[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_sqring_offsets {
	u32 head;
	u32 tail;
	u32 ring_mask;
	u32 ring_entries;
	u32 flags;
	u32 array;
};

struct aio_cqring_offsets {
	u32 head;
	u32 tail;
	u32 ring_mask;
	u32 ring_entries;
	u32 overflow;
	u32 events;
};

struct aio_uring_params {
	u32 sq_entries;
	u32 cq_entries;
	u32 flags;
	u16 sq_thread_cpu;
	u16 resv[9];
	struct aio_sqring_offsets sq_off;
	struct aio_cqring_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.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 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_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_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
ac122fea	65	struct aio_sqring_offsets {
9a2d78b3 JA	66	u32 head;
	67	u32 tail;
	68	u32 ring_mask;
	69	u32 ring_entries;
	70	u32 flags;
ac122fea JA	71	u32 array;
	72	};
	73
	74	struct aio_cqring_offsets {
	75	u32 head;
	76	u32 tail;
	77	u32 ring_mask;
	78	u32 ring_entries;
	79	u32 overflow;
	80	u32 events;
9a2d78b3 JA	81	};
	82
	83	struct aio_uring_params {
	84	u32 sq_entries;
	85	u32 cq_entries;
	86	u32 flags;
	87	u16 sq_thread_cpu;
	88	u16 resv[9];
ac122fea JA	89	struct aio_sqring_offsets sq_off;
ac122fea JA	90	struct aio_cqring_offsets cq_off;
9a2d78b3 JA	91	};
9a2d78b3 JA	92
52885fa2	93	struct aio_sq_ring {
9a2d78b3 JA	94	u32 *head;
	95	u32 *tail;
	96	u32 *ring_mask;
	97	u32 *ring_entries;
	98	u32 *flags;
	99	u32 *array;
52885fa2 JA	100	};
	101
	102	struct aio_cq_ring {
9a2d78b3 JA	103	u32 *head;
	104	u32 *tail;
	105	u32 *ring_mask;
	106	u32 *ring_entries;
	107	struct io_event *events;
	108	};
	109
	110	struct aioring_mmap {
	111	void *ptr;
	112	size_t len;
52885fa2 JA	113	};
	114
	115	struct aioring_data {
9a2d78b3 JA	116	int ring_fd;
9a2d78b3 JA	117
52885fa2 JA	118	struct io_u **io_us;
	119	struct io_u **io_u_index;
	120
9a2d78b3	121	struct aio_sq_ring sq_ring;
52885fa2	122	struct iocb *iocbs;
9a2d78b3	123	struct iovec *iovecs;
b87aa01a	124	unsigned sq_ring_mask;
52885fa2	125
9a2d78b3	126	struct aio_cq_ring cq_ring;
52885fa2	127	struct io_event *events;
b87aa01a	128	unsigned cq_ring_mask;
52885fa2 JA	129
	130	int queued;
	131	int cq_ring_off;
b87aa01a	132	unsigned iodepth;
96563db9 JA	133
	134	uint64_t cachehit;
	135	uint64_t cachemiss;
9a2d78b3 JA	136
9a2d78b3 JA	137	struct aioring_mmap mmap[3];
52885fa2 JA	138	};
	139
	140	struct aioring_options {
	141	void *pad;
	142	unsigned int hipri;
	143	unsigned int fixedbufs;
a90cd050 JA	144	unsigned int sqthread;
a90cd050 JA	145	unsigned int sqthread_set;
771c9901	146	unsigned int sqthread_poll;
a90cd050	147	unsigned int sqwq;
52885fa2 JA	148	};
52885fa2 JA	149
a90cd050 JA	150	static int fio_aioring_sqthread_cb(void *data,
	151	unsigned long long *val)
	152	{
	153	struct aioring_options *o = data;
	154
	155	o->sqthread = *val;
	156	o->sqthread_set = 1;
	157	return 0;
	158	}
	159
52885fa2 JA	160	static struct fio_option options[] = {
	161	{
	162	.name = "hipri",
	163	.lname = "High Priority",
	164	.type = FIO_OPT_STR_SET,
	165	.off1 = offsetof(struct aioring_options, hipri),
	166	.help = "Use polled IO completions",
	167	.category = FIO_OPT_C_ENGINE,
	168	.group = FIO_OPT_G_LIBAIO,
	169	},
	170	{
	171	.name = "fixedbufs",
	172	.lname = "Fixed (pre-mapped) IO buffers",
	173	.type = FIO_OPT_STR_SET,
	174	.off1 = offsetof(struct aioring_options, fixedbufs),
	175	.help = "Pre map IO buffers",
	176	.category = FIO_OPT_C_ENGINE,
	177	.group = FIO_OPT_G_LIBAIO,
	178	},
a90cd050 JA	179	{
	180	.name = "sqthread",
	181	.lname = "Use kernel SQ thread on this CPU",
	182	.type = FIO_OPT_INT,
	183	.cb = fio_aioring_sqthread_cb,
	184	.help = "Offload submission to kernel thread",
	185	.category = FIO_OPT_C_ENGINE,
	186	.group = FIO_OPT_G_LIBAIO,
	187	},
771c9901 JA	188	{
	189	.name = "sqthread_poll",
	190	.lname = "Kernel SQ thread should poll",
	191	.type = FIO_OPT_STR_SET,
	192	.off1 = offsetof(struct aioring_options, sqthread_poll),
	193	.help = "Used with sqthread, enables kernel side polling",
	194	.category = FIO_OPT_C_ENGINE,
	195	.group = FIO_OPT_G_LIBAIO,
	196	},
a90cd050 JA	197	{
	198	.name = "sqwq",
	199	.lname = "Offload submission to kernel workqueue",
	200	.type = FIO_OPT_STR_SET,
	201	.off1 = offsetof(struct aioring_options, sqwq),
	202	.help = "Offload submission to kernel workqueue",
	203	.category = FIO_OPT_C_ENGINE,
	204	.group = FIO_OPT_G_LIBAIO,
	205	},
52885fa2 JA	206	{
	207	.name = NULL,
	208	},
	209	};
	210
9a2d78b3	211	static int io_uring_enter(struct aioring_data *ld, unsigned int to_submit,
52885fa2 JA	212	unsigned int min_complete, unsigned int flags)
52885fa2 JA	213	{
9a2d78b3 JA	214	return syscall(__NR_sys_io_uring_enter, ld->ring_fd, to_submit,
9a2d78b3 JA	215	min_complete, flags);
52885fa2 JA	216	}
	217
	218	static int fio_aioring_prep(struct thread_data td, struct io_u io_u)
	219	{
	220	struct aioring_data *ld = td->io_ops_data;
	221	struct fio_file *f = io_u->file;
52885fa2 JA	222	struct iocb *iocb;
	223
	224	iocb = &ld->iocbs[io_u->index];
	225
e3970057 JA	226	if (io_u->ddir == DDIR_READ \|\| io_u->ddir == DDIR_WRITE) {
e3970057 JA	227	if (io_u->ddir == DDIR_READ)
52885fa2	228	iocb->aio_lio_opcode = IO_CMD_PREAD;
e3970057	229	else
52885fa2	230	iocb->aio_lio_opcode = IO_CMD_PWRITE;
e3970057 JA	231	iocb->aio_reqprio = 0;
	232	iocb->aio_fildes = f->fd;
	233	iocb->u.c.buf = io_u->xfer_buf;
	234	iocb->u.c.nbytes = io_u->xfer_buflen;
	235	iocb->u.c.offset = io_u->offset;
ac4f3d4e	236	iocb->u.c.flags = 0;
52885fa2 JA	237	} else if (ddir_sync(io_u->ddir))
	238	io_prep_fsync(iocb, f->fd);
	239
	240	iocb->data = io_u;
	241	return 0;
	242	}
	243
	244	static struct io_u fio_aioring_event(struct thread_data td, int event)
	245	{
	246	struct aioring_data *ld = td->io_ops_data;
	247	struct io_event *ev;
	248	struct io_u *io_u;
b87aa01a	249	unsigned index;
52885fa2	250
b87aa01a	251	index = (event + ld->cq_ring_off) & ld->cq_ring_mask;
52885fa2	252
9a2d78b3	253	ev = &ld->cq_ring.events[index];
52885fa2 JA	254	io_u = ev->data;
	255
	256	if (ev->res != io_u->xfer_buflen) {
	257	if (ev->res > io_u->xfer_buflen)
	258	io_u->error = -ev->res;
	259	else
	260	io_u->resid = io_u->xfer_buflen - ev->res;
	261	} else
	262	io_u->error = 0;
	263
96563db9 JA	264	if (io_u->ddir == DDIR_READ) {
	265	if (ev->res2 & IOEV_RES2_CACHEHIT)
	266	ld->cachehit++;
	267	else
	268	ld->cachemiss++;
	269	}
	270
52885fa2 JA	271	return io_u;
	272	}
	273
	274	static int fio_aioring_cqring_reap(struct thread_data *td, unsigned int events,
	275	unsigned int max)
	276	{
	277	struct aioring_data *ld = td->io_ops_data;
9a2d78b3	278	struct aio_cq_ring *ring = &ld->cq_ring;
52885fa2 JA	279	u32 head, reaped = 0;
52885fa2 JA	280
9a2d78b3	281	head = *ring->head;
52885fa2 JA	282	do {
52885fa2 JA	283	read_barrier();
9a2d78b3	284	if (head == *ring->tail)
52885fa2 JA	285	break;
	286	reaped++;
	287	head++;
52885fa2 JA	288	} while (reaped + events < max);
52885fa2 JA	289
9a2d78b3	290	*ring->head = head;
52885fa2 JA	291	write_barrier();
	292	return reaped;
	293	}
	294
	295	static int fio_aioring_getevents(struct thread_data *td, unsigned int min,
	296	unsigned int max, const struct timespec *t)
	297	{
	298	struct aioring_data *ld = td->io_ops_data;
	299	unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
771c9901	300	struct aioring_options *o = td->eo;
9a2d78b3	301	struct aio_cq_ring *ring = &ld->cq_ring;
b87aa01a JA	302	unsigned events = 0;
b87aa01a JA	303	int r;
52885fa2	304
9a2d78b3	305	ld->cq_ring_off = *ring->head;
52885fa2 JA	306	do {
	307	r = fio_aioring_cqring_reap(td, events, max);
	308	if (r) {
	309	events += r;
	310	continue;
	311	}
	312
771c9901	313	if (!o->sqthread_poll) {
9a2d78b3 JA	314	r = io_uring_enter(ld, 0, actual_min,
9a2d78b3 JA	315	IORING_ENTER_GETEVENTS);
771c9901 JA	316	if (r < 0) {
	317	if (errno == EAGAIN)
	318	continue;
9a2d78b3	319	td_verror(td, errno, "io_uring_enter");
771c9901 JA	320	break;
771c9901 JA	321	}
52885fa2 JA	322	}
	323	} while (events < min);
	324
	325	return r < 0 ? r : events;
	326	}
	327
	328	static enum fio_q_status fio_aioring_queue(struct thread_data *td,
	329	struct io_u *io_u)
	330	{
	331	struct aioring_data *ld = td->io_ops_data;
9a2d78b3	332	struct aio_sq_ring *ring = &ld->sq_ring;
52885fa2 JA	333	unsigned tail, next_tail;
	334
	335	fio_ro_check(td, io_u);
	336
b87aa01a	337	if (ld->queued == ld->iodepth)
52885fa2 JA	338	return FIO_Q_BUSY;
52885fa2 JA	339
52885fa2 JA	340	if (io_u->ddir == DDIR_TRIM) {
	341	if (ld->queued)
	342	return FIO_Q_BUSY;
	343
	344	do_io_u_trim(td, io_u);
	345	io_u_mark_submit(td, 1);
	346	io_u_mark_complete(td, 1);
	347	return FIO_Q_COMPLETED;
	348	}
	349
9a2d78b3	350	tail = *ring->tail;
52885fa2	351	next_tail = tail + 1;
52885fa2	352	read_barrier();
9a2d78b3	353	if (next_tail == *ring->head)
52885fa2 JA	354	return FIO_Q_BUSY;
52885fa2 JA	355
b87aa01a	356	ring->array[tail & ld->sq_ring_mask] = io_u->index;
9a2d78b3	357	*ring->tail = next_tail;
52885fa2 JA	358	write_barrier();
	359
	360	ld->queued++;
	361	return FIO_Q_QUEUED;
	362	}
	363
	364	static void fio_aioring_queued(struct thread_data *td, int start, int nr)
	365	{
	366	struct aioring_data *ld = td->io_ops_data;
	367	struct timespec now;
	368
	369	if (!fio_fill_issue_time(td))
	370	return;
	371
	372	fio_gettime(&now, NULL);
	373
	374	while (nr--) {
9a2d78b3 JA	375	struct aio_sq_ring *ring = &ld->sq_ring;
9a2d78b3 JA	376	int index = ring->array[start & ld->sq_ring_mask];
f8289afc	377	struct io_u *io_u = ld->io_u_index[index];
52885fa2 JA	378
	379	memcpy(&io_u->issue_time, &now, sizeof(now));
	380	io_u_queued(td, io_u);
	381
	382	start++;
52885fa2 JA	383	}
	384	}
	385
	386	static int fio_aioring_commit(struct thread_data *td)
	387	{
	388	struct aioring_data *ld = td->io_ops_data;
771c9901	389	struct aioring_options *o = td->eo;
52885fa2 JA	390	int ret;
	391
	392	if (!ld->queued)
	393	return 0;
	394
771c9901 JA	395	/* Nothing to do */
771c9901 JA	396	if (o->sqthread_poll) {
9a2d78b3	397	struct aio_sq_ring *ring = &ld->sq_ring;
4cdbc048	398
9a2d78b3 JA	399	if (*ring->flags & IORING_SQ_NEED_WAKEUP)
9a2d78b3 JA	400	io_uring_enter(ld, ld->queued, 0, 0);
771c9901 JA	401	ld->queued = 0;
	402	return 0;
	403	}
	404
52885fa2	405	do {
9a2d78b3	406	unsigned start = *ld->sq_ring.head;
52885fa2 JA	407	long nr = ld->queued;
52885fa2 JA	408
9a2d78b3	409	ret = io_uring_enter(ld, nr, 0, IORING_ENTER_GETEVENTS);
52885fa2 JA	410	if (ret > 0) {
	411	fio_aioring_queued(td, start, ret);
	412	io_u_mark_submit(td, ret);
	413
	414	ld->queued -= ret;
	415	ret = 0;
a90cd050 JA	416	} else if (!ret) {
a90cd050 JA	417	io_u_mark_submit(td, ret);
52885fa2	418	continue;
a90cd050 JA	419	} else {
	420	if (errno == EAGAIN) {
	421	ret = fio_aioring_cqring_reap(td, 0, ld->queued);
	422	if (ret)
	423	continue;
	424	/* Shouldn't happen */
	425	usleep(1);
	426	continue;
52885fa2	427	}
9a2d78b3	428	td_verror(td, errno, "io_uring_enter submit");
52885fa2	429	break;
a90cd050	430	}
52885fa2 JA	431	} while (ld->queued);
	432
	433	return ret;
	434	}
	435
9a2d78b3	436	static void fio_aioring_unmap(struct aioring_data *ld)
52885fa2	437	{
9a2d78b3	438	int i;
52885fa2	439
9a2d78b3 JA	440	for (i = 0; i < ARRAY_SIZE(ld->mmap); i++)
	441	munmap(ld->mmap[i].ptr, ld->mmap[i].len);
	442	close(ld->ring_fd);
b87aa01a JA	443	}
b87aa01a JA	444
52885fa2 JA	445	static void fio_aioring_cleanup(struct thread_data *td)
	446	{
	447	struct aioring_data *ld = td->io_ops_data;
	448
	449	if (ld) {
96563db9 JA	450	td->ts.cachehit += ld->cachehit;
	451	td->ts.cachemiss += ld->cachemiss;
	452
52885fa2 JA	453	/*
	454	* Work-around to avoid huge RCU stalls at exit time. If we
	455	* don't do this here, then it'll be torn down by exit_aio().
	456	* But for that case we can parallellize the freeing, thus
	457	* speeding it up a lot.
	458	*/
	459	if (!(td->flags & TD_F_CHILD))
9a2d78b3 JA	460	fio_aioring_unmap(ld);
9a2d78b3 JA	461
52885fa2 JA	462	free(ld->io_u_index);
52885fa2 JA	463	free(ld->io_us);
9a2d78b3	464	free(ld->iovecs);
52885fa2 JA	465	free(ld);
	466	}
	467	}
	468
9a2d78b3 JA	469	static int fio_aioring_mmap(struct aioring_data ld, struct aio_uring_params p)
	470	{
	471	struct aio_sq_ring *sring = &ld->sq_ring;
	472	struct aio_cq_ring *cring = &ld->cq_ring;
	473	void *ptr;
	474
ac122fea	475	ld->mmap[0].len = p->sq_off.array + p->sq_entries * sizeof(u32);
9a2d78b3 JA	476	ptr = mmap(0, ld->mmap[0].len, PROT_READ \| PROT_WRITE,
	477	MAP_SHARED \| MAP_POPULATE, ld->ring_fd,
	478	IORING_OFF_SQ_RING);
	479	ld->mmap[0].ptr = ptr;
	480	sring->head = ptr + p->sq_off.head;
	481	sring->tail = ptr + p->sq_off.tail;
	482	sring->ring_mask = ptr + p->sq_off.ring_mask;
	483	sring->ring_entries = ptr + p->sq_off.ring_entries;
	484	sring->flags = ptr + p->sq_off.flags;
ac122fea	485	sring->array = ptr + p->sq_off.array;
9a2d78b3 JA	486	ld->sq_ring_mask = *sring->ring_mask;
	487
	488	ld->mmap[1].len = p->sq_entries * sizeof(struct iocb);
	489	ld->iocbs = mmap(0, ld->mmap[1].len, PROT_READ \| PROT_WRITE,
	490	MAP_SHARED \| MAP_POPULATE, ld->ring_fd,
	491	IORING_OFF_IOCB);
	492	ld->mmap[1].ptr = ld->iocbs;
	493
ac122fea	494	ld->mmap[2].len = p->cq_off.events +
9a2d78b3 JA	495	p->cq_entries * sizeof(struct io_event);
	496	ptr = mmap(0, ld->mmap[2].len, PROT_READ \| PROT_WRITE,
	497	MAP_SHARED \| MAP_POPULATE, ld->ring_fd,
	498	IORING_OFF_CQ_RING);
	499	ld->mmap[2].ptr = ptr;
	500	cring->head = ptr + p->cq_off.head;
	501	cring->tail = ptr + p->cq_off.tail;
	502	cring->ring_mask = ptr + p->cq_off.ring_mask;
	503	cring->ring_entries = ptr + p->cq_off.ring_entries;
ac122fea	504	cring->events = ptr + p->cq_off.events;
9a2d78b3 JA	505	ld->cq_ring_mask = *cring->ring_mask;
	506	return 0;
	507	}
	508
52885fa2 JA	509	static int fio_aioring_queue_init(struct thread_data *td)
52885fa2 JA	510	{
52885fa2 JA	511	struct aioring_data *ld = td->io_ops_data;
52885fa2 JA	512	struct aioring_options *o = td->eo;
52885fa2	513	int depth = td->o.iodepth;
9a2d78b3 JA	514	struct aio_uring_params p;
	515	int ret;
	516
	517	memset(&p, 0, sizeof(p));
	518	p.flags = IOCTX_FLAG_SCQRING;
52885fa2 JA	519
52885fa2 JA	520	if (o->hipri)
9a2d78b3	521	p.flags \|= IOCTX_FLAG_IOPOLL;
a90cd050	522	if (o->sqthread_set) {
9a2d78b3 JA	523	p.sq_thread_cpu = o->sqthread;
9a2d78b3 JA	524	p.flags \|= IOCTX_FLAG_SQTHREAD;
771c9901	525	if (o->sqthread_poll)
9a2d78b3	526	p.flags \|= IOCTX_FLAG_SQPOLL;
f635f1fb JA	527	}
f635f1fb JA	528	if (o->sqwq)
9a2d78b3	529	p.flags \|= IOCTX_FLAG_SQWQ;
a90cd050	530
52885fa2 JA	531	if (o->fixedbufs) {
	532	struct rlimit rlim = {
	533	.rlim_cur = RLIM_INFINITY,
	534	.rlim_max = RLIM_INFINITY,
	535	};
	536
	537	setrlimit(RLIMIT_MEMLOCK, &rlim);
9a2d78b3	538	p.flags \|= IOCTX_FLAG_FIXEDBUFS;
52885fa2 JA	539	}
52885fa2 JA	540
9a2d78b3 JA	541	ret = syscall(__NR_sys_io_uring_setup, depth, ld->iovecs, &p);
	542	if (ret < 0)
	543	return ret;
	544
	545	ld->ring_fd = ret;
	546	return fio_aioring_mmap(ld, &p);
52885fa2 JA	547	}
	548
	549	static int fio_aioring_post_init(struct thread_data *td)
	550	{
	551	struct aioring_data *ld = td->io_ops_data;
	552	struct aioring_options *o = td->eo;
	553	struct io_u *io_u;
9a2d78b3	554	int err;
52885fa2 JA	555
	556	if (o->fixedbufs) {
	557	int i;
	558
	559	for (i = 0; i < td->o.iodepth; i++) {
9a2d78b3 JA	560	struct iovec *iov = &ld->iovecs[i];
9a2d78b3 JA	561
52885fa2	562	io_u = ld->io_u_index[i];
9a2d78b3 JA	563	iov->iov_base = io_u->buf;
9a2d78b3 JA	564	iov->iov_len = td_max_bs(td);
52885fa2 JA	565	}
	566	}
	567
	568	err = fio_aioring_queue_init(td);
	569	if (err) {
d63a472d	570	td_verror(td, errno, "io_queue_init");
52885fa2 JA	571	return 1;
	572	}
	573
	574	return 0;
	575	}
	576
9a2d78b3 JA	577	static unsigned roundup_pow2(unsigned depth)
	578	{
	579	return 1UL << __fls(depth - 1);
	580	}
	581
52885fa2 JA	582	static int fio_aioring_init(struct thread_data *td)
	583	{
	584	struct aioring_data *ld;
	585
52885fa2 JA	586	ld = calloc(1, sizeof(*ld));
52885fa2 JA	587
b87aa01a JA	588	/* ring depth must be a power-of-2 */
	589	ld->iodepth = td->o.iodepth;
	590	td->o.iodepth = roundup_pow2(td->o.iodepth);
	591
52885fa2 JA	592	/* io_u index */
	593	ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *));
	594	ld->io_us = calloc(td->o.iodepth, sizeof(struct io_u *));
	595
9a2d78b3	596	ld->iovecs = calloc(td->o.iodepth, sizeof(struct iovec));
52885fa2 JA	597
	598	td->io_ops_data = ld;
	599	return 0;
	600	}
	601
	602	static int fio_aioring_io_u_init(struct thread_data td, struct io_u io_u)
	603	{
	604	struct aioring_data *ld = td->io_ops_data;
	605
	606	ld->io_u_index[io_u->index] = io_u;
	607	return 0;
	608	}
	609
	610	static struct ioengine_ops ioengine = {
	611	.name = "aio-ring",
	612	.version = FIO_IOOPS_VERSION,
	613	.init = fio_aioring_init,
	614	.post_init = fio_aioring_post_init,
	615	.io_u_init = fio_aioring_io_u_init,
	616	.prep = fio_aioring_prep,
	617	.queue = fio_aioring_queue,
	618	.commit = fio_aioring_commit,
	619	.getevents = fio_aioring_getevents,
	620	.event = fio_aioring_event,
	621	.cleanup = fio_aioring_cleanup,
	622	.open_file = generic_open_file,
	623	.close_file = generic_close_file,
	624	.get_file_size = generic_get_file_size,
	625	.options = options,
	626	.option_struct_size = sizeof(struct aioring_options),
	627	};
	628
	629	static void fio_init fio_aioring_register(void)
	630	{
52885fa2	631	register_ioengine(&ioengine);
52885fa2 JA	632	}
	633
	634	static void fio_exit fio_aioring_unregister(void)
	635	{
52885fa2	636	unregister_ioengine(&ioengine);
52885fa2	637	}
1f90e9bb	638	#endif