[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 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 (io_u->ddir == DDIR_READ)
			sqe->opcode = IORING_OP_READV;
		else
			sqe->opcode = IORING_OP_WRITEV;

		if (o->fixedbufs) {
			sqe->flags |= IOSQE_FIXED_BUFFER;
			sqe->addr = io_u->xfer_buf;
			sqe->len = io_u->xfer_buflen;
			sqe->buf_index = io_u->index;
		} else {
			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->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->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->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->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;
3d7d00a3	76	unsigned int sqpoll_thread;
2ea53ca3 JA	77	unsigned int sqpoll_set;
2ea53ca3 JA	78	unsigned int sqpoll_cpu;
52885fa2 JA	79	};
52885fa2 JA	80
2ea53ca3	81	static int fio_ioring_sqpoll_cb(void data, unsigned long long val)
a90cd050	82	{
bffad86f	83	struct ioring_options *o = data;
a90cd050	84
2ea53ca3 JA	85	o->sqpoll_cpu = *val;
2ea53ca3 JA	86	o->sqpoll_set = 1;
a90cd050 JA	87	return 0;
	88	}
	89
52885fa2 JA	90	static struct fio_option options[] = {
	91	{
	92	.name = "hipri",
	93	.lname = "High Priority",
	94	.type = FIO_OPT_STR_SET,
bffad86f	95	.off1 = offsetof(struct ioring_options, hipri),
52885fa2 JA	96	.help = "Use polled IO completions",
	97	.category = FIO_OPT_C_ENGINE,
	98	.group = FIO_OPT_G_LIBAIO,
	99	},
	100	{
	101	.name = "fixedbufs",
	102	.lname = "Fixed (pre-mapped) IO buffers",
	103	.type = FIO_OPT_STR_SET,
bffad86f	104	.off1 = offsetof(struct ioring_options, fixedbufs),
52885fa2 JA	105	.help = "Pre map IO buffers",
	106	.category = FIO_OPT_C_ENGINE,
	107	.group = FIO_OPT_G_LIBAIO,
	108	},
771c9901 JA	109	{
771c9901 JA	110	.name = "sqthread_poll",
3d7d00a3 JA	111	.lname = "Kernel SQ thread polling",
	112	.type = FIO_OPT_INT,
	113	.off1 = offsetof(struct ioring_options, sqpoll_thread),
	114	.help = "Offload submission/completion to kernel thread",
	115	.category = FIO_OPT_C_ENGINE,
	116	.group = FIO_OPT_G_LIBAIO,
	117	},
	118	{
	119	.name = "sqthread_poll_cpu",
	120	.lname = "SQ Thread Poll CPU",
2ea53ca3 JA	121	.type = FIO_OPT_INT,
2ea53ca3 JA	122	.cb = fio_ioring_sqpoll_cb,
3d7d00a3	123	.help = "What CPU to run SQ thread polling on",
a90cd050 JA	124	.category = FIO_OPT_C_ENGINE,
	125	.group = FIO_OPT_G_LIBAIO,
	126	},
52885fa2 JA	127	{
	128	.name = NULL,
	129	},
	130	};
	131
bffad86f	132	static int io_uring_enter(struct ioring_data *ld, unsigned int to_submit,
52885fa2 JA	133	unsigned int min_complete, unsigned int flags)
52885fa2 JA	134	{
9a2d78b3 JA	135	return syscall(__NR_sys_io_uring_enter, ld->ring_fd, to_submit,
9a2d78b3 JA	136	min_complete, flags);
52885fa2 JA	137	}
52885fa2 JA	138
bffad86f	139	static int fio_ioring_prep(struct thread_data td, struct io_u io_u)
52885fa2	140	{
bffad86f	141	struct ioring_data *ld = td->io_ops_data;
cfcc8564	142	struct ioring_options *o = td->eo;
52885fa2	143	struct fio_file *f = io_u->file;
f0403f94	144	struct io_uring_sqe *sqe;
52885fa2	145
f0403f94 JA	146	sqe = &ld->sqes[io_u->index];
	147	sqe->fd = f->fd;
	148	sqe->flags = 0;
	149	sqe->ioprio = 0;
2ea53ca3	150	sqe->buf_index = 0;
52885fa2	151
e3970057	152	if (io_u->ddir == DDIR_READ \|\| io_u->ddir == DDIR_WRITE) {
2ea53ca3 JA	153	if (io_u->ddir == DDIR_READ)
	154	sqe->opcode = IORING_OP_READV;
	155	else
	156	sqe->opcode = IORING_OP_WRITEV;
	157
f0403f94	158	if (o->fixedbufs) {
2ea53ca3	159	sqe->flags \|= IOSQE_FIXED_BUFFER;
f0403f94 JA	160	sqe->addr = io_u->xfer_buf;
f0403f94 JA	161	sqe->len = io_u->xfer_buflen;
2ea53ca3	162	sqe->buf_index = io_u->index;
cfcc8564	163	} else {
f0403f94 JA	164	sqe->addr = &ld->iovecs[io_u->index];
f0403f94 JA	165	sqe->len = 1;
cfcc8564	166	}
f0403f94	167	sqe->off = io_u->offset;
52885fa2	168	} else if (ddir_sync(io_u->ddir))
f0403f94	169	sqe->opcode = IORING_OP_FSYNC;
52885fa2	170
a7086591	171	sqe->data = (unsigned long) io_u;
52885fa2 JA	172	return 0;
	173	}
	174
bffad86f	175	static struct io_u fio_ioring_event(struct thread_data td, int event)
52885fa2	176	{
bffad86f	177	struct ioring_data *ld = td->io_ops_data;
f0403f94	178	struct io_uring_cqe *cqe;
52885fa2	179	struct io_u *io_u;
b87aa01a	180	unsigned index;
52885fa2	181
b87aa01a	182	index = (event + ld->cq_ring_off) & ld->cq_ring_mask;
52885fa2	183
f0403f94	184	cqe = &ld->cq_ring.cqes[index];
a7086591	185	io_u = (struct io_u *) cqe->data;
52885fa2	186
f0403f94 JA	187	if (cqe->res != io_u->xfer_buflen) {
	188	if (cqe->res > io_u->xfer_buflen)
	189	io_u->error = -cqe->res;
52885fa2	190	else
f0403f94	191	io_u->resid = io_u->xfer_buflen - cqe->res;
52885fa2 JA	192	} else
	193	io_u->error = 0;
	194
96563db9	195	if (io_u->ddir == DDIR_READ) {
f0403f94	196	if (cqe->flags & IOCQE_FLAG_CACHEHIT)
96563db9 JA	197	ld->cachehit++;
	198	else
	199	ld->cachemiss++;
	200	}
	201
52885fa2 JA	202	return io_u;
	203	}
	204
bffad86f	205	static int fio_ioring_cqring_reap(struct thread_data *td, unsigned int events,
52885fa2 JA	206	unsigned int max)
52885fa2 JA	207	{
bffad86f JA	208	struct ioring_data *ld = td->io_ops_data;
bffad86f JA	209	struct io_cq_ring *ring = &ld->cq_ring;
e2239016	210	unsigned head, reaped = 0;
52885fa2	211
9a2d78b3	212	head = *ring->head;
52885fa2 JA	213	do {
52885fa2 JA	214	read_barrier();
9a2d78b3	215	if (head == *ring->tail)
52885fa2 JA	216	break;
	217	reaped++;
	218	head++;
52885fa2 JA	219	} while (reaped + events < max);
52885fa2 JA	220
9a2d78b3	221	*ring->head = head;
52885fa2 JA	222	write_barrier();
	223	return reaped;
	224	}
	225
bffad86f JA	226	static int fio_ioring_getevents(struct thread_data *td, unsigned int min,
bffad86f JA	227	unsigned int max, const struct timespec *t)
52885fa2	228	{
bffad86f	229	struct ioring_data *ld = td->io_ops_data;
52885fa2	230	unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
bffad86f JA	231	struct ioring_options *o = td->eo;
bffad86f JA	232	struct io_cq_ring *ring = &ld->cq_ring;
b87aa01a JA	233	unsigned events = 0;
b87aa01a JA	234	int r;
52885fa2	235
9a2d78b3	236	ld->cq_ring_off = *ring->head;
52885fa2	237	do {
bffad86f	238	r = fio_ioring_cqring_reap(td, events, max);
52885fa2 JA	239	if (r) {
	240	events += r;
	241	continue;
	242	}
	243
3d7d00a3	244	if (!o->sqpoll_thread) {
9a2d78b3 JA	245	r = io_uring_enter(ld, 0, actual_min,
9a2d78b3 JA	246	IORING_ENTER_GETEVENTS);
771c9901 JA	247	if (r < 0) {
	248	if (errno == EAGAIN)
	249	continue;
9a2d78b3	250	td_verror(td, errno, "io_uring_enter");
771c9901 JA	251	break;
771c9901 JA	252	}
52885fa2 JA	253	}
	254	} while (events < min);
	255
	256	return r < 0 ? r : events;
	257	}
	258
bffad86f JA	259	static enum fio_q_status fio_ioring_queue(struct thread_data *td,
bffad86f JA	260	struct io_u *io_u)
52885fa2	261	{
bffad86f JA	262	struct ioring_data *ld = td->io_ops_data;
bffad86f JA	263	struct io_sq_ring *ring = &ld->sq_ring;
52885fa2 JA	264	unsigned tail, next_tail;
	265
	266	fio_ro_check(td, io_u);
	267
b87aa01a	268	if (ld->queued == ld->iodepth)
52885fa2 JA	269	return FIO_Q_BUSY;
52885fa2 JA	270
52885fa2 JA	271	if (io_u->ddir == DDIR_TRIM) {
	272	if (ld->queued)
	273	return FIO_Q_BUSY;
	274
	275	do_io_u_trim(td, io_u);
	276	io_u_mark_submit(td, 1);
	277	io_u_mark_complete(td, 1);
	278	return FIO_Q_COMPLETED;
	279	}
	280
9a2d78b3	281	tail = *ring->tail;
52885fa2	282	next_tail = tail + 1;
52885fa2	283	read_barrier();
9a2d78b3	284	if (next_tail == *ring->head)
52885fa2 JA	285	return FIO_Q_BUSY;
52885fa2 JA	286
b87aa01a	287	ring->array[tail & ld->sq_ring_mask] = io_u->index;
9a2d78b3	288	*ring->tail = next_tail;
52885fa2 JA	289	write_barrier();
	290
	291	ld->queued++;
	292	return FIO_Q_QUEUED;
	293	}
	294
bffad86f	295	static void fio_ioring_queued(struct thread_data *td, int start, int nr)
52885fa2	296	{
bffad86f	297	struct ioring_data *ld = td->io_ops_data;
52885fa2 JA	298	struct timespec now;
	299
	300	if (!fio_fill_issue_time(td))
	301	return;
	302
	303	fio_gettime(&now, NULL);
	304
	305	while (nr--) {
bffad86f	306	struct io_sq_ring *ring = &ld->sq_ring;
9a2d78b3	307	int index = ring->array[start & ld->sq_ring_mask];
f8289afc	308	struct io_u *io_u = ld->io_u_index[index];
52885fa2 JA	309
	310	memcpy(&io_u->issue_time, &now, sizeof(now));
	311	io_u_queued(td, io_u);
	312
	313	start++;
52885fa2 JA	314	}
	315	}
	316
bffad86f	317	static int fio_ioring_commit(struct thread_data *td)
52885fa2	318	{
bffad86f JA	319	struct ioring_data *ld = td->io_ops_data;
bffad86f JA	320	struct ioring_options *o = td->eo;
52885fa2 JA	321	int ret;
	322
	323	if (!ld->queued)
	324	return 0;
	325
3d7d00a3 JA	326	/*
	327	* Kernel side does submission. just need to check if the ring is
	328	* flagged as needing a kick, if so, call io_uring_enter(). This
	329	* only happens if we've been idle too long.
	330	*/
	331	if (o->sqpoll_thread) {
bffad86f	332	struct io_sq_ring *ring = &ld->sq_ring;
4cdbc048	333
2ea53ca3	334	read_barrier();
9a2d78b3 JA	335	if (*ring->flags & IORING_SQ_NEED_WAKEUP)
9a2d78b3 JA	336	io_uring_enter(ld, ld->queued, 0, 0);
771c9901 JA	337	ld->queued = 0;
	338	return 0;
	339	}
	340
52885fa2	341	do {
9a2d78b3	342	unsigned start = *ld->sq_ring.head;
52885fa2 JA	343	long nr = ld->queued;
52885fa2 JA	344
9a2d78b3	345	ret = io_uring_enter(ld, nr, 0, IORING_ENTER_GETEVENTS);
52885fa2	346	if (ret > 0) {
bffad86f	347	fio_ioring_queued(td, start, ret);
52885fa2 JA	348	io_u_mark_submit(td, ret);
	349
	350	ld->queued -= ret;
	351	ret = 0;
a90cd050 JA	352	} else if (!ret) {
a90cd050 JA	353	io_u_mark_submit(td, ret);
52885fa2	354	continue;
a90cd050 JA	355	} else {
a90cd050 JA	356	if (errno == EAGAIN) {
bffad86f	357	ret = fio_ioring_cqring_reap(td, 0, ld->queued);
a90cd050 JA	358	if (ret)
	359	continue;
	360	/* Shouldn't happen */
	361	usleep(1);
	362	continue;
52885fa2	363	}
9a2d78b3	364	td_verror(td, errno, "io_uring_enter submit");
52885fa2	365	break;
a90cd050	366	}
52885fa2 JA	367	} while (ld->queued);
	368
	369	return ret;
	370	}
	371
bffad86f	372	static void fio_ioring_unmap(struct ioring_data *ld)
52885fa2	373	{
9a2d78b3	374	int i;
52885fa2	375
9a2d78b3 JA	376	for (i = 0; i < ARRAY_SIZE(ld->mmap); i++)
	377	munmap(ld->mmap[i].ptr, ld->mmap[i].len);
	378	close(ld->ring_fd);
b87aa01a JA	379	}
b87aa01a JA	380
bffad86f	381	static void fio_ioring_cleanup(struct thread_data *td)
52885fa2	382	{
bffad86f	383	struct ioring_data *ld = td->io_ops_data;
52885fa2 JA	384
52885fa2 JA	385	if (ld) {
96563db9 JA	386	td->ts.cachehit += ld->cachehit;
	387	td->ts.cachemiss += ld->cachemiss;
	388
52885fa2	389	if (!(td->flags & TD_F_CHILD))
bffad86f	390	fio_ioring_unmap(ld);
9a2d78b3	391
52885fa2 JA	392	free(ld->io_u_index);
52885fa2 JA	393	free(ld->io_us);
9a2d78b3	394	free(ld->iovecs);
52885fa2 JA	395	free(ld);
	396	}
	397	}
	398
bffad86f	399	static int fio_ioring_mmap(struct ioring_data ld, struct io_uring_params p)
9a2d78b3	400	{
bffad86f JA	401	struct io_sq_ring *sring = &ld->sq_ring;
bffad86f JA	402	struct io_cq_ring *cring = &ld->cq_ring;
9a2d78b3 JA	403	void *ptr;
9a2d78b3 JA	404
e2239016	405	ld->mmap[0].len = p->sq_off.array + p->sq_entries * sizeof(__u32);
9a2d78b3 JA	406	ptr = mmap(0, ld->mmap[0].len, PROT_READ \| PROT_WRITE,
	407	MAP_SHARED \| MAP_POPULATE, ld->ring_fd,
	408	IORING_OFF_SQ_RING);
	409	ld->mmap[0].ptr = ptr;
	410	sring->head = ptr + p->sq_off.head;
	411	sring->tail = ptr + p->sq_off.tail;
	412	sring->ring_mask = ptr + p->sq_off.ring_mask;
	413	sring->ring_entries = ptr + p->sq_off.ring_entries;
	414	sring->flags = ptr + p->sq_off.flags;
ac122fea	415	sring->array = ptr + p->sq_off.array;
9a2d78b3 JA	416	ld->sq_ring_mask = *sring->ring_mask;
9a2d78b3 JA	417
f0403f94 JA	418	ld->mmap[1].len = p->sq_entries * sizeof(struct io_uring_sqe);
f0403f94 JA	419	ld->sqes = mmap(0, ld->mmap[1].len, PROT_READ \| PROT_WRITE,
9a2d78b3	420	MAP_SHARED \| MAP_POPULATE, ld->ring_fd,
f0403f94 JA	421	IORING_OFF_SQES);
f0403f94 JA	422	ld->mmap[1].ptr = ld->sqes;
9a2d78b3	423
f0403f94 JA	424	ld->mmap[2].len = p->cq_off.cqes +
f0403f94 JA	425	p->cq_entries * sizeof(struct io_uring_cqe);
9a2d78b3 JA	426	ptr = mmap(0, ld->mmap[2].len, PROT_READ \| PROT_WRITE,
	427	MAP_SHARED \| MAP_POPULATE, ld->ring_fd,
	428	IORING_OFF_CQ_RING);
	429	ld->mmap[2].ptr = ptr;
	430	cring->head = ptr + p->cq_off.head;
	431	cring->tail = ptr + p->cq_off.tail;
	432	cring->ring_mask = ptr + p->cq_off.ring_mask;
	433	cring->ring_entries = ptr + p->cq_off.ring_entries;
f0403f94	434	cring->cqes = ptr + p->cq_off.cqes;
9a2d78b3 JA	435	ld->cq_ring_mask = *cring->ring_mask;
	436	return 0;
	437	}
	438
bffad86f	439	static int fio_ioring_queue_init(struct thread_data *td)
52885fa2	440	{
bffad86f JA	441	struct ioring_data *ld = td->io_ops_data;
bffad86f JA	442	struct ioring_options *o = td->eo;
52885fa2	443	int depth = td->o.iodepth;
bffad86f	444	struct io_uring_params p;
9a2d78b3 JA	445	int ret;
	446
	447	memset(&p, 0, sizeof(p));
52885fa2 JA	448
52885fa2 JA	449	if (o->hipri)
bffad86f	450	p.flags \|= IORING_SETUP_IOPOLL;
3d7d00a3 JA	451	if (o->sqpoll_thread) {
	452	p.flags \|= IORING_SETUP_SQPOLL;
	453	if (o->sqpoll_set) {
	454	p.flags \|= IORING_SETUP_SQ_AFF;
	455	p.sq_thread_cpu = o->sqpoll_cpu;
	456	}
f635f1fb	457	}
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);
52885fa2 JA	466	}
52885fa2 JA	467
2ea53ca3	468	ret = syscall(__NR_sys_io_uring_setup, depth, &p);
9a2d78b3 JA	469	if (ret < 0)
	470	return ret;
	471
	472	ld->ring_fd = ret;
2ea53ca3 JA	473
	474	if (o->fixedbufs) {
	475	struct io_uring_register_buffers reg = {
	476	.iovecs = ld->iovecs,
	477	.nr_iovecs = depth
	478	};
	479
	480	ret = syscall(__NR_sys_io_uring_register, ld->ring_fd,
	481	IORING_REGISTER_BUFFERS, &reg);
	482	if (ret < 0)
	483	return ret;
	484	}
	485
bffad86f	486	return fio_ioring_mmap(ld, &p);
52885fa2 JA	487	}
52885fa2 JA	488
bffad86f	489	static int fio_ioring_post_init(struct thread_data *td)
52885fa2	490	{
bffad86f	491	struct ioring_data *ld = td->io_ops_data;
52885fa2	492	struct io_u *io_u;
650346e1	493	int err, i;
52885fa2	494
650346e1 JA	495	for (i = 0; i < td->o.iodepth; i++) {
650346e1 JA	496	struct iovec *iov = &ld->iovecs[i];
9a2d78b3	497
650346e1 JA	498	io_u = ld->io_u_index[i];
	499	iov->iov_base = io_u->buf;
	500	iov->iov_len = td_max_bs(td);
52885fa2 JA	501	}
52885fa2 JA	502
bffad86f	503	err = fio_ioring_queue_init(td);
52885fa2	504	if (err) {
d63a472d	505	td_verror(td, errno, "io_queue_init");
52885fa2 JA	506	return 1;
	507	}
	508
	509	return 0;
	510	}
	511
9a2d78b3 JA	512	static unsigned roundup_pow2(unsigned depth)
	513	{
	514	return 1UL << __fls(depth - 1);
	515	}
	516
bffad86f	517	static int fio_ioring_init(struct thread_data *td)
52885fa2	518	{
bffad86f	519	struct ioring_data *ld;
52885fa2	520
52885fa2 JA	521	ld = calloc(1, sizeof(*ld));
52885fa2 JA	522
b87aa01a JA	523	/* ring depth must be a power-of-2 */
	524	ld->iodepth = td->o.iodepth;
	525	td->o.iodepth = roundup_pow2(td->o.iodepth);
	526
52885fa2 JA	527	/* io_u index */
	528	ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *));
	529	ld->io_us = calloc(td->o.iodepth, sizeof(struct io_u *));
650346e1	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