qemu-nbd: Use SOMAXCONN for socket listen() backlog

Our default of a backlog of 1 connection is rather puny; it gets in
the way when we are explicitly allowing multiple clients (such as
qemu-nbd -e N [--shared], or nbd-server-start with its default
"max-connections":0 for unlimited), but is even a problem when we
stick to qemu-nbd's default of only 1 active client but use -t
[--persistent] where a second client can start using the server once
the first finishes.  While the effects are less noticeable on TCP
sockets (since the client can poll() to learn when the server is ready
again), it is definitely observable on Unix sockets, where on Linux, a
client will fail with EAGAIN and no recourse but to sleep an arbitrary
amount of time before retrying if the server backlog is already full.

Since QMP nbd-server-start is always persistent, it now always
requests a backlog of SOMAXCONN; meanwhile, qemu-nbd will request
SOMAXCONN if persistent, otherwise its backlog should be based on the
expected number of clients.

See https://bugzilla.redhat.com/1925045 for a demonstration of where
our low backlog prevents libnbd from connecting as many parallel
clients as it wants.

Reported-by: Richard W.M. Jones <rjones@redhat.com>
Signed-off-by: Eric Blake <eblake@redhat.com>
CC: qemu-stable@nongnu.org
Message-Id: <20210209152759.209074-2-eblake@redhat.com>
Tested-by: Richard W.M. Jones <rjones@redhat.com>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
Signed-off-by: Eric Blake <eblake@redhat.com>

diff --git a/blockdev-nbd.c b/blockdev-nbd.c
index d8443d235b7338949a4e6e10dec5a8cc88fa9f32..b264620b98d8c024b147872ce089e5371d66dd1d 100644 (file)
--- a/blockdev-nbd.c
+++ b/blockdev-nbd.c
@@ -134,7 +134,12 @@ void nbd_server_start(SocketAddress *addr, const char *tls_creds,
     qio_net_listener_set_name(nbd_server->listener,
                               "nbd-listener");
 
-    if (qio_net_listener_open_sync(nbd_server->listener, addr, 1, errp) < 0) {
+    /*
+     * Because this server is persistent, a backlog of SOMAXCONN is
+     * better than trying to size it to max_connections.
+     */
+    if (qio_net_listener_open_sync(nbd_server->listener, addr, SOMAXCONN,
+                                   errp) < 0) {
         goto error;
     }
 

diff --git a/qemu-nbd.c b/qemu-nbd.c
index 608c63e82a255e080b6e28377c4f17d2b851ab0d..1a340ea4858d59b031147a40ee8a89ba482e29c0 100644 (file)
--- a/qemu-nbd.c
+++ b/qemu-nbd.c
@@ -964,8 +964,16 @@ int main(int argc, char **argv)
 
     server = qio_net_listener_new();
     if (socket_activation == 0) {
+        int backlog;
+
+        if (persistent) {
+            backlog = SOMAXCONN;
+        } else {
+            backlog = MIN(shared, SOMAXCONN);
+        }
         saddr = nbd_build_socket_address(sockpath, bindto, port);
-        if (qio_net_listener_open_sync(server, saddr, 1, &local_err) < 0) {
+        if (qio_net_listener_open_sync(server, saddr, backlog,
+                                       &local_err) < 0) {
             object_unref(OBJECT(server));
             error_report_err(local_err);
             exit(EXIT_FAILURE);