On Sat, Sep 03, 2022 at 03:56:38AM +0300, Vitaliy Makkoveev wrote:
> Or may be better to check `pr_type' with the ability of splicing?
I don't think that makes sense. Reasonable splicing does not depend
on the type of socket, not on the protocol family, but on the
protocol. That is the reason I used the usrreq pointer before. As
it involves sending packets, I suggested pru_send now.
You can splice tcp to tcp, and udp to udp.
You cannot splice udp to rip although both use pr_type SOCK_RAW.
I am not convinced that introducing sepecial IPv6 functions pointers
was a good idea. That means a lot of code duplication.
Whatever. With the current feature set your diff is equivalent to
any other solution discussed. I have not implemented splicing for
rip or rip6 yet.
Regress is broken, fix it somehow and quickly. OK bluhm@
> Index: sys/kern/uipc_socket.c
> ===================================================================
> RCS file: /cvs/src/sys/kern/uipc_socket.c,v
> retrieving revision 1.286
> diff -u -p -r1.286 uipc_socket.c
> --- sys/kern/uipc_socket.c 28 Aug 2022 18:43:12 -0000 1.286
> +++ sys/kern/uipc_socket.c 3 Sep 2022 00:53:39 -0000
> @@ -1293,7 +1293,8 @@ sosplice(struct socket *so, int fd, off_
> if ((error = getsock(curproc, fd, &fp)) != 0)
> return (error);
> sosp = fp->f_data;
> - if (sosp->so_proto->pr_usrreqs != so->so_proto->pr_usrreqs) {
> + if ((sosp->so_proto->pr_flags & PR_SPLICE) == 0 ||
> + (sosp->so_proto->pr_type != so->so_proto->pr_type)) {
> error = EPROTONOSUPPORT;
> goto frele;
> }
