORING_ENTER_GETEVENTS);
52885fa2	350	if (ret > 0) {
bffad86f	351	fio_ioring_queued(td, start, ret);
52885fa2 JA	352	io_u_mark_submit(td, ret);
	353
	354	ld->queued -= ret;
	355	ret = 0;
a90cd050 JA	356	} else if (!ret) {
a90cd050 JA	357	io_u_mark_submit(td, ret);
52885fa2	358	continue;
a90cd050 JA	359	} else {
a90cd050 JA	360	if (errno == EAGAIN) {
bffad86f	361	ret = fio_ioring_cqring_reap(td, 0, ld->queued);
a90cd050 JA	362	if (ret)
	363	continue;
	364	/* Shouldn't happen */
	365	usleep(1);
	366	continue;
52885fa2	367	}
9a2d78b3	368	td_verror(td, errno, "io_uring_enter submit");
52885fa2	369	break;
a90cd050	370	}
52885fa2 JA	371	} while (ld->queued);
	372
	373	return ret;
	374	}
	375
bffad86f	376	static void fio_ioring_unmap(struct ioring_data *ld)
52885fa2	377	{
9a2d78b3	378	int i;
52885fa2	379
9a2d78b3 JA	380	for (i = 0; i < ARRAY_SIZE(ld->mmap); i++)
	381	munmap(ld->mmap[i].ptr, ld->mmap[i].len);
	382	close(ld->ring_fd);
b87aa01a JA	383	}
b87aa01a JA	384
bffad86f	385	static void fio_ioring_cleanup(struct thread_data *td)
52885fa2	386	{
bffad86f	387	struct ioring_data *ld = td->io_ops_data;
52885fa2 JA	388
52885fa2 JA	389	if (ld) {
96563db9 JA	390	td->ts.cachehit += ld->cachehit;
	391	td->ts.cachemiss += ld->cachemiss;
	392
52885fa2	393	if (!(td->flags & TD_F_CHILD))
bffad86f	394	fio_ioring_unmap(ld);
9a2d78b3	395
52885fa2 JA	396	free(ld->io_u_index);
52885fa2 JA	397	free(ld->io_us);
9a2d78b3	398	free(ld->iovecs);
52885fa2 JA	399	free(ld);
	400	}
	401	}
	402
bffad86f	403	static int fio_ioring_mmap(struct ioring_data ld, struct io_uring_params p)
9a2d78b3	404	{
bffad86f JA	405	struct io_sq_ring *sring = &ld->sq_ring;
bffad86f JA	406	struct io_cq_ring *cring = &ld->cq_ring;
9a2d78b3 JA	407	void *ptr;
9a2d78b3 JA	408
e2239016	409	ld->mmap[0].len = p->sq_off.array + p->sq_entries * sizeof(__u32);
9a2d78b3 JA	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;
ac122fea	419	sring->array = ptr + p->sq_off.array;
9a2d78b3 JA	420	ld->sq_ring_mask = *sring->ring_mask;
9a2d78b3 JA	421
f0403f94 JA	422	ld->mmap[1].len = p->sq_entries * sizeof(struct io_uring_sqe);
f0403f94 JA	423	ld->sqes = mmap(0, ld->mmap[1].len, PROT_READ \| PROT_WRITE,
9a2d78b3	424	MAP_SHARED \| MAP_POPULATE, ld->ring_fd,
f0403f94 JA	425	IORING_OFF_SQES);
f0403f94 JA	426	ld->mmap[1].ptr = ld->sqes;
9a2d78b3	427
f0403f94 JA	428	ld->mmap[2].len = p->cq_off.cqes +
f0403f94 JA	429	p->cq_entries * sizeof(struct io_uring_cqe);
9a2d78b3 JA	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;
f0403f94	438	cring->cqes = ptr + p->cq_off.cqes;
9a2d78b3 JA	439	ld->cq_ring_mask = *cring->ring_mask;
	440	return 0;
	441	}
	442
bffad86f	443	static int fio_ioring_queue_init(struct thread_data *td)
52885fa2	444	{
bffad86f JA	445	struct ioring_data *ld = td->io_ops_data;
bffad86f JA	446	struct ioring_options *o = td->eo;
52885fa2	447	int depth = td->o.iodepth;
bffad86f	448	struct io_uring_params p;
9a2d78b3 JA	449	int ret;
	450
	451	memset(&p, 0, sizeof(p));
52885fa2 JA	452
52885fa2 JA	453	if (o->hipri)
bffad86f	454	p.flags \|= IORING_SETUP_IOPOLL;
a90cd050	455	if (o->sqthread_set) {
9a2d78b3	456	p.sq_thread_cpu = o->sqthread;
bffad86f	457	p.flags \|= IORING_SETUP_SQTHREAD;
771c9901	458	if (o->sqthread_poll)
bffad86f	459	p.flags \|= IORING_SETUP_SQPOLL;
f635f1fb JA	460	}
f635f1fb JA	461	if (o->sqwq)
bffad86f	462	p.flags \|= IORING_SETUP_SQWQ;
a90cd050	463
52885fa2 JA	464	if (o->fixedbufs) {
	465	struct rlimit rlim = {
	466	.rlim_cur = RLIM_INFINITY,
	467	.rlim_max = RLIM_INFINITY,
	468	};
	469
	470	setrlimit(RLIMIT_MEMLOCK, &rlim);
52885fa2 JA	471	}
52885fa2 JA	472
9a2d78b3 JA	473	ret = syscall(__NR_sys_io_uring_setup, depth, ld->iovecs, &p);
	474	if (ret < 0)
	475	return ret;
	476
	477	ld->ring_fd = ret;
bffad86f	478	return fio_ioring_mmap(ld, &p);
52885fa2 JA	479	}
52885fa2 JA	480
bffad86f	481	static int fio_ioring_post_init(struct thread_data *td)
52885fa2	482	{
bffad86f JA	483	struct ioring_data *ld = td->io_ops_data;
bffad86f JA	484	struct ioring_options *o = td->eo;
52885fa2	485	struct io_u *io_u;
9a2d78b3	486	int err;
52885fa2 JA	487
	488	if (o->fixedbufs) {
	489	int i;
	490
	491	for (i = 0; i < td->o.iodepth; i++) {
9a2d78b3 JA	492	struct iovec *iov = &ld->iovecs[i];
9a2d78b3 JA	493
52885fa2	494	io_u = ld->io_u_index[i];
9a2d78b3 JA	495	iov->iov_base = io_u->buf;
9a2d78b3 JA	496	iov->iov_len = td_max_bs(td);
52885fa2 JA	497	}
	498	}
	499
bffad86f	500	err = fio_ioring_queue_init(td);
52885fa2	501	if (err) {
d63a472d	502	td_verror(td, errno, "io_queue_init");
52885fa2 JA	503	return 1;
	504	}
	505
	506	return 0;
	507	}
	508
9a2d78b3 JA	509	static unsigned roundup_pow2(unsigned depth)
	510	{
	511	return 1UL << __fls(depth - 1);
	512	}
	513
bffad86f	514	static int fio_ioring_init(struct thread_data *td)
52885fa2	515	{
f0403f94	516	struct ioring_options *o = td->eo;
bffad86f	517	struct ioring_data *ld;
52885fa2	518
52885fa2 JA	519	ld = calloc(1, sizeof(*ld));
52885fa2 JA	520
b87aa01a JA	521	/* ring depth must be a power-of-2 */
	522	ld->iodepth = td->o.iodepth;
	523	td->o.iodepth = roundup_pow2(td->o.iodepth);
	524
52885fa2 JA	525	/* io_u index */
	526	ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *));
	527	ld->io_us = calloc(td->o.iodepth, sizeof(struct io_u *));
	528
f0403f94 JA	529	if (o->fixedbufs)
f0403f94 JA	530	ld->iovecs = calloc(td->o.iodepth, sizeof(struct iovec));
52885fa2 JA	531
	532	td->io_ops_data = ld;
	533	return 0;
	534	}
	535
bffad86f	536	static int fio_ioring_io_u_init(struct thread_data td, struct io_u io_u)
52885fa2	537	{
bffad86f	538	struct ioring_data *ld = td->io_ops_data;
52885fa2 JA	539
	540	ld->io_u_index[io_u->index] = io_u;
	541	return 0;
	542	}
	543
	544	static struct ioengine_ops ioengine = {
bffad86f	545	.name = "io_uring",
52885fa2	546	.version = FIO_IOOPS_VERSION,
bffad86f JA	547	.init = fio_ioring_init,
	548	.post_init = fio_ioring_post_init,
	549	.io_u_init = fio_ioring_io_u_init,
	550	.prep = fio_ioring_prep,
	551	.queue = fio_ioring_queue,
	552	.commit = fio_ioring_commit,
	553	.getevents = fio_ioring_getevents,
	554	.event = fio_ioring_event,
	555	.cleanup = fio_ioring_cleanup,
52885fa2 JA	556	.open_file = generic_open_file,
	557	.close_file = generic_close_file,
	558	.get_file_size = generic_get_file_size,
	559	.options = options,
bffad86f	560	.option_struct_size = sizeof(struct ioring_options),
52885fa2 JA	561	};
52885fa2 JA	562
bffad86f	563	static void fio_init fio_ioring_register(void)
52885fa2	564	{
52885fa2	565	register_ioengine(&ioengine);
52885fa2 JA	566	}
52885fa2 JA	567
bffad86f	568	static void fio_exit fio_ioring_unregister(void)
52885fa2	569	{
52885fa2	570	unregister_ioengine(&ioengine);
52885fa2	571	}
1f90e9bb	572	#endif