Re: [Qemu-devel] [PATCH] RFC: net/socket: learn to talk with a unix dgram socket

[Jason Wang]

On 2018/11/14 下午9:01, Marc-André Lureau wrote:
Hi

On Wed, Nov 14, 2018 at 7:46 AM Jason Wang <jasow...@redhat.com> wrote:

On 2018/11/10 上午3:56, Marc-André Lureau wrote:
-net socket has a fd argument, and may be passed pre-opened sockets.

TCP sockets use framing.
UDP sockets have datagram boundaries.

When given a unix dgram socket, it will be able to read from it, but
will attempt to send on the dgram_dst, which is unset. The other end
will not receive the data.

Let's teach -net socket to recognize a UNIX DGRAM socket, and use the
regular send() command (without dgram_dst).

This makes running slirp out-of-process possible that
way (python pseudo-code):

a, b = socket.socketpair(socket.AF_UNIX, socket.SOCK_DGRAM)

subprocess.Popen('qemu -net socket,fd=%d -net user' % a.fileno(), shell=True)
subprocess.Popen('qemu ... -net nic -net socket,fd=%d' % b.fileno(), shell=True)

(to make slirp a seperate project altogether, we would have to have
some compatibility code and/or deprecate various options & HMP
commands for dynamic port forwarding etc - but this looks like a
reachable goal)

Signed-off-by: Marc-André Lureau <marcandre.lur...@redhat.com>

I believe instead of supporting unnamed sockets, we should also support
named one through cli?
This could be a later patch, I have no need for it yet. Perhaps it
should be a chardev then?


I mean something like: -socket id=ud0,path=/tmp/XXX



---
   net/socket.c | 25 +++++++++++++++++++++----
   1 file changed, 21 insertions(+), 4 deletions(-)

diff --git a/net/socket.c b/net/socket.c
index 7095eb749f..8a9c30892d 100644
--- a/net/socket.c
+++ b/net/socket.c
@@ -119,9 +119,13 @@ static ssize_t net_socket_receive_dgram(NetClientState 
*nc, const uint8_t *buf,
       ssize_t ret;

       do {
-        ret = qemu_sendto(s->fd, buf, size, 0,
-                          (struct sockaddr *)&s->dgram_dst,
-                          sizeof(s->dgram_dst));
+        if (s->dgram_dst.sin_family != AF_UNIX) {
+            ret = qemu_sendto(s->fd, buf, size, 0,
+                              (struct sockaddr *)&s->dgram_dst,
+                              sizeof(s->dgram_dst));
+        } else {
+            ret = send(s->fd, buf, size, 0);
+        }

Any reason that send is a must here? send(2) said:
         call

             send(sockfd, buf, len, flags);

         is equivalent to

             sendto(sockfd, buf, len, flags, NULL, 0);
Yes they should be equivalent, but then we need to add ?: operators
for the dest arguments. I preferred to have an if() instead.

thanks


One possible issue here is I'm not sure there's a equivalent send() in 
e.g non POSIX system.

Thanks



       } while (ret == -1 && errno == EINTR);

       if (ret == -1 && errno == EAGAIN) {
@@ -322,6 +326,15 @@ static NetSocketState 
*net_socket_fd_init_dgram(NetClientState *peer,
       int newfd;
       NetClientState *nc;
       NetSocketState *s;
+    SocketAddress *sa;
+    SocketAddressType sa_type;
+
+    sa = socket_local_address(fd, errp);
+    if (!sa) {
+        return NULL;
+    }
+    sa_type = sa->type;
+    qapi_free_SocketAddress(sa);

       /* fd passed: multicast: "learn" dgram_dst address from bound address 
and save it
        * Because this may be "shared" socket from a "master" process, 
datagrams would be recv()
@@ -365,8 +378,12 @@ static NetSocketState 
*net_socket_fd_init_dgram(NetClientState *peer,
                    "socket: fd=%d (cloned mcast=%s:%d)",
                    fd, inet_ntoa(saddr.sin_addr), ntohs(saddr.sin_port));
       } else {
+        if (sa_type == SOCKET_ADDRESS_TYPE_UNIX) {
+            s->dgram_dst.sin_family = AF_UNIX;
+        }
+
           snprintf(nc->info_str, sizeof(nc->info_str),
-                 "socket: fd=%d", fd);
+                 "socket: fd=%d %s", fd, SocketAddressType_str(sa_type));
       }

       return s;