On Tue, 23 Sep 1997, der Mouse wrote:

: > - rewriting sys_open() into two segments that does a vfs lookup to a
: >   locked vnode beforehand in the nameidata struct, then calls
: >   vn_open(), which has been modified to check for the presence of an
: >   already locked vnode in the nameidata struct... but skips calling
: >   vn_open() and opens a socket instead for a vnode of type VSOCK

Yecch.  This doesn't work either, really.  If a vnode is in nameidata, it's
assumed a namei lookup is in progress and it's aborted, so that structure
element is unusable as a "trapdoor".

Lessee, another flag bit?  Too much expense to use on this application.

Short of duplicating a modified vn_open() inside of sys_open(), this is much
more painful than I thought.  Ick!  The only alternative I can see is this:

- config option to turn on openable AF_LOCAL/AF_UNIX sockets and compile in 
the required couple lines in vn_open() and the uipc_vnops.c file;

- a set of vnode v_ops in a uipc_vnops.c that translate vnode operations
into socket operations.  The socket would then, after opening, look like a
file instead of a socket, which could be a plus.  Even internal vnode
operations in the kernel would then work on them, and they could refer to
the parent directory and name of the actual listening socket;

- possibly have sys_open() detect the special case of a VSOCK returned by
vn_open(), and drop the VNODE while placing the socket directly into the
file descriptor.  I'm leaning away from this because of the advantage of
making the socket look like a file if open()ed;

- possibly in the future to implement pseudo seek/trunc/etc. that would be
"controlled" by the stream's creator.

Stoke the brainstorm fire some more if desired.  <grin>

=====
== Todd Vierling (Personal tv@pobox.com; Business tv@lucent.com)
== I know you like the Internet, Bobby.  Now go eat your Frosted Flakes.