[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"

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;
	char *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 *map;
	char *p;

	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 ret = 0;

	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 if (io_u->ddir == DDIR_TRIM)
		ret = do_io_u_trim(td, io_u);
	else
		ret = do_io_u_sync(td, io_u);

	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.mem_align = 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,
};

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 JA	17
2866c82d JA	18	struct spliceio_data {
2866c82d	19	int pipe[2];
f24254e1	20	int vmsplice_to_user;
8b850243	21	int vmsplice_to_user_map;
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	30	{
565e784d	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;
2866c82d JA	34	off_t offset;
c049b36e	35	char *p;
2866c82d JA	36
2866c82d JA	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	{
565e784d	75	struct spliceio_data *sd = td->io_ops_data;
f24254e1 JA	76	struct fio_file *f = io_u->file;
f24254e1 JA	77	struct iovec iov;
8b850243	78	int ret , buflen, mmap_len;
f24254e1	79	off_t offset;
c049b36e TK	80	void *map;
c049b36e TK	81	char *p;
f24254e1	82
8b850243	83	ret = 0;
f24254e1	84	offset = io_u->offset;
81887d5d JA	85	mmap_len = buflen = io_u->xfer_buflen;
81887d5d JA	86
8b850243 JA	87	if (sd->vmsplice_to_user_map) {
	88	map = mmap(io_u->xfer_buf, buflen, PROT_READ, MAP_PRIVATE\|OS_MAP_ANON, 0, 0);
	89	if (map == MAP_FAILED) {
	90	td_verror(td, errno, "mmap io_u");
	91	return -1;
	92	}
	93
	94	p = map;
	95	} else {
	96	map = NULL;
	97	p = io_u->xfer_buf;
81887d5d JA	98	}
81887d5d JA	99
f24254e1 JA	100	while (buflen) {
f24254e1 JA	101	int this_len = buflen;
81887d5d	102	int flags = 0;
f24254e1	103
81887d5d	104	if (this_len > SPLICE_DEF_SIZE) {
f24254e1	105	this_len = SPLICE_DEF_SIZE;
81887d5d JA	106	flags = SPLICE_F_MORE;
81887d5d JA	107	}
f24254e1	108
81887d5d	109	ret = splice(f->fd, &offset, sd->pipe[1], NULL, this_len,flags);
f24254e1 JA	110	if (ret < 0) {
	111	if (errno == ENODATA \|\| errno == EAGAIN)
	112	continue;
	113
81887d5d JA	114	td_verror(td, errno, "splice-from-fd");
81887d5d JA	115	break;
f24254e1 JA	116	}
	117
	118	buflen -= ret;
	119	iov.iov_base = p;
	120	iov.iov_len = ret;
f24254e1 JA	121
	122	while (iov.iov_len) {
	123	ret = vmsplice(sd->pipe[0], &iov, 1, SPLICE_F_MOVE);
81887d5d	124	if (ret < 0) {
cd98801f JA	125	if (errno == EFAULT &&
cd98801f JA	126	sd->vmsplice_to_user_map) {
8b850243 JA	127	sd->vmsplice_to_user_map = 0;
8b850243 JA	128	munmap(map, mmap_len);
cd98801f JA	129	map = NULL;
	130	p = io_u->xfer_buf;
	131	iov.iov_base = p;
	132	continue;
8b850243	133	}
3a6d267d JA	134	if (errno == EBADF) {
	135	ret = -EBADF;
	136	break;
	137	}
81887d5d JA	138	td_verror(td, errno, "vmsplice");
	139	break;
	140	} else if (!ret) {
	141	td_verror(td, ENODATA, "vmsplice");
	142	ret = -1;
	143	break;
	144	}
f24254e1 JA	145
	146	iov.iov_len -= ret;
	147	iov.iov_base += ret;
cd98801f	148	p += ret;
f24254e1	149	}
81887d5d JA	150	if (ret < 0)
	151	break;
	152	}
	153
8b850243	154	if (sd->vmsplice_to_user_map && munmap(map, mmap_len) < 0) {
81887d5d JA	155	td_verror(td, errno, "munnap io_u");
81887d5d JA	156	return -1;
f24254e1	157	}
81887d5d JA	158	if (ret < 0)
81887d5d JA	159	return ret;
f24254e1 JA	160
	161	return io_u->xfer_buflen;
	162	}
	163
2866c82d JA	164	/*
	165	* For splice writing, we can vmsplice our data buffer directly into a
	166	* pipe and then splice that to a file.
	167	*/
	168	static int fio_splice_write(struct thread_data td, struct io_u io_u)
	169	{
565e784d	170	struct spliceio_data *sd = td->io_ops_data;
f24254e1 JA	171	struct iovec iov = {
	172	.iov_base = io_u->xfer_buf,
	173	.iov_len = io_u->xfer_buflen,
2866c82d JA	174	};
2866c82d JA	175	struct pollfd pfd = { .fd = sd->pipe[1], .events = POLLOUT, };
53cdc686	176	struct fio_file *f = io_u->file;
2866c82d JA	177	off_t off = io_u->offset;
	178	int ret, ret2;
	179
f24254e1	180	while (iov.iov_len) {
2866c82d JA	181	if (poll(&pfd, 1, -1) < 0)
	182	return errno;
	183
f24254e1	184	ret = vmsplice(sd->pipe[1], &iov, 1, SPLICE_F_NONBLOCK);
2866c82d	185	if (ret < 0)
b4ba5f30	186	return -errno;
2866c82d	187
f24254e1 JA	188	iov.iov_len -= ret;
f24254e1 JA	189	iov.iov_base += ret;
2866c82d JA	190
2866c82d JA	191	while (ret) {
53cdc686	192	ret2 = splice(sd->pipe[0], NULL, f->fd, &off, ret, 0);
2866c82d	193	if (ret2 < 0)
b4ba5f30	194	return -errno;
2866c82d JA	195
	196	ret -= ret2;
	197	}
	198	}
	199
cec6b55d	200	return io_u->xfer_buflen;
2866c82d JA	201	}
	202
	203	static int fio_spliceio_queue(struct thread_data td, struct io_u io_u)
	204	{
565e784d	205	struct spliceio_data *sd = td->io_ops_data;
24d23ca7	206	int ret = 0;
2866c82d	207
7101d9c2 JA	208	fio_ro_check(td, io_u);
7101d9c2 JA	209
f24254e1	210	if (io_u->ddir == DDIR_READ) {
43688575	211	if (sd->vmsplice_to_user) {
f24254e1	212	ret = fio_splice_read(td, io_u);
43688575 JA	213	/*
	214	* This kernel doesn't support vmsplice to user
	215	* space. Reset the vmsplice_to_user flag, so that
	216	* we retry below and don't hit this path again.
	217	*/
	218	if (ret == -EBADF)
	219	sd->vmsplice_to_user = 0;
	220	}
	221	if (!sd->vmsplice_to_user)
f24254e1 JA	222	ret = fio_splice_read_old(td, io_u);
f24254e1 JA	223	} else if (io_u->ddir == DDIR_WRITE)
2866c82d	224	ret = fio_splice_write(td, io_u);
a5f3027c JA	225	else if (io_u->ddir == DDIR_TRIM)
a5f3027c JA	226	ret = do_io_u_trim(td, io_u);
87dc1ab1	227	else
f011531e	228	ret = do_io_u_sync(td, io_u);
2866c82d	229
cec6b55d	230	if (ret != (int) io_u->xfer_buflen) {
22819ec2	231	if (ret >= 0) {
cec6b55d JA	232	io_u->resid = io_u->xfer_buflen - ret;
cec6b55d JA	233	io_u->error = 0;
36167d82	234	return FIO_Q_COMPLETED;
2866c82d JA	235	} else
	236	io_u->error = errno;
	237	}
	238
0aa417a9	239	if (io_u->error) {
e1161c32	240	td_verror(td, io_u->error, "xfer");
0aa417a9 JA	241	if (io_u->error == EINVAL)
	242	log_err("fio: looks like splice doesn't work on this"
	243	" file system\n");
	244	}
2866c82d	245
36167d82	246	return FIO_Q_COMPLETED;
2866c82d JA	247	}
	248
	249	static void fio_spliceio_cleanup(struct thread_data *td)
	250	{
565e784d	251	struct spliceio_data *sd = td->io_ops_data;
2866c82d JA	252
	253	if (sd) {
	254	close(sd->pipe[0]);
	255	close(sd->pipe[1]);
	256	free(sd);
2866c82d JA	257	}
	258	}
	259
	260	static int fio_spliceio_init(struct thread_data *td)
	261	{
	262	struct spliceio_data sd = malloc(sizeof(sd));
	263
2866c82d	264	if (pipe(sd->pipe) < 0) {
e1161c32	265	td_verror(td, errno, "pipe");
2866c82d JA	266	free(sd);
	267	return 1;
	268	}
	269
f24254e1	270	/*
43688575	271	* Assume this work, we'll reset this if it doesn't
f24254e1	272	*/
43688575	273	sd->vmsplice_to_user = 1;
f24254e1	274
8b850243 JA	275	/*
	276	* Works with "real" vmsplice to user, eg mapping pages directly.
	277	* Reset if we fail.
	278	*/
	279	sd->vmsplice_to_user_map = 1;
	280
81887d5d JA	281	/*
	282	* And if vmsplice_to_user works, we definitely need aligned
	283	* buffers. Just set ->odirect to force that.
	284	*/
	285	if (td_read(td))
76a3179e	286	td->o.mem_align = 1;
81887d5d	287
565e784d	288	td->io_ops_data = sd;
2866c82d JA	289	return 0;
	290	}
	291
5f350952	292	static struct ioengine_ops ioengine = {
2866c82d JA	293	.name = "splice",
	294	.version = FIO_IOOPS_VERSION,
	295	.init = fio_spliceio_init,
	296	.queue = fio_spliceio_queue,
2866c82d	297	.cleanup = fio_spliceio_cleanup,
b5af8293 JA	298	.open_file = generic_open_file,
b5af8293 JA	299	.close_file = generic_close_file,
df9c26b1	300	.get_file_size = generic_get_file_size,
9c0d2241	301	.flags = FIO_SYNCIO \| FIO_PIPEIO,
2866c82d	302	};
34cfcdaf	303
5f350952 JA	304	static void fio_init fio_spliceio_register(void)
	305	{
	306	register_ioengine(&ioengine);
	307	}
	308
	309	static void fio_exit fio_spliceio_unregister(void)
	310	{
	311	unregister_ioengine(&ioengine);
	312	}