[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

typedef uint64_t u64;
typedef uint32_t u32;
typedef int32_t s32;
typedef uint16_t u16;
typedef uint8_t u8;

#include "../os/io_uring.h"

struct io_sq_ring {
	u32 *head;
	u32 *tail;
	u32 *ring_mask;
	u32 *ring_entries;
	u32 *flags;
	u32 *array;
};

struct io_cq_ring {
	u32 *head;
	u32 *tail;
	u32 *ring_mask;
	u32 *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;
	u32 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 JA	22
	23	typedef uint64_t u64;
	24	typedef uint32_t u32;
bffad86f	25	typedef int32_t s32;
52885fa2	26	typedef uint16_t u16;
bffad86f	27	typedef uint8_t u8;
52885fa2	28
bffad86f	29	#include "../os/io_uring.h"
9a2d78b3	30
bffad86f	31	struct io_sq_ring {
9a2d78b3 JA	32	u32 *head;
	33	u32 *tail;
	34	u32 *ring_mask;
	35	u32 *ring_entries;
	36	u32 *flags;
	37	u32 *array;
52885fa2 JA	38	};
52885fa2 JA	39
bffad86f	40	struct io_cq_ring {
9a2d78b3 JA	41	u32 *head;
	42	u32 *tail;
	43	u32 *ring_mask;
	44	u32 *ring_entries;
bffad86f	45	struct io_uring_event *events;
9a2d78b3 JA	46	};
9a2d78b3 JA	47
bffad86f	48	struct ioring_mmap {
9a2d78b3 JA	49	void *ptr;
9a2d78b3 JA	50	size_t len;
52885fa2 JA	51	};
52885fa2 JA	52
bffad86f	53	struct ioring_data {
9a2d78b3 JA	54	int ring_fd;
9a2d78b3 JA	55
52885fa2 JA	56	struct io_u **io_us;
	57	struct io_u **io_u_index;
	58
bffad86f JA	59	struct io_sq_ring sq_ring;
bffad86f JA	60	struct io_uring_iocb *iocbs;
9a2d78b3	61	struct iovec *iovecs;
b87aa01a	62	unsigned sq_ring_mask;
52885fa2	63
bffad86f	64	struct io_cq_ring cq_ring;
b87aa01a	65	unsigned cq_ring_mask;
52885fa2 JA	66
	67	int queued;
	68	int cq_ring_off;
b87aa01a	69	unsigned iodepth;
96563db9 JA	70
	71	uint64_t cachehit;
	72	uint64_t cachemiss;
9a2d78b3	73
bffad86f	74	struct ioring_mmap mmap[3];
52885fa2 JA	75	};
52885fa2 JA	76
bffad86f	77	struct ioring_options {
52885fa2 JA	78	void *pad;
	79	unsigned int hipri;
	80	unsigned int fixedbufs;
a90cd050 JA	81	unsigned int sqthread;
a90cd050 JA	82	unsigned int sqthread_set;
771c9901	83	unsigned int sqthread_poll;
a90cd050	84	unsigned int sqwq;
52885fa2 JA	85	};
52885fa2 JA	86
bffad86f	87	static int fio_ioring_sqthread_cb(void data, unsigned long long val)
a90cd050	88	{
bffad86f	89	struct ioring_options *o = data;
a90cd050 JA	90
	91	o->sqthread = *val;
	92	o->sqthread_set = 1;
	93	return 0;
	94	}
	95
52885fa2 JA	96	static struct fio_option options[] = {
	97	{
	98	.name = "hipri",
	99	.lname = "High Priority",
	100	.type = FIO_OPT_STR_SET,
bffad86f	101	.off1 = offsetof(struct ioring_options, hipri),
52885fa2 JA	102	.help = "Use polled IO completions",
	103	.category = FIO_OPT_C_ENGINE,
	104	.group = FIO_OPT_G_LIBAIO,
	105	},
	106	{
	107	.name = "fixedbufs",
	108	.lname = "Fixed (pre-mapped) IO buffers",
	109	.type = FIO_OPT_STR_SET,
bffad86f	110	.off1 = offsetof(struct ioring_options, fixedbufs),
52885fa2 JA	111	.help = "Pre map IO buffers",
	112	.category = FIO_OPT_C_ENGINE,
	113	.group = FIO_OPT_G_LIBAIO,
	114	},
a90cd050 JA	115	{
	116	.name = "sqthread",
	117	.lname = "Use kernel SQ thread on this CPU",
	118	.type = FIO_OPT_INT,
bffad86f	119	.cb = fio_ioring_sqthread_cb,
a90cd050 JA	120	.help = "Offload submission to kernel thread",
	121	.category = FIO_OPT_C_ENGINE,
	122	.group = FIO_OPT_G_LIBAIO,
	123	},
771c9901 JA	124	{
	125	.name = "sqthread_poll",
	126	.lname = "Kernel SQ thread should poll",
	127	.type = FIO_OPT_STR_SET,
bffad86f	128	.off1 = offsetof(struct ioring_options, sqthread_poll),
771c9901 JA	129	.help = "Used with sqthread, enables kernel side polling",
	130	.category = FIO_OPT_C_ENGINE,
	131	.group = FIO_OPT_G_LIBAIO,
	132	},
a90cd050 JA	133	{
	134	.name = "sqwq",
	135	.lname = "Offload submission to kernel workqueue",
	136	.type = FIO_OPT_STR_SET,
bffad86f	137	.off1 = offsetof(struct ioring_options, sqwq),
a90cd050 JA	138	.help = "Offload submission to kernel workqueue",
	139	.category = FIO_OPT_C_ENGINE,
	140	.group = FIO_OPT_G_LIBAIO,
	141	},
52885fa2 JA	142	{
	143	.name = NULL,
	144	},
	145	};
	146
bffad86f	147	static int io_uring_enter(struct ioring_data *ld, unsigned int to_submit,
52885fa2 JA	148	unsigned int min_complete, unsigned int flags)
52885fa2 JA	149	{
9a2d78b3 JA	150	return syscall(__NR_sys_io_uring_enter, ld->ring_fd, to_submit,
9a2d78b3 JA	151	min_complete, flags);
52885fa2 JA	152	}
52885fa2 JA	153
bffad86f	154	static int fio_ioring_prep(struct thread_data td, struct io_u io_u)
52885fa2	155	{
bffad86f	156	struct ioring_data *ld = td->io_ops_data;
52885fa2	157	struct fio_file *f = io_u->file;
bffad86f	158	struct io_uring_iocb *iocb;
52885fa2 JA	159
52885fa2 JA	160	iocb = &ld->iocbs[io_u->index];
bffad86f JA	161	iocb->fd = f->fd;
	162	iocb->flags = 0;
	163	iocb->ioprio = 0;
52885fa2	164
e3970057 JA	165	if (io_u->ddir == DDIR_READ \|\| io_u->ddir == DDIR_WRITE) {
e3970057 JA	166	if (io_u->ddir == DDIR_READ)
bffad86f	167	iocb->opcode = IORING_OP_READ;
e3970057	168	else
bffad86f JA	169	iocb->opcode = IORING_OP_WRITE;
	170	iocb->off = io_u->offset;
	171	iocb->addr = io_u->xfer_buf;
	172	iocb->len = io_u->xfer_buflen;
52885fa2	173	} else if (ddir_sync(io_u->ddir))
bffad86f	174	iocb->opcode = IORING_OP_FSYNC;
52885fa2	175
52885fa2 JA	176	return 0;
	177	}
	178
bffad86f	179	static struct io_u fio_ioring_event(struct thread_data td, int event)
52885fa2	180	{
bffad86f JA	181	struct ioring_data *ld = td->io_ops_data;
bffad86f JA	182	struct io_uring_event *ev;
52885fa2	183	struct io_u *io_u;
b87aa01a	184	unsigned index;
52885fa2	185
b87aa01a	186	index = (event + ld->cq_ring_off) & ld->cq_ring_mask;
52885fa2	187
9a2d78b3	188	ev = &ld->cq_ring.events[index];
bffad86f	189	io_u = ld->io_u_index[ev->index];
52885fa2 JA	190
	191	if (ev->res != io_u->xfer_buflen) {
	192	if (ev->res > io_u->xfer_buflen)
	193	io_u->error = -ev->res;
	194	else
	195	io_u->resid = io_u->xfer_buflen - ev->res;
	196	} else
	197	io_u->error = 0;
	198
96563db9	199	if (io_u->ddir == DDIR_READ) {
bffad86f	200	if (ev->flags & IOEV_FLAG_CACHEHIT)
96563db9 JA	201	ld->cachehit++;
	202	else
	203	ld->cachemiss++;
	204	}
	205
52885fa2 JA	206	return io_u;
	207	}
	208
bffad86f	209	static int fio_ioring_cqring_reap(struct thread_data *td, unsigned int events,
52885fa2 JA	210	unsigned int max)
52885fa2 JA	211	{
bffad86f JA	212	struct ioring_data *ld = td->io_ops_data;
bffad86f JA	213	struct io_cq_ring *ring = &ld->cq_ring;
52885fa2 JA	214	u32 head, reaped = 0;
52885fa2 JA	215
9a2d78b3	216	head = *ring->head;
52885fa2 JA	217	do {
52885fa2 JA	218	read_barrier();
9a2d78b3	219	if (head == *ring->tail)
52885fa2 JA	220	break;
	221	reaped++;
	222	head++;
52885fa2 JA	223	} while (reaped + events < max);
52885fa2 JA	224
9a2d78b3	225	*ring->head = head;
52885fa2 JA	226	write_barrier();
	227	return reaped;
	228	}
	229
bffad86f JA	230	static int fio_ioring_getevents(struct thread_data *td, unsigned int min,
bffad86f JA	231	unsigned int max, const struct timespec *t)
52885fa2	232	{
bffad86f	233	struct ioring_data *ld = td->io_ops_data;
52885fa2	234	unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
bffad86f JA	235	struct ioring_options *o = td->eo;
bffad86f JA	236	struct io_cq_ring *ring = &ld->cq_ring;
b87aa01a JA	237	unsigned events = 0;
b87aa01a JA	238	int r;
52885fa2	239
9a2d78b3	240	ld->cq_ring_off = *ring->head;
52885fa2	241	do {
bffad86f	242	r = fio_ioring_cqring_reap(td, events, max);
52885fa2 JA	243	if (r) {
	244	events += r;
	245	continue;
	246	}
	247
771c9901	248	if (!o->sqthread_poll) {
9a2d78b3 JA	249	r = io_uring_enter(ld, 0, actual_min,
9a2d78b3 JA	250	IORING_ENTER_GETEVENTS);
771c9901 JA	251	if (r < 0) {
	252	if (errno == EAGAIN)
	253	continue;
9a2d78b3	254	td_verror(td, errno, "io_uring_enter");
771c9901 JA	255	break;
771c9901 JA	256	}
52885fa2 JA	257	}
	258	} while (events < min);
	259
	260	return r < 0 ? r : events;
	261	}
	262
bffad86f JA	263	static enum fio_q_status fio_ioring_queue(struct thread_data *td,
bffad86f JA	264	struct io_u *io_u)
52885fa2	265	{
bffad86f JA	266	struct ioring_data *ld = td->io_ops_data;
bffad86f JA	267	struct io_sq_ring *ring = &ld->sq_ring;
52885fa2 JA	268	unsigned tail, next_tail;
	269
	270	fio_ro_check(td, io_u);
	271
b87aa01a	272	if (ld->queued == ld->iodepth)
52885fa2 JA	273	return FIO_Q_BUSY;
52885fa2 JA	274
52885fa2 JA	275	if (io_u->ddir == DDIR_TRIM) {
	276	if (ld->queued)
	277	return FIO_Q_BUSY;
	278
	279	do_io_u_trim(td, io_u);
	280	io_u_mark_submit(td, 1);
	281	io_u_mark_complete(td, 1);
	282	return FIO_Q_COMPLETED;
	283	}
	284
9a2d78b3	285	tail = *ring->tail;
52885fa2	286	next_tail = tail + 1;
52885fa2	287	read_barrier();
9a2d78b3	288	if (next_tail == *ring->head)
52885fa2 JA	289	return FIO_Q_BUSY;
52885fa2 JA	290
b87aa01a	291	ring->array[tail & ld->sq_ring_mask] = io_u->index;
9a2d78b3	292	*ring->tail = next_tail;
52885fa2 JA	293	write_barrier();
	294
	295	ld->queued++;
	296	return FIO_Q_QUEUED;
	297	}
	298
bffad86f	299	static void fio_ioring_queued(struct thread_data *td, int start, int nr)
52885fa2	300	{
bffad86f	301	struct ioring_data *ld = td->io_ops_data;
52885fa2 JA	302	struct timespec now;
	303
	304	if (!fio_fill_issue_time(td))
	305	return;
	306
	307	fio_gettime(&now, NULL);
	308
	309	while (nr--) {
bffad86f	310	struct io_sq_ring *ring = &ld->sq_ring;
9a2d78b3	311	int index = ring->array[start & ld->sq_ring_mask];
f8289afc	312	struct io_u *io_u = ld->io_u_index[index];
52885fa2 JA	313
	314	memcpy(&io_u->issue_time, &now, sizeof(now));
	315	io_u_queued(td, io_u);
	316
	317	start++;
52885fa2 JA	318	}
	319	}
	320
bffad86f	321	static int fio_ioring_commit(struct thread_data *td)
52885fa2	322	{
bffad86f JA	323	struct ioring_data *ld = td->io_ops_data;
bffad86f JA	324	struct ioring_options *o = td->eo;
52885fa2 JA	325	int ret;
	326
	327	if (!ld->queued)
	328	return 0;
	329
771c9901 JA	330	/* Nothing to do */
771c9901 JA	331	if (o->sqthread_poll) {
bffad86f	332	struct io_sq_ring *ring = &ld->sq_ring;
4cdbc048	333
9a2d78b3 JA	334	if (*ring->flags & IORING_SQ_NEED_WAKEUP)
9a2d78b3 JA	335	io_uring_enter(ld, ld->queued, 0, 0);
771c9901 JA	336	ld->queued = 0;
	337	return 0;
	338	}
	339
52885fa2	340	do {
9a2d78b3	341	unsigned start = *ld->sq_ring.head;
52885fa2 JA	342	long nr = ld->queued;
52885fa2 JA	343
9a2d78b3	344	ret = io_uring_enter(ld, nr, 0, IORING_ENTER_GETEVENTS);
52885fa2	345	if (ret > 0) {
bffad86f	346	fio_ioring_queued(td, start, ret);
52885fa2 JA	347	io_u_mark_submit(td, ret);
	348
	349	ld->queued -= ret;
	350	ret = 0;
a90cd050 JA	351	} else if (!ret) {
a90cd050 JA	352	io_u_mark_submit(td, ret);
52885fa2	353	continue;
a90cd050 JA	354	} else {
a90cd050 JA	355	if (errno == EAGAIN) {
bffad86f	356	ret = fio_ioring_cqring_reap(td, 0, ld->queued);
a90cd050 JA	357	if (ret)
	358	continue;
	359	/* Shouldn't happen */
	360	usleep(1);
	361	continue;
52885fa2	362	}
9a2d78b3	363	td_verror(td, errno, "io_uring_enter submit");
52885fa2	364	break;
a90cd050	365	}
52885fa2 JA	366	} while (ld->queued);
	367
	368	return ret;
	369	}
	370
bffad86f	371	static void fio_ioring_unmap(struct ioring_data *ld)
52885fa2	372	{
9a2d78b3	373	int i;
52885fa2	374
9a2d78b3 JA	375	for (i = 0; i < ARRAY_SIZE(ld->mmap); i++)
	376	munmap(ld->mmap[i].ptr, ld->mmap[i].len);
	377	close(ld->ring_fd);
b87aa01a JA	378	}
b87aa01a JA	379
bffad86f	380	static void fio_ioring_cleanup(struct thread_data *td)
52885fa2	381	{
bffad86f	382	struct ioring_data *ld = td->io_ops_data;
52885fa2 JA	383
52885fa2 JA	384	if (ld) {
96563db9 JA	385	td->ts.cachehit += ld->cachehit;
	386	td->ts.cachemiss += ld->cachemiss;
	387
52885fa2	388	if (!(td->flags & TD_F_CHILD))
bffad86f	389	fio_ioring_unmap(ld);
9a2d78b3	390
52885fa2 JA	391	free(ld->io_u_index);
52885fa2 JA	392	free(ld->io_us);
9a2d78b3	393	free(ld->iovecs);
52885fa2 JA	394	free(ld);
	395	}
	396	}
	397
bffad86f	398	static int fio_ioring_mmap(struct ioring_data ld, struct io_uring_params p)
9a2d78b3	399	{
bffad86f JA	400	struct io_sq_ring *sring = &ld->sq_ring;
bffad86f JA	401	struct io_cq_ring *cring = &ld->cq_ring;
9a2d78b3 JA	402	void *ptr;
9a2d78b3 JA	403
ac122fea	404	ld->mmap[0].len = p->sq_off.array + p->sq_entries * sizeof(u32);
9a2d78b3 JA	405	ptr = mmap(0, ld->mmap[0].len, PROT_READ \| PROT_WRITE,
	406	MAP_SHARED \| MAP_POPULATE, ld->ring_fd,
	407	IORING_OFF_SQ_RING);
	408	ld->mmap[0].ptr = ptr;
	409	sring->head = ptr + p->sq_off.head;
	410	sring->tail = ptr + p->sq_off.tail;
	411	sring->ring_mask = ptr + p->sq_off.ring_mask;
	412	sring->ring_entries = ptr + p->sq_off.ring_entries;
	413	sring->flags = ptr + p->sq_off.flags;
ac122fea	414	sring->array = ptr + p->sq_off.array;
9a2d78b3 JA	415	ld->sq_ring_mask = *sring->ring_mask;
9a2d78b3 JA	416
bffad86f	417	ld->mmap[1].len = p->sq_entries * sizeof(struct io_uring_iocb);
9a2d78b3 JA	418	ld->iocbs = mmap(0, ld->mmap[1].len, PROT_READ \| PROT_WRITE,
	419	MAP_SHARED \| MAP_POPULATE, ld->ring_fd,
	420	IORING_OFF_IOCB);
	421	ld->mmap[1].ptr = ld->iocbs;
	422
ac122fea	423	ld->mmap[2].len = p->cq_off.events +
bffad86f	424	p->cq_entries * sizeof(struct io_uring_event);
9a2d78b3 JA	425	ptr = mmap(0, ld->mmap[2].len, PROT_READ \| PROT_WRITE,
	426	MAP_SHARED \| MAP_POPULATE, ld->ring_fd,
	427	IORING_OFF_CQ_RING);
	428	ld->mmap[2].ptr = ptr;
	429	cring->head = ptr + p->cq_off.head;
	430	cring->tail = ptr + p->cq_off.tail;
	431	cring->ring_mask = ptr + p->cq_off.ring_mask;
	432	cring->ring_entries = ptr + p->cq_off.ring_entries;
ac122fea	433	cring->events = ptr + p->cq_off.events;
9a2d78b3 JA	434	ld->cq_ring_mask = *cring->ring_mask;
	435	return 0;
	436	}
	437
bffad86f	438	static int fio_ioring_queue_init(struct thread_data *td)
52885fa2	439	{
bffad86f JA	440	struct ioring_data *ld = td->io_ops_data;
bffad86f JA	441	struct ioring_options *o = td->eo;
52885fa2	442	int depth = td->o.iodepth;
bffad86f	443	struct io_uring_params p;
9a2d78b3 JA	444	int ret;
	445
	446	memset(&p, 0, sizeof(p));
52885fa2 JA	447
52885fa2 JA	448	if (o->hipri)
bffad86f	449	p.flags \|= IORING_SETUP_IOPOLL;
a90cd050	450	if (o->sqthread_set) {
9a2d78b3	451	p.sq_thread_cpu = o->sqthread;
bffad86f	452	p.flags \|= IORING_SETUP_SQTHREAD;
771c9901	453	if (o->sqthread_poll)
bffad86f	454	p.flags \|= IORING_SETUP_SQPOLL;
f635f1fb JA	455	}
f635f1fb JA	456	if (o->sqwq)
bffad86f	457	p.flags \|= IORING_SETUP_SQWQ;
a90cd050	458
52885fa2 JA	459	if (o->fixedbufs) {
	460	struct rlimit rlim = {
	461	.rlim_cur = RLIM_INFINITY,
	462	.rlim_max = RLIM_INFINITY,
	463	};
	464
	465	setrlimit(RLIMIT_MEMLOCK, &rlim);
bffad86f	466	p.flags \|= IORING_SETUP_FIXEDBUFS;
52885fa2 JA	467	}
52885fa2 JA	468
9a2d78b3 JA	469	ret = syscall(__NR_sys_io_uring_setup, depth, ld->iovecs, &p);
	470	if (ret < 0)
	471	return ret;
	472
	473	ld->ring_fd = ret;
bffad86f	474	return fio_ioring_mmap(ld, &p);
52885fa2 JA	475	}
52885fa2 JA	476
bffad86f	477	static int fio_ioring_post_init(struct thread_data *td)
52885fa2	478	{
bffad86f JA	479	struct ioring_data *ld = td->io_ops_data;
bffad86f JA	480	struct ioring_options *o = td->eo;
52885fa2	481	struct io_u *io_u;
9a2d78b3	482	int err;
52885fa2 JA	483
	484	if (o->fixedbufs) {
	485	int i;
	486
	487	for (i = 0; i < td->o.iodepth; i++) {
9a2d78b3 JA	488	struct iovec *iov = &ld->iovecs[i];
9a2d78b3 JA	489
52885fa2	490	io_u = ld->io_u_index[i];
9a2d78b3 JA	491	iov->iov_base = io_u->buf;
9a2d78b3 JA	492	iov->iov_len = td_max_bs(td);
52885fa2 JA	493	}
	494	}
	495
bffad86f	496	err = fio_ioring_queue_init(td);
52885fa2	497	if (err) {
d63a472d	498	td_verror(td, errno, "io_queue_init");
52885fa2 JA	499	return 1;
	500	}
	501
	502	return 0;
	503	}
	504
9a2d78b3 JA	505	static unsigned roundup_pow2(unsigned depth)
	506	{
	507	return 1UL << __fls(depth - 1);
	508	}
	509
bffad86f	510	static int fio_ioring_init(struct thread_data *td)
52885fa2	511	{
bffad86f	512	struct ioring_data *ld;
52885fa2	513
52885fa2 JA	514	ld = calloc(1, sizeof(*ld));
52885fa2 JA	515
b87aa01a JA	516	/* ring depth must be a power-of-2 */
	517	ld->iodepth = td->o.iodepth;
	518	td->o.iodepth = roundup_pow2(td->o.iodepth);
	519
52885fa2 JA	520	/* io_u index */
	521	ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *));
	522	ld->io_us = calloc(td->o.iodepth, sizeof(struct io_u *));
	523
9a2d78b3	524	ld->iovecs = calloc(td->o.iodepth, sizeof(struct iovec));
52885fa2 JA	525
	526	td->io_ops_data = ld;
	527	return 0;
	528	}
	529
bffad86f	530	static int fio_ioring_io_u_init(struct thread_data td, struct io_u io_u)
52885fa2	531	{
bffad86f	532	struct ioring_data *ld = td->io_ops_data;
52885fa2 JA	533
	534	ld->io_u_index[io_u->index] = io_u;
	535	return 0;
	536	}
	537
	538	static struct ioengine_ops ioengine = {
bffad86f	539	.name = "io_uring",
52885fa2	540	.version = FIO_IOOPS_VERSION,
bffad86f JA	541	.init = fio_ioring_init,
	542	.post_init = fio_ioring_post_init,
	543	.io_u_init = fio_ioring_io_u_init,
	544	.prep = fio_ioring_prep,
	545	.queue = fio_ioring_queue,
	546	.commit = fio_ioring_commit,
	547	.getevents = fio_ioring_getevents,
	548	.event = fio_ioring_event,
	549	.cleanup = fio_ioring_cleanup,
52885fa2 JA	550	.open_file = generic_open_file,
	551	.close_file = generic_close_file,
	552	.get_file_size = generic_get_file_size,
	553	.options = options,
bffad86f	554	.option_struct_size = sizeof(struct ioring_options),
52885fa2 JA	555	};
52885fa2 JA	556
bffad86f	557	static void fio_init fio_ioring_register(void)
52885fa2	558	{
52885fa2	559	register_ioengine(&ioengine);
52885fa2 JA	560	}
52885fa2 JA	561
bffad86f	562	static void fio_exit fio_ioring_unregister(void)
52885fa2	563	{
52885fa2	564	unregister_ioengine(&ioengine);
52885fa2	565	}
1f90e9bb	566	#endif