4 * IO engine that transfers data by doing splices to/from pipes and
18 struct spliceio_data {
21 int vmsplice_to_user_map;
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.
29 static int fio_splice_read_old(struct thread_data *td, struct io_u *io_u)
31 struct spliceio_data *sd = td->io_ops_data;
32 struct fio_file *f = io_u->file;
33 int ret, ret2, buflen;
37 offset = io_u->offset;
38 buflen = io_u->xfer_buflen;
41 int this_len = buflen;
43 if (this_len > SPLICE_DEF_SIZE)
44 this_len = SPLICE_DEF_SIZE;
46 ret = splice(f->fd, &offset, sd->pipe[1], NULL, this_len, SPLICE_F_MORE);
48 if (errno == ENODATA || errno == EAGAIN)
57 ret2 = read(sd->pipe[0], p, ret);
66 return io_u->xfer_buflen;
70 * We can now vmsplice into userspace, so do the transfer by splicing into
71 * a pipe and vmsplicing that into userspace.
73 static int fio_splice_read(struct thread_data *td, struct io_u *io_u)
75 struct spliceio_data *sd = td->io_ops_data;
76 struct fio_file *f = io_u->file;
78 int ret , buflen, mmap_len;
83 offset = io_u->offset;
84 mmap_len = buflen = io_u->xfer_buflen;
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");
100 int this_len = buflen;
103 if (this_len > SPLICE_DEF_SIZE) {
104 this_len = SPLICE_DEF_SIZE;
105 flags = SPLICE_F_MORE;
108 ret = splice(f->fd, &offset, sd->pipe[1], NULL, this_len,flags);
110 if (errno == ENODATA || errno == EAGAIN)
113 td_verror(td, errno, "splice-from-fd");
121 while (iov.iov_len) {
122 ret = vmsplice(sd->pipe[0], &iov, 1, SPLICE_F_MOVE);
124 if (errno == EFAULT &&
125 sd->vmsplice_to_user_map) {
126 sd->vmsplice_to_user_map = 0;
127 munmap(map, mmap_len);
133 if (errno == EBADF) {
137 td_verror(td, errno, "vmsplice");
140 td_verror(td, ENODATA, "vmsplice");
153 if (sd->vmsplice_to_user_map && munmap(map, mmap_len) < 0) {
154 td_verror(td, errno, "munnap io_u");
160 return io_u->xfer_buflen;
164 * For splice writing, we can vmsplice our data buffer directly into a
165 * pipe and then splice that to a file.
167 static int fio_splice_write(struct thread_data *td, struct io_u *io_u)
169 struct spliceio_data *sd = td->io_ops_data;
171 .iov_base = io_u->xfer_buf,
172 .iov_len = io_u->xfer_buflen,
174 struct pollfd pfd = { .fd = sd->pipe[1], .events = POLLOUT, };
175 struct fio_file *f = io_u->file;
176 off_t off = io_u->offset;
179 while (iov.iov_len) {
180 if (poll(&pfd, 1, -1) < 0)
183 ret = vmsplice(sd->pipe[1], &iov, 1, SPLICE_F_NONBLOCK);
191 ret2 = splice(sd->pipe[0], NULL, f->fd, &off, ret, 0);
199 return io_u->xfer_buflen;
202 static int fio_spliceio_queue(struct thread_data *td, struct io_u *io_u)
204 struct spliceio_data *sd = td->io_ops_data;
207 fio_ro_check(td, io_u);
209 if (io_u->ddir == DDIR_READ) {
210 if (sd->vmsplice_to_user) {
211 ret = fio_splice_read(td, io_u);
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.
218 sd->vmsplice_to_user = 0;
220 if (!sd->vmsplice_to_user)
221 ret = fio_splice_read_old(td, io_u);
222 } else if (io_u->ddir == DDIR_WRITE)
223 ret = fio_splice_write(td, io_u);
224 else if (io_u->ddir == DDIR_TRIM)
225 ret = do_io_u_trim(td, io_u);
227 ret = do_io_u_sync(td, io_u);
229 if (ret != (int) io_u->xfer_buflen) {
231 io_u->resid = io_u->xfer_buflen - ret;
233 return FIO_Q_COMPLETED;
239 td_verror(td, io_u->error, "xfer");
240 if (io_u->error == EINVAL)
241 log_err("fio: looks like splice doesn't work on this"
245 return FIO_Q_COMPLETED;
248 static void fio_spliceio_cleanup(struct thread_data *td)
250 struct spliceio_data *sd = td->io_ops_data;
259 static int fio_spliceio_init(struct thread_data *td)
261 struct spliceio_data *sd = malloc(sizeof(*sd));
263 if (pipe(sd->pipe) < 0) {
264 td_verror(td, errno, "pipe");
270 * Assume this work, we'll reset this if it doesn't
272 sd->vmsplice_to_user = 1;
275 * Works with "real" vmsplice to user, eg mapping pages directly.
278 sd->vmsplice_to_user_map = 1;
281 * And if vmsplice_to_user works, we definitely need aligned
282 * buffers. Just set ->odirect to force that.
287 td->io_ops_data = sd;
291 static struct ioengine_ops ioengine = {
293 .version = FIO_IOOPS_VERSION,
294 .init = fio_spliceio_init,
295 .queue = fio_spliceio_queue,
296 .cleanup = fio_spliceio_cleanup,
297 .open_file = generic_open_file,
298 .close_file = generic_close_file,
299 .get_file_size = generic_get_file_size,
300 .flags = FIO_SYNCIO | FIO_PIPEIO,
303 static void fio_init fio_spliceio_register(void)
305 register_ioengine(&ioengine);
308 static void fio_exit fio_spliceio_unregister(void)
310 unregister_ioengine(&ioengine);