[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 || 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;
	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;
	}
	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);

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

	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;

	/* 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
e3970057 JA	200	if (io_u->ddir == DDIR_READ \|\| io_u->ddir == DDIR_WRITE) {
e3970057 JA	201	if (io_u->ddir == DDIR_READ)
52885fa2	202	iocb->aio_lio_opcode = IO_CMD_PREAD;
e3970057	203	else
52885fa2	204	iocb->aio_lio_opcode = IO_CMD_PWRITE;
e3970057 JA	205	iocb->aio_reqprio = 0;
	206	iocb->aio_fildes = f->fd;
	207	iocb->u.c.buf = io_u->xfer_buf;
	208	iocb->u.c.nbytes = io_u->xfer_buflen;
	209	iocb->u.c.offset = io_u->offset;
	210	if (o->hipri)
	211	iocb->u.c.flags \|= IOCB_FLAG_HIPRI;
	212	else
	213	iocb->u.c.flags = 0;
52885fa2 JA	214	} else if (ddir_sync(io_u->ddir))
	215	io_prep_fsync(iocb, f->fd);
	216
	217	iocb->data = io_u;
	218	return 0;
	219	}
	220
	221	static struct io_u fio_aioring_event(struct thread_data td, int event)
	222	{
	223	struct aioring_data *ld = td->io_ops_data;
	224	struct io_event *ev;
	225	struct io_u *io_u;
	226	int index;
	227
	228	index = event + ld->cq_ring_off;
	229	if (index >= ld->cq_ring->nr_events)
	230	index -= ld->cq_ring->nr_events;
	231
	232	ev = &ld->cq_ring->events[index];
	233	io_u = ev->data;
	234
	235	if (ev->res != io_u->xfer_buflen) {
	236	if (ev->res > io_u->xfer_buflen)
	237	io_u->error = -ev->res;
	238	else
	239	io_u->resid = io_u->xfer_buflen - ev->res;
	240	} else
	241	io_u->error = 0;
	242
96563db9 JA	243	if (io_u->ddir == DDIR_READ) {
	244	if (ev->res2 & IOEV_RES2_CACHEHIT)
	245	ld->cachehit++;
	246	else
	247	ld->cachemiss++;
	248	}
	249
52885fa2 JA	250	return io_u;
	251	}
	252
	253	static int fio_aioring_cqring_reap(struct thread_data *td, unsigned int events,
	254	unsigned int max)
	255	{
	256	struct aioring_data *ld = td->io_ops_data;
	257	struct aio_cq_ring *ring = ld->cq_ring;
	258	u32 head, reaped = 0;
	259
	260	head = ring->head;
	261	do {
	262	read_barrier();
	263	if (head == ring->tail)
	264	break;
	265	reaped++;
	266	head++;
	267	if (head == ring->nr_events)
	268	head = 0;
	269	} while (reaped + events < max);
	270
	271	ring->head = head;
	272	write_barrier();
	273	return reaped;
	274	}
	275
	276	static int fio_aioring_getevents(struct thread_data *td, unsigned int min,
	277	unsigned int max, const struct timespec *t)
	278	{
	279	struct aioring_data *ld = td->io_ops_data;
	280	unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
771c9901	281	struct aioring_options *o = td->eo;
52885fa2 JA	282	struct aio_cq_ring *ring = ld->cq_ring;
	283	int r, events = 0;
	284
	285	ld->cq_ring_off = ring->head;
	286	do {
	287	r = fio_aioring_cqring_reap(td, events, max);
	288	if (r) {
	289	events += r;
	290	continue;
	291	}
	292
771c9901 JA	293	if (!o->sqthread_poll) {
	294	r = io_ring_enter(ld->aio_ctx, 0, actual_min,
	295	IORING_FLAG_GETEVENTS);
	296	if (r < 0) {
	297	if (errno == EAGAIN)
	298	continue;
	299	td_verror(td, errno, "io_ring_enter get");
	300	break;
	301	}
52885fa2 JA	302	}
	303	} while (events < min);
	304
	305	return r < 0 ? r : events;
	306	}
	307
	308	static enum fio_q_status fio_aioring_queue(struct thread_data *td,
	309	struct io_u *io_u)
	310	{
	311	struct aioring_data *ld = td->io_ops_data;
	312	struct aio_sq_ring *ring = ld->sq_ring;
	313	unsigned tail, next_tail;
	314
	315	fio_ro_check(td, io_u);
	316
	317	if (ld->queued == td->o.iodepth)
	318	return FIO_Q_BUSY;
	319
52885fa2 JA	320	if (io_u->ddir == DDIR_TRIM) {
	321	if (ld->queued)
	322	return FIO_Q_BUSY;
	323
	324	do_io_u_trim(td, io_u);
	325	io_u_mark_submit(td, 1);
	326	io_u_mark_complete(td, 1);
	327	return FIO_Q_COMPLETED;
	328	}
	329
	330	tail = ring->tail;
	331	next_tail = tail + 1;
	332	if (next_tail == ring->nr_events)
	333	next_tail = 0;
	334	read_barrier();
	335	if (next_tail == ring->head)
	336	return FIO_Q_BUSY;
	337
	338	ring->array[tail] = io_u->index;
	339	ring->tail = next_tail;
	340	write_barrier();
	341
	342	ld->queued++;
	343	return FIO_Q_QUEUED;
	344	}
	345
	346	static void fio_aioring_queued(struct thread_data *td, int start, int nr)
	347	{
	348	struct aioring_data *ld = td->io_ops_data;
	349	struct timespec now;
	350
	351	if (!fio_fill_issue_time(td))
	352	return;
	353
	354	fio_gettime(&now, NULL);
	355
	356	while (nr--) {
	357	int index = ld->sq_ring->array[start];
	358	struct io_u *io_u = io_u = ld->io_u_index[index];
	359
	360	memcpy(&io_u->issue_time, &now, sizeof(now));
	361	io_u_queued(td, io_u);
	362
	363	start++;
	364	if (start == ld->sq_ring->nr_events)
	365	start = 0;
	366	}
	367	}
	368
	369	static int fio_aioring_commit(struct thread_data *td)
	370	{
	371	struct aioring_data *ld = td->io_ops_data;
771c9901	372	struct aioring_options *o = td->eo;
52885fa2 JA	373	int ret;
	374
	375	if (!ld->queued)
	376	return 0;
	377
771c9901 JA	378	/* Nothing to do */
771c9901 JA	379	if (o->sqthread_poll) {
4cdbc048 JA	380	struct aio_sq_ring *ring = ld->sq_ring;
	381
	382	if (ring->kflags & IORING_SQ_NEED_WAKEUP)
	383	io_ring_enter(ld->aio_ctx, ld->queued, 0, IORING_FLAG_SUBMIT);
771c9901 JA	384	ld->queued = 0;
	385	return 0;
	386	}
	387
52885fa2 JA	388	do {
	389	int start = ld->sq_ring->head;
	390	long nr = ld->queued;
	391
	392	ret = io_ring_enter(ld->aio_ctx, nr, 0, IORING_FLAG_SUBMIT \|
	393	IORING_FLAG_GETEVENTS);
52885fa2 JA	394	if (ret > 0) {
	395	fio_aioring_queued(td, start, ret);
	396	io_u_mark_submit(td, ret);
	397
	398	ld->queued -= ret;
	399	ret = 0;
a90cd050 JA	400	} else if (!ret) {
a90cd050 JA	401	io_u_mark_submit(td, ret);
52885fa2	402	continue;
a90cd050 JA	403	} else {
	404	if (errno == EAGAIN) {
	405	ret = fio_aioring_cqring_reap(td, 0, ld->queued);
	406	if (ret)
	407	continue;
	408	/* Shouldn't happen */
	409	usleep(1);
	410	continue;
52885fa2	411	}
a90cd050	412	td_verror(td, errno, "io_ring_enter sumit");
52885fa2	413	break;
a90cd050	414	}
52885fa2 JA	415	} while (ld->queued);
	416
	417	return ret;
	418	}
	419
	420	static size_t aioring_cq_size(struct thread_data *td)
	421	{
	422	return sizeof(struct aio_cq_ring) + 2 * td->o.iodepth * sizeof(struct io_event);
	423	}
	424
	425	static size_t aioring_sq_iocb(struct thread_data *td)
	426	{
	427	return sizeof(struct iocb) * td->o.iodepth;
	428	}
	429
	430	static size_t aioring_sq_size(struct thread_data *td)
	431	{
	432	return sizeof(struct aio_sq_ring) + td->o.iodepth * sizeof(u32);
	433	}
	434
	435	static void fio_aioring_cleanup(struct thread_data *td)
	436	{
	437	struct aioring_data *ld = td->io_ops_data;
	438
	439	if (ld) {
96563db9 JA	440	td->ts.cachehit += ld->cachehit;
	441	td->ts.cachemiss += ld->cachemiss;
	442
a90cd050 JA	443	/* Bump depth to match init depth */
	444	td->o.iodepth++;
	445
52885fa2 JA	446	/*
	447	* Work-around to avoid huge RCU stalls at exit time. If we
	448	* don't do this here, then it'll be torn down by exit_aio().
	449	* But for that case we can parallellize the freeing, thus
	450	* speeding it up a lot.
	451	*/
	452	if (!(td->flags & TD_F_CHILD))
	453	io_destroy(ld->aio_ctx);
	454	free(ld->io_u_index);
	455	free(ld->io_us);
	456	fio_memfree(ld->sq_ring, aioring_sq_size(td), false);
	457	fio_memfree(ld->iocbs, aioring_sq_iocb(td), false);
	458	fio_memfree(ld->cq_ring, aioring_cq_size(td), false);
	459	free(ld);
	460	}
	461	}
	462
	463	static int fio_aioring_queue_init(struct thread_data *td)
	464	{
52885fa2 JA	465	struct aioring_data *ld = td->io_ops_data;
	466	struct aioring_options *o = td->eo;
	467	int flags = IOCTX_FLAG_SCQRING;
	468	int depth = td->o.iodepth;
	469
	470	if (o->hipri)
	471	flags \|= IOCTX_FLAG_IOPOLL;
a90cd050 JA	472	if (o->sqthread_set) {
	473	ld->sq_ring->sq_thread_cpu = o->sqthread;
	474	flags \|= IOCTX_FLAG_SQTHREAD;
771c9901 JA	475	if (o->sqthread_poll)
771c9901 JA	476	flags \|= IOCTX_FLAG_SQPOLL;
f635f1fb JA	477	}
f635f1fb JA	478	if (o->sqwq)
a90cd050 JA	479	flags \|= IOCTX_FLAG_SQWQ;
a90cd050 JA	480
52885fa2 JA	481	if (o->fixedbufs) {
	482	struct rlimit rlim = {
	483	.rlim_cur = RLIM_INFINITY,
	484	.rlim_max = RLIM_INFINITY,
	485	};
	486
	487	setrlimit(RLIMIT_MEMLOCK, &rlim);
	488	flags \|= IOCTX_FLAG_FIXEDBUFS;
	489	}
	490
	491	return syscall(__NR_sys_io_setup2, depth, flags,
	492	ld->sq_ring, ld->cq_ring, &ld->aio_ctx);
52885fa2 JA	493	}
	494
	495	static int fio_aioring_post_init(struct thread_data *td)
	496	{
	497	struct aioring_data *ld = td->io_ops_data;
	498	struct aioring_options *o = td->eo;
	499	struct io_u *io_u;
	500	struct iocb *iocb;
	501	int err = 0;
	502
	503	if (o->fixedbufs) {
	504	int i;
	505
	506	for (i = 0; i < td->o.iodepth; i++) {
	507	io_u = ld->io_u_index[i];
	508	iocb = &ld->iocbs[i];
	509	iocb->u.c.buf = io_u->buf;
	510	iocb->u.c.nbytes = td_max_bs(td);
	511
	512	if (o->hipri)
	513	iocb->u.c.flags \|= IOCB_FLAG_HIPRI;
	514	}
	515	}
	516
	517	err = fio_aioring_queue_init(td);
d63a472d JA	518
	519	/* Adjust depth back again */
	520	td->o.iodepth--;
	521
52885fa2	522	if (err) {
d63a472d	523	td_verror(td, errno, "io_queue_init");
52885fa2 JA	524	return 1;
	525	}
	526
	527	return 0;
	528	}
	529
	530	static int fio_aioring_init(struct thread_data *td)
	531	{
	532	struct aioring_data *ld;
	533
3b617146 JA	534	/* ring needs an extra entry, add one to achieve QD set */
3b617146 JA	535	td->o.iodepth++;
52885fa2 JA	536
	537	ld = calloc(1, sizeof(*ld));
	538
	539	/* io_u index */
	540	ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *));
	541	ld->io_us = calloc(td->o.iodepth, sizeof(struct io_u *));
	542
	543	ld->iocbs = fio_memalign(page_size, aioring_sq_iocb(td), false);
	544	memset(ld->iocbs, 0, aioring_sq_iocb(td));
	545
	546	ld->sq_ring = fio_memalign(page_size, aioring_sq_size(td), false);
	547	memset(ld->sq_ring, 0, aioring_sq_size(td));
	548	ld->sq_ring->nr_events = td->o.iodepth;
1f90e9bb	549	ld->sq_ring->iocbs = (u64) (uintptr_t) ld->iocbs;
52885fa2 JA	550
	551	ld->cq_ring = fio_memalign(page_size, aioring_cq_size(td), false);
	552	memset(ld->cq_ring, 0, aioring_cq_size(td));
	553	ld->cq_ring->nr_events = td->o.iodepth * 2;
	554
	555	td->io_ops_data = ld;
	556	return 0;
	557	}
	558
	559	static int fio_aioring_io_u_init(struct thread_data td, struct io_u io_u)
	560	{
	561	struct aioring_data *ld = td->io_ops_data;
	562
	563	ld->io_u_index[io_u->index] = io_u;
	564	return 0;
	565	}
	566
	567	static struct ioengine_ops ioengine = {
	568	.name = "aio-ring",
	569	.version = FIO_IOOPS_VERSION,
	570	.init = fio_aioring_init,
	571	.post_init = fio_aioring_post_init,
	572	.io_u_init = fio_aioring_io_u_init,
	573	.prep = fio_aioring_prep,
	574	.queue = fio_aioring_queue,
	575	.commit = fio_aioring_commit,
	576	.getevents = fio_aioring_getevents,
	577	.event = fio_aioring_event,
	578	.cleanup = fio_aioring_cleanup,
	579	.open_file = generic_open_file,
	580	.close_file = generic_close_file,
	581	.get_file_size = generic_get_file_size,
	582	.options = options,
	583	.option_struct_size = sizeof(struct aioring_options),
	584	};
	585
	586	static void fio_init fio_aioring_register(void)
	587	{
52885fa2	588	register_ioengine(&ioengine);
52885fa2 JA	589	}
	590
	591	static void fio_exit fio_aioring_unregister(void)
	592	{
52885fa2	593	unregister_ioengine(&ioengine);
52885fa2	594	}
1f90e9bb	595	#endif