[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"

#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_SUBMIT
#define IORING_FLAG_SUBMIT	(1 << 0)
#endif
#ifndef IORING_FLAG_GETEVENTS
#define IORING_FLAG_GETEVENTS	(1 << 1)
#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;

	struct aio_cq_ring *cq_ring;
	struct io_event *events;

	int queued;
	int cq_ring_off;

	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) {
		if (o->fixedbufs) {
			iocb->aio_fildes = f->fd;
			iocb->aio_lio_opcode = IO_CMD_PREAD;
			iocb->u.c.offset = io_u->offset;
		} else {
			io_prep_pread(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
			if (o->hipri)
				iocb->u.c.flags |= IOCB_FLAG_HIPRI;
		}
	} else if (io_u->ddir == DDIR_WRITE) {
		if (o->fixedbufs) {
			iocb->aio_fildes = f->fd;
			iocb->aio_lio_opcode = IO_CMD_PWRITE;
			iocb->u.c.offset = io_u->offset;
		} else {
			io_prep_pwrite(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
			if (o->hipri)
				iocb->u.c.flags |= IOCB_FLAG_HIPRI;
		}
	} 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;
	int index;

	index = event + ld->cq_ring_off;
	if (index >= ld->cq_ring->nr_events)
		index -= ld->cq_ring->nr_events;

	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++;
		if (head == ring->nr_events)
			head = 0;
	} 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;
	int r, events = 0;

	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 == td->o.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;
	if (next_tail == ring->nr_events)
		next_tail = 0;
	read_barrier();
	if (next_tail == ring->head)
		return FIO_Q_BUSY;

	ring->array[tail] = 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];
		struct io_u *io_u = io_u = ld->io_u_index[index];

		memcpy(&io_u->issue_time, &now, sizeof(now));
		io_u_queued(td, io_u);

		start++;
		if (start == ld->sq_ring->nr_events)
			start = 0;
	}
}

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, IORING_FLAG_SUBMIT);
		ld->queued = 0;
		return 0;
	}

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

		ret = io_ring_enter(ld->aio_ctx, nr, 0, IORING_FLAG_SUBMIT |
						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 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;

		/* Bump depth to match init depth */
		td->o.iodepth++;

		/*
		 * 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;
	} else 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, -err, "io_queue_init");
		return 1;
	}

	/* Adjust depth back again */
	td->o.iodepth--;
	return 0;
}

static int fio_aioring_init(struct thread_data *td)
{
	struct aioring_options *o = td->eo;
	struct aioring_data *ld;

	if (o->sqthread_set && o->sqwq) {
		log_err("fio: aioring sqthread and sqwq are mutually exclusive\n");
		return 1;
	}

	/* ring needs an extra entry, add one to achieve QD set */
	td->o.iodepth++;

	ld = calloc(1, sizeof(*ld));

	/* 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->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;

	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"
	20
a30e63cf JA	21	#ifdef ARCH_HAVE_AIORING
a30e63cf JA	22
52885fa2 JA	23	#ifndef IOCB_FLAG_HIPRI
	24	#define IOCB_FLAG_HIPRI (1 << 2)
	25	#endif
	26
	27	/*
	28	* io_setup2(2) flags
	29	*/
	30	#ifndef IOCTX_FLAG_IOPOLL
	31	#define IOCTX_FLAG_IOPOLL (1 << 0)
	32	#endif
	33	#ifndef IOCTX_FLAG_SCQRING
	34	#define IOCTX_FLAG_SCQRING (1 << 1)
	35	#endif
	36	#ifndef IOCTX_FLAG_FIXEDBUFS
	37	#define IOCTX_FLAG_FIXEDBUFS (1 << 2)
	38	#endif
	39	#ifndef IOCTX_FLAG_SQTHREAD
	40	#define IOCTX_FLAG_SQTHREAD (1 << 3)
	41	#endif
	42	#ifndef IOCTX_FLAG_SQWQ
	43	#define IOCTX_FLAG_SQWQ (1 << 4)
	44	#endif
771c9901 JA	45	#ifndef IOCTX_FLAG_SQPOLL
	46	#define IOCTX_FLAG_SQPOLL (1 << 5)
	47	#endif
	48
52885fa2 JA	49
	50	/*
	51	* io_ring_enter(2) flags
	52	*/
	53	#ifndef IORING_FLAG_SUBMIT
	54	#define IORING_FLAG_SUBMIT (1 << 0)
	55	#endif
	56	#ifndef IORING_FLAG_GETEVENTS
	57	#define IORING_FLAG_GETEVENTS (1 << 1)
	58	#endif
	59
	60	typedef uint64_t u64;
	61	typedef uint32_t u32;
	62	typedef uint16_t u16;
	63
4cdbc048 JA	64	#define IORING_SQ_NEED_WAKEUP (1 << 0)
4cdbc048 JA	65
96563db9 JA	66	#define IOEV_RES2_CACHEHIT (1 << 0)
96563db9 JA	67
52885fa2 JA	68	struct aio_sq_ring {
	69	union {
	70	struct {
	71	u32 head;
	72	u32 tail;
	73	u32 nr_events;
	74	u16 sq_thread_cpu;
4cdbc048	75	u16 kflags;
52885fa2 JA	76	u64 iocbs;
	77	};
	78	u32 pad[16];
	79	};
	80	u32 array[0];
	81	};
	82
	83	struct aio_cq_ring {
	84	union {
	85	struct {
	86	u32 head;
	87	u32 tail;
	88	u32 nr_events;
	89	};
	90	struct io_event pad;
	91	};
	92	struct io_event events[0];
	93	};
	94
	95	struct aioring_data {
	96	io_context_t aio_ctx;
	97	struct io_u **io_us;
	98	struct io_u **io_u_index;
	99
	100	struct aio_sq_ring *sq_ring;
	101	struct iocb *iocbs;
	102
	103	struct aio_cq_ring *cq_ring;
	104	struct io_event *events;
	105
	106	int queued;
	107	int cq_ring_off;
96563db9 JA	108
	109	uint64_t cachehit;
	110	uint64_t cachemiss;
52885fa2 JA	111	};
	112
	113	struct aioring_options {
	114	void *pad;
	115	unsigned int hipri;
	116	unsigned int fixedbufs;
a90cd050 JA	117	unsigned int sqthread;
a90cd050 JA	118	unsigned int sqthread_set;
771c9901	119	unsigned int sqthread_poll;
a90cd050	120	unsigned int sqwq;
52885fa2 JA	121	};
52885fa2 JA	122
a90cd050 JA	123	static int fio_aioring_sqthread_cb(void *data,
	124	unsigned long long *val)
	125	{
	126	struct aioring_options *o = data;
	127
	128	o->sqthread = *val;
	129	o->sqthread_set = 1;
	130	return 0;
	131	}
	132
52885fa2 JA	133	static struct fio_option options[] = {
	134	{
	135	.name = "hipri",
	136	.lname = "High Priority",
	137	.type = FIO_OPT_STR_SET,
	138	.off1 = offsetof(struct aioring_options, hipri),
	139	.help = "Use polled IO completions",
	140	.category = FIO_OPT_C_ENGINE,
	141	.group = FIO_OPT_G_LIBAIO,
	142	},
	143	{
	144	.name = "fixedbufs",
	145	.lname = "Fixed (pre-mapped) IO buffers",
	146	.type = FIO_OPT_STR_SET,
	147	.off1 = offsetof(struct aioring_options, fixedbufs),
	148	.help = "Pre map IO buffers",
	149	.category = FIO_OPT_C_ENGINE,
	150	.group = FIO_OPT_G_LIBAIO,
	151	},
a90cd050 JA	152	{
	153	.name = "sqthread",
	154	.lname = "Use kernel SQ thread on this CPU",
	155	.type = FIO_OPT_INT,
	156	.cb = fio_aioring_sqthread_cb,
	157	.help = "Offload submission to kernel thread",
	158	.category = FIO_OPT_C_ENGINE,
	159	.group = FIO_OPT_G_LIBAIO,
	160	},
771c9901 JA	161	{
	162	.name = "sqthread_poll",
	163	.lname = "Kernel SQ thread should poll",
	164	.type = FIO_OPT_STR_SET,
	165	.off1 = offsetof(struct aioring_options, sqthread_poll),
	166	.help = "Used with sqthread, enables kernel side polling",
	167	.category = FIO_OPT_C_ENGINE,
	168	.group = FIO_OPT_G_LIBAIO,
	169	},
a90cd050 JA	170	{
	171	.name = "sqwq",
	172	.lname = "Offload submission to kernel workqueue",
	173	.type = FIO_OPT_STR_SET,
	174	.off1 = offsetof(struct aioring_options, sqwq),
	175	.help = "Offload submission to kernel workqueue",
	176	.category = FIO_OPT_C_ENGINE,
	177	.group = FIO_OPT_G_LIBAIO,
	178	},
52885fa2 JA	179	{
	180	.name = NULL,
	181	},
	182	};
	183
52885fa2 JA	184	static int io_ring_enter(io_context_t ctx, unsigned int to_submit,
	185	unsigned int min_complete, unsigned int flags)
	186	{
52885fa2 JA	187	return syscall(__NR_sys_io_ring_enter, ctx, to_submit, min_complete,
52885fa2 JA	188	flags);
52885fa2 JA	189	}
	190
	191	static int fio_aioring_prep(struct thread_data td, struct io_u io_u)
	192	{
	193	struct aioring_data *ld = td->io_ops_data;
	194	struct fio_file *f = io_u->file;
	195	struct aioring_options *o = td->eo;
	196	struct iocb *iocb;
	197
	198	iocb = &ld->iocbs[io_u->index];
	199
	200	if (io_u->ddir == DDIR_READ) {
	201	if (o->fixedbufs) {
	202	iocb->aio_fildes = f->fd;
	203	iocb->aio_lio_opcode = IO_CMD_PREAD;
	204	iocb->u.c.offset = io_u->offset;
	205	} else {
	206	io_prep_pread(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
	207	if (o->hipri)
	208	iocb->u.c.flags \|= IOCB_FLAG_HIPRI;
	209	}
	210	} else if (io_u->ddir == DDIR_WRITE) {
	211	if (o->fixedbufs) {
	212	iocb->aio_fildes = f->fd;
	213	iocb->aio_lio_opcode = IO_CMD_PWRITE;
	214	iocb->u.c.offset = io_u->offset;
	215	} else {
	216	io_prep_pwrite(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
	217	if (o->hipri)
	218	iocb->u.c.flags \|= IOCB_FLAG_HIPRI;
	219	}
	220	} else if (ddir_sync(io_u->ddir))
	221	io_prep_fsync(iocb, f->fd);
	222
	223	iocb->data = io_u;
	224	return 0;
	225	}
	226
	227	static struct io_u fio_aioring_event(struct thread_data td, int event)
	228	{
	229	struct aioring_data *ld = td->io_ops_data;
	230	struct io_event *ev;
	231	struct io_u *io_u;
	232	int index;
	233
	234	index = event + ld->cq_ring_off;
	235	if (index >= ld->cq_ring->nr_events)
	236	index -= ld->cq_ring->nr_events;
	237
	238	ev = &ld->cq_ring->events[index];
	239	io_u = ev->data;
	240
	241	if (ev->res != io_u->xfer_buflen) {
	242	if (ev->res > io_u->xfer_buflen)
	243	io_u->error = -ev->res;
	244	else
	245	io_u->resid = io_u->xfer_buflen - ev->res;
	246	} else
	247	io_u->error = 0;
	248
96563db9 JA	249	if (io_u->ddir == DDIR_READ) {
	250	if (ev->res2 & IOEV_RES2_CACHEHIT)
	251	ld->cachehit++;
	252	else
	253	ld->cachemiss++;
	254	}
	255
52885fa2 JA	256	return io_u;
	257	}
	258
	259	static int fio_aioring_cqring_reap(struct thread_data *td, unsigned int events,
	260	unsigned int max)
	261	{
	262	struct aioring_data *ld = td->io_ops_data;
	263	struct aio_cq_ring *ring = ld->cq_ring;
	264	u32 head, reaped = 0;
	265
	266	head = ring->head;
	267	do {
	268	read_barrier();
	269	if (head == ring->tail)
	270	break;
	271	reaped++;
	272	head++;
	273	if (head == ring->nr_events)
	274	head = 0;
	275	} while (reaped + events < max);
	276
	277	ring->head = head;
	278	write_barrier();
	279	return reaped;
	280	}
	281
	282	static int fio_aioring_getevents(struct thread_data *td, unsigned int min,
	283	unsigned int max, const struct timespec *t)
	284	{
	285	struct aioring_data *ld = td->io_ops_data;
	286	unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
771c9901	287	struct aioring_options *o = td->eo;
52885fa2 JA	288	struct aio_cq_ring *ring = ld->cq_ring;
	289	int r, events = 0;
	290
	291	ld->cq_ring_off = ring->head;
	292	do {
	293	r = fio_aioring_cqring_reap(td, events, max);
	294	if (r) {
	295	events += r;
	296	continue;
	297	}
	298
771c9901 JA	299	if (!o->sqthread_poll) {
	300	r = io_ring_enter(ld->aio_ctx, 0, actual_min,
	301	IORING_FLAG_GETEVENTS);
	302	if (r < 0) {
	303	if (errno == EAGAIN)
	304	continue;
	305	td_verror(td, errno, "io_ring_enter get");
	306	break;
	307	}
52885fa2 JA	308	}
	309	} while (events < min);
	310
	311	return r < 0 ? r : events;
	312	}
	313
	314	static enum fio_q_status fio_aioring_queue(struct thread_data *td,
	315	struct io_u *io_u)
	316	{
	317	struct aioring_data *ld = td->io_ops_data;
	318	struct aio_sq_ring *ring = ld->sq_ring;
	319	unsigned tail, next_tail;
	320
	321	fio_ro_check(td, io_u);
	322
	323	if (ld->queued == td->o.iodepth)
	324	return FIO_Q_BUSY;
	325
52885fa2 JA	326	if (io_u->ddir == DDIR_TRIM) {
	327	if (ld->queued)
	328	return FIO_Q_BUSY;
	329
	330	do_io_u_trim(td, io_u);
	331	io_u_mark_submit(td, 1);
	332	io_u_mark_complete(td, 1);
	333	return FIO_Q_COMPLETED;
	334	}
	335
	336	tail = ring->tail;
	337	next_tail = tail + 1;
	338	if (next_tail == ring->nr_events)
	339	next_tail = 0;
	340	read_barrier();
	341	if (next_tail == ring->head)
	342	return FIO_Q_BUSY;
	343
	344	ring->array[tail] = io_u->index;
	345	ring->tail = next_tail;
	346	write_barrier();
	347
	348	ld->queued++;
	349	return FIO_Q_QUEUED;
	350	}
	351
	352	static void fio_aioring_queued(struct thread_data *td, int start, int nr)
	353	{
	354	struct aioring_data *ld = td->io_ops_data;
	355	struct timespec now;
	356
	357	if (!fio_fill_issue_time(td))
	358	return;
	359
	360	fio_gettime(&now, NULL);
	361
	362	while (nr--) {
	363	int index = ld->sq_ring->array[start];
	364	struct io_u *io_u = io_u = ld->io_u_index[index];
	365
	366	memcpy(&io_u->issue_time, &now, sizeof(now));
	367	io_u_queued(td, io_u);
	368
	369	start++;
	370	if (start == ld->sq_ring->nr_events)
	371	start = 0;
	372	}
	373	}
	374
	375	static int fio_aioring_commit(struct thread_data *td)
	376	{
	377	struct aioring_data *ld = td->io_ops_data;
771c9901	378	struct aioring_options *o = td->eo;
52885fa2 JA	379	int ret;
	380
	381	if (!ld->queued)
	382	return 0;
	383
771c9901 JA	384	/* Nothing to do */
771c9901 JA	385	if (o->sqthread_poll) {
4cdbc048 JA	386	struct aio_sq_ring *ring = ld->sq_ring;
	387
	388	if (ring->kflags & IORING_SQ_NEED_WAKEUP)
	389	io_ring_enter(ld->aio_ctx, ld->queued, 0, IORING_FLAG_SUBMIT);
771c9901 JA	390	ld->queued = 0;
	391	return 0;
	392	}
	393
52885fa2 JA	394	do {
	395	int start = ld->sq_ring->head;
	396	long nr = ld->queued;
	397
	398	ret = io_ring_enter(ld->aio_ctx, nr, 0, IORING_FLAG_SUBMIT \|
	399	IORING_FLAG_GETEVENTS);
52885fa2 JA	400	if (ret > 0) {
	401	fio_aioring_queued(td, start, ret);
	402	io_u_mark_submit(td, ret);
	403
	404	ld->queued -= ret;
	405	ret = 0;
a90cd050 JA	406	} else if (!ret) {
a90cd050 JA	407	io_u_mark_submit(td, ret);
52885fa2	408	continue;
a90cd050 JA	409	} else {
	410	if (errno == EAGAIN) {
	411	ret = fio_aioring_cqring_reap(td, 0, ld->queued);
	412	if (ret)
	413	continue;
	414	/* Shouldn't happen */
	415	usleep(1);
	416	continue;
52885fa2	417	}
a90cd050	418	td_verror(td, errno, "io_ring_enter sumit");
52885fa2	419	break;
a90cd050	420	}
52885fa2 JA	421	} while (ld->queued);
	422
	423	return ret;
	424	}
	425
	426	static size_t aioring_cq_size(struct thread_data *td)
	427	{
	428	return sizeof(struct aio_cq_ring) + 2 * td->o.iodepth * sizeof(struct io_event);
	429	}
	430
	431	static size_t aioring_sq_iocb(struct thread_data *td)
	432	{
	433	return sizeof(struct iocb) * td->o.iodepth;
	434	}
	435
	436	static size_t aioring_sq_size(struct thread_data *td)
	437	{
	438	return sizeof(struct aio_sq_ring) + td->o.iodepth * sizeof(u32);
	439	}
	440
	441	static void fio_aioring_cleanup(struct thread_data *td)
	442	{
	443	struct aioring_data *ld = td->io_ops_data;
	444
	445	if (ld) {
96563db9 JA	446	td->ts.cachehit += ld->cachehit;
	447	td->ts.cachemiss += ld->cachemiss;
	448
a90cd050 JA	449	/* Bump depth to match init depth */
	450	td->o.iodepth++;
	451
52885fa2 JA	452	/*
	453	* Work-around to avoid huge RCU stalls at exit time. If we
	454	* don't do this here, then it'll be torn down by exit_aio().
	455	* But for that case we can parallellize the freeing, thus
	456	* speeding it up a lot.
	457	*/
	458	if (!(td->flags & TD_F_CHILD))
	459	io_destroy(ld->aio_ctx);
	460	free(ld->io_u_index);
	461	free(ld->io_us);
	462	fio_memfree(ld->sq_ring, aioring_sq_size(td), false);
	463	fio_memfree(ld->iocbs, aioring_sq_iocb(td), false);
	464	fio_memfree(ld->cq_ring, aioring_cq_size(td), false);
	465	free(ld);
	466	}
	467	}
	468
	469	static int fio_aioring_queue_init(struct thread_data *td)
	470	{
52885fa2 JA	471	struct aioring_data *ld = td->io_ops_data;
	472	struct aioring_options *o = td->eo;
	473	int flags = IOCTX_FLAG_SCQRING;
	474	int depth = td->o.iodepth;
	475
	476	if (o->hipri)
	477	flags \|= IOCTX_FLAG_IOPOLL;
a90cd050 JA	478	if (o->sqthread_set) {
	479	ld->sq_ring->sq_thread_cpu = o->sqthread;
	480	flags \|= IOCTX_FLAG_SQTHREAD;
771c9901 JA	481	if (o->sqthread_poll)
771c9901 JA	482	flags \|= IOCTX_FLAG_SQPOLL;
a90cd050 JA	483	} else if (o->sqwq)
	484	flags \|= IOCTX_FLAG_SQWQ;
	485
52885fa2 JA	486	if (o->fixedbufs) {
	487	struct rlimit rlim = {
	488	.rlim_cur = RLIM_INFINITY,
	489	.rlim_max = RLIM_INFINITY,
	490	};
	491
	492	setrlimit(RLIMIT_MEMLOCK, &rlim);
	493	flags \|= IOCTX_FLAG_FIXEDBUFS;
	494	}
	495
	496	return syscall(__NR_sys_io_setup2, depth, flags,
	497	ld->sq_ring, ld->cq_ring, &ld->aio_ctx);
52885fa2 JA	498	}
	499
	500	static int fio_aioring_post_init(struct thread_data *td)
	501	{
	502	struct aioring_data *ld = td->io_ops_data;
	503	struct aioring_options *o = td->eo;
	504	struct io_u *io_u;
	505	struct iocb *iocb;
	506	int err = 0;
	507
	508	if (o->fixedbufs) {
	509	int i;
	510
	511	for (i = 0; i < td->o.iodepth; i++) {
	512	io_u = ld->io_u_index[i];
	513	iocb = &ld->iocbs[i];
	514	iocb->u.c.buf = io_u->buf;
	515	iocb->u.c.nbytes = td_max_bs(td);
	516
	517	if (o->hipri)
	518	iocb->u.c.flags \|= IOCB_FLAG_HIPRI;
	519	}
	520	}
	521
	522	err = fio_aioring_queue_init(td);
	523	if (err) {
	524	td_verror(td, -err, "io_queue_init");
	525	return 1;
	526	}
	527
a90cd050 JA	528	/* Adjust depth back again */
a90cd050 JA	529	td->o.iodepth--;
52885fa2 JA	530	return 0;
	531	}
	532
	533	static int fio_aioring_init(struct thread_data *td)
	534	{
a90cd050	535	struct aioring_options *o = td->eo;
52885fa2 JA	536	struct aioring_data *ld;
52885fa2 JA	537
a90cd050 JA	538	if (o->sqthread_set && o->sqwq) {
	539	log_err("fio: aioring sqthread and sqwq are mutually exclusive\n");
	540	return 1;
	541	}
	542
3b617146 JA	543	/* ring needs an extra entry, add one to achieve QD set */
3b617146 JA	544	td->o.iodepth++;
52885fa2 JA	545
	546	ld = calloc(1, sizeof(*ld));
	547
	548	/* io_u index */
	549	ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *));
	550	ld->io_us = calloc(td->o.iodepth, sizeof(struct io_u *));
	551
	552	ld->iocbs = fio_memalign(page_size, aioring_sq_iocb(td), false);
	553	memset(ld->iocbs, 0, aioring_sq_iocb(td));
	554
	555	ld->sq_ring = fio_memalign(page_size, aioring_sq_size(td), false);
	556	memset(ld->sq_ring, 0, aioring_sq_size(td));
	557	ld->sq_ring->nr_events = td->o.iodepth;
1f90e9bb	558	ld->sq_ring->iocbs = (u64) (uintptr_t) ld->iocbs;
52885fa2 JA	559
	560	ld->cq_ring = fio_memalign(page_size, aioring_cq_size(td), false);
	561	memset(ld->cq_ring, 0, aioring_cq_size(td));
	562	ld->cq_ring->nr_events = td->o.iodepth * 2;
	563
	564	td->io_ops_data = ld;
	565	return 0;
	566	}
	567
	568	static int fio_aioring_io_u_init(struct thread_data td, struct io_u io_u)
	569	{
	570	struct aioring_data *ld = td->io_ops_data;
	571
	572	ld->io_u_index[io_u->index] = io_u;
	573	return 0;
	574	}
	575
	576	static struct ioengine_ops ioengine = {
	577	.name = "aio-ring",
	578	.version = FIO_IOOPS_VERSION,
	579	.init = fio_aioring_init,
	580	.post_init = fio_aioring_post_init,
	581	.io_u_init = fio_aioring_io_u_init,
	582	.prep = fio_aioring_prep,
	583	.queue = fio_aioring_queue,
	584	.commit = fio_aioring_commit,
	585	.getevents = fio_aioring_getevents,
	586	.event = fio_aioring_event,
	587	.cleanup = fio_aioring_cleanup,
	588	.open_file = generic_open_file,
	589	.close_file = generic_close_file,
	590	.get_file_size = generic_get_file_size,
	591	.options = options,
	592	.option_struct_size = sizeof(struct aioring_options),
	593	};
	594
	595	static void fio_init fio_aioring_register(void)
	596	{
52885fa2	597	register_ioengine(&ioengine);
52885fa2 JA	598	}
	599
	600	static void fio_exit fio_aioring_unregister(void)
	601	{
52885fa2	602	unregister_ioengine(&ioengine);
52885fa2	603	}
1f90e9bb	604	#endif