[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 <sys/mman.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 = 0 , buflen, mmap_len;
	off_t offset;
	void *p, *map;

	offset = io_u->offset;
	mmap_len = buflen = io_u->xfer_buflen;

	map = mmap(io_u->xfer_buf, buflen, PROT_READ, MAP_PRIVATE|OS_MAP_ANON, 0, 0);
	if (map == MAP_FAILED) {
		td_verror(td, errno, "mmap io_u");
		return -1;
	}

	p = map;
	while (buflen) {
		int this_len = buflen;
		int flags = 0;

		if (this_len > SPLICE_DEF_SIZE) {
			this_len = SPLICE_DEF_SIZE;
			flags = SPLICE_F_MORE;
		}

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

			td_verror(td, errno, "splice-from-fd");
			break;
		}

		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) {
				td_verror(td, errno, "vmsplice");
				break;
			} else if (!ret) {
				td_verror(td, ENODATA, "vmsplice");
				ret = -1;
				break;
			}

			iov.iov_len -= ret;
			iov.iov_base += ret;
		}
		if (ret < 0)
			break;
	}

	if (munmap(map, mmap_len) < 0) {
		td_verror(td, errno, "munnap io_u");
		return -1;
	}
	if (ret < 0)
		return ret;

	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 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;

	/*
	 * And if vmsplice_to_user works, we definitely need aligned
	 * buffers. Just set ->odirect to force that.
	 */
	if (td_read(td))
		td->o.odirect = 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>
81887d5d	14	#include <sys/mman.h>
5f350952 JA	15
5f350952 JA	16	#include "../fio.h"
2866c82d	17
34cfcdaf JA	18	#ifdef FIO_HAVE_SPLICE
34cfcdaf JA	19
2866c82d	20	struct spliceio_data {
2866c82d	21	int pipe[2];
f24254e1	22	int vmsplice_to_user;
2866c82d JA	23	};
2866c82d JA	24
2866c82d	25	/*
f24254e1 JA	26	* vmsplice didn't use to support splicing to user space, this is the old
	27	* variant of getting that job done. Doesn't make a lot of sense, but it
	28	* uses splices to move data from the source into a pipe.
2866c82d	29	*/
f24254e1	30	static int fio_splice_read_old(struct thread_data td, struct io_u io_u)
2866c82d JA	31	{
2866c82d JA	32	struct spliceio_data *sd = td->io_ops->data;
53cdc686	33	struct fio_file *f = io_u->file;
2866c82d JA	34	int ret, ret2, buflen;
	35	off_t offset;
	36	void *p;
	37
	38	offset = io_u->offset;
cec6b55d JA	39	buflen = io_u->xfer_buflen;
cec6b55d JA	40	p = io_u->xfer_buf;
2866c82d JA	41	while (buflen) {
	42	int this_len = buflen;
	43
	44	if (this_len > SPLICE_DEF_SIZE)
	45	this_len = SPLICE_DEF_SIZE;
	46
53cdc686	47	ret = splice(f->fd, &offset, sd->pipe[1], NULL, this_len, SPLICE_F_MORE);
2866c82d JA	48	if (ret < 0) {
	49	if (errno == ENODATA \|\| errno == EAGAIN)
	50	continue;
	51
b4ba5f30	52	return -errno;
2866c82d JA	53	}
	54
	55	buflen -= ret;
	56
	57	while (ret) {
	58	ret2 = read(sd->pipe[0], p, ret);
	59	if (ret2 < 0)
b4ba5f30	60	return -errno;
2866c82d JA	61
	62	ret -= ret2;
	63	p += ret2;
	64	}
	65	}
	66
cec6b55d	67	return io_u->xfer_buflen;
2866c82d JA	68	}
2866c82d JA	69
f24254e1 JA	70	/*
	71	* We can now vmsplice into userspace, so do the transfer by splicing into
	72	* a pipe and vmsplicing that into userspace.
	73	*/
	74	static int fio_splice_read(struct thread_data td, struct io_u io_u)
	75	{
	76	struct spliceio_data *sd = td->io_ops->data;
	77	struct fio_file *f = io_u->file;
	78	struct iovec iov;
81887d5d	79	int ret = 0 , buflen, mmap_len;
f24254e1	80	off_t offset;
81887d5d	81	void p, map;
f24254e1 JA	82
f24254e1 JA	83	offset = io_u->offset;
81887d5d JA	84	mmap_len = buflen = io_u->xfer_buflen;
	85
	86	map = mmap(io_u->xfer_buf, buflen, PROT_READ, MAP_PRIVATE\|OS_MAP_ANON, 0, 0);
	87	if (map == MAP_FAILED) {
	88	td_verror(td, errno, "mmap io_u");
	89	return -1;
	90	}
	91
	92	p = map;
f24254e1 JA	93	while (buflen) {
f24254e1 JA	94	int this_len = buflen;
81887d5d	95	int flags = 0;
f24254e1	96
81887d5d	97	if (this_len > SPLICE_DEF_SIZE) {
f24254e1	98	this_len = SPLICE_DEF_SIZE;
81887d5d JA	99	flags = SPLICE_F_MORE;
81887d5d JA	100	}
f24254e1	101
81887d5d	102	ret = splice(f->fd, &offset, sd->pipe[1], NULL, this_len,flags);
f24254e1 JA	103	if (ret < 0) {
	104	if (errno == ENODATA \|\| errno == EAGAIN)
	105	continue;
	106
81887d5d JA	107	td_verror(td, errno, "splice-from-fd");
81887d5d JA	108	break;
f24254e1 JA	109	}
	110
	111	buflen -= ret;
	112	iov.iov_base = p;
	113	iov.iov_len = ret;
	114	p += ret;
	115
	116	while (iov.iov_len) {
	117	ret = vmsplice(sd->pipe[0], &iov, 1, SPLICE_F_MOVE);
81887d5d JA	118	if (ret < 0) {
	119	td_verror(td, errno, "vmsplice");
	120	break;
	121	} else if (!ret) {
	122	td_verror(td, ENODATA, "vmsplice");
	123	ret = -1;
	124	break;
	125	}
f24254e1 JA	126
	127	iov.iov_len -= ret;
	128	iov.iov_base += ret;
	129	}
81887d5d JA	130	if (ret < 0)
	131	break;
	132	}
	133
	134	if (munmap(map, mmap_len) < 0) {
	135	td_verror(td, errno, "munnap io_u");
	136	return -1;
f24254e1	137	}
81887d5d JA	138	if (ret < 0)
81887d5d JA	139	return ret;
f24254e1 JA	140
	141	return io_u->xfer_buflen;
	142	}
	143
2866c82d JA	144	/*
	145	* For splice writing, we can vmsplice our data buffer directly into a
	146	* pipe and then splice that to a file.
	147	*/
	148	static int fio_splice_write(struct thread_data td, struct io_u io_u)
	149	{
	150	struct spliceio_data *sd = td->io_ops->data;
f24254e1 JA	151	struct iovec iov = {
	152	.iov_base = io_u->xfer_buf,
	153	.iov_len = io_u->xfer_buflen,
2866c82d JA	154	};
2866c82d JA	155	struct pollfd pfd = { .fd = sd->pipe[1], .events = POLLOUT, };
53cdc686	156	struct fio_file *f = io_u->file;
2866c82d JA	157	off_t off = io_u->offset;
	158	int ret, ret2;
	159
f24254e1	160	while (iov.iov_len) {
2866c82d JA	161	if (poll(&pfd, 1, -1) < 0)
	162	return errno;
	163
f24254e1	164	ret = vmsplice(sd->pipe[1], &iov, 1, SPLICE_F_NONBLOCK);
2866c82d	165	if (ret < 0)
b4ba5f30	166	return -errno;
2866c82d	167
f24254e1 JA	168	iov.iov_len -= ret;
f24254e1 JA	169	iov.iov_base += ret;
2866c82d JA	170
2866c82d JA	171	while (ret) {
53cdc686	172	ret2 = splice(sd->pipe[0], NULL, f->fd, &off, ret, 0);
2866c82d	173	if (ret2 < 0)
b4ba5f30	174	return -errno;
2866c82d JA	175
	176	ret -= ret2;
	177	}
	178	}
	179
cec6b55d	180	return io_u->xfer_buflen;
2866c82d JA	181	}
	182
	183	static int fio_spliceio_queue(struct thread_data td, struct io_u io_u)
	184	{
f24254e1	185	struct spliceio_data *sd = td->io_ops->data;
cec6b55d	186	int ret;
2866c82d	187
f24254e1	188	if (io_u->ddir == DDIR_READ) {
43688575	189	if (sd->vmsplice_to_user) {
f24254e1	190	ret = fio_splice_read(td, io_u);
43688575 JA	191	/*
	192	* This kernel doesn't support vmsplice to user
	193	* space. Reset the vmsplice_to_user flag, so that
	194	* we retry below and don't hit this path again.
	195	*/
	196	if (ret == -EBADF)
	197	sd->vmsplice_to_user = 0;
	198	}
	199	if (!sd->vmsplice_to_user)
f24254e1 JA	200	ret = fio_splice_read_old(td, io_u);
f24254e1 JA	201	} else if (io_u->ddir == DDIR_WRITE)
2866c82d	202	ret = fio_splice_write(td, io_u);
87dc1ab1 JA	203	else
87dc1ab1 JA	204	ret = fsync(io_u->file->fd);
2866c82d	205
cec6b55d	206	if (ret != (int) io_u->xfer_buflen) {
22819ec2	207	if (ret >= 0) {
cec6b55d JA	208	io_u->resid = io_u->xfer_buflen - ret;
cec6b55d JA	209	io_u->error = 0;
36167d82	210	return FIO_Q_COMPLETED;
2866c82d JA	211	} else
	212	io_u->error = errno;
	213	}
	214
36167d82	215	if (io_u->error)
e1161c32	216	td_verror(td, io_u->error, "xfer");
2866c82d	217
36167d82	218	return FIO_Q_COMPLETED;
2866c82d JA	219	}
	220
	221	static void fio_spliceio_cleanup(struct thread_data *td)
	222	{
	223	struct spliceio_data *sd = td->io_ops->data;
	224
	225	if (sd) {
	226	close(sd->pipe[0]);
	227	close(sd->pipe[1]);
	228	free(sd);
	229	td->io_ops->data = NULL;
	230	}
	231	}
	232
	233	static int fio_spliceio_init(struct thread_data *td)
	234	{
	235	struct spliceio_data sd = malloc(sizeof(sd));
	236
2866c82d	237	if (pipe(sd->pipe) < 0) {
e1161c32	238	td_verror(td, errno, "pipe");
2866c82d JA	239	free(sd);
	240	return 1;
	241	}
	242
f24254e1	243	/*
43688575	244	* Assume this work, we'll reset this if it doesn't
f24254e1	245	*/
43688575	246	sd->vmsplice_to_user = 1;
f24254e1	247
81887d5d JA	248	/*
	249	* And if vmsplice_to_user works, we definitely need aligned
	250	* buffers. Just set ->odirect to force that.
	251	*/
	252	if (td_read(td))
	253	td->o.odirect = 1;
	254
2866c82d JA	255	td->io_ops->data = sd;
	256	return 0;
	257	}
	258
5f350952	259	static struct ioengine_ops ioengine = {
2866c82d JA	260	.name = "splice",
	261	.version = FIO_IOOPS_VERSION,
	262	.init = fio_spliceio_init,
	263	.queue = fio_spliceio_queue,
2866c82d	264	.cleanup = fio_spliceio_cleanup,
b5af8293 JA	265	.open_file = generic_open_file,
b5af8293 JA	266	.close_file = generic_close_file,
2866c82d JA	267	.flags = FIO_SYNCIO,
2866c82d JA	268	};
34cfcdaf JA	269
	270	#else /* FIO_HAVE_SPLICE */
	271
	272	/*
	273	* When we have a proper configure system in place, we simply wont build
	274	* and install this io engine. For now install a crippled version that
	275	* just complains and fails to load.
	276	*/
	277	static int fio_spliceio_init(struct thread_data fio_unused *td)
	278	{
	279	fprintf(stderr, "fio: splice not available\n");
	280	return 1;
	281	}
	282
5f350952	283	static struct ioengine_ops ioengine = {
34cfcdaf JA	284	.name = "splice",
	285	.version = FIO_IOOPS_VERSION,
	286	.init = fio_spliceio_init,
	287	};
	288
	289	#endif
5f350952 JA	290
	291	static void fio_init fio_spliceio_register(void)
	292	{
	293	register_ioengine(&ioengine);
	294	}
	295
	296	static void fio_exit fio_spliceio_unregister(void)
	297	{
	298	unregister_ioengine(&ioengine);
	299	}