[fio.git] / engines / aioring.c

/*
 * aioring engine
 *
 * IO engine using the new native Linux libaio ring interface. See:
 *
 * http://git.kernel.dk/cgit/linux-block/log/?h=aio-poll
 *
 */
#include <stdlib.h>
#include <unistd.h>
#include <errno.h>
#include <libaio.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_AIORING

/*
 * io_uring_setup(2) flags
 */
#ifndef IOCTX_FLAG_SCQRING
#define IOCTX_FLAG_SCQRING	(1 << 0)
#endif
#ifndef IOCTX_FLAG_IOPOLL
#define IOCTX_FLAG_IOPOLL	(1 << 1)
#endif
#ifndef IOCTX_FLAG_FIXEDBUFS
#define IOCTX_FLAG_FIXEDBUFS	(1 << 2)
#endif
#ifndef IOCTX_FLAG_SQTHREAD
#define IOCTX_FLAG_SQTHREAD	(1 << 3)
#endif
#ifndef IOCTX_FLAG_SQWQ
#define IOCTX_FLAG_SQWQ		(1 << 4)
#endif
#ifndef IOCTX_FLAG_SQPOLL
#define IOCTX_FLAG_SQPOLL	(1 << 5)
#endif

#define IORING_OFF_SQ_RING	0ULL
#define IORING_OFF_CQ_RING	0x8000000ULL
#define IORING_OFF_IOCB		0x10000000ULL

/*
 * io_uring_enter(2) flags
 */
#ifndef IORING_ENTER_GETEVENTS
#define IORING_ENTER_GETEVENTS	(1 << 0)
#endif

typedef uint64_t u64;
typedef uint32_t u32;
typedef uint16_t u16;

#define IORING_SQ_NEED_WAKEUP	(1 << 0)

#define IOEV_RES2_CACHEHIT	(1 << 0)

struct aio_uring_offsets {
	u32 head;
	u32 tail;
	u32 ring_mask;
	u32 ring_entries;
	u32 flags;
	u32 elems;
};

struct aio_uring_params {
	u32 sq_entries;
	u32 cq_entries;
	u32 flags;
	u16 sq_thread_cpu;
	u16 resv[9];
	struct aio_uring_offsets sq_off;
	struct aio_uring_offsets cq_off;
};

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

struct aio_cq_ring {
	u32 *head;
	u32 *tail;
	u32 *ring_mask;
	u32 *ring_entries;
	struct io_event *events;
};

struct aioring_mmap {
	void *ptr;
	size_t len;
};

struct aioring_data {
	int ring_fd;

	struct io_u **io_us;
	struct io_u **io_u_index;

	struct aio_sq_ring sq_ring;
	struct iocb *iocbs;
	struct iovec *iovecs;
	unsigned sq_ring_mask;

	struct aio_cq_ring cq_ring;
	struct io_event *events;
	unsigned cq_ring_mask;

	int queued;
	int cq_ring_off;
	unsigned iodepth;

	uint64_t cachehit;
	uint64_t cachemiss;

	struct aioring_mmap mmap[3];
};

struct aioring_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_aioring_sqthread_cb(void *data,
				   unsigned long long *val)
{
	struct aioring_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 aioring_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 aioring_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_aioring_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 aioring_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 aioring_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 aioring_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_aioring_prep(struct thread_data *td, struct io_u *io_u)
{
	struct aioring_data *ld = td->io_ops_data;
	struct fio_file *f = io_u->file;
	struct iocb *iocb;

	iocb = &ld->iocbs[io_u->index];

	if (io_u->ddir == DDIR_READ || io_u->ddir == DDIR_WRITE) {
		if (io_u->ddir == DDIR_READ)
			iocb->aio_lio_opcode = IO_CMD_PREAD;
		else
			iocb->aio_lio_opcode = IO_CMD_PWRITE;
		iocb->aio_reqprio = 0;
		iocb->aio_fildes = f->fd;
		iocb->u.c.buf = io_u->xfer_buf;
		iocb->u.c.nbytes = io_u->xfer_buflen;
		iocb->u.c.offset = io_u->offset;
		iocb->u.c.flags = 0;
	} else if (ddir_sync(io_u->ddir))
		io_prep_fsync(iocb, f->fd);

	iocb->data = io_u;
	return 0;
}

static struct io_u *fio_aioring_event(struct thread_data *td, int event)
{
	struct aioring_data *ld = td->io_ops_data;
	struct io_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 = ev->data;

	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->res2 & IOEV_RES2_CACHEHIT)
			ld->cachehit++;
		else
			ld->cachemiss++;
	}

	return io_u;
}

static int fio_aioring_cqring_reap(struct thread_data *td, unsigned int events,
				   unsigned int max)
{
	struct aioring_data *ld = td->io_ops_data;
	struct aio_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_aioring_getevents(struct thread_data *td, unsigned int min,
				 unsigned int max, const struct timespec *t)
{
	struct aioring_data *ld = td->io_ops_data;
	unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
	struct aioring_options *o = td->eo;
	struct aio_cq_ring *ring = &ld->cq_ring;
	unsigned events = 0;
	int r;

	ld->cq_ring_off = *ring->head;
	do {
		r = fio_aioring_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_aioring_queue(struct thread_data *td,
					   struct io_u *io_u)
{
	struct aioring_data *ld = td->io_ops_data;
	struct aio_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_aioring_queued(struct thread_data *td, int start, int nr)
{
	struct aioring_data *ld = td->io_ops_data;
	struct timespec now;

	if (!fio_fill_issue_time(td))
		return;

	fio_gettime(&now, NULL);

	while (nr--) {
		struct aio_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_aioring_commit(struct thread_data *td)
{
	struct aioring_data *ld = td->io_ops_data;
	struct aioring_options *o = td->eo;
	int ret;

	if (!ld->queued)
		return 0;

	/* Nothing to do */
	if (o->sqthread_poll) {
		struct aio_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_aioring_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_aioring_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_aioring_unmap(struct aioring_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_aioring_cleanup(struct thread_data *td)
{
	struct aioring_data *ld = td->io_ops_data;

	if (ld) {
		td->ts.cachehit += ld->cachehit;
		td->ts.cachemiss += ld->cachemiss;

		/*
		 * Work-around to avoid huge RCU stalls at exit time. If we
		 * don't do this here, then it'll be torn down by exit_aio().
		 * But for that case we can parallellize the freeing, thus
		 * speeding it up a lot.
		 */
		if (!(td->flags & TD_F_CHILD))
			fio_aioring_unmap(ld);

		free(ld->io_u_index);
		free(ld->io_us);
		free(ld->iovecs);
		free(ld);
	}
}

static int fio_aioring_mmap(struct aioring_data *ld, struct aio_uring_params *p)
{
	struct aio_sq_ring *sring = &ld->sq_ring;
	struct aio_cq_ring *cring = &ld->cq_ring;
	void *ptr;

	ld->mmap[0].len = p->sq_off.elems + 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.elems;
	ld->sq_ring_mask = *sring->ring_mask;

	ld->mmap[1].len = p->sq_entries * sizeof(struct 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.elems +
				p->cq_entries * sizeof(struct io_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.elems;
	ld->cq_ring_mask = *cring->ring_mask;
	return 0;
}

static int fio_aioring_queue_init(struct thread_data *td)
{
	struct aioring_data *ld = td->io_ops_data;
	struct aioring_options *o = td->eo;
	int depth = td->o.iodepth;
	struct aio_uring_params p;
	int ret;

	memset(&p, 0, sizeof(p));
	p.flags = IOCTX_FLAG_SCQRING;

	if (o->hipri)
		p.flags |= IOCTX_FLAG_IOPOLL;
	if (o->sqthread_set) {
		p.sq_thread_cpu = o->sqthread;
		p.flags |= IOCTX_FLAG_SQTHREAD;
		if (o->sqthread_poll)
			p.flags |= IOCTX_FLAG_SQPOLL;
	}
	if (o->sqwq)
		p.flags |= IOCTX_FLAG_SQWQ;

	if (o->fixedbufs) {
		struct rlimit rlim = {
			.rlim_cur = RLIM_INFINITY,
			.rlim_max = RLIM_INFINITY,
		};

		setrlimit(RLIMIT_MEMLOCK, &rlim);
		p.flags |= IOCTX_FLAG_FIXEDBUFS;
	}

	ret = syscall(__NR_sys_io_uring_setup, depth, ld->iovecs, &p);
	if (ret < 0)
		return ret;

	ld->ring_fd = ret;
	return fio_aioring_mmap(ld, &p);
}

static int fio_aioring_post_init(struct thread_data *td)
{
	struct aioring_data *ld = td->io_ops_data;
	struct aioring_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_aioring_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_aioring_init(struct thread_data *td)
{
	struct aioring_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_aioring_io_u_init(struct thread_data *td, struct io_u *io_u)
{
	struct aioring_data *ld = td->io_ops_data;

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

static struct ioengine_ops ioengine = {
	.name			= "aio-ring",
	.version		= FIO_IOOPS_VERSION,
	.init			= fio_aioring_init,
	.post_init		= fio_aioring_post_init,
	.io_u_init		= fio_aioring_io_u_init,
	.prep			= fio_aioring_prep,
	.queue			= fio_aioring_queue,
	.commit			= fio_aioring_commit,
	.getevents		= fio_aioring_getevents,
	.event			= fio_aioring_event,
	.cleanup		= fio_aioring_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 aioring_options),
};

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

static void fio_exit fio_aioring_unregister(void)
{
	unregister_ioengine(&ioengine);
}
#endif
Commit	Line	Data
52885fa2 JA	1	/*
	2	* aioring engine
	3	*
a90cd050 JA	4	* IO engine using the new native Linux libaio ring interface. See:
	5	*
	6	* http://git.kernel.dk/cgit/linux-block/log/?h=aio-poll
52885fa2 JA	7	*
	8	*/
	9	#include <stdlib.h>
	10	#include <unistd.h>
	11	#include <errno.h>
	12	#include <libaio.h>
	13	#include <sys/time.h>
	14	#include <sys/resource.h>
	15
	16	#include "../fio.h"
	17	#include "../lib/pow2.h"
	18	#include "../optgroup.h"
	19	#include "../lib/memalign.h"
b87aa01a	20	#include "../lib/fls.h"
52885fa2	21
a30e63cf JA	22	#ifdef ARCH_HAVE_AIORING
a30e63cf JA	23
52885fa2	24	/*
9a2d78b3	25	* io_uring_setup(2) flags
52885fa2	26	*/
52885fa2	27	#ifndef IOCTX_FLAG_SCQRING
9a2d78b3 JA	28	#define IOCTX_FLAG_SCQRING (1 << 0)
	29	#endif
	30	#ifndef IOCTX_FLAG_IOPOLL
	31	#define IOCTX_FLAG_IOPOLL (1 << 1)
52885fa2 JA	32	#endif
	33	#ifndef IOCTX_FLAG_FIXEDBUFS
	34	#define IOCTX_FLAG_FIXEDBUFS (1 << 2)
	35	#endif
	36	#ifndef IOCTX_FLAG_SQTHREAD
	37	#define IOCTX_FLAG_SQTHREAD (1 << 3)
	38	#endif
	39	#ifndef IOCTX_FLAG_SQWQ
	40	#define IOCTX_FLAG_SQWQ (1 << 4)
	41	#endif
771c9901 JA	42	#ifndef IOCTX_FLAG_SQPOLL
	43	#define IOCTX_FLAG_SQPOLL (1 << 5)
	44	#endif
	45
9a2d78b3 JA	46	#define IORING_OFF_SQ_RING 0ULL
	47	#define IORING_OFF_CQ_RING 0x8000000ULL
	48	#define IORING_OFF_IOCB 0x10000000ULL
52885fa2 JA	49
52885fa2 JA	50	/*
9a2d78b3	51	* io_uring_enter(2) flags
52885fa2	52	*/
9a2d78b3 JA	53	#ifndef IORING_ENTER_GETEVENTS
9a2d78b3 JA	54	#define IORING_ENTER_GETEVENTS (1 << 0)
52885fa2 JA	55	#endif
	56
	57	typedef uint64_t u64;
	58	typedef uint32_t u32;
	59	typedef uint16_t u16;
	60
4cdbc048 JA	61	#define IORING_SQ_NEED_WAKEUP (1 << 0)
4cdbc048 JA	62
96563db9 JA	63	#define IOEV_RES2_CACHEHIT (1 << 0)
96563db9 JA	64
9a2d78b3 JA	65	struct aio_uring_offsets {
	66	u32 head;
	67	u32 tail;
	68	u32 ring_mask;
	69	u32 ring_entries;
	70	u32 flags;
	71	u32 elems;
	72	};
	73
	74	struct aio_uring_params {
	75	u32 sq_entries;
	76	u32 cq_entries;
	77	u32 flags;
	78	u16 sq_thread_cpu;
	79	u16 resv[9];
	80	struct aio_uring_offsets sq_off;
	81	struct aio_uring_offsets cq_off;
	82	};
	83
52885fa2	84	struct aio_sq_ring {
9a2d78b3 JA	85	u32 *head;
	86	u32 *tail;
	87	u32 *ring_mask;
	88	u32 *ring_entries;
	89	u32 *flags;
	90	u32 *array;
52885fa2 JA	91	};
	92
	93	struct aio_cq_ring {
9a2d78b3 JA	94	u32 *head;
	95	u32 *tail;
	96	u32 *ring_mask;
	97	u32 *ring_entries;
	98	struct io_event *events;
	99	};
	100
	101	struct aioring_mmap {
	102	void *ptr;
	103	size_t len;
52885fa2 JA	104	};
	105
	106	struct aioring_data {
9a2d78b3 JA	107	int ring_fd;
9a2d78b3 JA	108
52885fa2 JA	109	struct io_u **io_us;
	110	struct io_u **io_u_index;
	111
9a2d78b3	112	struct aio_sq_ring sq_ring;
52885fa2	113	struct iocb *iocbs;
9a2d78b3	114	struct iovec *iovecs;
b87aa01a	115	unsigned sq_ring_mask;
52885fa2	116
9a2d78b3	117	struct aio_cq_ring cq_ring;
52885fa2	118	struct io_event *events;
b87aa01a	119	unsigned cq_ring_mask;
52885fa2 JA	120
	121	int queued;
	122	int cq_ring_off;
b87aa01a	123	unsigned iodepth;
96563db9 JA	124
	125	uint64_t cachehit;
	126	uint64_t cachemiss;
9a2d78b3 JA	127
9a2d78b3 JA	128	struct aioring_mmap mmap[3];
52885fa2 JA	129	};
	130
	131	struct aioring_options {
	132	void *pad;
	133	unsigned int hipri;
	134	unsigned int fixedbufs;
a90cd050 JA	135	unsigned int sqthread;
a90cd050 JA	136	unsigned int sqthread_set;
771c9901	137	unsigned int sqthread_poll;
a90cd050	138	unsigned int sqwq;
52885fa2 JA	139	};
52885fa2 JA	140
a90cd050 JA	141	static int fio_aioring_sqthread_cb(void *data,
	142	unsigned long long *val)
	143	{
	144	struct aioring_options *o = data;
	145
	146	o->sqthread = *val;
	147	o->sqthread_set = 1;
	148	return 0;
	149	}
	150
52885fa2 JA	151	static struct fio_option options[] = {
	152	{
	153	.name = "hipri",
	154	.lname = "High Priority",
	155	.type = FIO_OPT_STR_SET,
	156	.off1 = offsetof(struct aioring_options, hipri),
	157	.help = "Use polled IO completions",
	158	.category = FIO_OPT_C_ENGINE,
	159	.group = FIO_OPT_G_LIBAIO,
	160	},
	161	{
	162	.name = "fixedbufs",
	163	.lname = "Fixed (pre-mapped) IO buffers",
	164	.type = FIO_OPT_STR_SET,
	165	.off1 = offsetof(struct aioring_options, fixedbufs),
	166	.help = "Pre map IO buffers",
	167	.category = FIO_OPT_C_ENGINE,
	168	.group = FIO_OPT_G_LIBAIO,
	169	},
a90cd050 JA	170	{
	171	.name = "sqthread",
	172	.lname = "Use kernel SQ thread on this CPU",
	173	.type = FIO_OPT_INT,
	174	.cb = fio_aioring_sqthread_cb,
	175	.help = "Offload submission to kernel thread",
	176	.category = FIO_OPT_C_ENGINE,
	177	.group = FIO_OPT_G_LIBAIO,
	178	},
771c9901 JA	179	{
	180	.name = "sqthread_poll",
	181	.lname = "Kernel SQ thread should poll",
	182	.type = FIO_OPT_STR_SET,
	183	.off1 = offsetof(struct aioring_options, sqthread_poll),
	184	.help = "Used with sqthread, enables kernel side polling",
	185	.category = FIO_OPT_C_ENGINE,
	186	.group = FIO_OPT_G_LIBAIO,
	187	},
a90cd050 JA	188	{
	189	.name = "sqwq",
	190	.lname = "Offload submission to kernel workqueue",
	191	.type = FIO_OPT_STR_SET,
	192	.off1 = offsetof(struct aioring_options, sqwq),
	193	.help = "Offload submission to kernel workqueue",
	194	.category = FIO_OPT_C_ENGINE,
	195	.group = FIO_OPT_G_LIBAIO,
	196	},
52885fa2 JA	197	{
	198	.name = NULL,
	199	},
	200	};
	201
9a2d78b3	202	static int io_uring_enter(struct aioring_data *ld, unsigned int to_submit,
52885fa2 JA	203	unsigned int min_complete, unsigned int flags)
52885fa2 JA	204	{
9a2d78b3 JA	205	return syscall(__NR_sys_io_uring_enter, ld->ring_fd, to_submit,
9a2d78b3 JA	206	min_complete, flags);
52885fa2 JA	207	}
	208
	209	static int fio_aioring_prep(struct thread_data td, struct io_u io_u)
	210	{
	211	struct aioring_data *ld = td->io_ops_data;
	212	struct fio_file *f = io_u->file;
52885fa2 JA	213	struct iocb *iocb;
	214
	215	iocb = &ld->iocbs[io_u->index];
	216
e3970057 JA	217	if (io_u->ddir == DDIR_READ \|\| io_u->ddir == DDIR_WRITE) {
e3970057 JA	218	if (io_u->ddir == DDIR_READ)
52885fa2	219	iocb->aio_lio_opcode = IO_CMD_PREAD;
e3970057	220	else
52885fa2	221	iocb->aio_lio_opcode = IO_CMD_PWRITE;
e3970057 JA	222	iocb->aio_reqprio = 0;
	223	iocb->aio_fildes = f->fd;
	224	iocb->u.c.buf = io_u->xfer_buf;
	225	iocb->u.c.nbytes = io_u->xfer_buflen;
	226	iocb->u.c.offset = io_u->offset;
ac4f3d4e	227	iocb->u.c.flags = 0;
52885fa2 JA	228	} else if (ddir_sync(io_u->ddir))
	229	io_prep_fsync(iocb, f->fd);
	230
	231	iocb->data = io_u;
	232	return 0;
	233	}
	234
	235	static struct io_u fio_aioring_event(struct thread_data td, int event)
	236	{
	237	struct aioring_data *ld = td->io_ops_data;
	238	struct io_event *ev;
	239	struct io_u *io_u;
b87aa01a	240	unsigned index;
52885fa2	241
b87aa01a	242	index = (event + ld->cq_ring_off) & ld->cq_ring_mask;
52885fa2	243
9a2d78b3	244	ev = &ld->cq_ring.events[index];
52885fa2 JA	245	io_u = ev->data;
	246
	247	if (ev->res != io_u->xfer_buflen) {
	248	if (ev->res > io_u->xfer_buflen)
	249	io_u->error = -ev->res;
	250	else
	251	io_u->resid = io_u->xfer_buflen - ev->res;
	252	} else
	253	io_u->error = 0;
	254
96563db9 JA	255	if (io_u->ddir == DDIR_READ) {
	256	if (ev->res2 & IOEV_RES2_CACHEHIT)
	257	ld->cachehit++;
	258	else
	259	ld->cachemiss++;
	260	}
	261
52885fa2 JA	262	return io_u;
	263	}
	264
	265	static int fio_aioring_cqring_reap(struct thread_data *td, unsigned int events,
	266	unsigned int max)
	267	{
	268	struct aioring_data *ld = td->io_ops_data;
9a2d78b3	269	struct aio_cq_ring *ring = &ld->cq_ring;
52885fa2 JA	270	u32 head, reaped = 0;
52885fa2 JA	271
9a2d78b3	272	head = *ring->head;
52885fa2 JA	273	do {
52885fa2 JA	274	read_barrier();
9a2d78b3	275	if (head == *ring->tail)
52885fa2 JA	276	break;
	277	reaped++;
	278	head++;
52885fa2 JA	279	} while (reaped + events < max);
52885fa2 JA	280
9a2d78b3	281	*ring->head = head;
52885fa2 JA	282	write_barrier();
	283	return reaped;
	284	}
	285
	286	static int fio_aioring_getevents(struct thread_data *td, unsigned int min,
	287	unsigned int max, const struct timespec *t)
	288	{
	289	struct aioring_data *ld = td->io_ops_data;
	290	unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
771c9901	291	struct aioring_options *o = td->eo;
9a2d78b3	292	struct aio_cq_ring *ring = &ld->cq_ring;
b87aa01a JA	293	unsigned events = 0;
b87aa01a JA	294	int r;
52885fa2	295
9a2d78b3	296	ld->cq_ring_off = *ring->head;
52885fa2 JA	297	do {
	298	r = fio_aioring_cqring_reap(td, events, max);
	299	if (r) {
	300	events += r;
	301	continue;
	302	}
	303
771c9901	304	if (!o->sqthread_poll) {
9a2d78b3 JA	305	r = io_uring_enter(ld, 0, actual_min,
9a2d78b3 JA	306	IORING_ENTER_GETEVENTS);
771c9901 JA	307	if (r < 0) {
	308	if (errno == EAGAIN)
	309	continue;
9a2d78b3	310	td_verror(td, errno, "io_uring_enter");
771c9901 JA	311	break;
771c9901 JA	312	}
52885fa2 JA	313	}
	314	} while (events < min);
	315
	316	return r < 0 ? r : events;
	317	}
	318
	319	static enum fio_q_status fio_aioring_queue(struct thread_data *td,
	320	struct io_u *io_u)
	321	{
	322	struct aioring_data *ld = td->io_ops_data;
9a2d78b3	323	struct aio_sq_ring *ring = &ld->sq_ring;
52885fa2 JA	324	unsigned tail, next_tail;
	325
	326	fio_ro_check(td, io_u);
	327
b87aa01a	328	if (ld->queued == ld->iodepth)
52885fa2 JA	329	return FIO_Q_BUSY;
52885fa2 JA	330
52885fa2 JA	331	if (io_u->ddir == DDIR_TRIM) {
	332	if (ld->queued)
	333	return FIO_Q_BUSY;
	334
	335	do_io_u_trim(td, io_u);
	336	io_u_mark_submit(td, 1);
	337	io_u_mark_complete(td, 1);
	338	return FIO_Q_COMPLETED;
	339	}
	340
9a2d78b3	341	tail = *ring->tail;
52885fa2	342	next_tail = tail + 1;
52885fa2	343	read_barrier();
9a2d78b3	344	if (next_tail == *ring->head)
52885fa2 JA	345	return FIO_Q_BUSY;
52885fa2 JA	346
b87aa01a	347	ring->array[tail & ld->sq_ring_mask] = io_u->index;
9a2d78b3	348	*ring->tail = next_tail;
52885fa2 JA	349	write_barrier();
	350
	351	ld->queued++;
	352	return FIO_Q_QUEUED;
	353	}
	354
	355	static void fio_aioring_queued(struct thread_data *td, int start, int nr)
	356	{
	357	struct aioring_data *ld = td->io_ops_data;
	358	struct timespec now;
	359
	360	if (!fio_fill_issue_time(td))
	361	return;
	362
	363	fio_gettime(&now, NULL);
	364
	365	while (nr--) {
9a2d78b3 JA	366	struct aio_sq_ring *ring = &ld->sq_ring;
9a2d78b3 JA	367	int index = ring->array[start & ld->sq_ring_mask];
f8289afc	368	struct io_u *io_u = ld->io_u_index[index];
52885fa2 JA	369
	370	memcpy(&io_u->issue_time, &now, sizeof(now));
	371	io_u_queued(td, io_u);
	372
	373	start++;
52885fa2 JA	374	}
	375	}
	376
	377	static int fio_aioring_commit(struct thread_data *td)
	378	{
	379	struct aioring_data *ld = td->io_ops_data;
771c9901	380	struct aioring_options *o = td->eo;
52885fa2 JA	381	int ret;
	382
	383	if (!ld->queued)
	384	return 0;
	385
771c9901 JA	386	/* Nothing to do */
771c9901 JA	387	if (o->sqthread_poll) {
9a2d78b3	388	struct aio_sq_ring *ring = &ld->sq_ring;
4cdbc048	389
9a2d78b3 JA	390	if (*ring->flags & IORING_SQ_NEED_WAKEUP)
9a2d78b3 JA	391	io_uring_enter(ld, ld->queued, 0, 0);
771c9901 JA	392	ld->queued = 0;
	393	return 0;
	394	}
	395
52885fa2	396	do {
9a2d78b3	397	unsigned start = *ld->sq_ring.head;
52885fa2 JA	398	long nr = ld->queued;
52885fa2 JA	399
9a2d78b3	400	ret = io_uring_enter(ld, nr, 0, IORING_ENTER_GETEVENTS);
52885fa2 JA	401	if (ret > 0) {
	402	fio_aioring_queued(td, start, ret);
	403	io_u_mark_submit(td, ret);
	404
	405	ld->queued -= ret;
	406	ret = 0;
a90cd050 JA	407	} else if (!ret) {
a90cd050 JA	408	io_u_mark_submit(td, ret);
52885fa2	409	continue;
a90cd050 JA	410	} else {
	411	if (errno == EAGAIN) {
	412	ret = fio_aioring_cqring_reap(td, 0, ld->queued);
	413	if (ret)
	414	continue;
	415	/* Shouldn't happen */
	416	usleep(1);
	417	continue;
52885fa2	418	}
9a2d78b3	419	td_verror(td, errno, "io_uring_enter submit");
52885fa2	420	break;
a90cd050	421	}
52885fa2 JA	422	} while (ld->queued);
	423
	424	return ret;
	425	}
	426
9a2d78b3	427	static void fio_aioring_unmap(struct aioring_data *ld)
52885fa2	428	{
9a2d78b3	429	int i;
52885fa2	430
9a2d78b3 JA	431	for (i = 0; i < ARRAY_SIZE(ld->mmap); i++)
	432	munmap(ld->mmap[i].ptr, ld->mmap[i].len);
	433	close(ld->ring_fd);
b87aa01a JA	434	}
b87aa01a JA	435
52885fa2 JA	436	static void fio_aioring_cleanup(struct thread_data *td)
	437	{
	438	struct aioring_data *ld = td->io_ops_data;
	439
	440	if (ld) {
96563db9 JA	441	td->ts.cachehit += ld->cachehit;
	442	td->ts.cachemiss += ld->cachemiss;
	443
52885fa2 JA	444	/*
	445	* Work-around to avoid huge RCU stalls at exit time. If we
	446	* don't do this here, then it'll be torn down by exit_aio().
	447	* But for that case we can parallellize the freeing, thus
	448	* speeding it up a lot.
	449	*/
	450	if (!(td->flags & TD_F_CHILD))
9a2d78b3 JA	451	fio_aioring_unmap(ld);
9a2d78b3 JA	452
52885fa2 JA	453	free(ld->io_u_index);
52885fa2 JA	454	free(ld->io_us);
9a2d78b3	455	free(ld->iovecs);
52885fa2 JA	456	free(ld);
	457	}
	458	}
	459
9a2d78b3 JA	460	static int fio_aioring_mmap(struct aioring_data ld, struct aio_uring_params p)
	461	{
	462	struct aio_sq_ring *sring = &ld->sq_ring;
	463	struct aio_cq_ring *cring = &ld->cq_ring;
	464	void *ptr;
	465
	466	ld->mmap[0].len = p->sq_off.elems + p->sq_entries * sizeof(u32);
	467	ptr = mmap(0, ld->mmap[0].len, PROT_READ \| PROT_WRITE,
	468	MAP_SHARED \| MAP_POPULATE, ld->ring_fd,
	469	IORING_OFF_SQ_RING);
	470	ld->mmap[0].ptr = ptr;
	471	sring->head = ptr + p->sq_off.head;
	472	sring->tail = ptr + p->sq_off.tail;
	473	sring->ring_mask = ptr + p->sq_off.ring_mask;
	474	sring->ring_entries = ptr + p->sq_off.ring_entries;
	475	sring->flags = ptr + p->sq_off.flags;
	476	sring->array = ptr + p->sq_off.elems;
	477	ld->sq_ring_mask = *sring->ring_mask;
	478
	479	ld->mmap[1].len = p->sq_entries * sizeof(struct iocb);
	480	ld->iocbs = mmap(0, ld->mmap[1].len, PROT_READ \| PROT_WRITE,
	481	MAP_SHARED \| MAP_POPULATE, ld->ring_fd,
	482	IORING_OFF_IOCB);
	483	ld->mmap[1].ptr = ld->iocbs;
	484
	485	ld->mmap[2].len = p->cq_off.elems +
	486	p->cq_entries * sizeof(struct io_event);
	487	ptr = mmap(0, ld->mmap[2].len, PROT_READ \| PROT_WRITE,
	488	MAP_SHARED \| MAP_POPULATE, ld->ring_fd,
	489	IORING_OFF_CQ_RING);
	490	ld->mmap[2].ptr = ptr;
	491	cring->head = ptr + p->cq_off.head;
	492	cring->tail = ptr + p->cq_off.tail;
	493	cring->ring_mask = ptr + p->cq_off.ring_mask;
	494	cring->ring_entries = ptr + p->cq_off.ring_entries;
	495	cring->events = ptr + p->cq_off.elems;
	496	ld->cq_ring_mask = *cring->ring_mask;
	497	return 0;
	498	}
	499
52885fa2 JA	500	static int fio_aioring_queue_init(struct thread_data *td)
52885fa2 JA	501	{
52885fa2 JA	502	struct aioring_data *ld = td->io_ops_data;
52885fa2 JA	503	struct aioring_options *o = td->eo;
52885fa2	504	int depth = td->o.iodepth;
9a2d78b3 JA	505	struct aio_uring_params p;
	506	int ret;
	507
	508	memset(&p, 0, sizeof(p));
	509	p.flags = IOCTX_FLAG_SCQRING;
52885fa2 JA	510
52885fa2 JA	511	if (o->hipri)
9a2d78b3	512	p.flags \|= IOCTX_FLAG_IOPOLL;
a90cd050	513	if (o->sqthread_set) {
9a2d78b3 JA	514	p.sq_thread_cpu = o->sqthread;
9a2d78b3 JA	515	p.flags \|= IOCTX_FLAG_SQTHREAD;
771c9901	516	if (o->sqthread_poll)
9a2d78b3	517	p.flags \|= IOCTX_FLAG_SQPOLL;
f635f1fb JA	518	}
f635f1fb JA	519	if (o->sqwq)
9a2d78b3	520	p.flags \|= IOCTX_FLAG_SQWQ;
a90cd050	521
52885fa2 JA	522	if (o->fixedbufs) {
	523	struct rlimit rlim = {
	524	.rlim_cur = RLIM_INFINITY,
	525	.rlim_max = RLIM_INFINITY,
	526	};
	527
	528	setrlimit(RLIMIT_MEMLOCK, &rlim);
9a2d78b3	529	p.flags \|= IOCTX_FLAG_FIXEDBUFS;
52885fa2 JA	530	}
52885fa2 JA	531
9a2d78b3 JA	532	ret = syscall(__NR_sys_io_uring_setup, depth, ld->iovecs, &p);
	533	if (ret < 0)
	534	return ret;
	535
	536	ld->ring_fd = ret;
	537	return fio_aioring_mmap(ld, &p);
52885fa2 JA	538	}
	539
	540	static int fio_aioring_post_init(struct thread_data *td)
	541	{
	542	struct aioring_data *ld = td->io_ops_data;
	543	struct aioring_options *o = td->eo;
	544	struct io_u *io_u;
9a2d78b3	545	int err;
52885fa2 JA	546
	547	if (o->fixedbufs) {
	548	int i;
	549
	550	for (i = 0; i < td->o.iodepth; i++) {
9a2d78b3 JA	551	struct iovec *iov = &ld->iovecs[i];
9a2d78b3 JA	552
52885fa2	553	io_u = ld->io_u_index[i];
9a2d78b3 JA	554	iov->iov_base = io_u->buf;
9a2d78b3 JA	555	iov->iov_len = td_max_bs(td);
52885fa2 JA	556	}
	557	}
	558
	559	err = fio_aioring_queue_init(td);
	560	if (err) {
d63a472d	561	td_verror(td, errno, "io_queue_init");
52885fa2 JA	562	return 1;
	563	}
	564
	565	return 0;
	566	}
	567
9a2d78b3 JA	568	static unsigned roundup_pow2(unsigned depth)
	569	{
	570	return 1UL << __fls(depth - 1);
	571	}
	572
52885fa2 JA	573	static int fio_aioring_init(struct thread_data *td)
	574	{
	575	struct aioring_data *ld;
	576
52885fa2 JA	577	ld = calloc(1, sizeof(*ld));
52885fa2 JA	578
b87aa01a JA	579	/* ring depth must be a power-of-2 */
	580	ld->iodepth = td->o.iodepth;
	581	td->o.iodepth = roundup_pow2(td->o.iodepth);
	582
52885fa2 JA	583	/* io_u index */
	584	ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *));
	585	ld->io_us = calloc(td->o.iodepth, sizeof(struct io_u *));
	586
9a2d78b3	587	ld->iovecs = calloc(td->o.iodepth, sizeof(struct iovec));
52885fa2 JA	588
	589	td->io_ops_data = ld;
	590	return 0;
	591	}
	592
	593	static int fio_aioring_io_u_init(struct thread_data td, struct io_u io_u)
	594	{
	595	struct aioring_data *ld = td->io_ops_data;
	596
	597	ld->io_u_index[io_u->index] = io_u;
	598	return 0;
	599	}
	600
	601	static struct ioengine_ops ioengine = {
	602	.name = "aio-ring",
	603	.version = FIO_IOOPS_VERSION,
	604	.init = fio_aioring_init,
	605	.post_init = fio_aioring_post_init,
	606	.io_u_init = fio_aioring_io_u_init,
	607	.prep = fio_aioring_prep,
	608	.queue = fio_aioring_queue,
	609	.commit = fio_aioring_commit,
	610	.getevents = fio_aioring_getevents,
	611	.event = fio_aioring_event,
	612	.cleanup = fio_aioring_cleanup,
	613	.open_file = generic_open_file,
	614	.close_file = generic_close_file,
	615	.get_file_size = generic_get_file_size,
	616	.options = options,
	617	.option_struct_size = sizeof(struct aioring_options),
	618	};
	619
	620	static void fio_init fio_aioring_register(void)
	621	{
52885fa2	622	register_ioengine(&ioengine);
52885fa2 JA	623	}
	624
	625	static void fio_exit fio_aioring_unregister(void)
	626	{
52885fa2	627	unregister_ioengine(&ioengine);
52885fa2	628	}
1f90e9bb	629	#endif