#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;
};
/*
- * For splice reading, we unfortunately cannot (yet) vmsplice the other way.
- * So just splice the data from the file into the pipe, and use regular
- * read to fill the buffer. Doesn't make a lot of sense, but...
+ * 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(struct thread_data *td, struct io_u *io_u)
+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;
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;
+
+restart:
+ 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;
+ p += 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);
+ goto restart;
+ }
+ 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;
+ }
+ 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[1] = {
- {
- .iov_base = io_u->xfer_buf,
- .iov_len = io_u->xfer_buflen,
- }
+ 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[0].iov_len) {
+ while (iov.iov_len) {
if (poll(&pfd, 1, -1) < 0)
return errno;
- ret = vmsplice(sd->pipe[1], iov, 1, SPLICE_F_NONBLOCK);
+ ret = vmsplice(sd->pipe[1], &iov, 1, SPLICE_F_NONBLOCK);
if (ret < 0)
return -errno;
- iov[0].iov_len -= ret;
- iov[0].iov_base += ret;
+ iov.iov_len -= ret;
+ iov.iov_base += ret;
while (ret) {
ret2 = splice(sd->pipe[0], NULL, f->fd, &off, ret, 0);
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)
- ret = fio_splice_read(td, io_u);
- else if (io_u->ddir == DDIR_WRITE)
+ 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);
io_u->error = errno;
}
- if (io_u->error)
+ 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;
}
close(sd->pipe[0]);
close(sd->pipe[1]);
free(sd);
- td->io_ops->data = NULL;
}
}
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;
}
projects / fio.git / blobdiff