— splice two sockets for
zero-copy data transfer
socket *so, int fd,
struct timeval *tv);
socket *so, int
is used to splice together a source and a drain socket. The source socket is
passed as the so argument; the file descriptor of the
drain is passed in fd. If fd is
negative, an existing splicing gets dissolved. If max
is positive, at most that many bytes will get transferred. If
tv is not NULL, a
timeout(9) is scheduled to dissolve splicing in the case when no data
can be transferred for the specified period of time. Socket splicing can be
invoked from userland via the
setsockopt(2) system-call at the
SOL_SOCKET level with the socket option
Before connecting both sockets, several checks are executed. See the ERRORS section for possible failures. The connection between both sockets is implemented by setting these additional fields in struct socket:
- struct socket *so_splice links from the source to the drain socket.
- struct socket *so_spliceback links back from the drain to the source socket.
- off_t so_splicelen counts the number of bytes spliced so far from this socket.
- off_t so_splicemax specifies the maximum number of bytes to splice from this socket if non-zero.
- struct timeval so_idletv specifies the maximum idle time if non-zero.
- struct timeout so_idleto provides storage for the kernel timeout if idle time is used.
After connecting both sockets,
somove() to transfer the mbufs already in the
source receive buffer to the drain send buffer. Finally the socket buffer
SB_SPLICE is set on both socket buffers, to
indicate that the protocol layer has to call
somove() whenever data or space is available.
transfers data from the source's receive buffer to the drain's send buffer.
It must be called at
splsoftnet(9) and so must be a spliced
source socket. It may be necessary to split an mbuf to handle out-of-band
data inline or when the maximum splice length has been reached. If
mbufs will always succeed. For
out-of-band property might get lost or a short splice might happen. In the
latter case, less than the given maximum number of bytes are transferred and
userland has to cope with this. Note that a short splice cannot happen if
somove() was called by
sosplice(). So a second
setsockopt(2) after a short splice pointing to the same
maximum will always succeed.
Before transferring data,
checks both sockets for errors and that the drain socket is connected. If
the drain cannot send anymore, an
EPIPE error is set
on the source socket. The data length to move is limited by the optional
maximum splice length and the space in the drain's send socket buffer. Up to
this amount of data is taken out of the source's receive socket buffer. To
avoid splicing loops created by userland, the number of times an mbuf may be
moved between sockets is limited to 128.
For atomic protocols, either one complete packet is taken out, or
nothing is taken at all if: the packet is bigger than the drain's send
buffer size, in which case the splicing gets aborted with an
EMSGSIZE error; the packet does not fit into the
drain's current send buffer space, in which case it is left in the source's
receive buffer for later processing; or the maximum splice length is located
within a packet, in which case splicing gets dissolved like a short splice.
All address or control mbufs associated with the taken packet are
If the maximum splice length has been reached, an mbuf may get
split for non-atomic protocols. Otherwise an mbuf is either moved completely
to the send buffer or left in the receive buffer for later processing. If
SO_OOBINLINE is set, out-of-band data will get moved as such although this
might not be reliable. The data is sent out to the drain socket via the
protocol function. If that fails and the drain socket cannot send anymore,
EPIPE error is set on the source socket.
For packet oriented protocols
iterates over the next packet queue.
If a maximum splice length was specified and at least
this amount of data has been received from the drain socket, splicing gets
dissolved. In this case, an
EFBIG error is set on
the source socket if the maximum amount of data has been transferred.
Userland can process this error to distinguish the full splice from a short
splice or to react to the completed maximum splice immediately. If an idle
timeout was specified and no data has been transferred for that period of
time, the handler
dissolves splicing and sets an
ETIMEDOUT error on
the source socket.
is called to dissolve the socket splicing if the source socket cannot
receive anymore and its receive buffer is empty; or if the drain socket
cannot send anymore; or if the maximum has been reached; or if an error
occurred; or if the idle timeout has fired.
If the socket buffer flag
SB_SPLICE is set, the functions
somove() to trigger the transfer when new
data or buffer space is available. While socket splicing is active, any
read(2) from the source socket will block and the wakeup will not be
delivered to the file descriptor. A read event or a socket error is signaled
to userland after dissolving.
sosplice() returns 0 on success and
otherwise the error number.
somove() returns 0 if
socket splicing has been finished and 1 if it continues.
sosplice() will succeed unless:
- The given file descriptor fd is not an active descriptor.
- The source or the drain socket is already spliced.
- The given maximum value max is negative.
- The source socket requires a connection and is neither connected nor in the process of connecting to a peer.
- The drain socket is neither connected nor in the process of connecting to a peer.
- The given file descriptor fd is not a socket.
- The source or the drain socket is a listen socket.
- The source socket's protocol layer does not have the
PR_SPLICEflag set. Only TCP and UDP socket splicing is supported.
- The drain socket's protocol does not have the same pr_usrreq function as the source.
- The source socket is non-blocking and the receive buffer is already locked.
setsockopt(2), options(4), timeout(9)
Socket splicing for TCP first appeared in OpenBSD 4.9; support for UDP was added in OpenBSD 5.3.
The idea for socket splicing originally came from Markus Friedl <email@example.com>, and Alexander Bluhm <firstname.lastname@example.org> implemented it. Mike Belopuhov <email@example.com> added the timeout feature.