[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

#ifndef IOCB_FLAG_HIPRI
#define IOCB_FLAG_HIPRI	(1 << 2)
#endif

/*
 * io_setup2(2) flags
 */
#ifndef IOCTX_FLAG_IOPOLL
#define IOCTX_FLAG_IOPOLL	(1 << 0)
#endif
#ifndef IOCTX_FLAG_SCQRING
#define IOCTX_FLAG_SCQRING	(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


/*
 * io_ring_enter(2) flags
 */
#ifndef IORING_FLAG_GETEVENTS
#define IORING_FLAG_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_sq_ring {
	union {
		struct {
			u32 head;
			u32 tail;
			u32 nr_events;
			u16 sq_thread_cpu;
			u16 kflags;
			u64 iocbs;
		};
		u32 pad[16];
	};
	u32 array[0];
};

struct aio_cq_ring {
	union {
		struct {
			u32 head;
			u32 tail;
			u32 nr_events;
		};
		struct io_event pad;
	};
	struct io_event events[0];
};

struct aioring_data {
	io_context_t aio_ctx;
	struct io_u **io_us;
	struct io_u **io_u_index;

	struct aio_sq_ring *sq_ring;
	struct iocb *iocbs;
	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_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_ring_enter(io_context_t ctx, unsigned int to_submit,
			 unsigned int min_complete, unsigned int flags)
{
	return syscall(__NR_sys_io_ring_enter, ctx, 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 aioring_options *o = td->eo;
	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;
		if (o->hipri)
			iocb->u.c.flags |= IOCB_FLAG_HIPRI;
		else
			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_ring_enter(ld->aio_ctx, 0, actual_min,
						IORING_FLAG_GETEVENTS);
			if (r < 0) {
				if (errno == EAGAIN)
					continue;
				td_verror(td, errno, "io_ring_enter get");
				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--) {
		int index = ld->sq_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->kflags & IORING_SQ_NEED_WAKEUP)
			io_ring_enter(ld->aio_ctx, ld->queued, 0, 0);
		ld->queued = 0;
		return 0;
	}

	do {
		unsigned start = ld->sq_ring->head;
		long nr = ld->queued;

		ret = io_ring_enter(ld->aio_ctx, nr, 0, IORING_FLAG_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_ring_enter sumit");
			break;
		}
	} while (ld->queued);

	return ret;
}

static size_t aioring_cq_size(struct thread_data *td)
{
	return sizeof(struct aio_cq_ring) + 2 * td->o.iodepth * sizeof(struct io_event);
}

static size_t aioring_sq_iocb(struct thread_data *td)
{
	return sizeof(struct iocb) * td->o.iodepth;
}

static size_t aioring_sq_size(struct thread_data *td)
{
	return sizeof(struct aio_sq_ring) + td->o.iodepth * sizeof(u32);
}

static unsigned roundup_pow2(unsigned depth)
{
	return 1UL << __fls(depth - 1);
}

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))
			io_destroy(ld->aio_ctx);
		free(ld->io_u_index);
		free(ld->io_us);
		fio_memfree(ld->sq_ring, aioring_sq_size(td), false);
		fio_memfree(ld->iocbs, aioring_sq_iocb(td), false);
		fio_memfree(ld->cq_ring, aioring_cq_size(td), false);
		free(ld);
	}
}

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 flags = IOCTX_FLAG_SCQRING;
	int depth = td->o.iodepth;

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

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

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

	return syscall(__NR_sys_io_setup2, depth, flags,
			ld->sq_ring, ld->cq_ring, &ld->aio_ctx);
}

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;
	struct iocb *iocb;
	int err = 0;

	if (o->fixedbufs) {
		int i;

		for (i = 0; i < td->o.iodepth; i++) {
			io_u = ld->io_u_index[i];
			iocb = &ld->iocbs[i];
			iocb->u.c.buf = io_u->buf;
			iocb->u.c.nbytes = td_max_bs(td);

			if (o->hipri)
				iocb->u.c.flags |= IOCB_FLAG_HIPRI;
		}
	}

	err = fio_aioring_queue_init(td);

	if (err) {
		td_verror(td, errno, "io_queue_init");
		return 1;
	}

	return 0;
}

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->iocbs = fio_memalign(page_size, aioring_sq_iocb(td), false);
	memset(ld->iocbs, 0, aioring_sq_iocb(td));

	ld->sq_ring = fio_memalign(page_size, aioring_sq_size(td), false);
	memset(ld->sq_ring, 0, aioring_sq_size(td));
	ld->sq_ring->nr_events = td->o.iodepth;
	ld->sq_ring->iocbs = (u64) (uintptr_t) ld->iocbs;
	ld->sq_ring_mask = td->o.iodepth - 1;

	ld->cq_ring = fio_memalign(page_size, aioring_cq_size(td), false);
	memset(ld->cq_ring, 0, aioring_cq_size(td));
	ld->cq_ring->nr_events = td->o.iodepth * 2;
	ld->cq_ring_mask = (2 * td->o.iodepth) - 1;

	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 JA	24	#ifndef IOCB_FLAG_HIPRI
	25	#define IOCB_FLAG_HIPRI (1 << 2)
	26	#endif
	27
	28	/*
	29	* io_setup2(2) flags
	30	*/
	31	#ifndef IOCTX_FLAG_IOPOLL
	32	#define IOCTX_FLAG_IOPOLL (1 << 0)
	33	#endif
	34	#ifndef IOCTX_FLAG_SCQRING
	35	#define IOCTX_FLAG_SCQRING (1 << 1)
	36	#endif
	37	#ifndef IOCTX_FLAG_FIXEDBUFS
	38	#define IOCTX_FLAG_FIXEDBUFS (1 << 2)
	39	#endif
	40	#ifndef IOCTX_FLAG_SQTHREAD
	41	#define IOCTX_FLAG_SQTHREAD (1 << 3)
	42	#endif
	43	#ifndef IOCTX_FLAG_SQWQ
	44	#define IOCTX_FLAG_SQWQ (1 << 4)
	45	#endif
771c9901 JA	46	#ifndef IOCTX_FLAG_SQPOLL
	47	#define IOCTX_FLAG_SQPOLL (1 << 5)
	48	#endif
	49
52885fa2 JA	50
	51	/*
	52	* io_ring_enter(2) flags
	53	*/
52885fa2	54	#ifndef IORING_FLAG_GETEVENTS
605d3b6a	55	#define IORING_FLAG_GETEVENTS (1 << 0)
52885fa2 JA	56	#endif
	57
	58	typedef uint64_t u64;
	59	typedef uint32_t u32;
	60	typedef uint16_t u16;
	61
4cdbc048 JA	62	#define IORING_SQ_NEED_WAKEUP (1 << 0)
4cdbc048 JA	63
96563db9 JA	64	#define IOEV_RES2_CACHEHIT (1 << 0)
96563db9 JA	65
52885fa2 JA	66	struct aio_sq_ring {
	67	union {
	68	struct {
	69	u32 head;
	70	u32 tail;
	71	u32 nr_events;
	72	u16 sq_thread_cpu;
4cdbc048	73	u16 kflags;
52885fa2 JA	74	u64 iocbs;
	75	};
	76	u32 pad[16];
	77	};
	78	u32 array[0];
	79	};
	80
	81	struct aio_cq_ring {
	82	union {
	83	struct {
	84	u32 head;
	85	u32 tail;
	86	u32 nr_events;
	87	};
	88	struct io_event pad;
	89	};
	90	struct io_event events[0];
	91	};
	92
	93	struct aioring_data {
	94	io_context_t aio_ctx;
	95	struct io_u **io_us;
	96	struct io_u **io_u_index;
	97
	98	struct aio_sq_ring *sq_ring;
	99	struct iocb *iocbs;
b87aa01a	100	unsigned sq_ring_mask;
52885fa2 JA	101
	102	struct aio_cq_ring *cq_ring;
	103	struct io_event *events;
b87aa01a	104	unsigned cq_ring_mask;
52885fa2 JA	105
	106	int queued;
	107	int cq_ring_off;
b87aa01a	108	unsigned iodepth;
96563db9 JA	109
	110	uint64_t cachehit;
	111	uint64_t cachemiss;
52885fa2 JA	112	};
	113
	114	struct aioring_options {
	115	void *pad;
	116	unsigned int hipri;
	117	unsigned int fixedbufs;
a90cd050 JA	118	unsigned int sqthread;
a90cd050 JA	119	unsigned int sqthread_set;
771c9901	120	unsigned int sqthread_poll;
a90cd050	121	unsigned int sqwq;
52885fa2 JA	122	};
52885fa2 JA	123
a90cd050 JA	124	static int fio_aioring_sqthread_cb(void *data,
	125	unsigned long long *val)
	126	{
	127	struct aioring_options *o = data;
	128
	129	o->sqthread = *val;
	130	o->sqthread_set = 1;
	131	return 0;
	132	}
	133
52885fa2 JA	134	static struct fio_option options[] = {
	135	{
	136	.name = "hipri",
	137	.lname = "High Priority",
	138	.type = FIO_OPT_STR_SET,
	139	.off1 = offsetof(struct aioring_options, hipri),
	140	.help = "Use polled IO completions",
	141	.category = FIO_OPT_C_ENGINE,
	142	.group = FIO_OPT_G_LIBAIO,
	143	},
	144	{
	145	.name = "fixedbufs",
	146	.lname = "Fixed (pre-mapped) IO buffers",
	147	.type = FIO_OPT_STR_SET,
	148	.off1 = offsetof(struct aioring_options, fixedbufs),
	149	.help = "Pre map IO buffers",
	150	.category = FIO_OPT_C_ENGINE,
	151	.group = FIO_OPT_G_LIBAIO,
	152	},
a90cd050 JA	153	{
	154	.name = "sqthread",
	155	.lname = "Use kernel SQ thread on this CPU",
	156	.type = FIO_OPT_INT,
	157	.cb = fio_aioring_sqthread_cb,
	158	.help = "Offload submission to kernel thread",
	159	.category = FIO_OPT_C_ENGINE,
	160	.group = FIO_OPT_G_LIBAIO,
	161	},
771c9901 JA	162	{
	163	.name = "sqthread_poll",
	164	.lname = "Kernel SQ thread should poll",
	165	.type = FIO_OPT_STR_SET,
	166	.off1 = offsetof(struct aioring_options, sqthread_poll),
	167	.help = "Used with sqthread, enables kernel side polling",
	168	.category = FIO_OPT_C_ENGINE,
	169	.group = FIO_OPT_G_LIBAIO,
	170	},
a90cd050 JA	171	{
	172	.name = "sqwq",
	173	.lname = "Offload submission to kernel workqueue",
	174	.type = FIO_OPT_STR_SET,
	175	.off1 = offsetof(struct aioring_options, sqwq),
	176	.help = "Offload submission to kernel workqueue",
	177	.category = FIO_OPT_C_ENGINE,
	178	.group = FIO_OPT_G_LIBAIO,
	179	},
52885fa2 JA	180	{
	181	.name = NULL,
	182	},
	183	};
	184
52885fa2 JA	185	static int io_ring_enter(io_context_t ctx, unsigned int to_submit,
	186	unsigned int min_complete, unsigned int flags)
	187	{
52885fa2 JA	188	return syscall(__NR_sys_io_ring_enter, ctx, to_submit, min_complete,
52885fa2 JA	189	flags);
52885fa2 JA	190	}
	191
	192	static int fio_aioring_prep(struct thread_data td, struct io_u io_u)
	193	{
	194	struct aioring_data *ld = td->io_ops_data;
	195	struct fio_file *f = io_u->file;
	196	struct aioring_options *o = td->eo;
	197	struct iocb *iocb;
	198
	199	iocb = &ld->iocbs[io_u->index];
	200
e3970057 JA	201	if (io_u->ddir == DDIR_READ \|\| io_u->ddir == DDIR_WRITE) {
e3970057 JA	202	if (io_u->ddir == DDIR_READ)
52885fa2	203	iocb->aio_lio_opcode = IO_CMD_PREAD;
e3970057	204	else
52885fa2	205	iocb->aio_lio_opcode = IO_CMD_PWRITE;
e3970057 JA	206	iocb->aio_reqprio = 0;
	207	iocb->aio_fildes = f->fd;
	208	iocb->u.c.buf = io_u->xfer_buf;
	209	iocb->u.c.nbytes = io_u->xfer_buflen;
	210	iocb->u.c.offset = io_u->offset;
	211	if (o->hipri)
	212	iocb->u.c.flags \|= IOCB_FLAG_HIPRI;
	213	else
	214	iocb->u.c.flags = 0;
52885fa2 JA	215	} else if (ddir_sync(io_u->ddir))
	216	io_prep_fsync(iocb, f->fd);
	217
	218	iocb->data = io_u;
	219	return 0;
	220	}
	221
	222	static struct io_u fio_aioring_event(struct thread_data td, int event)
	223	{
	224	struct aioring_data *ld = td->io_ops_data;
	225	struct io_event *ev;
	226	struct io_u *io_u;
b87aa01a	227	unsigned index;
52885fa2	228
b87aa01a	229	index = (event + ld->cq_ring_off) & ld->cq_ring_mask;
52885fa2 JA	230
	231	ev = &ld->cq_ring->events[index];
	232	io_u = ev->data;
	233
	234	if (ev->res != io_u->xfer_buflen) {
	235	if (ev->res > io_u->xfer_buflen)
	236	io_u->error = -ev->res;
	237	else
	238	io_u->resid = io_u->xfer_buflen - ev->res;
	239	} else
	240	io_u->error = 0;
	241
96563db9 JA	242	if (io_u->ddir == DDIR_READ) {
	243	if (ev->res2 & IOEV_RES2_CACHEHIT)
	244	ld->cachehit++;
	245	else
	246	ld->cachemiss++;
	247	}
	248
52885fa2 JA	249	return io_u;
	250	}
	251
	252	static int fio_aioring_cqring_reap(struct thread_data *td, unsigned int events,
	253	unsigned int max)
	254	{
	255	struct aioring_data *ld = td->io_ops_data;
	256	struct aio_cq_ring *ring = ld->cq_ring;
	257	u32 head, reaped = 0;
	258
	259	head = ring->head;
	260	do {
	261	read_barrier();
	262	if (head == ring->tail)
	263	break;
	264	reaped++;
	265	head++;
52885fa2 JA	266	} while (reaped + events < max);
	267
	268	ring->head = head;
	269	write_barrier();
	270	return reaped;
	271	}
	272
	273	static int fio_aioring_getevents(struct thread_data *td, unsigned int min,
	274	unsigned int max, const struct timespec *t)
	275	{
	276	struct aioring_data *ld = td->io_ops_data;
	277	unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
771c9901	278	struct aioring_options *o = td->eo;
52885fa2	279	struct aio_cq_ring *ring = ld->cq_ring;
b87aa01a JA	280	unsigned events = 0;
b87aa01a JA	281	int r;
52885fa2 JA	282
	283	ld->cq_ring_off = ring->head;
	284	do {
	285	r = fio_aioring_cqring_reap(td, events, max);
	286	if (r) {
	287	events += r;
	288	continue;
	289	}
	290
771c9901 JA	291	if (!o->sqthread_poll) {
	292	r = io_ring_enter(ld->aio_ctx, 0, actual_min,
	293	IORING_FLAG_GETEVENTS);
	294	if (r < 0) {
	295	if (errno == EAGAIN)
	296	continue;
	297	td_verror(td, errno, "io_ring_enter get");
	298	break;
	299	}
52885fa2 JA	300	}
	301	} while (events < min);
	302
	303	return r < 0 ? r : events;
	304	}
	305
	306	static enum fio_q_status fio_aioring_queue(struct thread_data *td,
	307	struct io_u *io_u)
	308	{
	309	struct aioring_data *ld = td->io_ops_data;
	310	struct aio_sq_ring *ring = ld->sq_ring;
	311	unsigned tail, next_tail;
	312
	313	fio_ro_check(td, io_u);
	314
b87aa01a	315	if (ld->queued == ld->iodepth)
52885fa2 JA	316	return FIO_Q_BUSY;
52885fa2 JA	317
52885fa2 JA	318	if (io_u->ddir == DDIR_TRIM) {
	319	if (ld->queued)
	320	return FIO_Q_BUSY;
	321
	322	do_io_u_trim(td, io_u);
	323	io_u_mark_submit(td, 1);
	324	io_u_mark_complete(td, 1);
	325	return FIO_Q_COMPLETED;
	326	}
	327
	328	tail = ring->tail;
	329	next_tail = tail + 1;
52885fa2 JA	330	read_barrier();
	331	if (next_tail == ring->head)
	332	return FIO_Q_BUSY;
	333
b87aa01a	334	ring->array[tail & ld->sq_ring_mask] = io_u->index;
52885fa2 JA	335	ring->tail = next_tail;
	336	write_barrier();
	337
	338	ld->queued++;
	339	return FIO_Q_QUEUED;
	340	}
	341
	342	static void fio_aioring_queued(struct thread_data *td, int start, int nr)
	343	{
	344	struct aioring_data *ld = td->io_ops_data;
	345	struct timespec now;
	346
	347	if (!fio_fill_issue_time(td))
	348	return;
	349
	350	fio_gettime(&now, NULL);
	351
	352	while (nr--) {
b87aa01a	353	int index = ld->sq_ring->array[start & ld->sq_ring_mask];
f8289afc	354	struct io_u *io_u = ld->io_u_index[index];
52885fa2 JA	355
	356	memcpy(&io_u->issue_time, &now, sizeof(now));
	357	io_u_queued(td, io_u);
	358
	359	start++;
52885fa2 JA	360	}
	361	}
	362
	363	static int fio_aioring_commit(struct thread_data *td)
	364	{
	365	struct aioring_data *ld = td->io_ops_data;
771c9901	366	struct aioring_options *o = td->eo;
52885fa2 JA	367	int ret;
	368
	369	if (!ld->queued)
	370	return 0;
	371
771c9901 JA	372	/* Nothing to do */
771c9901 JA	373	if (o->sqthread_poll) {
4cdbc048 JA	374	struct aio_sq_ring *ring = ld->sq_ring;
	375
	376	if (ring->kflags & IORING_SQ_NEED_WAKEUP)
605d3b6a	377	io_ring_enter(ld->aio_ctx, ld->queued, 0, 0);
771c9901 JA	378	ld->queued = 0;
	379	return 0;
	380	}
	381
52885fa2	382	do {
b87aa01a	383	unsigned start = ld->sq_ring->head;
52885fa2 JA	384	long nr = ld->queued;
52885fa2 JA	385
605d3b6a	386	ret = io_ring_enter(ld->aio_ctx, nr, 0, IORING_FLAG_GETEVENTS);
52885fa2 JA	387	if (ret > 0) {
	388	fio_aioring_queued(td, start, ret);
	389	io_u_mark_submit(td, ret);
	390
	391	ld->queued -= ret;
	392	ret = 0;
a90cd050 JA	393	} else if (!ret) {
a90cd050 JA	394	io_u_mark_submit(td, ret);
52885fa2	395	continue;
a90cd050 JA	396	} else {
	397	if (errno == EAGAIN) {
	398	ret = fio_aioring_cqring_reap(td, 0, ld->queued);
	399	if (ret)
	400	continue;
	401	/* Shouldn't happen */
	402	usleep(1);
	403	continue;
52885fa2	404	}
a90cd050	405	td_verror(td, errno, "io_ring_enter sumit");
52885fa2	406	break;
a90cd050	407	}
52885fa2 JA	408	} while (ld->queued);
	409
	410	return ret;
	411	}
	412
	413	static size_t aioring_cq_size(struct thread_data *td)
	414	{
	415	return sizeof(struct aio_cq_ring) + 2 * td->o.iodepth * sizeof(struct io_event);
	416	}
	417
	418	static size_t aioring_sq_iocb(struct thread_data *td)
	419	{
	420	return sizeof(struct iocb) * td->o.iodepth;
	421	}
	422
	423	static size_t aioring_sq_size(struct thread_data *td)
	424	{
	425	return sizeof(struct aio_sq_ring) + td->o.iodepth * sizeof(u32);
	426	}
	427
b87aa01a JA	428	static unsigned roundup_pow2(unsigned depth)
	429	{
	430	return 1UL << __fls(depth - 1);
	431	}
	432
52885fa2 JA	433	static void fio_aioring_cleanup(struct thread_data *td)
	434	{
	435	struct aioring_data *ld = td->io_ops_data;
	436
	437	if (ld) {
96563db9 JA	438	td->ts.cachehit += ld->cachehit;
	439	td->ts.cachemiss += ld->cachemiss;
	440
52885fa2 JA	441	/*
	442	* Work-around to avoid huge RCU stalls at exit time. If we
	443	* don't do this here, then it'll be torn down by exit_aio().
	444	* But for that case we can parallellize the freeing, thus
	445	* speeding it up a lot.
	446	*/
	447	if (!(td->flags & TD_F_CHILD))
	448	io_destroy(ld->aio_ctx);
	449	free(ld->io_u_index);
	450	free(ld->io_us);
	451	fio_memfree(ld->sq_ring, aioring_sq_size(td), false);
	452	fio_memfree(ld->iocbs, aioring_sq_iocb(td), false);
	453	fio_memfree(ld->cq_ring, aioring_cq_size(td), false);
	454	free(ld);
	455	}
	456	}
	457
	458	static int fio_aioring_queue_init(struct thread_data *td)
	459	{
52885fa2 JA	460	struct aioring_data *ld = td->io_ops_data;
	461	struct aioring_options *o = td->eo;
	462	int flags = IOCTX_FLAG_SCQRING;
	463	int depth = td->o.iodepth;
	464
	465	if (o->hipri)
	466	flags \|= IOCTX_FLAG_IOPOLL;
a90cd050 JA	467	if (o->sqthread_set) {
	468	ld->sq_ring->sq_thread_cpu = o->sqthread;
	469	flags \|= IOCTX_FLAG_SQTHREAD;
771c9901 JA	470	if (o->sqthread_poll)
771c9901 JA	471	flags \|= IOCTX_FLAG_SQPOLL;
f635f1fb JA	472	}
f635f1fb JA	473	if (o->sqwq)
a90cd050 JA	474	flags \|= IOCTX_FLAG_SQWQ;
a90cd050 JA	475
52885fa2 JA	476	if (o->fixedbufs) {
	477	struct rlimit rlim = {
	478	.rlim_cur = RLIM_INFINITY,
	479	.rlim_max = RLIM_INFINITY,
	480	};
	481
	482	setrlimit(RLIMIT_MEMLOCK, &rlim);
	483	flags \|= IOCTX_FLAG_FIXEDBUFS;
	484	}
	485
	486	return syscall(__NR_sys_io_setup2, depth, flags,
	487	ld->sq_ring, ld->cq_ring, &ld->aio_ctx);
52885fa2 JA	488	}
	489
	490	static int fio_aioring_post_init(struct thread_data *td)
	491	{
	492	struct aioring_data *ld = td->io_ops_data;
	493	struct aioring_options *o = td->eo;
	494	struct io_u *io_u;
	495	struct iocb *iocb;
	496	int err = 0;
	497
	498	if (o->fixedbufs) {
	499	int i;
	500
	501	for (i = 0; i < td->o.iodepth; i++) {
	502	io_u = ld->io_u_index[i];
	503	iocb = &ld->iocbs[i];
	504	iocb->u.c.buf = io_u->buf;
	505	iocb->u.c.nbytes = td_max_bs(td);
	506
	507	if (o->hipri)
	508	iocb->u.c.flags \|= IOCB_FLAG_HIPRI;
	509	}
	510	}
	511
	512	err = fio_aioring_queue_init(td);
d63a472d	513
52885fa2	514	if (err) {
d63a472d	515	td_verror(td, errno, "io_queue_init");
52885fa2 JA	516	return 1;
	517	}
	518
	519	return 0;
	520	}
	521
	522	static int fio_aioring_init(struct thread_data *td)
	523	{
	524	struct aioring_data *ld;
	525
52885fa2 JA	526	ld = calloc(1, sizeof(*ld));
52885fa2 JA	527
b87aa01a JA	528	/* ring depth must be a power-of-2 */
	529	ld->iodepth = td->o.iodepth;
	530	td->o.iodepth = roundup_pow2(td->o.iodepth);
	531
52885fa2 JA	532	/* io_u index */
	533	ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *));
	534	ld->io_us = calloc(td->o.iodepth, sizeof(struct io_u *));
	535
	536	ld->iocbs = fio_memalign(page_size, aioring_sq_iocb(td), false);
	537	memset(ld->iocbs, 0, aioring_sq_iocb(td));
	538
	539	ld->sq_ring = fio_memalign(page_size, aioring_sq_size(td), false);
	540	memset(ld->sq_ring, 0, aioring_sq_size(td));
	541	ld->sq_ring->nr_events = td->o.iodepth;
1f90e9bb	542	ld->sq_ring->iocbs = (u64) (uintptr_t) ld->iocbs;
b87aa01a	543	ld->sq_ring_mask = td->o.iodepth - 1;
52885fa2 JA	544
	545	ld->cq_ring = fio_memalign(page_size, aioring_cq_size(td), false);
	546	memset(ld->cq_ring, 0, aioring_cq_size(td));
	547	ld->cq_ring->nr_events = td->o.iodepth * 2;
b87aa01a	548	ld->cq_ring_mask = (2 * td->o.iodepth) - 1;
52885fa2 JA	549
	550	td->io_ops_data = ld;
	551	return 0;
	552	}
	553
	554	static int fio_aioring_io_u_init(struct thread_data td, struct io_u io_u)
	555	{
	556	struct aioring_data *ld = td->io_ops_data;
	557
	558	ld->io_u_index[io_u->index] = io_u;
	559	return 0;
	560	}
	561
	562	static struct ioengine_ops ioengine = {
	563	.name = "aio-ring",
	564	.version = FIO_IOOPS_VERSION,
	565	.init = fio_aioring_init,
	566	.post_init = fio_aioring_post_init,
	567	.io_u_init = fio_aioring_io_u_init,
	568	.prep = fio_aioring_prep,
	569	.queue = fio_aioring_queue,
	570	.commit = fio_aioring_commit,
	571	.getevents = fio_aioring_getevents,
	572	.event = fio_aioring_event,
	573	.cleanup = fio_aioring_cleanup,
	574	.open_file = generic_open_file,
	575	.close_file = generic_close_file,
	576	.get_file_size = generic_get_file_size,
	577	.options = options,
	578	.option_struct_size = sizeof(struct aioring_options),
	579	};
	580
	581	static void fio_init fio_aioring_register(void)
	582	{
52885fa2	583	register_ioengine(&ioengine);
52885fa2 JA	584	}
	585
	586	static void fio_exit fio_aioring_unregister(void)
	587	{
52885fa2	588	unregister_ioengine(&ioengine);
52885fa2	589	}
1f90e9bb	590	#endif