[fio.git] / engines / aioring.c

/*
 * aioring engine
 *
 * IO engine using the new native Linux libaio ring interface
 *
 */
#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

/*
 * 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;

struct aio_sq_ring {
	union {
		struct {
			u32 head;
			u32 tail;
			u32 nr_events;
			u16 sq_thread_cpu;
			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;
};

struct aioring_options {
	void *pad;
	unsigned int hipri;
	unsigned int fixedbufs;
};

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	= NULL,
	},
};

static int fio_aioring_commit(struct thread_data *td);

static int io_ring_enter(io_context_t ctx, unsigned int to_submit,
			 unsigned int min_complete, unsigned int flags)
{
#ifdef __NR_sys_io_ring_enter
	return syscall(__NR_sys_io_ring_enter, ctx, to_submit, min_complete,
			flags);
#else
	return -1;
#endif
}

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;

	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 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;
		}

		r = io_ring_enter(ld->aio_ctx, 0, actual_min,
					IORING_FLAG_GETEVENTS);
		if (r < 0) {
			if (errno == EAGAIN)
				continue;
			perror("ring 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 == td->o.iodepth)
		return FIO_Q_BUSY;

	/*
	 * fsync is tricky, since it can fail and we need to do it
	 * serialized with other io. the reason is that linux doesn't
	 * support aio fsync yet. So return busy for the case where we
	 * have pending io, to let fio complete those first.
	 */
	if (ddir_sync(io_u->ddir)) {
		if (ld->queued)
			return FIO_Q_BUSY;

		do_io_u_sync(td, io_u);
		return FIO_Q_COMPLETED;
	}

	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;
	int ret;

	if (!ld->queued)
		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 == -1)
			perror("io_ring_enter");
		if (ret > 0) {
			fio_aioring_queued(td, start, ret);
			io_u_mark_submit(td, ret);

			ld->queued -= ret;
			ret = 0;
		} else if (ret == -EINTR || !ret) {
			if (!ret)
				io_u_mark_submit(td, ret);
			continue;
		} else if (ret == -EAGAIN) {
			/*
			 * If we get EAGAIN, we should break out without
			 * error and let the upper layer reap some
			 * events for us. If we have no queued IO, we
			 * must loop here. If we loop for more than 30s,
			 * just error out, something must be buggy in the
			 * IO path.
			 */
			if (ld->queued) {
				ret = 0;
				break;
			}
			usleep(1);
			continue;
		} else if (ret == -ENOMEM) {
			/*
			 * If we get -ENOMEM, reap events if we can. If
			 * we cannot, treat it as a fatal event since there's
			 * nothing we can do about it.
			 */
			if (ld->queued)
				ret = 0;
			break;
		} else
			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) {
		/*
		 * 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)
{
#ifdef __NR_sys_io_setup2
	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->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);
#else
	return -1;
#endif
}

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;
	}

	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	*
	4	* IO engine using the new native Linux libaio ring interface
	5	*
	6	*/
	7	#include <stdlib.h>
	8	#include <unistd.h>
	9	#include <errno.h>
	10	#include <libaio.h>
	11	#include <sys/time.h>
	12	#include <sys/resource.h>
	13
	14	#include "../fio.h"
	15	#include "../lib/pow2.h"
	16	#include "../optgroup.h"
	17	#include "../lib/memalign.h"
	18
a30e63cf JA	19	#ifdef ARCH_HAVE_AIORING
a30e63cf JA	20
52885fa2 JA	21	#ifndef IOCB_FLAG_HIPRI
	22	#define IOCB_FLAG_HIPRI (1 << 2)
	23	#endif
	24
	25	/*
	26	* io_setup2(2) flags
	27	*/
	28	#ifndef IOCTX_FLAG_IOPOLL
	29	#define IOCTX_FLAG_IOPOLL (1 << 0)
	30	#endif
	31	#ifndef IOCTX_FLAG_SCQRING
	32	#define IOCTX_FLAG_SCQRING (1 << 1)
	33	#endif
	34	#ifndef IOCTX_FLAG_FIXEDBUFS
	35	#define IOCTX_FLAG_FIXEDBUFS (1 << 2)
	36	#endif
	37	#ifndef IOCTX_FLAG_SQTHREAD
	38	#define IOCTX_FLAG_SQTHREAD (1 << 3)
	39	#endif
	40	#ifndef IOCTX_FLAG_SQWQ
	41	#define IOCTX_FLAG_SQWQ (1 << 4)
	42	#endif
	43
	44	/*
	45	* io_ring_enter(2) flags
	46	*/
	47	#ifndef IORING_FLAG_SUBMIT
	48	#define IORING_FLAG_SUBMIT (1 << 0)
	49	#endif
	50	#ifndef IORING_FLAG_GETEVENTS
	51	#define IORING_FLAG_GETEVENTS (1 << 1)
	52	#endif
	53
	54	typedef uint64_t u64;
	55	typedef uint32_t u32;
	56	typedef uint16_t u16;
	57
	58	struct aio_sq_ring {
	59	union {
	60	struct {
	61	u32 head;
	62	u32 tail;
	63	u32 nr_events;
	64	u16 sq_thread_cpu;
	65	u64 iocbs;
	66	};
	67	u32 pad[16];
	68	};
	69	u32 array[0];
	70	};
	71
	72	struct aio_cq_ring {
	73	union {
	74	struct {
	75	u32 head;
	76	u32 tail;
	77	u32 nr_events;
	78	};
	79	struct io_event pad;
	80	};
	81	struct io_event events[0];
	82	};
	83
	84	struct aioring_data {
85	io_context_t aio_ctx;
86	struct io_u **io_us;
87	struct io_u **io_u_index;
88
89	struct aio_sq_ring *sq_ring;
90	struct iocb *iocbs;
91
92	struct aio_cq_ring *cq_ring;
93	struct io_event *events;
94
95	int queued;
96	int cq_ring_off;
97	};
98
99	struct aioring_options {
100	void *pad;
101	unsigned int hipri;
102	unsigned int fixedbufs;
103	};
104
105	static struct fio_option options[] = {
106	{
107	.name = "hipri",
108	.lname = "High Priority",
109	.type = FIO_OPT_STR_SET,
110	.off1 = offsetof(struct aioring_options, hipri),
111	.help = "Use polled IO completions",
112	.category = FIO_OPT_C_ENGINE,
113	.group = FIO_OPT_G_LIBAIO,
114	},
115	{
116	.name = "fixedbufs",
117	.lname = "Fixed (pre-mapped) IO buffers",
118	.type = FIO_OPT_STR_SET,
119	.off1 = offsetof(struct aioring_options, fixedbufs),
120	.help = "Pre map IO buffers",
121	.category = FIO_OPT_C_ENGINE,
122	.group = FIO_OPT_G_LIBAIO,
123	},
124	{
125	.name = NULL,
126	},
127	};
128
129	static int fio_aioring_commit(struct thread_data *td);
130
131	static int io_ring_enter(io_context_t ctx, unsigned int to_submit,
132	unsigned int min_complete, unsigned int flags)
133	{
134	#ifdef __NR_sys_io_ring_enter
135	return syscall(__NR_sys_io_ring_enter, ctx, to_submit, min_complete,
136	flags);
137	#else
138	return -1;
139	#endif
140	}
141
142	static int fio_aioring_prep(struct thread_data td, struct io_u io_u)
143	{
144	struct aioring_data *ld = td->io_ops_data;
145	struct fio_file *f = io_u->file;
146	struct aioring_options *o = td->eo;
147	struct iocb *iocb;
148
149	iocb = &ld->iocbs[io_u->index];
150
151	if (io_u->ddir == DDIR_READ) {
152	if (o->fixedbufs) {
153	iocb->aio_fildes = f->fd;
154	iocb->aio_lio_opcode = IO_CMD_PREAD;
155	iocb->u.c.offset = io_u->offset;
156	} else {
157	io_prep_pread(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
158	if (o->hipri)
159	iocb->u.c.flags \|= IOCB_FLAG_HIPRI;
160	}
161	} else if (io_u->ddir == DDIR_WRITE) {
162	if (o->fixedbufs) {
163	iocb->aio_fildes = f->fd;
164	iocb->aio_lio_opcode = IO_CMD_PWRITE;
165	iocb->u.c.offset = io_u->offset;
166	} else {
167	io_prep_pwrite(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
168	if (o->hipri)
169	iocb->u.c.flags \|= IOCB_FLAG_HIPRI;
170	}
171	} else if (ddir_sync(io_u->ddir))
172	io_prep_fsync(iocb, f->fd);
173
174	iocb->data = io_u;
175	return 0;
176	}
177
178	static struct io_u fio_aioring_event(struct thread_data td, int event)
179	{
180	struct aioring_data *ld = td->io_ops_data;
181	struct io_event *ev;
182	struct io_u *io_u;
183	int index;
184
185	index = event + ld->cq_ring_off;
186	if (index >= ld->cq_ring->nr_events)
187	index -= ld->cq_ring->nr_events;
188
189	ev = &ld->cq_ring->events[index];
190	io_u = ev->data;
191
192	if (ev->res != io_u->xfer_buflen) {
193	if (ev->res > io_u->xfer_buflen)
194	io_u->error = -ev->res;
195	else
196	io_u->resid = io_u->xfer_buflen - ev->res;
197	} else
198	io_u->error = 0;
199
200	return io_u;
201	}
202
203	static int fio_aioring_cqring_reap(struct thread_data *td, unsigned int events,
204	unsigned int max)
205	{
206	struct aioring_data *ld = td->io_ops_data;
207	struct aio_cq_ring *ring = ld->cq_ring;
208	u32 head, reaped = 0;
209
210	head = ring->head;
211	do {
212	read_barrier();
213	if (head == ring->tail)
214	break;
215	reaped++;
216	head++;
217	if (head == ring->nr_events)
218	head = 0;
219	} while (reaped + events < max);
220
221	ring->head = head;
222	write_barrier();
223	return reaped;
224	}
225
226	static int fio_aioring_getevents(struct thread_data *td, unsigned int min,
227	unsigned int max, const struct timespec *t)
228	{
229	struct aioring_data *ld = td->io_ops_data;
230	unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
231	struct aio_cq_ring *ring = ld->cq_ring;
232	int r, events = 0;
233
234	ld->cq_ring_off = ring->head;
235	do {
236	r = fio_aioring_cqring_reap(td, events, max);
237	if (r) {
238	events += r;
239	continue;
240	}
241
242	r = io_ring_enter(ld->aio_ctx, 0, actual_min,
243	IORING_FLAG_GETEVENTS);
244	if (r < 0) {
245	if (errno == EAGAIN)
246	continue;
247	perror("ring enter");
248	break;
249	}
250	} while (events < min);
251
252	return r < 0 ? r : events;
253	}
254
255	static enum fio_q_status fio_aioring_queue(struct thread_data *td,
256	struct io_u *io_u)
257	{
258	struct aioring_data *ld = td->io_ops_data;
259	struct aio_sq_ring *ring = ld->sq_ring;
260	unsigned tail, next_tail;
261
262	fio_ro_check(td, io_u);
263
264	if (ld->queued == td->o.iodepth)
265	return FIO_Q_BUSY;
266
267	/*
268	* fsync is tricky, since it can fail and we need to do it
269	* serialized with other io. the reason is that linux doesn't
270	* support aio fsync yet. So return busy for the case where we
271	* have pending io, to let fio complete those first.
272	*/
273	if (ddir_sync(io_u->ddir)) {
274	if (ld->queued)
275	return FIO_Q_BUSY;
276
277	do_io_u_sync(td, io_u);
278	return FIO_Q_COMPLETED;
279	}
280
281	if (io_u->ddir == DDIR_TRIM) {
282	if (ld->queued)
283	return FIO_Q_BUSY;
284
285	do_io_u_trim(td, io_u);
286	io_u_mark_submit(td, 1);
287	io_u_mark_complete(td, 1);
288	return FIO_Q_COMPLETED;
289	}
290
291	tail = ring->tail;
292	next_tail = tail + 1;
293	if (next_tail == ring->nr_events)
294	next_tail = 0;
295	read_barrier();
296	if (next_tail == ring->head)
297	return FIO_Q_BUSY;
298
299	ring->array[tail] = io_u->index;
300	ring->tail = next_tail;
301	write_barrier();
302
303	ld->queued++;
304	return FIO_Q_QUEUED;
305	}
306
307	static void fio_aioring_queued(struct thread_data *td, int start, int nr)
308	{
309	struct aioring_data *ld = td->io_ops_data;
310	struct timespec now;
311
312	if (!fio_fill_issue_time(td))
313	return;
314
315	fio_gettime(&now, NULL);
316
317	while (nr--) {
318	int index = ld->sq_ring->array[start];
319	struct io_u *io_u = io_u = ld->io_u_index[index];
320
321	memcpy(&io_u->issue_time, &now, sizeof(now));
322	io_u_queued(td, io_u);
323
324	start++;
325	if (start == ld->sq_ring->nr_events)
326	start = 0;
327	}
328	}
329
330	static int fio_aioring_commit(struct thread_data *td)
331	{
332	struct aioring_data *ld = td->io_ops_data;
333	int ret;
334
335	if (!ld->queued)
336	return 0;
337
338	do {
339	int start = ld->sq_ring->head;
340	long nr = ld->queued;
341
342	ret = io_ring_enter(ld->aio_ctx, nr, 0, IORING_FLAG_SUBMIT \|
343	IORING_FLAG_GETEVENTS);
344	if (ret == -1)
345	perror("io_ring_enter");
346	if (ret > 0) {
347	fio_aioring_queued(td, start, ret);
348	io_u_mark_submit(td, ret);
349
350	ld->queued -= ret;
351	ret = 0;
352	} else if (ret == -EINTR \|\| !ret) {
353	if (!ret)
354	io_u_mark_submit(td, ret);
355	continue;
356	} else if (ret == -EAGAIN) {
357	/*
358	* If we get EAGAIN, we should break out without
359	* error and let the upper layer reap some
360	* events for us. If we have no queued IO, we
361	* must loop here. If we loop for more than 30s,
362	* just error out, something must be buggy in the
363	* IO path.
364	*/
365	if (ld->queued) {
366	ret = 0;
367	break;
368	}
369	usleep(1);
370	continue;
371	} else if (ret == -ENOMEM) {
372	/*
373	* If we get -ENOMEM, reap events if we can. If
374	* we cannot, treat it as a fatal event since there's
375	* nothing we can do about it.
376	*/
377	if (ld->queued)
378	ret = 0;
379	break;
380	} else
381	break;
382	} while (ld->queued);
383
384	return ret;
385	}
386
387	static size_t aioring_cq_size(struct thread_data *td)
388	{
389	return sizeof(struct aio_cq_ring) + 2 * td->o.iodepth * sizeof(struct io_event);
390	}
391
392	static size_t aioring_sq_iocb(struct thread_data *td)
393	{
394	return sizeof(struct iocb) * td->o.iodepth;
395	}
396
397	static size_t aioring_sq_size(struct thread_data *td)
398	{
399	return sizeof(struct aio_sq_ring) + td->o.iodepth * sizeof(u32);
400	}
401
402	static void fio_aioring_cleanup(struct thread_data *td)
403	{
404	struct aioring_data *ld = td->io_ops_data;
405
406	if (ld) {
407	/*
408	* Work-around to avoid huge RCU stalls at exit time. If we
409	* don't do this here, then it'll be torn down by exit_aio().
410	* But for that case we can parallellize the freeing, thus
411	* speeding it up a lot.
412	*/
413	if (!(td->flags & TD_F_CHILD))
414	io_destroy(ld->aio_ctx);
415	free(ld->io_u_index);
416	free(ld->io_us);
417	fio_memfree(ld->sq_ring, aioring_sq_size(td), false);
418	fio_memfree(ld->iocbs, aioring_sq_iocb(td), false);
419	fio_memfree(ld->cq_ring, aioring_cq_size(td), false);
420	free(ld);
421	}
422	}
423
424	static int fio_aioring_queue_init(struct thread_data *td)
425	{
426	#ifdef __NR_sys_io_setup2
427	struct aioring_data *ld = td->io_ops_data;
428	struct aioring_options *o = td->eo;
429	int flags = IOCTX_FLAG_SCQRING;
430	int depth = td->o.iodepth;
431
432	if (o->hipri)
433	flags \|= IOCTX_FLAG_IOPOLL;
434	if (o->fixedbufs) {
435	struct rlimit rlim = {
436	.rlim_cur = RLIM_INFINITY,
437	.rlim_max = RLIM_INFINITY,
438	};
439
440	setrlimit(RLIMIT_MEMLOCK, &rlim);
441	flags \|= IOCTX_FLAG_FIXEDBUFS;
442	}
443
444	return syscall(__NR_sys_io_setup2, depth, flags,
445	ld->sq_ring, ld->cq_ring, &ld->aio_ctx);
446	#else
447	return -1;
448	#endif
449	}
450
451	static int fio_aioring_post_init(struct thread_data *td)
452	{
453	struct aioring_data *ld = td->io_ops_data;
454	struct aioring_options *o = td->eo;
455	struct io_u *io_u;
456	struct iocb *iocb;
457	int err = 0;
458
459	if (o->fixedbufs) {
460	int i;
461
462	for (i = 0; i < td->o.iodepth; i++) {
463	io_u = ld->io_u_index[i];
464	iocb = &ld->iocbs[i];
465	iocb->u.c.buf = io_u->buf;
466	iocb->u.c.nbytes = td_max_bs(td);
467
468	if (o->hipri)
469	iocb->u.c.flags \|= IOCB_FLAG_HIPRI;
470	}
471	}
472
473	err = fio_aioring_queue_init(td);
474	if (err) {
475	td_verror(td, -err, "io_queue_init");
476	return 1;
477	}
478
479	return 0;
480	}
481
482	static int fio_aioring_init(struct thread_data *td)
483	{
484	struct aioring_data *ld;
485
3b617146 JA	486	/* ring needs an extra entry, add one to achieve QD set */
3b617146 JA	487	td->o.iodepth++;
52885fa2 JA	488
	489	ld = calloc(1, sizeof(*ld));
	490
	491	/* io_u index */
	492	ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *));
	493	ld->io_us = calloc(td->o.iodepth, sizeof(struct io_u *));
	494
	495	ld->iocbs = fio_memalign(page_size, aioring_sq_iocb(td), false);
	496	memset(ld->iocbs, 0, aioring_sq_iocb(td));
	497
	498	ld->sq_ring = fio_memalign(page_size, aioring_sq_size(td), false);
	499	memset(ld->sq_ring, 0, aioring_sq_size(td));
	500	ld->sq_ring->nr_events = td->o.iodepth;
1f90e9bb	501	ld->sq_ring->iocbs = (u64) (uintptr_t) ld->iocbs;
52885fa2 JA	502
	503	ld->cq_ring = fio_memalign(page_size, aioring_cq_size(td), false);
	504	memset(ld->cq_ring, 0, aioring_cq_size(td));
	505	ld->cq_ring->nr_events = td->o.iodepth * 2;
	506
	507	td->io_ops_data = ld;
	508	return 0;
	509	}
	510
	511	static int fio_aioring_io_u_init(struct thread_data td, struct io_u io_u)
	512	{
	513	struct aioring_data *ld = td->io_ops_data;
	514
	515	ld->io_u_index[io_u->index] = io_u;
	516	return 0;
	517	}
	518
	519	static struct ioengine_ops ioengine = {
	520	.name = "aio-ring",
	521	.version = FIO_IOOPS_VERSION,
	522	.init = fio_aioring_init,
	523	.post_init = fio_aioring_post_init,
	524	.io_u_init = fio_aioring_io_u_init,
	525	.prep = fio_aioring_prep,
	526	.queue = fio_aioring_queue,
	527	.commit = fio_aioring_commit,
	528	.getevents = fio_aioring_getevents,
	529	.event = fio_aioring_event,
	530	.cleanup = fio_aioring_cleanup,
	531	.open_file = generic_open_file,
	532	.close_file = generic_close_file,
	533	.get_file_size = generic_get_file_size,
	534	.options = options,
	535	.option_struct_size = sizeof(struct aioring_options),
	536	};
	537
	538	static void fio_init fio_aioring_register(void)
	539	{
52885fa2	540	register_ioengine(&ioengine);
52885fa2 JA	541	}
	542
	543	static void fio_exit fio_aioring_unregister(void)
	544	{
52885fa2	545	unregister_ioengine(&ioengine);
52885fa2	546	}
1f90e9bb	547	#endif