[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;
	int vmsplice_to_user_map;
};

/*
 * 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, mmap_len;
	off_t offset;
	void *p, *map;

	ret = 0;
	offset = io_u->offset;
	mmap_len = buflen = io_u->xfer_buflen;

	if (sd->vmsplice_to_user_map) {
		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;
	} else {
		map = NULL;
		p = io_u->xfer_buf;
	}

	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;

		while (iov.iov_len) {
			ret = vmsplice(sd->pipe[0], &iov, 1, SPLICE_F_MOVE);
			if (ret < 0) {
				if (errno == EFAULT &&
				    sd->vmsplice_to_user_map) {
					sd->vmsplice_to_user_map = 0;
					munmap(map, mmap_len);
					map = NULL;
					p = io_u->xfer_buf;
					iov.iov_base = p;
					continue;
				}
				if (errno == EBADF) {
					ret = -EBADF;
					break;
				}
				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;
			p += ret;
		}
		if (ret < 0)
			break;
	}

	if (sd->vmsplice_to_user_map && 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 uninitialized_var(ret);

	fio_ro_check(td, io_u);

	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");
		if (io_u->error == EINVAL)
			log_err("fio: looks like splice doesn't work on this"
					" file system\n");
	}

	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);
	}
}

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;

	/*
	 * Works with "real" vmsplice to user, eg mapping pages directly.
	 * Reset if we fail.
	 */
	sd->vmsplice_to_user_map = 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,
	.get_file_size	= generic_get_file_size,
	.flags		= FIO_SYNCIO | FIO_PIPEIO,
};

#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;
8b850243	23	int vmsplice_to_user_map;
2866c82d JA	24	};
2866c82d JA	25
2866c82d	26	/*
f24254e1 JA	27	* vmsplice didn't use to support splicing to user space, this is the old
	28	* variant of getting that job done. Doesn't make a lot of sense, but it
	29	* uses splices to move data from the source into a pipe.
2866c82d	30	*/
f24254e1	31	static int fio_splice_read_old(struct thread_data td, struct io_u io_u)
2866c82d JA	32	{
2866c82d JA	33	struct spliceio_data *sd = td->io_ops->data;
53cdc686	34	struct fio_file *f = io_u->file;
2866c82d JA	35	int ret, ret2, buflen;
	36	off_t offset;
	37	void *p;
	38
	39	offset = io_u->offset;
cec6b55d JA	40	buflen = io_u->xfer_buflen;
cec6b55d JA	41	p = io_u->xfer_buf;
2866c82d JA	42	while (buflen) {
	43	int this_len = buflen;
	44
	45	if (this_len > SPLICE_DEF_SIZE)
	46	this_len = SPLICE_DEF_SIZE;
	47
53cdc686	48	ret = splice(f->fd, &offset, sd->pipe[1], NULL, this_len, SPLICE_F_MORE);
2866c82d JA	49	if (ret < 0) {
	50	if (errno == ENODATA \|\| errno == EAGAIN)
	51	continue;
	52
b4ba5f30	53	return -errno;
2866c82d JA	54	}
	55
	56	buflen -= ret;
	57
	58	while (ret) {
	59	ret2 = read(sd->pipe[0], p, ret);
	60	if (ret2 < 0)
b4ba5f30	61	return -errno;
2866c82d JA	62
	63	ret -= ret2;
	64	p += ret2;
	65	}
	66	}
	67
cec6b55d	68	return io_u->xfer_buflen;
2866c82d JA	69	}
2866c82d JA	70
f24254e1 JA	71	/*
	72	* We can now vmsplice into userspace, so do the transfer by splicing into
	73	* a pipe and vmsplicing that into userspace.
	74	*/
	75	static int fio_splice_read(struct thread_data td, struct io_u io_u)
	76	{
	77	struct spliceio_data *sd = td->io_ops->data;
	78	struct fio_file *f = io_u->file;
	79	struct iovec iov;
8b850243	80	int ret , buflen, mmap_len;
f24254e1	81	off_t offset;
81887d5d	82	void p, map;
f24254e1	83
8b850243	84	ret = 0;
f24254e1	85	offset = io_u->offset;
81887d5d JA	86	mmap_len = buflen = io_u->xfer_buflen;
81887d5d JA	87
8b850243 JA	88	if (sd->vmsplice_to_user_map) {
	89	map = mmap(io_u->xfer_buf, buflen, PROT_READ, MAP_PRIVATE\|OS_MAP_ANON, 0, 0);
	90	if (map == MAP_FAILED) {
	91	td_verror(td, errno, "mmap io_u");
	92	return -1;
	93	}
	94
	95	p = map;
	96	} else {
	97	map = NULL;
	98	p = io_u->xfer_buf;
81887d5d JA	99	}
81887d5d JA	100
f24254e1 JA	101	while (buflen) {
f24254e1 JA	102	int this_len = buflen;
81887d5d	103	int flags = 0;
f24254e1	104
81887d5d	105	if (this_len > SPLICE_DEF_SIZE) {
f24254e1	106	this_len = SPLICE_DEF_SIZE;
81887d5d JA	107	flags = SPLICE_F_MORE;
81887d5d JA	108	}
f24254e1	109
81887d5d	110	ret = splice(f->fd, &offset, sd->pipe[1], NULL, this_len,flags);
f24254e1 JA	111	if (ret < 0) {
	112	if (errno == ENODATA \|\| errno == EAGAIN)
	113	continue;
	114
81887d5d JA	115	td_verror(td, errno, "splice-from-fd");
81887d5d JA	116	break;
f24254e1 JA	117	}
	118
	119	buflen -= ret;
	120	iov.iov_base = p;
	121	iov.iov_len = ret;
f24254e1 JA	122
	123	while (iov.iov_len) {
	124	ret = vmsplice(sd->pipe[0], &iov, 1, SPLICE_F_MOVE);
81887d5d	125	if (ret < 0) {
cd98801f JA	126	if (errno == EFAULT &&
cd98801f JA	127	sd->vmsplice_to_user_map) {
8b850243 JA	128	sd->vmsplice_to_user_map = 0;
8b850243 JA	129	munmap(map, mmap_len);
cd98801f JA	130	map = NULL;
	131	p = io_u->xfer_buf;
	132	iov.iov_base = p;
	133	continue;
8b850243	134	}
3a6d267d JA	135	if (errno == EBADF) {
	136	ret = -EBADF;
	137	break;
	138	}
81887d5d JA	139	td_verror(td, errno, "vmsplice");
	140	break;
	141	} else if (!ret) {
	142	td_verror(td, ENODATA, "vmsplice");
	143	ret = -1;
	144	break;
	145	}
f24254e1 JA	146
	147	iov.iov_len -= ret;
	148	iov.iov_base += ret;
cd98801f	149	p += ret;
f24254e1	150	}
81887d5d JA	151	if (ret < 0)
	152	break;
	153	}
	154
8b850243	155	if (sd->vmsplice_to_user_map && munmap(map, mmap_len) < 0) {
81887d5d JA	156	td_verror(td, errno, "munnap io_u");
81887d5d JA	157	return -1;
f24254e1	158	}
81887d5d JA	159	if (ret < 0)
81887d5d JA	160	return ret;
f24254e1 JA	161
	162	return io_u->xfer_buflen;
	163	}
	164
2866c82d JA	165	/*
	166	* For splice writing, we can vmsplice our data buffer directly into a
	167	* pipe and then splice that to a file.
	168	*/
	169	static int fio_splice_write(struct thread_data td, struct io_u io_u)
	170	{
	171	struct spliceio_data *sd = td->io_ops->data;
f24254e1 JA	172	struct iovec iov = {
	173	.iov_base = io_u->xfer_buf,
	174	.iov_len = io_u->xfer_buflen,
2866c82d JA	175	};
2866c82d JA	176	struct pollfd pfd = { .fd = sd->pipe[1], .events = POLLOUT, };
53cdc686	177	struct fio_file *f = io_u->file;
2866c82d JA	178	off_t off = io_u->offset;
	179	int ret, ret2;
	180
f24254e1	181	while (iov.iov_len) {
2866c82d JA	182	if (poll(&pfd, 1, -1) < 0)
	183	return errno;
	184
f24254e1	185	ret = vmsplice(sd->pipe[1], &iov, 1, SPLICE_F_NONBLOCK);
2866c82d	186	if (ret < 0)
b4ba5f30	187	return -errno;
2866c82d	188
f24254e1 JA	189	iov.iov_len -= ret;
f24254e1 JA	190	iov.iov_base += ret;
2866c82d JA	191
2866c82d JA	192	while (ret) {
53cdc686	193	ret2 = splice(sd->pipe[0], NULL, f->fd, &off, ret, 0);
2866c82d	194	if (ret2 < 0)
b4ba5f30	195	return -errno;
2866c82d JA	196
	197	ret -= ret2;
	198	}
	199	}
	200
cec6b55d	201	return io_u->xfer_buflen;
2866c82d JA	202	}
	203
	204	static int fio_spliceio_queue(struct thread_data td, struct io_u io_u)
	205	{
f24254e1	206	struct spliceio_data *sd = td->io_ops->data;
f3f552b9	207	int uninitialized_var(ret);
2866c82d	208
7101d9c2 JA	209	fio_ro_check(td, io_u);
7101d9c2 JA	210
f24254e1	211	if (io_u->ddir == DDIR_READ) {
43688575	212	if (sd->vmsplice_to_user) {
f24254e1	213	ret = fio_splice_read(td, io_u);
43688575 JA	214	/*
	215	* This kernel doesn't support vmsplice to user
	216	* space. Reset the vmsplice_to_user flag, so that
	217	* we retry below and don't hit this path again.
	218	*/
	219	if (ret == -EBADF)
	220	sd->vmsplice_to_user = 0;
	221	}
	222	if (!sd->vmsplice_to_user)
f24254e1 JA	223	ret = fio_splice_read_old(td, io_u);
f24254e1 JA	224	} else if (io_u->ddir == DDIR_WRITE)
2866c82d	225	ret = fio_splice_write(td, io_u);
87dc1ab1 JA	226	else
87dc1ab1 JA	227	ret = fsync(io_u->file->fd);
2866c82d	228
cec6b55d	229	if (ret != (int) io_u->xfer_buflen) {
22819ec2	230	if (ret >= 0) {
cec6b55d JA	231	io_u->resid = io_u->xfer_buflen - ret;
cec6b55d JA	232	io_u->error = 0;
36167d82	233	return FIO_Q_COMPLETED;
2866c82d JA	234	} else
	235	io_u->error = errno;
	236	}
	237
0aa417a9	238	if (io_u->error) {
e1161c32	239	td_verror(td, io_u->error, "xfer");
0aa417a9 JA	240	if (io_u->error == EINVAL)
	241	log_err("fio: looks like splice doesn't work on this"
	242	" file system\n");
	243	}
2866c82d	244
36167d82	245	return FIO_Q_COMPLETED;
2866c82d JA	246	}
	247
	248	static void fio_spliceio_cleanup(struct thread_data *td)
	249	{
	250	struct spliceio_data *sd = td->io_ops->data;
	251
	252	if (sd) {
	253	close(sd->pipe[0]);
	254	close(sd->pipe[1]);
	255	free(sd);
2866c82d JA	256	}
	257	}
	258
	259	static int fio_spliceio_init(struct thread_data *td)
	260	{
	261	struct spliceio_data sd = malloc(sizeof(sd));
	262
2866c82d	263	if (pipe(sd->pipe) < 0) {
e1161c32	264	td_verror(td, errno, "pipe");
2866c82d JA	265	free(sd);
	266	return 1;
	267	}
	268
f24254e1	269	/*
43688575	270	* Assume this work, we'll reset this if it doesn't
f24254e1	271	*/
43688575	272	sd->vmsplice_to_user = 1;
f24254e1	273
8b850243 JA	274	/*
	275	* Works with "real" vmsplice to user, eg mapping pages directly.
	276	* Reset if we fail.
	277	*/
	278	sd->vmsplice_to_user_map = 1;
	279
81887d5d JA	280	/*
	281	* And if vmsplice_to_user works, we definitely need aligned
	282	* buffers. Just set ->odirect to force that.
	283	*/
	284	if (td_read(td))
	285	td->o.odirect = 1;
	286
2866c82d JA	287	td->io_ops->data = sd;
	288	return 0;
	289	}
	290
5f350952	291	static struct ioengine_ops ioengine = {
2866c82d JA	292	.name = "splice",
	293	.version = FIO_IOOPS_VERSION,
	294	.init = fio_spliceio_init,
	295	.queue = fio_spliceio_queue,
2866c82d	296	.cleanup = fio_spliceio_cleanup,
b5af8293 JA	297	.open_file = generic_open_file,
b5af8293 JA	298	.close_file = generic_close_file,
df9c26b1	299	.get_file_size = generic_get_file_size,
9c0d2241	300	.flags = FIO_SYNCIO \| FIO_PIPEIO,
2866c82d	301	};
34cfcdaf JA	302
	303	#else /* FIO_HAVE_SPLICE */
	304
	305	/*
	306	* When we have a proper configure system in place, we simply wont build
	307	* and install this io engine. For now install a crippled version that
	308	* just complains and fails to load.
	309	*/
	310	static int fio_spliceio_init(struct thread_data fio_unused *td)
	311	{
	312	fprintf(stderr, "fio: splice not available\n");
	313	return 1;
	314	}
	315
5f350952	316	static struct ioengine_ops ioengine = {
34cfcdaf JA	317	.name = "splice",
	318	.version = FIO_IOOPS_VERSION,
	319	.init = fio_spliceio_init,
	320	};
	321
	322	#endif
5f350952 JA	323
	324	static void fio_init fio_spliceio_register(void)
	325	{
	326	register_ioengine(&ioengine);
	327	}
	328
	329	static void fio_exit fio_spliceio_unregister(void)
	330	{
	331	unregister_ioengine(&ioengine);
	332	}