[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;
	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 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;
	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	{
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;
81887d5d	80	void p, map;
f24254e1	81
8b850243	82	ret = 0;
f24254e1	83	offset = io_u->offset;
81887d5d JA	84	mmap_len = buflen = io_u->xfer_buflen;
81887d5d JA	85
8b850243 JA	86	if (sd->vmsplice_to_user_map) {
	87	map = mmap(io_u->xfer_buf, buflen, PROT_READ, MAP_PRIVATE\|OS_MAP_ANON, 0, 0);
	88	if (map == MAP_FAILED) {
	89	td_verror(td, errno, "mmap io_u");
	90	return -1;
	91	}
	92
	93	p = map;
	94	} else {
	95	map = NULL;
	96	p = io_u->xfer_buf;
81887d5d JA	97	}
81887d5d JA	98
f24254e1 JA	99	while (buflen) {
f24254e1 JA	100	int this_len = buflen;
81887d5d	101	int flags = 0;
f24254e1	102
81887d5d	103	if (this_len > SPLICE_DEF_SIZE) {
f24254e1	104	this_len = SPLICE_DEF_SIZE;
81887d5d JA	105	flags = SPLICE_F_MORE;
81887d5d JA	106	}
f24254e1	107
81887d5d	108	ret = splice(f->fd, &offset, sd->pipe[1], NULL, this_len,flags);
f24254e1 JA	109	if (ret < 0) {
	110	if (errno == ENODATA \|\| errno == EAGAIN)
	111	continue;
	112
81887d5d JA	113	td_verror(td, errno, "splice-from-fd");
81887d5d JA	114	break;
f24254e1 JA	115	}
	116
	117	buflen -= ret;
	118	iov.iov_base = p;
	119	iov.iov_len = ret;
f24254e1 JA	120
	121	while (iov.iov_len) {
	122	ret = vmsplice(sd->pipe[0], &iov, 1, SPLICE_F_MOVE);
81887d5d	123	if (ret < 0) {
cd98801f JA	124	if (errno == EFAULT &&
cd98801f JA	125	sd->vmsplice_to_user_map) {
8b850243 JA	126	sd->vmsplice_to_user_map = 0;
8b850243 JA	127	munmap(map, mmap_len);
cd98801f JA	128	map = NULL;
	129	p = io_u->xfer_buf;
	130	iov.iov_base = p;
	131	continue;
8b850243	132	}
3a6d267d JA	133	if (errno == EBADF) {
	134	ret = -EBADF;
	135	break;
	136	}
81887d5d JA	137	td_verror(td, errno, "vmsplice");
	138	break;
	139	} else if (!ret) {
	140	td_verror(td, ENODATA, "vmsplice");
	141	ret = -1;
	142	break;
	143	}
f24254e1 JA	144
	145	iov.iov_len -= ret;
	146	iov.iov_base += ret;
cd98801f	147	p += ret;
f24254e1	148	}
81887d5d JA	149	if (ret < 0)
	150	break;
	151	}
	152
8b850243	153	if (sd->vmsplice_to_user_map && munmap(map, mmap_len) < 0) {
81887d5d JA	154	td_verror(td, errno, "munnap io_u");
81887d5d JA	155	return -1;
f24254e1	156	}
81887d5d JA	157	if (ret < 0)
81887d5d JA	158	return ret;
f24254e1 JA	159
	160	return io_u->xfer_buflen;
	161	}
	162
2866c82d JA	163	/*
	164	* For splice writing, we can vmsplice our data buffer directly into a
	165	* pipe and then splice that to a file.
	166	*/
	167	static int fio_splice_write(struct thread_data td, struct io_u io_u)
	168	{
565e784d	169	struct spliceio_data *sd = td->io_ops_data;
f24254e1 JA	170	struct iovec iov = {
	171	.iov_base = io_u->xfer_buf,
	172	.iov_len = io_u->xfer_buflen,
2866c82d JA	173	};
2866c82d JA	174	struct pollfd pfd = { .fd = sd->pipe[1], .events = POLLOUT, };
53cdc686	175	struct fio_file *f = io_u->file;
2866c82d JA	176	off_t off = io_u->offset;
	177	int ret, ret2;
	178
f24254e1	179	while (iov.iov_len) {
2866c82d JA	180	if (poll(&pfd, 1, -1) < 0)
	181	return errno;
	182
f24254e1	183	ret = vmsplice(sd->pipe[1], &iov, 1, SPLICE_F_NONBLOCK);
2866c82d	184	if (ret < 0)
b4ba5f30	185	return -errno;
2866c82d	186
f24254e1 JA	187	iov.iov_len -= ret;
f24254e1 JA	188	iov.iov_base += ret;
2866c82d JA	189
2866c82d JA	190	while (ret) {
53cdc686	191	ret2 = splice(sd->pipe[0], NULL, f->fd, &off, ret, 0);
2866c82d	192	if (ret2 < 0)
b4ba5f30	193	return -errno;
2866c82d JA	194
	195	ret -= ret2;
	196	}
	197	}
	198
cec6b55d	199	return io_u->xfer_buflen;
2866c82d JA	200	}
	201
	202	static int fio_spliceio_queue(struct thread_data td, struct io_u io_u)
	203	{
565e784d	204	struct spliceio_data *sd = td->io_ops_data;
24d23ca7	205	int ret = 0;
2866c82d	206
7101d9c2 JA	207	fio_ro_check(td, io_u);
7101d9c2 JA	208
f24254e1	209	if (io_u->ddir == DDIR_READ) {
43688575	210	if (sd->vmsplice_to_user) {
f24254e1	211	ret = fio_splice_read(td, io_u);
43688575 JA	212	/*
	213	* This kernel doesn't support vmsplice to user
	214	* space. Reset the vmsplice_to_user flag, so that
	215	* we retry below and don't hit this path again.
	216	*/
	217	if (ret == -EBADF)
	218	sd->vmsplice_to_user = 0;
	219	}
	220	if (!sd->vmsplice_to_user)
f24254e1 JA	221	ret = fio_splice_read_old(td, io_u);
f24254e1 JA	222	} else if (io_u->ddir == DDIR_WRITE)
2866c82d	223	ret = fio_splice_write(td, io_u);
a5f3027c JA	224	else if (io_u->ddir == DDIR_TRIM)
a5f3027c JA	225	ret = do_io_u_trim(td, io_u);
87dc1ab1	226	else
f011531e	227	ret = do_io_u_sync(td, io_u);
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	{
565e784d	250	struct spliceio_data *sd = td->io_ops_data;
2866c82d JA	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))
76a3179e	285	td->o.mem_align = 1;
81887d5d	286
565e784d	287	td->io_ops_data = sd;
2866c82d JA	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	302
5f350952 JA	303	static void fio_init fio_spliceio_register(void)
	304	{
	305	register_ioengine(&ioengine);
	306	}
	307
	308	static void fio_exit fio_spliceio_unregister(void)
	309	{
	310	unregister_ioengine(&ioengine);
	311	}