[fio.git] / engines / splice.c

/*
 * splice engine
 *
 * IO engine that transfers data by doing splices to/from pipes and
 * the files.
 *
 */
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <errno.h>
#include <assert.h>
#include <sys/poll.h>

#include "../fio.h"

#ifdef FIO_HAVE_SPLICE

struct spliceio_data {
	int pipe[2];
	int vmsplice_to_user;
};

/*
 * vmsplice didn't use to support splicing to user space, this is the old
 * variant of getting that job done. Doesn't make a lot of sense, but it
 * uses splices to move data from the source into a pipe.
 */
static int fio_splice_read_old(struct thread_data *td, struct io_u *io_u)
{
	struct spliceio_data *sd = td->io_ops->data;
	struct fio_file *f = io_u->file;
	int ret, ret2, buflen;
	off_t offset;
	void *p;

	offset = io_u->offset;
	buflen = io_u->xfer_buflen;
	p = io_u->xfer_buf;
	while (buflen) {
		int this_len = buflen;

		if (this_len > SPLICE_DEF_SIZE)
			this_len = SPLICE_DEF_SIZE;

		ret = splice(f->fd, &offset, sd->pipe[1], NULL, this_len, SPLICE_F_MORE);
		if (ret < 0) {
			if (errno == ENODATA || errno == EAGAIN)
				continue;

			return -errno;
		}

		buflen -= ret;

		while (ret) {
			ret2 = read(sd->pipe[0], p, ret);
			if (ret2 < 0)
				return -errno;

			ret -= ret2;
			p += ret2;
		}
	}

	return io_u->xfer_buflen;
}

/*
 * We can now vmsplice into userspace, so do the transfer by splicing into
 * a pipe and vmsplicing that into userspace.
 */
static int fio_splice_read(struct thread_data *td, struct io_u *io_u)
{
	struct spliceio_data *sd = td->io_ops->data;
	struct fio_file *f = io_u->file;
	struct iovec iov;
	int ret, buflen;
	off_t offset;
	void *p;

	offset = io_u->offset;
	buflen = io_u->xfer_buflen;
	p = io_u->xfer_buf;
	io_u->xfer_buf = NULL;
	while (buflen) {
		int this_len = buflen;

		if (this_len > SPLICE_DEF_SIZE)
			this_len = SPLICE_DEF_SIZE;

		ret = splice(f->fd, &offset, sd->pipe[1], NULL, this_len, SPLICE_F_MORE);
		if (ret < 0) {
			if (errno == ENODATA || errno == EAGAIN)
				continue;

			return -errno;
		}

		buflen -= ret;
		iov.iov_base = p;
		iov.iov_len = ret;
		p += ret;

		while (iov.iov_len) {
			ret = vmsplice(sd->pipe[0], &iov, 1, SPLICE_F_MOVE);
			if (ret < 0)
				return -errno;
			else if (!ret)
				return -ENODATA;

			if (!io_u->xfer_buf)
				io_u->xfer_buf = iov.iov_base;
			iov.iov_len -= ret;
			iov.iov_base += ret;
		}
	}

	io_u->unmap = splice_unmap_io_u;
	return io_u->xfer_buflen;
}

/*
 * For splice writing, we can vmsplice our data buffer directly into a
 * pipe and then splice that to a file.
 */
static int fio_splice_write(struct thread_data *td, struct io_u *io_u)
{
	struct spliceio_data *sd = td->io_ops->data;
	struct iovec iov = {
		.iov_base = io_u->xfer_buf,
		.iov_len = io_u->xfer_buflen,
	};
	struct pollfd pfd = { .fd = sd->pipe[1], .events = POLLOUT, };
	struct fio_file *f = io_u->file;
	off_t off = io_u->offset;
	int ret, ret2;

	while (iov.iov_len) {
		if (poll(&pfd, 1, -1) < 0)
			return errno;

		ret = vmsplice(sd->pipe[1], &iov, 1, SPLICE_F_NONBLOCK);
		if (ret < 0)
			return -errno;

		iov.iov_len -= ret;
		iov.iov_base += ret;

		while (ret) {
			ret2 = splice(sd->pipe[0], NULL, f->fd, &off, ret, 0);
			if (ret2 < 0)
				return -errno;

			ret -= ret2;
		}
	}

	return io_u->xfer_buflen;
}

static void splice_unmap_io_u(struct thread_data *td, struct io_u *io_u)
{
	struct spliceio_data *sd = td->io_ops->data;
	struct iovec iov = {
		.iov_base = io_u->xfer_buf,
		.iov_len = io_u->xfer_buflen,
	};

	vmsplice(sd->pipe[0], &iov, 1, SPLICE_F_UNMAP);
}

static int fio_spliceio_queue(struct thread_data *td, struct io_u *io_u)
{
	struct spliceio_data *sd = td->io_ops->data;
	int ret;

	if (io_u->ddir == DDIR_READ) {
		if (sd->vmsplice_to_user) {
			ret = fio_splice_read(td, io_u);
			/*
			 * This kernel doesn't support vmsplice to user
			 * space. Reset the vmsplice_to_user flag, so that
			 * we retry below and don't hit this path again.
			 */
			if (ret == -EBADF)
				sd->vmsplice_to_user = 0;
		}
		if (!sd->vmsplice_to_user)
			ret = fio_splice_read_old(td, io_u);
	} else if (io_u->ddir == DDIR_WRITE)
		ret = fio_splice_write(td, io_u);
	else
		ret = fsync(io_u->file->fd);

	if (ret != (int) io_u->xfer_buflen) {
		if (ret >= 0) {
			io_u->resid = io_u->xfer_buflen - ret;
			io_u->error = 0;
			return FIO_Q_COMPLETED;
		} else
			io_u->error = errno;
	}

	if (io_u->error)
		td_verror(td, io_u->error, "xfer");

	return FIO_Q_COMPLETED;
}

static void fio_spliceio_cleanup(struct thread_data *td)
{
	struct spliceio_data *sd = td->io_ops->data;

	if (sd) {
		close(sd->pipe[0]);
		close(sd->pipe[1]);
		free(sd);
		td->io_ops->data = NULL;
	}
}

static int fio_spliceio_init(struct thread_data *td)
{
	struct spliceio_data *sd = malloc(sizeof(*sd));

	if (pipe(sd->pipe) < 0) {
		td_verror(td, errno, "pipe");
		free(sd);
		return 1;
	}

	/*
	 * Assume this work, we'll reset this if it doesn't
	 */
	sd->vmsplice_to_user = 1;

	td->io_ops->data = sd;
	return 0;
}

static struct ioengine_ops ioengine = {
	.name		= "splice",
	.version	= FIO_IOOPS_VERSION,
	.init		= fio_spliceio_init,
	.queue		= fio_spliceio_queue,
	.cleanup	= fio_spliceio_cleanup,
	.open_file	= generic_open_file,
	.close_file	= generic_close_file,
	.flags		= FIO_SYNCIO,
};

#else /* FIO_HAVE_SPLICE */

/*
 * When we have a proper configure system in place, we simply wont build
 * and install this io engine. For now install a crippled version that
 * just complains and fails to load.
 */
static int fio_spliceio_init(struct thread_data fio_unused *td)
{
	fprintf(stderr, "fio: splice not available\n");
	return 1;
}

static struct ioengine_ops ioengine = {
	.name		= "splice",
	.version	= FIO_IOOPS_VERSION,
	.init		= fio_spliceio_init,
};

#endif

static void fio_init fio_spliceio_register(void)
{
	register_ioengine(&ioengine);
}

static void fio_exit fio_spliceio_unregister(void)
{
	unregister_ioengine(&ioengine);
}
Commit	Line	Data
2866c82d	1	/*
da751ca9 JA	2	* splice engine
	3	*
	4	* IO engine that transfers data by doing splices to/from pipes and
	5	* the files.
2866c82d JA	6	*
	7	*/
	8	#include <stdio.h>
	9	#include <stdlib.h>
	10	#include <unistd.h>
	11	#include <errno.h>
	12	#include <assert.h>
	13	#include <sys/poll.h>
5f350952 JA	14
5f350952 JA	15	#include "../fio.h"
2866c82d	16
34cfcdaf JA	17	#ifdef FIO_HAVE_SPLICE
34cfcdaf JA	18
2866c82d	19	struct spliceio_data {
2866c82d	20	int pipe[2];
f24254e1	21	int vmsplice_to_user;
2866c82d JA	22	};
2866c82d JA	23
2866c82d	24	/*
f24254e1 JA	25	* vmsplice didn't use to support splicing to user space, this is the old
	26	* variant of getting that job done. Doesn't make a lot of sense, but it
	27	* uses splices to move data from the source into a pipe.
2866c82d	28	*/
f24254e1	29	static int fio_splice_read_old(struct thread_data td, struct io_u io_u)
2866c82d JA	30	{
2866c82d JA	31	struct spliceio_data *sd = td->io_ops->data;
53cdc686	32	struct fio_file *f = io_u->file;
2866c82d JA	33	int ret, ret2, buflen;
	34	off_t offset;
	35	void *p;
	36
	37	offset = io_u->offset;
cec6b55d JA	38	buflen = io_u->xfer_buflen;
cec6b55d JA	39	p = io_u->xfer_buf;
2866c82d JA	40	while (buflen) {
	41	int this_len = buflen;
	42
	43	if (this_len > SPLICE_DEF_SIZE)
	44	this_len = SPLICE_DEF_SIZE;
	45
53cdc686	46	ret = splice(f->fd, &offset, sd->pipe[1], NULL, this_len, SPLICE_F_MORE);
2866c82d JA	47	if (ret < 0) {
	48	if (errno == ENODATA \|\| errno == EAGAIN)
	49	continue;
	50
b4ba5f30	51	return -errno;
2866c82d JA	52	}
	53
	54	buflen -= ret;
	55
	56	while (ret) {
	57	ret2 = read(sd->pipe[0], p, ret);
	58	if (ret2 < 0)
b4ba5f30	59	return -errno;
2866c82d JA	60
	61	ret -= ret2;
	62	p += ret2;
	63	}
	64	}
	65
cec6b55d	66	return io_u->xfer_buflen;
2866c82d JA	67	}
2866c82d JA	68
f24254e1 JA	69	/*
	70	* We can now vmsplice into userspace, so do the transfer by splicing into
	71	* a pipe and vmsplicing that into userspace.
	72	*/
	73	static int fio_splice_read(struct thread_data td, struct io_u io_u)
	74	{
	75	struct spliceio_data *sd = td->io_ops->data;
	76	struct fio_file *f = io_u->file;
	77	struct iovec iov;
	78	int ret, buflen;
	79	off_t offset;
	80	void *p;
	81
	82	offset = io_u->offset;
	83	buflen = io_u->xfer_buflen;
	84	p = io_u->xfer_buf;
bb446c11	85	io_u->xfer_buf = NULL;
f24254e1 JA	86	while (buflen) {
	87	int this_len = buflen;
	88
	89	if (this_len > SPLICE_DEF_SIZE)
	90	this_len = SPLICE_DEF_SIZE;
	91
	92	ret = splice(f->fd, &offset, sd->pipe[1], NULL, this_len, SPLICE_F_MORE);
	93	if (ret < 0) {
	94	if (errno == ENODATA \|\| errno == EAGAIN)
	95	continue;
	96
	97	return -errno;
	98	}
	99
	100	buflen -= ret;
	101	iov.iov_base = p;
	102	iov.iov_len = ret;
	103	p += ret;
	104
	105	while (iov.iov_len) {
	106	ret = vmsplice(sd->pipe[0], &iov, 1, SPLICE_F_MOVE);
	107	if (ret < 0)
	108	return -errno;
	109	else if (!ret)
	110	return -ENODATA;
	111
bb446c11 JA	112	if (!io_u->xfer_buf)
bb446c11 JA	113	io_u->xfer_buf = iov.iov_base;
f24254e1 JA	114	iov.iov_len -= ret;
	115	iov.iov_base += ret;
	116	}
	117	}
	118
bb446c11	119	io_u->unmap = splice_unmap_io_u;
f24254e1 JA	120	return io_u->xfer_buflen;
	121	}
	122
2866c82d JA	123	/*
	124	* For splice writing, we can vmsplice our data buffer directly into a
	125	* pipe and then splice that to a file.
	126	*/
	127	static int fio_splice_write(struct thread_data td, struct io_u io_u)
	128	{
	129	struct spliceio_data *sd = td->io_ops->data;
f24254e1 JA	130	struct iovec iov = {
	131	.iov_base = io_u->xfer_buf,
	132	.iov_len = io_u->xfer_buflen,
2866c82d JA	133	};
2866c82d JA	134	struct pollfd pfd = { .fd = sd->pipe[1], .events = POLLOUT, };
53cdc686	135	struct fio_file *f = io_u->file;
2866c82d JA	136	off_t off = io_u->offset;
	137	int ret, ret2;
	138
f24254e1	139	while (iov.iov_len) {
2866c82d JA	140	if (poll(&pfd, 1, -1) < 0)
	141	return errno;
	142
f24254e1	143	ret = vmsplice(sd->pipe[1], &iov, 1, SPLICE_F_NONBLOCK);
2866c82d	144	if (ret < 0)
b4ba5f30	145	return -errno;
2866c82d	146
f24254e1 JA	147	iov.iov_len -= ret;
f24254e1 JA	148	iov.iov_base += ret;
2866c82d JA	149
2866c82d JA	150	while (ret) {
53cdc686	151	ret2 = splice(sd->pipe[0], NULL, f->fd, &off, ret, 0);
2866c82d	152	if (ret2 < 0)
b4ba5f30	153	return -errno;
2866c82d JA	154
	155	ret -= ret2;
	156	}
	157	}
	158
cec6b55d	159	return io_u->xfer_buflen;
2866c82d JA	160	}
2866c82d JA	161
bb446c11 JA	162	static void splice_unmap_io_u(struct thread_data td, struct io_u io_u)
	163	{
	164	struct spliceio_data *sd = td->io_ops->data;
	165	struct iovec iov = {
	166	.iov_base = io_u->xfer_buf,
	167	.iov_len = io_u->xfer_buflen,
	168	};
	169
	170	vmsplice(sd->pipe[0], &iov, 1, SPLICE_F_UNMAP);
	171	}
	172
2866c82d JA	173	static int fio_spliceio_queue(struct thread_data td, struct io_u io_u)
2866c82d JA	174	{
f24254e1	175	struct spliceio_data *sd = td->io_ops->data;
cec6b55d	176	int ret;
2866c82d	177
f24254e1	178	if (io_u->ddir == DDIR_READ) {
43688575	179	if (sd->vmsplice_to_user) {
f24254e1	180	ret = fio_splice_read(td, io_u);
43688575 JA	181	/*
	182	* This kernel doesn't support vmsplice to user
	183	* space. Reset the vmsplice_to_user flag, so that
	184	* we retry below and don't hit this path again.
	185	*/
	186	if (ret == -EBADF)
	187	sd->vmsplice_to_user = 0;
	188	}
	189	if (!sd->vmsplice_to_user)
f24254e1 JA	190	ret = fio_splice_read_old(td, io_u);
f24254e1 JA	191	} else if (io_u->ddir == DDIR_WRITE)
2866c82d	192	ret = fio_splice_write(td, io_u);
87dc1ab1 JA	193	else
87dc1ab1 JA	194	ret = fsync(io_u->file->fd);
2866c82d	195
cec6b55d	196	if (ret != (int) io_u->xfer_buflen) {
22819ec2	197	if (ret >= 0) {
cec6b55d JA	198	io_u->resid = io_u->xfer_buflen - ret;
cec6b55d JA	199	io_u->error = 0;
36167d82	200	return FIO_Q_COMPLETED;
2866c82d JA	201	} else
	202	io_u->error = errno;
	203	}
	204
36167d82	205	if (io_u->error)
e1161c32	206	td_verror(td, io_u->error, "xfer");
2866c82d	207
36167d82	208	return FIO_Q_COMPLETED;
2866c82d JA	209	}
	210
	211	static void fio_spliceio_cleanup(struct thread_data *td)
	212	{
	213	struct spliceio_data *sd = td->io_ops->data;
	214
	215	if (sd) {
	216	close(sd->pipe[0]);
	217	close(sd->pipe[1]);
	218	free(sd);
	219	td->io_ops->data = NULL;
	220	}
	221	}
	222
	223	static int fio_spliceio_init(struct thread_data *td)
	224	{
	225	struct spliceio_data sd = malloc(sizeof(sd));
	226
2866c82d	227	if (pipe(sd->pipe) < 0) {
e1161c32	228	td_verror(td, errno, "pipe");
2866c82d JA	229	free(sd);
	230	return 1;
	231	}
	232
f24254e1	233	/*
43688575	234	* Assume this work, we'll reset this if it doesn't
f24254e1	235	*/
43688575	236	sd->vmsplice_to_user = 1;
f24254e1	237
2866c82d JA	238	td->io_ops->data = sd;
	239	return 0;
	240	}
	241
5f350952	242	static struct ioengine_ops ioengine = {
2866c82d JA	243	.name = "splice",
	244	.version = FIO_IOOPS_VERSION,
	245	.init = fio_spliceio_init,
	246	.queue = fio_spliceio_queue,
2866c82d	247	.cleanup = fio_spliceio_cleanup,
b5af8293 JA	248	.open_file = generic_open_file,
b5af8293 JA	249	.close_file = generic_close_file,
2866c82d JA	250	.flags = FIO_SYNCIO,
2866c82d JA	251	};
34cfcdaf JA	252
	253	#else /* FIO_HAVE_SPLICE */
	254
	255	/*
	256	* When we have a proper configure system in place, we simply wont build
	257	* and install this io engine. For now install a crippled version that
	258	* just complains and fails to load.
	259	*/
	260	static int fio_spliceio_init(struct thread_data fio_unused *td)
	261	{
	262	fprintf(stderr, "fio: splice not available\n");
	263	return 1;
	264	}
	265
5f350952	266	static struct ioengine_ops ioengine = {
34cfcdaf JA	267	.name = "splice",
	268	.version = FIO_IOOPS_VERSION,
	269	.init = fio_spliceio_init,
	270	};
	271
	272	#endif
5f350952 JA	273
	274	static void fio_init fio_spliceio_register(void)
	275	{
	276	register_ioengine(&ioengine);
	277	}
	278
	279	static void fio_exit fio_spliceio_unregister(void)
	280	{
	281	unregister_ioengine(&ioengine);
	282	}