On Wed, Mar 14, 2012 at 04:33:14PM +0800, Amos Kong wrote:
> On 14/03/12 00:39, Michael Roth wrote:
> >On Wed, Mar 07, 2012 at 06:47:45AM +0800, Amos Kong wrote:
> >>Introduce tcp_server_start() by moving original code in
> >>tcp_start_incoming_migration().
> >>
> >>Signed-off-by: Amos Kong<akong@xxxxxxxxxx>
> >>---
> >> net.c | 28 ++++++++++++++++++++++++++++
> >> qemu_socket.h | 2 ++
> >> 2 files changed, 30 insertions(+), 0 deletions(-)
> >>
> >>diff --git a/net.c b/net.c
> >>index c34474f..e90ff23 100644
> >>--- a/net.c
> >>+++ b/net.c
> >>@@ -99,6 +99,34 @@ static int get_str_sep(char *buf, int buf_size, const char **pp, int sep)
> >> return 0;
> >> }
> >>
> >>+int tcp_server_start(const char *str, int *fd)
> >>+{
> >>+ int val, ret;
> >>+ struct sockaddr_in saddr;
> >>+
> >>+ if (parse_host_port(&saddr, str)< 0) {
> >>+ error_report("invalid host/port combination: %s", str);
> >>+ return -EINVAL;
> >>+ }
> >>+
> >>+ *fd = qemu_socket(PF_INET, SOCK_STREAM, 0);
> >>+ if (fd< 0) {
> >>+ perror("socket");
> >>+ return -1;
> >>+ }
> >>+ socket_set_nonblock(*fd);
> >>+
> >>+ /* allow fast reuse */
> >>+ val = 1;
> >>+ setsockopt(*fd, SOL_SOCKET, SO_REUSEADDR, (const char *)&val, sizeof(val));
> >>+
> >>+ ret = bind(*fd, (struct sockaddr *)&saddr, sizeof(saddr));
> >>+ if (ret< 0) {
> >>+ closesocket(*fd);
> >>+ }
> >>+ return ret;
> >>+}
> >>+
> >
> >I would combine this with patch 2, since it provides context for why
> >this function is being added. Would also do the same for 3 and 4.
> >
> >I see client the client implementation you need to pass fd back by
> >reference since ret can be set to EINPROGRESS/EWOULDBLOCK on success,
>
> ret restores 0 or -socket_error()
> success: 0, -EINPROGRESS
> fail : ret < 0 && ret !=-EINTR && ret != -EWOULDBLOCK
>
> , it should be -EINPROGRESS
I see, I think I was confued by patch #4 where you do a
+ ret = tcp_client_start(host_port, &s->fd);
+ if (ret == -EINPROGRESS || ret == -EWOULDBLOCK) {
+ DPRINTF("connect in progress");
+ qemu_set_fd_handler2(s->fd, NULL, NULL, tcp_wait_for_connect, s);
If ret == EWOULDBLOCK is a failure (or if the call isn't supposed to
return EWOULDBLOCK), we should fail it there rather than passing it on to
tcp_wait_for_connect().
>
> + ret = connect(*fd, (struct sockaddr *)&saddr, sizeof(saddr));
> + if (ret < 0) {
> + ret = -socket_error();
> + if (ret == -EINPROGRESS) {
> + break;
>
>
> >but here it looks like fd is only value if ret=0, so might as well just
> >pass back the fd as the function return value.
>
> Passed *fd tells us if socket creation success
> 'ret' would return 0 or -socket_error()
>
> -socket_error() is the error of socket creation/connection, it would
> be used by other functions.
>
> eg: migration-tcp.c:
> int tcp_start_incoming_migration(
> ...
> ret = tcp_server_start(host_port, &s);
> if (ret < 0) {
> fprintf(stderr, "tcp_server_start: %s\n", strerror(-ret));
>
> We can also add a error msg in tcp_start_outgoing_migration() when
> tcp_client_start() fails.
>
> >Also, is there any reason we can't re-use
> >qemu-sockets.c:inet_listen()/qemu-sockets.c:inet_connect()? AFAICT they
> >serve the same purpose, and already include some of the work from your
> >PATCH #6.
>
> We could not directly use it, there are some difference,
> such as tcp_start_incoming_migration() doesn't set socket no-blocked,
> but net_socket_listen_init() sets socket no-blocked.
I think adding a common function with blocking/non-blocking flag and
having inet_listen_opts()/socket_listen_opts() call it with a wrapper
would be reasonable.
A lot of code is being introduced here to solve problems that are
already handled in qemu-sockets.c. inet_listen()/inet_connect()
already handles backeted-enclosed ipv6 addrs, getting port numbers when
there's more than one colon, getaddrinfo()-based connections, and most
importantly it's had ipv6 support from day 1.
Not 100% sure it'll work for what you're doing, but qemu-sockets.c was
specifically added for this type of use-case and is heavilly used
currently (vnc, nbd, Chardev users), so I think we should use it unless
there's a good reason not to.
>
> There are some repeated code, so I write a tcp_common_start(), and
> use it in net_socket_listen_init()/
> net_socket_connect_init() and tcp_start_incoming_migration()/
> tcp_start_outgoing_migration()
>
>
> >> int parse_host_port(struct sockaddr_in *saddr, const char *str)
> >> {
> >> char buf[512];
> >>diff --git a/qemu_socket.h b/qemu_socket.h
> >>index fe4cf6c..d612793 100644
> >>--- a/qemu_socket.h
> >>+++ b/qemu_socket.h
> >>@@ -54,6 +54,8 @@ int unix_listen(const char *path, char *ostr, int olen);
> >> int unix_connect_opts(QemuOpts *opts);
> >> int unix_connect(const char *path);
> >>
> >>+int tcp_server_start(const char *str, int *fd);
> >>+
> >> /* Old, ipv4 only bits. Don't use for new code. */
> >> int parse_host_port(struct sockaddr_in *saddr, const char *str);
> >> int socket_init(void);
> >>
> >>
> >
>
> --
> Amos.
>
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
